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_OXPenalty           = -1;      // Penalty for observed _owner change
 107 static int Knob_SpinSetSucc         = 1;       // spinners set the _succ field
 108 static int Knob_SpinEarly           = 1;
 109 static int Knob_SuccEnabled         = 1;       // futile wake throttling
 110 static int Knob_SuccRestrict        = 0;       // Limit successors + spinners to at-most-one
 111 static int Knob_MaxSpinners         = -1;      // Should be a function of # CPUs
 112 static int Knob_Bonus               = 100;     // spin success bonus
 113 static int Knob_BonusB              = 100;     // spin success bonus
 114 static int Knob_Penalty             = 200;     // spin failure penalty
 115 static int Knob_Poverty             = 1000;
 116 static int Knob_SpinAfterFutile     = 1;       // Spin after returning from park()
 117 static int Knob_FixedSpin           = 0;
 118 static int Knob_OState              = 3;       // Spinner checks thread state of _owner
 119 static int Knob_UsePause            = 1;
 120 static int Knob_ExitPolicy          = 0;
 121 static int Knob_PreSpin             = 10;      // 20-100 likely better
 122 static int Knob_ResetEvent          = 0;
 123 
 124 static int Knob_FastHSSEC           = 0;
 125 static int Knob_MoveNotifyee        = 2;       // notify() - disposition of notifyee
 126 static int Knob_QMode               = 0;       // EntryList-cxq policy - queue discipline
 127 static volatile int InitDone        = 0;
 128 
 129 // -----------------------------------------------------------------------------
 130 // Theory of operations -- Monitors lists, thread residency, etc:
 131 //
 132 // * A thread acquires ownership of a monitor by successfully
 133 //   CAS()ing the _owner field from null to non-null.
 134 //
 135 // * Invariant: A thread appears on at most one monitor list --
 136 //   cxq, EntryList or WaitSet -- at any one time.
 137 //
 138 // * Contending threads "push" themselves onto the cxq with CAS
 139 //   and then spin/park.
 140 //
 141 // * After a contending thread eventually acquires the lock it must
 142 //   dequeue itself from either the EntryList or the cxq.
 143 //
 144 // * The exiting thread identifies and unparks an "heir presumptive"
 145 //   tentative successor thread on the EntryList.  Critically, the
 146 //   exiting thread doesn't unlink the successor thread from the EntryList.
 147 //   After having been unparked, the wakee will recontend for ownership of
 148 //   the monitor.   The successor (wakee) will either acquire the lock or
 149 //   re-park itself.
 150 //
 151 //   Succession is provided for by a policy of competitive handoff.
 152 //   The exiting thread does _not_ grant or pass ownership to the
 153 //   successor thread.  (This is also referred to as "handoff" succession").
 154 //   Instead the exiting thread releases ownership and possibly wakes
 155 //   a successor, so the successor can (re)compete for ownership of the lock.
 156 //   If the EntryList is empty but the cxq is populated the exiting
 157 //   thread will drain the cxq into the EntryList.  It does so by
 158 //   by detaching the cxq (installing null with CAS) and folding
 159 //   the threads from the cxq into the EntryList.  The EntryList is
 160 //   doubly linked, while the cxq is singly linked because of the
 161 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 162 //
 163 // * Concurrency invariants:
 164 //
 165 //   -- only the monitor owner may access or mutate the EntryList.
 166 //      The mutex property of the monitor itself protects the EntryList
 167 //      from concurrent interference.
 168 //   -- Only the monitor owner may detach the cxq.
 169 //
 170 // * The monitor entry list operations avoid locks, but strictly speaking
 171 //   they're not lock-free.  Enter is lock-free, exit is not.
 172 //   For a description of 'Methods and apparatus providing non-blocking access
 173 //   to a resource,' see U.S. Pat. No. 7844973.
 174 //
 175 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 176 //   detaching thread.  This mechanism is immune from the ABA corruption.
 177 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 178 //
 179 // * Taken together, the cxq and the EntryList constitute or form a
 180 //   single logical queue of threads stalled trying to acquire the lock.
 181 //   We use two distinct lists to improve the odds of a constant-time
 182 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 183 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 184 //   A key desideratum is to minimize queue & monitor metadata manipulation
 185 //   that occurs while holding the monitor lock -- that is, we want to
 186 //   minimize monitor lock holds times.  Note that even a small amount of
 187 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 188 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 189 //   locks and monitor metadata.
 190 //
 191 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 192 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 193 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 194 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 195 //
 196 //   The EntryList is ordered by the prevailing queue discipline and
 197 //   can be organized in any convenient fashion, such as a doubly-linked list or
 198 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 199 //   to operate in constant-time.  If we need a priority queue then something akin
 200 //   to Solaris' sleepq would work nicely.  Viz.,
 201 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 202 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 203 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 204 //   EntryList accordingly.
 205 //
 206 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 207 //   somewhat similar to an elevator-scan.
 208 //
 209 // * The monitor synchronization subsystem avoids the use of native
 210 //   synchronization primitives except for the narrow platform-specific
 211 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 212 //   the semantics of park-unpark.  Put another way, this monitor implementation
 213 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 214 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 215 //   underlying OS manages the READY<->RUN transitions.
 216 //
 217 // * Waiting threads reside on the WaitSet list -- wait() puts
 218 //   the caller onto the WaitSet.
 219 //
 220 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 221 //   either the EntryList or cxq.  Subsequent exit() operations will
 222 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 223 //   it's likely the notifyee would simply impale itself on the lock held
 224 //   by the notifier.
 225 //
 226 // * An interesting alternative is to encode cxq as (List,LockByte) where
 227 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 228 //   variable, like _recursions, in the scheme.  The threads or Events that form
 229 //   the list would have to be aligned in 256-byte addresses.  A thread would
 230 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 231 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 232 //   Note that is is *not* word-tearing, but it does presume that full-word
 233 //   CAS operations are coherent with intermix with STB operations.  That's true
 234 //   on most common processors.
 235 //
 236 // * See also http://blogs.sun.com/dave
 237 
 238 
 239 void* ObjectMonitor::operator new (size_t size) throw() {
 240   return AllocateHeap(size, mtInternal);
 241 }
 242 void* ObjectMonitor::operator new[] (size_t size) throw() {
 243   return operator new (size);
 244 }
 245 void ObjectMonitor::operator delete(void* p) {
 246   FreeHeap(p);
 247 }
 248 void ObjectMonitor::operator delete[] (void *p) {
 249   operator delete(p);
 250 }
 251 
 252 // -----------------------------------------------------------------------------
 253 // Enter support
 254 
 255 void ObjectMonitor::enter(TRAPS) {
 256   // The following code is ordered to check the most common cases first
 257   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 258   Thread * const Self = THREAD;
 259 
 260   void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL);
 261   if (cur == NULL) {
 262     // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
 263     assert(_recursions == 0, "invariant");
 264     assert(_owner == Self, "invariant");
 265     return;
 266   }
 267 
 268   if (cur == Self) {
 269     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 270     _recursions++;
 271     return;
 272   }
 273 
 274   if (Self->is_lock_owned ((address)cur)) {
 275     assert(_recursions == 0, "internal state error");
 276     _recursions = 1;
 277     // Commute owner from a thread-specific on-stack BasicLockObject address to
 278     // a full-fledged "Thread *".
 279     _owner = Self;
 280     return;
 281   }
 282 
 283   // We've encountered genuine contention.
 284   assert(Self->_Stalled == 0, "invariant");
 285   Self->_Stalled = intptr_t(this);
 286 
 287   // Try one round of spinning *before* enqueueing Self
 288   // and before going through the awkward and expensive state
 289   // transitions.  The following spin is strictly optional ...
 290   // Note that if we acquire the monitor from an initial spin
 291   // we forgo posting JVMTI events and firing DTRACE probes.
 292   if (Knob_SpinEarly && TrySpin (Self) > 0) {
 293     assert(_owner == Self, "invariant");
 294     assert(_recursions == 0, "invariant");
 295     assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 296     Self->_Stalled = 0;
 297     return;
 298   }
 299 
 300   assert(_owner != Self, "invariant");
 301   assert(_succ != Self, "invariant");
 302   assert(Self->is_Java_thread(), "invariant");
 303   JavaThread * jt = (JavaThread *) Self;
 304   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 305   assert(jt->thread_state() != _thread_blocked, "invariant");
 306   assert(this->object() != NULL, "invariant");
 307   assert(_count >= 0, "invariant");
 308 
 309   // Prevent deflation at STW-time.  See deflate_idle_monitors() and is_busy().
 310   // Ensure the object-monitor relationship remains stable while there's contention.
 311   Atomic::inc(&_count);
 312 
 313   JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
 314   EventJavaMonitorEnter event;
 315   if (event.should_commit()) {
 316     event.set_monitorClass(((oop)this->object())->klass());
 317     event.set_address((uintptr_t)(this->object_addr()));
 318   }
 319 
 320   { // Change java thread status to indicate blocked on monitor enter.
 321     JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
 322 
 323     Self->set_current_pending_monitor(this);
 324 
 325     DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
 326     if (JvmtiExport::should_post_monitor_contended_enter()) {
 327       JvmtiExport::post_monitor_contended_enter(jt, this);
 328 
 329       // The current thread does not yet own the monitor and does not
 330       // yet appear on any queues that would get it made the successor.
 331       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 332       // handler cannot accidentally consume an unpark() meant for the
 333       // ParkEvent associated with this ObjectMonitor.
 334     }
 335 
 336     OSThreadContendState osts(Self->osthread());
 337     ThreadBlockInVM tbivm(jt);
 338 
 339     // TODO-FIXME: change the following for(;;) loop to straight-line code.
 340     for (;;) {
 341       jt->set_suspend_equivalent();
 342       // cleared by handle_special_suspend_equivalent_condition()
 343       // or java_suspend_self()
 344 
 345       EnterI(THREAD);
 346 
 347       if (!ExitSuspendEquivalent(jt)) break;
 348 
 349       // We have acquired the contended monitor, but while we were
 350       // waiting another thread suspended us. We don't want to enter
 351       // the monitor while suspended because that would surprise the
 352       // thread that suspended us.
