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