 353       //
 354       _recursions = 0;
 355       _succ = NULL;
 356       exit(false, Self);
 357 
 358       jt->java_suspend_self();
 359     }
 360     Self->set_current_pending_monitor(NULL);
 361 
 362     // We cleared the pending monitor info since we've just gotten past
 363     // the enter-check-for-suspend dance and we now own the monitor free
 364     // and clear, i.e., it is no longer pending. The ThreadBlockInVM
 365     // destructor can go to a safepoint at the end of this block. If we
 366     // do a thread dump during that safepoint, then this thread will show
 367     // as having "-locked" the monitor, but the OS and java.lang.Thread
 368     // states will still report that the thread is blocked trying to
 369     // acquire it.
 370   }
 371 
 372   Atomic::dec(&_count);
 373   assert(_count >= 0, "invariant");
 374   Self->_Stalled = 0;
 375 
 376   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 377   assert(_recursions == 0, "invariant");
 378   assert(_owner == Self, "invariant");
 379   assert(_succ != Self, "invariant");
 380   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 381 
 382   // The thread -- now the owner -- is back in vm mode.
 383   // Report the glorious news via TI,DTrace and jvmstat.
 384   // The probe effect is non-trivial.  All the reportage occurs
 385   // while we hold the monitor, increasing the length of the critical
 386   // section.  Amdahl's parallel speedup law comes vividly into play.
 387   //
 388   // Another option might be to aggregate the events (thread local or
 389   // per-monitor aggregation) and defer reporting until a more opportune
 390   // time -- such as next time some thread encounters contention but has
 391   // yet to acquire the lock.  While spinning that thread could
 392   // spinning we could increment JVMStat counters, etc.
 393 
 394   DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
 395   if (JvmtiExport::should_post_monitor_contended_entered()) {
 396     JvmtiExport::post_monitor_contended_entered(jt, this);
 397 
 398     // The current thread already owns the monitor and is not going to
 399     // call park() for the remainder of the monitor enter protocol. So
 400     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 401     // event handler consumed an unpark() issued by the thread that
 402     // just exited the monitor.
 403   }
 404   if (event.should_commit()) {
 405     event.set_previousOwner((uintptr_t)_previous_owner_tid);
 406     event.commit();
 407   }
 408   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 409 }
 410 
 411 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 412 // Callers must compensate as needed.
 413 
 414 int ObjectMonitor::TryLock(Thread * Self) {
 415   void * own = _owner;
 416   if (own != NULL) return 0;
 417   if (Atomic::replace_if_null(Self, &_owner)) {
 418     // Either guarantee _recursions == 0 or set _recursions = 0.
 419     assert(_recursions == 0, "invariant");
 420     assert(_owner == Self, "invariant");
 421     return 1;
 422   }
 423   // The lock had been free momentarily, but we lost the race to the lock.
 424   // Interference -- the CAS failed.
 425   // We can either return -1 or retry.
 426   // Retry doesn't make as much sense because the lock was just acquired.
 427   return -1;
 428 }
 429 
 430 #define MAX_RECHECK_INTERVAL 1000
 431 
 432 void ObjectMonitor::EnterI(TRAPS) {
 433   Thread * const Self = THREAD;
 434   assert(Self->is_Java_thread(), "invariant");
 435   assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
 436 
 437   // Try the lock - TATAS
 438   if (TryLock (Self) > 0) {
 439     assert(_succ != Self, "invariant");
 440     assert(_owner == Self, "invariant");
 441     assert(_Responsible != Self, "invariant");
 442     return;
 443   }
 444 
 445   DeferredInitialize();
 446 
 447   // We try one round of spinning *before* enqueueing Self.
 448   //
 449   // If the _owner is ready but OFFPROC we could use a YieldTo()
 450   // operation to donate the remainder of this thread's quantum
 451   // to the owner.  This has subtle but beneficial affinity
 452   // effects.
 453 
 454   if (TrySpin (Self) > 0) {
 455     assert(_owner == Self, "invariant");
 456     assert(_succ != Self, "invariant");
 457     assert(_Responsible != Self, "invariant");
 458     return;
 459   }
 460 
 461   // The Spin failed -- Enqueue and park the thread ...
 462   assert(_succ != Self, "invariant");
 463   assert(_owner != Self, "invariant");
 464   assert(_Responsible != Self, "invariant");
 465 
 466   // Enqueue "Self" on ObjectMonitor's _cxq.
 467   //
 468   // Node acts as a proxy for Self.
 469   // As an aside, if were to ever rewrite the synchronization code mostly
 470   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 471   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 472   // as well as eliminate a subset of ABA issues.
 473   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 474 
 475   ObjectWaiter node(Self);
 476   Self->_ParkEvent->reset();
 477   node._prev   = (ObjectWaiter *) 0xBAD;
 478   node.TState  = ObjectWaiter::TS_CXQ;
 479 
 480   // Push "Self" onto the front of the _cxq.
 481   // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
 482   // Note that spinning tends to reduce the rate at which threads
 483   // enqueue and dequeue on EntryList|cxq.
 484   ObjectWaiter * nxt;
 485   for (;;) {
 486     node._next = nxt = _cxq;
 487     if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break;
 488 
 489     // Interference - the CAS failed because _cxq changed.  Just retry.
 490     // As an optional optimization we retry the lock.
 491     if (TryLock (Self) > 0) {
 492       assert(_succ != Self, "invariant");
 493       assert(_owner == Self, "invariant");
 494       assert(_Responsible != Self, "invariant");
 495       return;
 496     }
 497   }
 498 
 499   // Check for cxq|EntryList edge transition to non-null.  This indicates
 500   // the onset of contention.  While contention persists exiting threads
 501   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 502   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 503   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 504   // arrange for one of the contending thread to use a timed park() operations
 505   // to detect and recover from the race.  (Stranding is form of progress failure
 506   // where the monitor is unlocked but all the contending threads remain parked).
 507   // That is, at least one of the contended threads will periodically poll _owner.
 508   // One of the contending threads will become the designated "Responsible" thread.
 509   // The Responsible thread uses a timed park instead of a normal indefinite park
 510   // operation -- it periodically wakes and checks for and recovers from potential
 511   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 512   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 513   // be responsible for a monitor.
 514   //
 515   // Currently, one of the contended threads takes on the added role of "Responsible".
 516   // A viable alternative would be to use a dedicated "stranding checker" thread
 517   // that periodically iterated over all the threads (or active monitors) and unparked
 518   // successors where there was risk of stranding.  This would help eliminate the
 519   // timer scalability issues we see on some platforms as we'd only have one thread
 520   // -- the checker -- parked on a timer.
 521 
 522   if (nxt == NULL && _EntryList == NULL) {
 523     // Try to assume the role of responsible thread for the monitor.
 524     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=Self }
 525     Atomic::replace_if_null(Self, &_Responsible);
 526   }
 527 
 528   // The lock might have been released while this thread was occupied queueing
 529   // itself onto _cxq.  To close the race and avoid "stranding" and
 530   // progress-liveness failure we must resample-retry _owner before parking.
 531   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 532   // In this case the ST-MEMBAR is accomplished with CAS().
 533   //
 534   // TODO: Defer all thread state transitions until park-time.
 535   // Since state transitions are heavy and inefficient we'd like
 536   // to defer the state transitions until absolutely necessary,
 537   // and in doing so avoid some transitions ...
 538 
 539   int nWakeups = 0;
 540   int recheckInterval = 1;
 541 
 542   for (;;) {
 543 
 544     if (TryLock(Self) > 0) break;
 545     assert(_owner != Self, "invariant");
 546 
 547     // park self
 548     if (_Responsible == Self) {
 549       Self->_ParkEvent->park((jlong) recheckInterval);
 550       // Increase the recheckInterval, but clamp the value.
 551       recheckInterval *= 8;
 552       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 553         recheckInterval = MAX_RECHECK_INTERVAL;
 554       }
 555     } else {
 556       Self->_ParkEvent->park();
 557     }
 558 
 559     if (TryLock(Self) > 0) break;
 560 
 561     // The lock is still contested.
 562     // Keep a tally of the # of futile wakeups.
 563     // Note that the counter is not protected by a lock or updated by atomics.
 564     // That is by design - we trade "lossy" counters which are exposed to
 565     // races during updates for a lower probe effect.
 566 
 567     // This PerfData object can be used in parallel with a safepoint.
 568     // See the work around in PerfDataManager::destroy().
 569     OM_PERFDATA_OP(FutileWakeups, inc());
 570     ++nWakeups;
 571 
 572     // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
 573     // We can defer clearing _succ until after the spin completes
 574     // TrySpin() must tolerate being called with _succ == Self.
 575     // Try yet another round of adaptive spinning.
 576     if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
 577 
 578     // We can find that we were unpark()ed and redesignated _succ while
 579     // we were spinning.  That's harmless.  If we iterate and call park(),
 580     // park() will consume the event and return immediately and we'll
 581     // just spin again.  This pattern can repeat, leaving _succ to simply
 582     // spin on a CPU.  Enable Knob_ResetEvent to clear pending unparks().
 583     // Alternately, we can sample fired() here, and if set, forgo spinning
 584     // in the next iteration.
 585 
 586     if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
 587       Self->_ParkEvent->reset();
 588       OrderAccess::fence();
 589     }
 590     if (_succ == Self) _succ = NULL;
 591 
 592     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 593     OrderAccess::fence();
 594   }
 595 
 596   // Egress :
 597   // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
 598   // Normally we'll find Self on the EntryList .
 599   // From the perspective of the lock owner (this thread), the
 600   // EntryList is stable and cxq is prepend-only.
 601   // The head of cxq is volatile but the interior is stable.
 602   // In addition, Self.TState is stable.
 603 
 604   assert(_owner == Self, "invariant");
 605   assert(object() != NULL, "invariant");
 606   // I'd like to write:
 607   //   guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 608   // but as we're at a safepoint that's not safe.
 609 
 610   UnlinkAfterAcquire(Self, &node);
 611   if (_succ == Self) _succ = NULL;
 612 
 613   assert(_succ != Self, "invariant");
 614   if (_Responsible == Self) {
 615     _Responsible = NULL;
 616     OrderAccess::fence(); // Dekker pivot-point
 617 
 618     // We may leave threads on cxq|EntryList without a designated
 619     // "Responsible" thread.  This is benign.  When this thread subsequently
 620     // exits the monitor it can "see" such preexisting "old" threads --
 621     // threads that arrived on the cxq|EntryList before the fence, above --
 622     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 623     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 624     // non-null and elect a new "Responsible" timer thread.
 625     //
 626     // This thread executes:
 627     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 628     //    LD cxq|EntryList               (in subsequent exit)
 629     //
 630     // Entering threads in the slow/contended path execute:
 631     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 632     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 633     //
 634     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 635     // exit operation from floating above the ST Responsible=null.
 636   }
 637 
 638   // We've acquired ownership with CAS().
 639   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 640   // But since the CAS() this thread may have also stored into _succ,
 641   // EntryList, cxq or Responsible.  These meta-data updates must be
 642   // visible __before this thread subsequently drops the lock.
 643   // Consider what could occur if we didn't enforce this constraint --
 644   // STs to monitor meta-data and user-data could reorder with (become
 645   // visible after) the ST in exit that drops ownership of the lock.
 646   // Some other thread could then acquire the lock, but observe inconsistent
 647   // or old monitor meta-data and heap data.  That violates the JMM.
 648   // To that end, the 1-0 exit() operation must have at least STST|LDST
 649   // "release" barrier semantics.  Specifically, there must be at least a
 650   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 651   // the lock.   The barrier ensures that changes to monitor meta-data and data
 652   // protected by the lock will be visible before we release the lock, and
 653   // therefore before some other thread (CPU) has a chance to acquire the lock.
 654   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 655   //
 656   // Critically, any prior STs to _succ or EntryList must be visible before
 657   // the ST of null into _owner in the *subsequent* (following) corresponding
 658   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 659   // execute a serializing instruction.
 660 
 661   return;
 662 }
 663 
 664 // ReenterI() is a specialized inline form of the latter half of the
 665 // contended slow-path from EnterI().  We use ReenterI() only for
 666 // monitor reentry in wait().
 667 //
 668 // In the future we should reconcile EnterI() and ReenterI(), adding
 669 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
 670 // loop accordingly.
 671 
 672 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
 673   assert(Self != NULL, "invariant");
 674   assert(SelfNode != NULL, "invariant");
 675   assert(SelfNode->_thread == Self, "invariant");
 676   assert(_waiters > 0, "invariant");
 677   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 678   assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
 679   JavaThread * jt = (JavaThread *) Self;
 680 
 681   int nWakeups = 0;
 682   for (;;) {
 683     ObjectWaiter::TStates v = SelfNode->TState;
 684     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
 685     assert(_owner != Self, "invariant");
 686 
 687     if (TryLock(Self) > 0) break;
 688     if (TrySpin(Self) > 0) break;
 689 
 690     // State transition wrappers around park() ...
 691     // ReenterI() wisely defers state transitions until
 692     // it's clear we must park the thread.
 693     {
 694       OSThreadContendState osts(Self->osthread());
 695       ThreadBlockInVM tbivm(jt);
 696 
 697       // cleared by handle_special_suspend_equivalent_condition()
 698       // or java_suspend_self()
 699       jt->set_suspend_equivalent();
 700       Self->_ParkEvent->park();
 701 
 702       // were we externally suspended while we were waiting?
 703       for (;;) {
 704         if (!ExitSuspendEquivalent(jt)) break;
 705         if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
 706         jt->java_suspend_self();
 707         jt->set_suspend_equivalent();
 708       }
 709     }
 710 
 711     // Try again, but just so we distinguish between futile wakeups and
 712     // successful wakeups.  The following test isn't algorithmically
 713     // necessary, but it helps us maintain sensible statistics.
 714     if (TryLock(Self) > 0) break;
 715 
 716     // The lock is still contested.
 717     // Keep a tally of the # of futile wakeups.
 718     // Note that the counter is not protected by a lock or updated by atomics.
 719     // That is by design - we trade "lossy" counters which are exposed to
 720     // races during updates for a lower probe effect.
 721     ++nWakeups;
 722 
 723     // Assuming this is not a spurious wakeup we'll normally
 724     // find that _succ == Self.
 725     if (_succ == Self) _succ = NULL;
 726 
 727     // Invariant: after clearing _succ a contending thread
 728     // *must* retry  _owner before parking.
 729     OrderAccess::fence();
 730 
 731     // This PerfData object can be used in parallel with a safepoint.
 732     // See the work around in PerfDataManager::destroy().
 733     OM_PERFDATA_OP(FutileWakeups, inc());
 734   }
 735 
 736   // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
 737   // Normally we'll find Self on the EntryList.
 738   // Unlinking from the EntryList is constant-time and atomic-free.
 739   // From the perspective of the lock owner (this thread), the
 740   // EntryList is stable and cxq is prepend-only.
 741   // The head of cxq is volatile but the interior is stable.
 742   // In addition, Self.TState is stable.
 743 
 744   assert(_owner == Self, "invariant");
 745   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 746   UnlinkAfterAcquire(Self, SelfNode);
 747   if (_succ == Self) _succ = NULL;
 748   assert(_succ != Self, "invariant");
 749   SelfNode->TState = ObjectWaiter::TS_RUN;
 750   OrderAccess::fence();      // see comments at the end of EnterI()
 751 }
 752 
 753 // By convention we unlink a contending thread from EntryList|cxq immediately
 754 // after the thread acquires the lock in ::enter().  Equally, we could defer
 755 // unlinking the thread until ::exit()-time.
 756 
 757 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
 758   assert(_owner == Self, "invariant");
 759   assert(SelfNode->_thread == Self, "invariant");
 760 
 761   if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
 762     // Normal case: remove Self from the DLL EntryList .
 763     // This is a constant-time operation.
 764     ObjectWaiter * nxt = SelfNode->_next;
 765     ObjectWaiter * prv = SelfNode->_prev;
 766     if (nxt != NULL) nxt->_prev = prv;
 767     if (prv != NULL) prv->_next = nxt;
 768     if (SelfNode == _EntryList) _EntryList = nxt;
 769     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
 770     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
 771   } else {
 772     assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
 773     // Inopportune interleaving -- Self is still on the cxq.
 774     // This usually means the enqueue of self raced an exiting thread.
 775     // Normally we'll find Self near the front of the cxq, so
 776     // dequeueing is typically fast.  If needbe we can accelerate
 777     // this with some MCS/CHL-like bidirectional list hints and advisory
 778     // back-links so dequeueing from the interior will normally operate
 779     // in constant-time.
 780     // Dequeue Self from either the head (with CAS) or from the interior
 781     // with a linear-time scan and normal non-atomic memory operations.
 782     // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
 783     // and then unlink Self from EntryList.  We have to drain eventually,
 784     // so it might as well be now.
 785 
 786     ObjectWaiter * v = _cxq;
 787     assert(v != NULL, "invariant");
 788     if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) {
 789       // The CAS above can fail from interference IFF a "RAT" arrived.
 790       // In that case Self must be in the interior and can no longer be
 791       // at the head of cxq.
 792       if (v == SelfNode) {
 793         assert(_cxq != v, "invariant");
 794         v = _cxq;          // CAS above failed - start scan at head of list
 795       }
 796       ObjectWaiter * p;
 797       ObjectWaiter * q = NULL;
 798       for (p = v; p != NULL && p != SelfNode; p = p->_next) {
 799         q = p;
 800         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
 801       }
 802       assert(v != SelfNode, "invariant");
 803       assert(p == SelfNode, "Node not found on cxq");
 804       assert(p != _cxq, "invariant");
 805       assert(q != NULL, "invariant");
 806       assert(q->_next == p, "invariant");
 807       q->_next = p->_next;
 808     }
 809   }
 810 
 811 #ifdef ASSERT
 812   // Diagnostic hygiene ...
 813   SelfNode->_prev  = (ObjectWaiter *) 0xBAD;
 814   SelfNode->_next  = (ObjectWaiter *) 0xBAD;
 815   SelfNode->TState = ObjectWaiter::TS_RUN;
 816 #endif
 817 }
 818 
 819 // -----------------------------------------------------------------------------
 820 // Exit support
 821 //
 822 // exit()
 823 // ~~~~~~
 824 // Note that the collector can't reclaim the objectMonitor or deflate
 825 // the object out from underneath the thread calling ::exit() as the
 826 // thread calling ::exit() never transitions to a stable state.
 827 // This inhibits GC, which in turn inhibits asynchronous (and
 828 // inopportune) reclamation of "this".
 829 //
 830 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
 831 // There's one exception to the claim above, however.  EnterI() can call
 832 // exit() to drop a lock if the acquirer has been externally suspended.
 833 // In that case exit() is called with _thread_state as _thread_blocked,
 834 // but the monitor's _count field is > 0, which inhibits reclamation.
 835 //
 836 // 1-0 exit
 837 // ~~~~~~~~
 838 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
 839 // the fast-path operators have been optimized so the common ::exit()
 840 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
 841 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
 842 // greatly improves latency -- MEMBAR and CAS having considerable local
 843 // latency on modern processors -- but at the cost of "stranding".  Absent the
 844 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
 845 // ::enter() path, resulting in the entering thread being stranding
 846 // and a progress-liveness failure.   Stranding is extremely rare.
 847 // We use timers (timed park operations) & periodic polling to detect
 848 // and recover from stranding.  Potentially stranded threads periodically
 849 // wake up and poll the lock.  See the usage of the _Responsible variable.
 850 //
 851 // The CAS() in enter provides for safety and exclusion, while the CAS or
 852 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
 853 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
 854 // We detect and recover from stranding with timers.
 855 //
 856 // If a thread transiently strands it'll park until (a) another
 857 // thread acquires the lock and then drops the lock, at which time the
 858 // exiting thread will notice and unpark the stranded thread, or, (b)
 859 // the timer expires.  If the lock is high traffic then the stranding latency
 860 // will be low due to (a).  If the lock is low traffic then the odds of
 861 // stranding are lower, although the worst-case stranding latency
 862 // is longer.  Critically, we don't want to put excessive load in the
 863 // platform's timer subsystem.  We want to minimize both the timer injection
 864 // rate (timers created/sec) as well as the number of timers active at
 865 // any one time.  (more precisely, we want to minimize timer-seconds, which is
 866 // the integral of the # of active timers at any instant over time).
 867 // Both impinge on OS scalability.  Given that, at most one thread parked on
 868 // a monitor will use a timer.
 869 //
 870 // There is also the risk of a futile wake-up. If we drop the lock
 871 // another thread can reacquire the lock immediately, and we can
 872 // then wake a thread unnecessarily. This is benign, and we've
 873 // structured the code so the windows are short and the frequency
 874 // of such futile wakups is low.
 875 
 876 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
 877   Thread * const Self = THREAD;
 878   if (THREAD != _owner) {
 879     if (THREAD->is_lock_owned((address) _owner)) {
 880       // Transmute _owner from a BasicLock pointer to a Thread address.
 881       // We don't need to hold _mutex for this transition.
 882       // Non-null to Non-null is safe as long as all readers can
 883       // tolerate either flavor.
 884       assert(_recursions == 0, "invariant");
 885       _owner = THREAD;
 886       _recursions = 0;
 887     } else {
 888       // Apparent unbalanced locking ...
 889       // Naively we'd like to throw IllegalMonitorStateException.
 890       // As a practical matter we can neither allocate nor throw an
 891       // exception as ::exit() can be called from leaf routines.
 892       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
 893       // Upon deeper reflection, however, in a properly run JVM the only
 894       // way we should encounter this situation is in the presence of
 895       // unbalanced JNI locking. TODO: CheckJNICalls.
 896       // See also: CR4414101
 897       assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
 898       return;
 899     }
 900   }
 901 
 902   if (_recursions != 0) {
 903     _recursions--;        // this is simple recursive enter
 904     return;
 905   }
 906 
 907   // Invariant: after setting Responsible=null an thread must execute
 908   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
 909   _Responsible = NULL;
 910 
 911 #if INCLUDE_JFR
 912   // get the owner's thread id for the MonitorEnter event
 913   // if it is enabled and the thread isn't suspended
 914   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
 915     _previous_owner_tid = JFR_THREAD_ID(Self);
 916   }
 917 #endif
 918 
 919   for (;;) {
 920     assert(THREAD == _owner, "invariant");
 921 
 922     if (Knob_ExitPolicy == 0) {
 923       // release semantics: prior loads and stores from within the critical section
 924       // must not float (reorder) past the following store that drops the lock.
 925       // On SPARC that requires MEMBAR #loadstore|#storestore.
 926       // But of course in TSO #loadstore|#storestore is not required.
 927       // I'd like to write one of the following:
 928       // A.  OrderAccess::release() ; _owner = NULL
 929       // B.  OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
 930       // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
 931       // store into a _dummy variable.  That store is not needed, but can result
 932       // in massive wasteful coherency traffic on classic SMP systems.
 933       // Instead, I use release_store(), which is implemented as just a simple
 934       // ST on x64, x86 and SPARC.
 935       OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 936       OrderAccess::storeload();                        // See if we need to wake a successor
 937       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 938         return;
 939       }
 940       // Other threads are blocked trying to acquire the lock.
 941 
 942       // Normally the exiting thread is responsible for ensuring succession,
 943       // but if other successors are ready or other entering threads are spinning
 944       // then this thread can simply store NULL into _owner and exit without
 945       // waking a successor.  The existence of spinners or ready successors
 946       // guarantees proper succession (liveness).  Responsibility passes to the
 947       // ready or running successors.  The exiting thread delegates the duty.
 948       // More precisely, if a successor already exists this thread is absolved
 949       // of the responsibility of waking (unparking) one.
 950       //
 951       // The _succ variable is critical to reducing futile wakeup frequency.
 952       // _succ identifies the "heir presumptive" thread that has been made
 953       // ready (unparked) but that has not yet run.  We need only one such
 954       // successor thread to guarantee progress.
 955       // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 956       // section 3.3 "Futile Wakeup Throttling" for details.
 957       //
 958       // Note that spinners in Enter() also set _succ non-null.
 959       // In the current implementation spinners opportunistically set
 960       // _succ so that exiting threads might avoid waking a successor.
 961       // Another less appealing alternative would be for the exiting thread
 962       // to drop the lock and then spin briefly to see if a spinner managed
 963       // to acquire the lock.  If so, the exiting thread could exit
 964       // immediately without waking a successor, otherwise the exiting
 965       // thread would need to dequeue and wake a successor.
 966       // (Note that we'd need to make the post-drop spin short, but no
 967       // shorter than the worst-case round-trip cache-line migration time.
 968       // The dropped lock needs to become visible to the spinner, and then
 969       // the acquisition of the lock by the spinner must become visible to
 970       // the exiting thread).
 971 
 972       // It appears that an heir-presumptive (successor) must be made ready.
 973       // Only the current lock owner can manipulate the EntryList or
 974       // drain _cxq, so we need to reacquire the lock.  If we fail
 975       // to reacquire the lock the responsibility for ensuring succession
 976       // falls to the new owner.
 977       //
 978       if (!Atomic::replace_if_null(THREAD, &_owner)) {
 979         return;
 980       }
 981     } else {
 982       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 983         OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 984         OrderAccess::storeload();
 985         // Ratify the previously observed values.
 986         if (_cxq == NULL || _succ != NULL) {
 987           return;
 988         }
 989 
 990         // inopportune interleaving -- the exiting thread (this thread)
 991         // in the fast-exit path raced an entering thread in the slow-enter
 992         // path.
 993         // We have two choices:
 994         // A.  Try to reacquire the lock.
 995         //     If the CAS() fails return immediately, otherwise
 996         //     we either restart/rerun the exit operation, or simply
 997         //     fall-through into the code below which wakes a successor.
 998         // B.  If the elements forming the EntryList|cxq are TSM
 999         //     we could simply unpark() the lead thread and return
1000         //     without having set _succ.
1001         if (!Atomic::replace_if_null(THREAD, &_owner)) {
1002           return;
1003         }
1004       }
1005     }
1006 
1007     guarantee(_owner == THREAD, "invariant");
1008 
1009     ObjectWaiter * w = NULL;
1010     int QMode = Knob_QMode;
1011 
1012     if (QMode == 2 && _cxq != NULL) {
1013       // QMode == 2 : cxq has precedence over EntryList.
1014       // Try to directly wake a successor from the cxq.
1015       // If successful, the successor will need to unlink itself from cxq.
1016       w = _cxq;
1017       assert(w != NULL, "invariant");
1018       assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1019       ExitEpilog(Self, w);
1020       return;
1021     }
1022 
1023     if (QMode == 3 && _cxq != NULL) {
1024       // Aggressively drain cxq into EntryList at the first opportunity.
1025       // This policy ensure that recently-run threads live at the head of EntryList.
1026       // Drain _cxq into EntryList - bulk transfer.
1027       // First, detach _cxq.
1028       // The following loop is tantamount to: w = swap(&cxq, NULL)
1029       w = _cxq;
1030       for (;;) {
1031         assert(w != NULL, "Invariant");
1032         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1033         if (u == w) break;
1034         w = u;
1035       }
1036       assert(w != NULL, "invariant");
1037 
1038       ObjectWaiter * q = NULL;
1039       ObjectWaiter * p;
1040       for (p = w; p != NULL; p = p->_next) {
1041         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1042         p->TState = ObjectWaiter::TS_ENTER;
1043         p->_prev = q;
1044         q = p;
1045       }
1046 
1047       // Append the RATs to the EntryList
1048       // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1049       ObjectWaiter * Tail;
1050       for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1051            Tail = Tail->_next)
1052         /* empty */;
1053       if (Tail == NULL) {
1054         _EntryList = w;
1055       } else {
1056         Tail->_next = w;
1057         w->_prev = Tail;
1058       }
1059 
1060       // Fall thru into code that tries to wake a successor from EntryList
1061     }
1062 
1063     if (QMode == 4 && _cxq != NULL) {
1064       // Aggressively drain cxq into EntryList at the first opportunity.
1065       // This policy ensure that recently-run threads live at the head of EntryList.
1066 
1067       // Drain _cxq into EntryList - bulk transfer.
1068       // First, detach _cxq.
1069       // The following loop is tantamount to: w = swap(&cxq, NULL)
1070       w = _cxq;
1071       for (;;) {
1072         assert(w != NULL, "Invariant");
1073         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1074         if (u == w) break;
1075         w = u;
1076       }
1077       assert(w != NULL, "invariant");
1078 
1079       ObjectWaiter * q = NULL;
1080       ObjectWaiter * p;
1081       for (p = w; p != NULL; p = p->_next) {
1082         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1083         p->TState = ObjectWaiter::TS_ENTER;
1084         p->_prev = q;
1085         q = p;
1086       }
1087 
1088       // Prepend the RATs to the EntryList
1089       if (_EntryList != NULL) {
1090         q->_next = _EntryList;
1091         _EntryList->_prev = q;
1092       }
1093       _EntryList = w;
1094 
1095       // Fall thru into code that tries to wake a successor from EntryList
1096     }
1097 
1098     w = _EntryList;
1099     if (w != NULL) {
1100       // I'd like to write: guarantee (w->_thread != Self).
1101       // But in practice an exiting thread may find itself on the EntryList.
1102       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1103       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1104       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1105       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1106       // release the lock "O".  T2 resumes immediately after the ST of null into
1107       // _owner, above.  T2 notices that the EntryList is populated, so it
1108       // reacquires the lock and then finds itself on the EntryList.
1109       // Given all that, we have to tolerate the circumstance where "w" is
1110       // associated with Self.
1111       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1112       ExitEpilog(Self, w);
1113       return;
1114     }
1115 
1116     // If we find that both _cxq and EntryList are null then just
1117     // re-run the exit protocol from the top.
1118     w = _cxq;
1119     if (w == NULL) continue;
1120 
1121     // Drain _cxq into EntryList - bulk transfer.
1122     // First, detach _cxq.
1123     // The following loop is tantamount to: w = swap(&cxq, NULL)
1124     for (;;) {
1125       assert(w != NULL, "Invariant");
1126       ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1127       if (u == w) break;
1128       w = u;
1129     }
1130 
1131     assert(w != NULL, "invariant");
1132     assert(_EntryList == NULL, "invariant");
1133 
1134     // Convert the LIFO SLL anchored by _cxq into a DLL.
1135     // The list reorganization step operates in O(LENGTH(w)) time.
1136     // It's critical that this step operate quickly as
1137     // "Self" still holds the outer-lock, restricting parallelism
1138     // and effectively lengthening the critical section.
1139     // Invariant: s chases t chases u.
1140     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1141     // we have faster access to the tail.
1142 
1143     if (QMode == 1) {
1144       // QMode == 1 : drain cxq to EntryList, reversing order
1145       // We also reverse the order of the list.
1146       ObjectWaiter * s = NULL;
1147       ObjectWaiter * t = w;
1148       ObjectWaiter * u = NULL;
1149       while (t != NULL) {
1150         guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1151         t->TState = ObjectWaiter::TS_ENTER;
1152         u = t->_next;
1153         t->_prev = u;
1154         t->_next = s;
1155         s = t;
1156         t = u;
1157       }
1158       _EntryList  = s;
1159       assert(s != NULL, "invariant");
1160     } else {
1161       // QMode == 0 or QMode == 2
1162       _EntryList = w;
1163       ObjectWaiter * q = NULL;
1164       ObjectWaiter * p;
1165       for (p = w; p != NULL; p = p->_next) {
1166         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1167         p->TState = ObjectWaiter::TS_ENTER;
1168         p->_prev = q;
1169         q = p;
1170       }
1171     }
1172 
1173     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1174     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1175 
1176     // See if we can abdicate to a spinner instead of waking a thread.
1177     // A primary goal of the implementation is to reduce the
1178     // context-switch rate.
1179     if (_succ != NULL) continue;
1180 
1181     w = _EntryList;
1182     if (w != NULL) {
1183       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1184       ExitEpilog(Self, w);
1185       return;
1186     }
1187   }
1188 }
1189 
1190 // ExitSuspendEquivalent:
1191 // A faster alternate to handle_special_suspend_equivalent_condition()
1192 //
1193 // handle_special_suspend_equivalent_condition() unconditionally
1194 // acquires the SR_lock.  On some platforms uncontended MutexLocker()
1195 // operations have high latency.  Note that in ::enter() we call HSSEC
1196 // while holding the monitor, so we effectively lengthen the critical sections.
1197 //
1198 // There are a number of possible solutions:
1199 //
1200 // A.  To ameliorate the problem we might also defer state transitions
1201 //     to as late as possible -- just prior to parking.
1202 //     Given that, we'd call HSSEC after having returned from park(),
1203 //     but before attempting to acquire the monitor.  This is only a
1204 //     partial solution.  It avoids calling HSSEC while holding the
1205 //     monitor (good), but it still increases successor reacquisition latency --
1206 //     the interval between unparking a successor and the time the successor
1207 //     resumes and retries the lock.  See ReenterI(), which defers state transitions.
1208 //     If we use this technique we can also avoid EnterI()-exit() loop
1209 //     in ::enter() where we iteratively drop the lock and then attempt
1210 //     to reacquire it after suspending.
1211 //
1212 // B.  In the future we might fold all the suspend bits into a
1213 //     composite per-thread suspend flag and then update it with CAS().
1214 //     Alternately, a Dekker-like mechanism with multiple variables
1215 //     would suffice:
1216 //       ST Self->_suspend_equivalent = false
1217 //       MEMBAR
1218 //       LD Self_>_suspend_flags
1219 //
1220 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1221 // with a more efficient implementation, the need to use "FastHSSEC" has
1222 // decreased. - Dave
1223 
1224 
1225 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1226   const int Mode = Knob_FastHSSEC;
1227   if (Mode && !jSelf->is_external_suspend()) {
1228     assert(jSelf->is_suspend_equivalent(), "invariant");
1229     jSelf->clear_suspend_equivalent();
1230     if (2 == Mode) OrderAccess::storeload();
1231     if (!jSelf->is_external_suspend()) return false;
1232     // We raced a suspension -- fall thru into the slow path
1233     jSelf->set_suspend_equivalent();
1234   }
1235   return jSelf->handle_special_suspend_equivalent_condition();
1236 }
1237 
1238 
1239 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1240   assert(_owner == Self, "invariant");
1241 
1242   // Exit protocol:
1243   // 1. ST _succ = wakee
1244   // 2. membar #loadstore|#storestore;
1245   // 2. ST _owner = NULL
1246   // 3. unpark(wakee)
1247 
1248   _succ = Knob_SuccEnabled ? Wakee->_thread : NULL;
1249   ParkEvent * Trigger = Wakee->_event;
1250 
1251   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1252   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1253   // out-of-scope (non-extant).
1254   Wakee  = NULL;
1255 
1256   // Drop the lock
1257   OrderAccess::release_store(&_owner, (void*)NULL);
1258   OrderAccess::fence();                               // ST _owner vs LD in unpark()
1259 
1260   DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1261   Trigger->unpark();
1262 
1263   // Maintain stats and report events to JVMTI
1264   OM_PERFDATA_OP(Parks, inc());
1265 }
1266 
1267 
1268 // -----------------------------------------------------------------------------
1269 // Class Loader deadlock handling.
1270 //
1271 // complete_exit exits a lock returning recursion count
1272 // complete_exit/reenter operate as a wait without waiting
1273 // complete_exit requires an inflated monitor
1274 // The _owner field is not always the Thread addr even with an
1275 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1276 // thread due to contention.
1277 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1278   Thread * const Self = THREAD;
1279   assert(Self->is_Java_thread(), "Must be Java thread!");
1280   JavaThread *jt = (JavaThread *)THREAD;
1281 
1282   DeferredInitialize();
1283 
1284   if (THREAD != _owner) {
1285     if (THREAD->is_lock_owned ((address)_owner)) {
1286       assert(_recursions == 0, "internal state error");
1287       _owner = THREAD;   // Convert from basiclock addr to Thread addr
1288       _recursions = 0;
1289     }
1290   }
1291 
1292   guarantee(Self == _owner, "complete_exit not owner");
1293   intptr_t save = _recursions; // record the old recursion count
1294   _recursions = 0;        // set the recursion level to be 0
1295   exit(true, Self);           // exit the monitor
1296   guarantee(_owner != Self, "invariant");
1297   return save;
1298 }
1299 
1300 // reenter() enters a lock and sets recursion count
1301 // complete_exit/reenter operate as a wait without waiting
1302 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1303   Thread * const Self = THREAD;
1304   assert(Self->is_Java_thread(), "Must be Java thread!");
1305   JavaThread *jt = (JavaThread *)THREAD;
1306 
1307   guarantee(_owner != Self, "reenter already owner");
1308   enter(THREAD);       // enter the monitor
1309   guarantee(_recursions == 0, "reenter recursion");
1310   _recursions = recursions;
1311   return;
1312 }
1313 
1314 
1315 // -----------------------------------------------------------------------------
1316 // A macro is used below because there may already be a pending
1317 // exception which should not abort the execution of the routines
1318 // which use this (which is why we don't put this into check_slow and
1319 // call it with a CHECK argument).
1320 
1321 #define CHECK_OWNER()                                                       \
1322   do {                                                                      \
1323     if (THREAD != _owner) {                                                 \
1324       if (THREAD->is_lock_owned((address) _owner)) {                        \
1325         _owner = THREAD;  /* Convert from basiclock addr to Thread addr */  \
1326         _recursions = 0;                                                    \
1327       } else {                                                              \
1328         THROW(vmSymbols::java_lang_IllegalMonitorStateException());         \
1329       }                                                                     \
1330     }                                                                       \
1331   } while (false)
1332 
1333 // check_slow() is a misnomer.  It's called to simply to throw an IMSX exception.
1334 // TODO-FIXME: remove check_slow() -- it's likely dead.
1335 
1336 void ObjectMonitor::check_slow(TRAPS) {
1337   assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1338   THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1339 }
1340 
1341 static int Adjust(volatile int * adr, int dx) {
1342   int v;
1343   for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1344   return v;
1345 }
1346 
1347 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1348                                     ObjectMonitor* monitor,
1349                                     jlong notifier_tid,
1350                                     jlong timeout,
1351                                     bool timedout) {
1352   assert(event != NULL, "invariant");
1353   assert(monitor != NULL, "invariant");
1354   event->set_monitorClass(((oop)monitor->object())->klass());
1355   event->set_timeout(timeout);
1356   event->set_address((uintptr_t)monitor->object_addr());
1357   event->set_notifier(notifier_tid);
1358   event->set_timedOut(timedout);
1359   event->commit();
1360 }
1361 
1362 // -----------------------------------------------------------------------------
1363 // Wait/Notify/NotifyAll
1364 //
1365 // Note: a subset of changes to ObjectMonitor::wait()
1366 // will need to be replicated in complete_exit
1367 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1368   Thread * const Self = THREAD;
1369   assert(Self->is_Java_thread(), "Must be Java thread!");
1370   JavaThread *jt = (JavaThread *)THREAD;
1371 
1372   DeferredInitialize();
1373 
1374   // Throw IMSX or IEX.
1375   CHECK_OWNER();
1376 
1377   EventJavaMonitorWait event;
1378 
1379   // check for a pending interrupt
1380   if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1381     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1382     if (JvmtiExport::should_post_monitor_waited()) {
1383       // Note: 'false' parameter is passed here because the
1384       // wait was not timed out due to thread interrupt.
1385       JvmtiExport::post_monitor_waited(jt, this, false);
1386 
1387       // In this short circuit of the monitor wait protocol, the
1388       // current thread never drops ownership of the monitor and
1389       // never gets added to the wait queue so the current thread
1390       // cannot be made the successor. This means that the
1391       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1392       // consume an unpark() meant for the ParkEvent associated with
1393       // this ObjectMonitor.
1394     }
1395     if (event.should_commit()) {
1396       post_monitor_wait_event(&event, this, 0, millis, false);
1397     }
1398     THROW(vmSymbols::java_lang_InterruptedException());
1399     return;
1400   }
1401 
1402   assert(Self->_Stalled == 0, "invariant");
1403   Self->_Stalled = intptr_t(this);
1404   jt->set_current_waiting_monitor(this);
1405 
1406   // create a node to be put into the queue
1407   // Critically, after we reset() the event but prior to park(), we must check
1408   // for a pending interrupt.
1409   ObjectWaiter node(Self);
1410   node.TState = ObjectWaiter::TS_WAIT;
1411   Self->_ParkEvent->reset();
1412   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1413 
1414   // Enter the waiting queue, which is a circular doubly linked list in this case
1415   // but it could be a priority queue or any data structure.
1416   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1417   // by the the owner of the monitor *except* in the case where park()
1418   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1419   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1420 
1421   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1422   AddWaiter(&node);
1423   Thread::SpinRelease(&_WaitSetLock);
1424 
1425   _Responsible = NULL;
1426 
1427   intptr_t save = _recursions; // record the old recursion count
1428   _waiters++;                  // increment the number of waiters
1429   _recursions = 0;             // set the recursion level to be 1
1430   exit(true, Self);                    // exit the monitor
1431   guarantee(_owner != Self, "invariant");
1432 
1433   // The thread is on the WaitSet list - now park() it.
1434   // On MP systems it's conceivable that a brief spin before we park
1435   // could be profitable.
1436   //
1437   // TODO-FIXME: change the following logic to a loop of the form
1438   //   while (!timeout && !interrupted && _notified == 0) park()
1439 
1440   int ret = OS_OK;
1441   int WasNotified = 0;
1442   { // State transition wrappers
1443     OSThread* osthread = Self->osthread();
1444     OSThreadWaitState osts(osthread, true);
1445     {
1446       ThreadBlockInVM tbivm(jt);
1447       // Thread is in thread_blocked state and oop access is unsafe.
1448       jt->set_suspend_equivalent();
1449 
1450       if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1451         // Intentionally empty
1452       } else if (node._notified == 0) {
1453         if (millis <= 0) {
1454           Self->_ParkEvent->park();
1455         } else {
1456           ret = Self->_ParkEvent->park(millis);
1457         }
1458       }
1459 
1460       // were we externally suspended while we were waiting?
1461       if (ExitSuspendEquivalent (jt)) {
1462         // TODO-FIXME: add -- if succ == Self then succ = null.
1463         jt->java_suspend_self();
1464       }
1465 
1466     } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1467 
1468     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1469     // from the WaitSet to the EntryList.
1470     // See if we need to remove Node from the WaitSet.
1471     // We use double-checked locking to avoid grabbing _WaitSetLock
1472     // if the thread is not on the wait queue.
1473     //
1474     // Note that we don't need a fence before the fetch of TState.
1475     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1476     // written by the is thread. (perhaps the fetch might even be satisfied
1477     // by a look-aside into the processor's own store buffer, although given
1478     // the length of the code path between the prior ST and this load that's
1479     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1480     // then we'll acquire the lock and then re-fetch a fresh TState value.
1481     // That is, we fail toward safety.
1482 
1483     if (node.TState == ObjectWaiter::TS_WAIT) {
1484       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1485       if (node.TState == ObjectWaiter::TS_WAIT) {
1486         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1487         assert(node._notified == 0, "invariant");
1488         node.TState = ObjectWaiter::TS_RUN;
1489       }
1490       Thread::SpinRelease(&_WaitSetLock);
1491     }
1492 
1493     // The thread is now either on off-list (TS_RUN),
1494     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1495     // The Node's TState variable is stable from the perspective of this thread.
1496     // No other threads will asynchronously modify TState.
1497     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1498     OrderAccess::loadload();
1499     if (_succ == Self) _succ = NULL;
1500     WasNotified = node._notified;
1501 
1502     // Reentry phase -- reacquire the monitor.
1503     // re-enter contended monitor after object.wait().
1504     // retain OBJECT_WAIT state until re-enter successfully completes
1505     // Thread state is thread_in_vm and oop access is again safe,
1506     // although the raw address of the object may have changed.
1507     // (Don't cache naked oops over safepoints, of course).
1508 
1509     // post monitor waited event. Note that this is past-tense, we are done waiting.
1510     if (JvmtiExport::should_post_monitor_waited()) {
1511       JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1512 
1513       if (node._notified != 0 && _succ == Self) {
1514         // In this part of the monitor wait-notify-reenter protocol it
1515         // is possible (and normal) for another thread to do a fastpath
1516         // monitor enter-exit while this thread is still trying to get
1517         // to the reenter portion of the protocol.
1518         //
1519         // The ObjectMonitor was notified and the current thread is
1520         // the successor which also means that an unpark() has already
1521         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1522         // consume the unpark() that was done when the successor was
1523         // set because the same ParkEvent is shared between Java
1524         // monitors and JVM/TI RawMonitors (for now).
1525         //
1526         // We redo the unpark() to ensure forward progress, i.e., we
1527         // don't want all pending threads hanging (parked) with none
1528         // entering the unlocked monitor.
1529         node._event->unpark();
1530       }
1531     }
1532 
1533     if (event.should_commit()) {
1534       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1535     }
1536 
1537     OrderAccess::fence();
1538 
1539     assert(Self->_Stalled != 0, "invariant");
1540     Self->_Stalled = 0;
1541 
1542     assert(_owner != Self, "invariant");
1543     ObjectWaiter::TStates v = node.TState;
1544     if (v == ObjectWaiter::TS_RUN) {
1545       enter(Self);
1546     } else {
1547       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1548       ReenterI(Self, &node);
1549       node.wait_reenter_end(this);
1550     }
1551 
1552     // Self has reacquired the lock.
1553     // Lifecycle - the node representing Self must not appear on any queues.
1554     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1555     // want residual elements associated with this thread left on any lists.
1556     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1557     assert(_owner == Self, "invariant");
1558     assert(_succ != Self, "invariant");
1559   } // OSThreadWaitState()
1560 
1561   jt->set_current_waiting_monitor(NULL);
1562 
1563   guarantee(_recursions == 0, "invariant");
1564   _recursions = save;     // restore the old recursion count
1565   _waiters--;             // decrement the number of waiters
1566 
1567   // Verify a few postconditions
1568   assert(_owner == Self, "invariant");
1569   assert(_succ != Self, "invariant");
1570   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1571 
1572   // check if the notification happened
1573   if (!WasNotified) {
1574     // no, it could be timeout or Thread.interrupt() or both
1575     // check for interrupt event, otherwise it is timeout
1576     if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1577       THROW(vmSymbols::java_lang_InterruptedException());
1578     }
1579   }
1580 
1581   // NOTE: Spurious wake up will be consider as timeout.
1582   // Monitor notify has precedence over thread interrupt.
1583 }
1584 
1585 
1586 // Consider:
1587 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1588 // then instead of transferring a thread from the WaitSet to the EntryList
1589 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1590 
1591 void ObjectMonitor::INotify(Thread * Self) {
1592   const int policy = Knob_MoveNotifyee;
1593 
1594   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1595   ObjectWaiter * iterator = DequeueWaiter();
1596   if (iterator != NULL) {
1597     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1598     guarantee(iterator->_notified == 0, "invariant");
1599     // Disposition - what might we do with iterator ?
1600     // a.  add it directly to the EntryList - either tail (policy == 1)
1601     //     or head (policy == 0).
1602     // b.  push it onto the front of the _cxq (policy == 2).
1603     // For now we use (b).
1604     if (policy != 4) {
1605       iterator->TState = ObjectWaiter::TS_ENTER;
1606     }
1607     iterator->_notified = 1;
1608     iterator->_notifier_tid = JFR_THREAD_ID(Self);
1609 
1610     ObjectWaiter * list = _EntryList;
1611     if (list != NULL) {
1612       assert(list->_prev == NULL, "invariant");
1613       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1614       assert(list != iterator, "invariant");
1615     }
1616 
1617     if (policy == 0) {       // prepend to EntryList
1618       if (list == NULL) {
1619         iterator->_next = iterator->_prev = NULL;
1620         _EntryList = iterator;
1621       } else {
1622         list->_prev = iterator;
1623         iterator->_next = list;
1624         iterator->_prev = NULL;
1625         _EntryList = iterator;
1626       }
1627     } else if (policy == 1) {      // append to EntryList
1628       if (list == NULL) {
1629         iterator->_next = iterator->_prev = NULL;
1630         _EntryList = iterator;
1631       } else {
1632         // CONSIDER:  finding the tail currently requires a linear-time walk of
1633         // the EntryList.  We can make tail access constant-time by converting to
1634         // a CDLL instead of using our current DLL.
1635         ObjectWaiter * tail;
1636         for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1637         assert(tail != NULL && tail->_next == NULL, "invariant");
1638         tail->_next = iterator;
1639         iterator->_prev = tail;
1640         iterator->_next = NULL;
1641       }
1642     } else if (policy == 2) {      // prepend to cxq
1643       if (list == NULL) {
1644         iterator->_next = iterator->_prev = NULL;
1645         _EntryList = iterator;
1646       } else {
1647         iterator->TState = ObjectWaiter::TS_CXQ;
1648         for (;;) {
1649           ObjectWaiter * front = _cxq;
1650           iterator->_next = front;
1651           if (Atomic::cmpxchg(iterator, &_cxq, front) == front) {
1652             break;
1653           }
1654         }
1655       }
1656     } else if (policy == 3) {      // append to cxq
1657       iterator->TState = ObjectWaiter::TS_CXQ;
1658       for (;;) {
1659         ObjectWaiter * tail = _cxq;
1660         if (tail == NULL) {
1661           iterator->_next = NULL;
1662           if (Atomic::replace_if_null(iterator, &_cxq)) {
1663             break;
1664           }
1665         } else {
1666           while (tail->_next != NULL) tail = tail->_next;
1667           tail->_next = iterator;
1668           iterator->_prev = tail;
1669           iterator->_next = NULL;
1670           break;
1671         }
1672       }
1673     } else {
1674       ParkEvent * ev = iterator->_event;
1675       iterator->TState = ObjectWaiter::TS_RUN;
1676       OrderAccess::fence();
1677       ev->unpark();
1678     }
1679 
1680     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1681     // move the add-to-EntryList operation, above, outside the critical section
1682     // protected by _WaitSetLock.  In practice that's not useful.  With the
1683     // exception of  wait() timeouts and interrupts the monitor owner
1684     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1685     // on _WaitSetLock so it's not profitable to reduce the length of the
1686     // critical section.
1687 
1688     if (policy < 4) {
1689       iterator->wait_reenter_begin(this);
1690     }
1691   }
1692   Thread::SpinRelease(&_WaitSetLock);
1693 }
1694 
1695 // Consider: a not-uncommon synchronization bug is to use notify() when
1696 // notifyAll() is more appropriate, potentially resulting in stranded
1697 // threads; this is one example of a lost wakeup. A useful diagnostic
1698 // option is to force all notify() operations to behave as notifyAll().
1699 //
1700 // Note: We can also detect many such problems with a "minimum wait".
1701 // When the "minimum wait" is set to a small non-zero timeout value
1702 // and the program does not hang whereas it did absent "minimum wait",
1703 // that suggests a lost wakeup bug.
1704 
1705 void ObjectMonitor::notify(TRAPS) {
1706   CHECK_OWNER();
1707   if (_WaitSet == NULL) {
1708     return;
1709   }
1710   DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1711   INotify(THREAD);
1712   OM_PERFDATA_OP(Notifications, inc(1));
1713 }
1714 
1715 
1716 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1717 // from the waitset to the EntryList. This could be done more efficiently with a
1718 // single bulk transfer but in practice it's not time-critical. Beware too,
1719 // that in prepend-mode we invert the order of the waiters. Let's say that the
1720 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1721 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1722 
1723 void ObjectMonitor::notifyAll(TRAPS) {
1724   CHECK_OWNER();
1725   if (_WaitSet == NULL) {
1726     return;
1727   }
1728 
1729   DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1730   int tally = 0;
1731   while (_WaitSet != NULL) {
1732     tally++;
1733     INotify(THREAD);
1734   }
1735 
1736   OM_PERFDATA_OP(Notifications, inc(tally));
1737 }
1738 
1739 // -----------------------------------------------------------------------------
1740 // Adaptive Spinning Support
1741 //
1742 // Adaptive spin-then-block - rational spinning
1743 //
1744 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1745 // algorithm.  On high order SMP systems it would be better to start with
1746 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1747 // a contending thread could enqueue itself on the cxq and then spin locally
1748 // on a thread-specific variable such as its ParkEvent._Event flag.
1749 // That's left as an exercise for the reader.  Note that global spinning is
1750 // not problematic on Niagara, as the L2 cache serves the interconnect and
1751 // has both low latency and massive bandwidth.
1752 //
1753 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1754 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1755 // (duration) or we can fix the count at approximately the duration of
1756 // a context switch and vary the frequency.   Of course we could also
1757 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1758 // For a description of 'Adaptive spin-then-block mutual exclusion in
1759 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1760 //
1761 // This implementation varies the duration "D", where D varies with
1762 // the success rate of recent spin attempts. (D is capped at approximately
1763 // length of a round-trip context switch).  The success rate for recent
1764 // spin attempts is a good predictor of the success rate of future spin
1765 // attempts.  The mechanism adapts automatically to varying critical
1766 // section length (lock modality), system load and degree of parallelism.
1767 // D is maintained per-monitor in _SpinDuration and is initialized
1768 // optimistically.  Spin frequency is fixed at 100%.
1769 //
1770 // Note that _SpinDuration is volatile, but we update it without locks
1771 // or atomics.  The code is designed so that _SpinDuration stays within
1772 // a reasonable range even in the presence of races.  The arithmetic
1773 // operations on _SpinDuration are closed over the domain of legal values,
1774 // so at worst a race will install and older but still legal value.
1775 // At the very worst this introduces some apparent non-determinism.
1776 // We might spin when we shouldn't or vice-versa, but since the spin
1777 // count are relatively short, even in the worst case, the effect is harmless.
1778 //
1779 // Care must be taken that a low "D" value does not become an
1780 // an absorbing state.  Transient spinning failures -- when spinning
1781 // is overall profitable -- should not cause the system to converge
1782 // on low "D" values.  We want spinning to be stable and predictable
1783 // and fairly responsive to change and at the same time we don't want
1784 // it to oscillate, become metastable, be "too" non-deterministic,
1785 // or converge on or enter undesirable stable absorbing states.
1786 //
1787 // We implement a feedback-based control system -- using past behavior
1788 // to predict future behavior.  We face two issues: (a) if the
1789 // input signal is random then the spin predictor won't provide optimal
1790 // results, and (b) if the signal frequency is too high then the control
1791 // system, which has some natural response lag, will "chase" the signal.
1792 // (b) can arise from multimodal lock hold times.  Transient preemption
1793 // can also result in apparent bimodal lock hold times.
1794 // Although sub-optimal, neither condition is particularly harmful, as
1795 // in the worst-case we'll spin when we shouldn't or vice-versa.
1796 // The maximum spin duration is rather short so the failure modes aren't bad.
1797 // To be conservative, I've tuned the gain in system to bias toward
1798 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1799 // "rings" or oscillates between spinning and not spinning.  This happens
1800 // when spinning is just on the cusp of profitability, however, so the
1801 // situation is not dire.  The state is benign -- there's no need to add
1802 // hysteresis control to damp the transition rate between spinning and
1803 // not spinning.
1804 
1805 // Spinning: Fixed frequency (100%), vary duration
1806 int ObjectMonitor::TrySpin(Thread * Self) {
1807   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1808   int ctr = Knob_FixedSpin;
1809   if (ctr != 0) {
1810     while (--ctr >= 0) {
1811       if (TryLock(Self) > 0) return 1;
1812       SpinPause();
1813     }
1814     return 0;
1815   }
1816 
1817   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1818     if (TryLock(Self) > 0) {
1819       // Increase _SpinDuration ...
1820       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1821       // Raising _SpurDuration to the poverty line is key.
1822       int x = _SpinDuration;
1823       if (x < Knob_SpinLimit) {
1824         if (x < Knob_Poverty) x = Knob_Poverty;
1825         _SpinDuration = x + Knob_BonusB;
1826       }
1827       return 1;
1828     }
1829     SpinPause();
1830   }
1831 
1832   // Admission control - verify preconditions for spinning
1833   //
1834   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1835   // becoming an absorbing state.  Put another way, we spin briefly to
1836   // sample, just in case the system load, parallelism, contention, or lock
1837   // modality changed.
1838   //
1839   // Consider the following alternative:
1840   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1841   // spin attempt.  "Periodically" might mean after a tally of
1842   // the # of failed spin attempts (or iterations) reaches some threshold.
1843   // This takes us into the realm of 1-out-of-N spinning, where we
1844   // hold the duration constant but vary the frequency.
1845 
1846   ctr = _SpinDuration;
1847   if (ctr <= 0) return 0;
1848 
1849   if (Knob_SuccRestrict && _succ != NULL) return 0;
1850   if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1851     return 0;
1852   }
1853 
1854   int MaxSpin = Knob_MaxSpinners;
1855   if (MaxSpin >= 0) {
1856     if (_Spinner > MaxSpin) {
1857       return 0;
1858     }
1859     // Slightly racy, but benign ...
1860     Adjust(&_Spinner, 1);
1861   }
1862 
1863   // We're good to spin ... spin ingress.
1864   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1865   // when preparing to LD...CAS _owner, etc and the CAS is likely
1866   // to succeed.
1867   int oxpty   = Knob_OXPenalty;
1868   int sss     = Knob_SpinSetSucc;
1869   if (sss && _succ == NULL) _succ = Self;
1870   Thread * prv = NULL;
1871 
1872   // There are three ways to exit the following loop:
1873   // 1.  A successful spin where this thread has acquired the lock.
1874   // 2.  Spin failure with prejudice
1875   // 3.  Spin failure without prejudice
1876 
1877   while (--ctr >= 0) {
1878 
1879     // Periodic polling -- Check for pending GC
1880     // Threads may spin while they're unsafe.
1881     // We don't want spinning threads to delay the JVM from reaching
1882     // a stop-the-world safepoint or to steal cycles from GC.
1883     // If we detect a pending safepoint we abort in order that
1884     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1885     // this thread, if safe, doesn't steal cycles from GC.
1886     // This is in keeping with the "no loitering in runtime" rule.
1887     // We periodically check to see if there's a safepoint pending.
1888     if ((ctr & 0xFF) == 0) {
1889       if (SafepointMechanism::poll(Self)) {
1890         goto Abort;           // abrupt spin egress
1891       }
1892       if (Knob_UsePause & 1) SpinPause();
1893     }
1894 
1895     if (Knob_UsePause & 2) SpinPause();
1896 
1897     // Probe _owner with TATAS
1898     // If this thread observes the monitor transition or flicker
1899     // from locked to unlocked to locked, then the odds that this
1900     // thread will acquire the lock in this spin attempt go down
1901     // considerably.  The same argument applies if the CAS fails
1902     // or if we observe _owner change from one non-null value to
1903     // another non-null value.   In such cases we might abort
1904     // the spin without prejudice or apply a "penalty" to the
1905     // spin count-down variable "ctr", reducing it by 100, say.
1906 
1907     Thread * ox = (Thread *) _owner;
1908     if (ox == NULL) {
1909       ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL);
1910       if (ox == NULL) {
1911         // The CAS succeeded -- this thread acquired ownership
1912         // Take care of some bookkeeping to exit spin state.
1913         if (sss && _succ == Self) {
1914           _succ = NULL;
1915         }
1916         if (MaxSpin > 0) Adjust(&_Spinner, -1);
1917 
1918         // Increase _SpinDuration :
1919         // The spin was successful (profitable) so we tend toward
1920         // longer spin attempts in the future.
1921         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1922         // If we acquired the lock early in the spin cycle it
1923         // makes sense to increase _SpinDuration proportionally.
1924         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1925         int x = _SpinDuration;
1926         if (x < Knob_SpinLimit) {
1927           if (x < Knob_Poverty) x = Knob_Poverty;
1928           _SpinDuration = x + Knob_Bonus;
1929         }
1930         return 1;
1931       }
1932 
1933       // The CAS failed ... we can take any of the following actions:
1934       // * penalize: ctr -= CASPenalty
1935       // * exit spin with prejudice -- goto Abort;
1936       // * exit spin without prejudice.
1937       // * Since CAS is high-latency, retry again immediately.
1938       prv = ox;
1939       goto Abort;
1940     }
1941 
1942     // Did lock ownership change hands ?
1943     if (ox != prv && prv != NULL) {
1944       if (oxpty == -2) break;
1945       if (oxpty == -1) goto Abort;
1946       ctr -= oxpty;
1947     }
1948     prv = ox;
1949 
1950     // Abort the spin if the owner is not executing.
1951     // The owner must be executing in order to drop the lock.
1952     // Spinning while the owner is OFFPROC is idiocy.
1953     // Consider: ctr -= RunnablePenalty ;
1954     if (Knob_OState && NotRunnable (Self, ox)) {
1955       goto Abort;
1956     }
1957     if (sss && _succ == NULL) _succ = Self;
1958   }
1959 
1960   // Spin failed with prejudice -- reduce _SpinDuration.
1961   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1962   // AIMD is globally stable.
1963   {
1964     int x = _SpinDuration;
1965     if (x > 0) {
1966       // Consider an AIMD scheme like: x -= (x >> 3) + 100
1967       // This is globally sample and tends to damp the response.
1968       x -= Knob_Penalty;
1969       if (x < 0) x = 0;
1970       _SpinDuration = x;
1971     }
1972   }
1973 
1974  Abort:
1975   if (MaxSpin >= 0) Adjust(&_Spinner, -1);
1976   if (sss && _succ == Self) {
1977     _succ = NULL;
1978     // Invariant: after setting succ=null a contending thread
1979     // must recheck-retry _owner before parking.  This usually happens
1980     // in the normal usage of TrySpin(), but it's safest
1981     // to make TrySpin() as foolproof as possible.
1982     OrderAccess::fence();
1983     if (TryLock(Self) > 0) return 1;
1984   }
1985   return 0;
1986 }
1987 
1988 // NotRunnable() -- informed spinning
1989 //
1990 // Don't bother spinning if the owner is not eligible to drop the lock.
1991 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
1992 // The thread must be runnable in order to drop the lock in timely fashion.
1993 // If the _owner is not runnable then spinning will not likely be
1994 // successful (profitable).
1995 //
1996 // Beware -- the thread referenced by _owner could have died
1997 // so a simply fetch from _owner->_thread_state might trap.
1998 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
1999 // Because of the lifecycle issues, the _thread_state values
2000 // observed by NotRunnable() might be garbage.  NotRunnable must
2001 // tolerate this and consider the observed _thread_state value
2002 // as advisory.
2003 //
2004 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2005 // a thread pointer.
2006 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2007 // with the LSB of _owner.  Another option would be to probabilistically probe
2008 // the putative _owner->TypeTag value.
2009 //
2010 // Checking _thread_state isn't perfect.  Even if the thread is
2011 // in_java it might be blocked on a page-fault or have been preempted
2012 // and sitting on a ready/dispatch queue.
2013 //
2014 // The return value from NotRunnable() is *advisory* -- the
2015 // result is based on sampling and is not necessarily coherent.
2016 // The caller must tolerate false-negative and false-positive errors.
2017 // Spinning, in general, is probabilistic anyway.
2018 
2019 
2020 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2021   // Check ox->TypeTag == 2BAD.
2022   if (ox == NULL) return 0;
2023 
2024   // Avoid transitive spinning ...
2025   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2026   // Immediately after T1 acquires L it's possible that T2, also
2027   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2028   // This occurs transiently after T1 acquired L but before
2029   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2030   // its spin in this circumstance.
2031   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2032 
2033   if (BlockedOn == 1) return 1;
2034   if (BlockedOn != 0) {
2035     return BlockedOn != intptr_t(this) && _owner == ox;
2036   }
2037 
2038   assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2039   int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2040   // consider also: jst != _thread_in_Java -- but that's overspecific.
2041   return jst == _thread_blocked || jst == _thread_in_native;
2042 }
2043 
2044 
2045 // -----------------------------------------------------------------------------
2046 // WaitSet management ...
2047 
2048 ObjectWaiter::ObjectWaiter(Thread* thread) {
2049   _next     = NULL;
2050   _prev     = NULL;
2051   _notified = 0;
2052   _notifier_tid = 0;
2053   TState    = TS_RUN;
2054   _thread   = thread;
2055   _event    = thread->_ParkEvent;
2056   _active   = false;
2057   assert(_event != NULL, "invariant");
2058 }
2059 
2060 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2061   JavaThread *jt = (JavaThread *)this->_thread;
2062   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2063 }
2064 
2065 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2066   JavaThread *jt = (JavaThread *)this->_thread;
2067   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2068 }
2069 
2070 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2071   assert(node != NULL, "should not add NULL node");
2072   assert(node->_prev == NULL, "node already in list");
2073   assert(node->_next == NULL, "node already in list");
2074   // put node at end of queue (circular doubly linked list)
2075   if (_WaitSet == NULL) {
2076     _WaitSet = node;
2077     node->_prev = node;
2078     node->_next = node;
2079   } else {
2080     ObjectWaiter* head = _WaitSet;
2081     ObjectWaiter* tail = head->_prev;
2082     assert(tail->_next == head, "invariant check");
2083     tail->_next = node;
2084     head->_prev = node;
2085     node->_next = head;
2086     node->_prev = tail;
2087   }
2088 }
2089 
2090 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2091   // dequeue the very first waiter
2092   ObjectWaiter* waiter = _WaitSet;
2093   if (waiter) {
2094     DequeueSpecificWaiter(waiter);
2095   }
2096   return waiter;
2097 }
2098 
2099 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2100   assert(node != NULL, "should not dequeue NULL node");
2101   assert(node->_prev != NULL, "node already removed from list");
2102   assert(node->_next != NULL, "node already removed from list");
2103   // when the waiter has woken up because of interrupt,
2104   // timeout or other spurious wake-up, dequeue the
2105   // waiter from waiting list
2106   ObjectWaiter* next = node->_next;
2107   if (next == node) {
2108     assert(node->_prev == node, "invariant check");
2109     _WaitSet = NULL;
2110   } else {
2111     ObjectWaiter* prev = node->_prev;
2112     assert(prev->_next == node, "invariant check");
2113     assert(next->_prev == node, "invariant check");
2114     next->_prev = prev;
2115     prev->_next = next;
2116     if (_WaitSet == node) {
2117       _WaitSet = next;
2118     }
2119   }
2120   node->_next = NULL;
2121   node->_prev = NULL;
2122 }
2123 
2124 // -----------------------------------------------------------------------------
2125 // PerfData support
2126 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2127 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2128 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2129 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2130 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2131 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2132 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2133 
2134 // One-shot global initialization for the sync subsystem.
2135 // We could also defer initialization and initialize on-demand
2136 // the first time we call inflate().  Initialization would
2137 // be protected - like so many things - by the MonitorCache_lock.
2138 
2139 void ObjectMonitor::Initialize() {
2140   static int InitializationCompleted = 0;
2141   assert(InitializationCompleted == 0, "invariant");
2142   InitializationCompleted = 1;
2143   if (UsePerfData) {
2144     EXCEPTION_MARK;
2145 #define NEWPERFCOUNTER(n)                                                \
2146   {                                                                      \
2147     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2148                                         CHECK);                          \
2149   }
2150 #define NEWPERFVARIABLE(n)                                                \
2151   {                                                                       \
2152     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2153                                          CHECK);                          \
2154   }
2155     NEWPERFCOUNTER(_sync_Inflations);
2156     NEWPERFCOUNTER(_sync_Deflations);
2157     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2158     NEWPERFCOUNTER(_sync_FutileWakeups);
2159     NEWPERFCOUNTER(_sync_Parks);
2160     NEWPERFCOUNTER(_sync_Notifications);
2161     NEWPERFVARIABLE(_sync_MonExtant);
2162 #undef NEWPERFCOUNTER
2163 #undef NEWPERFVARIABLE
2164   }
2165 }
2166 
2167 void ObjectMonitor::DeferredInitialize() {
2168   if (InitDone > 0) return;
2169   if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2170     while (InitDone != 1) /* empty */;
2171     return;
2172   }
2173 
2174   // One-shot global initialization ...
2175   // The initialization is idempotent, so we don't need locks.
2176   // In the future consider doing this via os::init_2().
2177 
2178   if (!os::is_MP()) {
2179     Knob_SpinLimit = 0;
2180     Knob_PreSpin   = 0;
2181     Knob_FixedSpin = -1;
2182   }
2183 
2184   OrderAccess::fence();
2185   InitDone = 1;
2186 }
2187