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_ExitRelease  = 0;
 105 int ObjectMonitor::Knob_InlineNotify = 1;
 106 int ObjectMonitor::Knob_Verbose      = 0;
 107 int ObjectMonitor::Knob_VerifyInUse  = 0;
 108 int ObjectMonitor::Knob_VerifyMatch  = 0;
 109 int ObjectMonitor::Knob_SpinLimit    = 5000;    // derived by an external tool -
 110 
 111 static int Knob_SpinBase            = 0;       // Floor AKA SpinMin
 112 static int Knob_CASPenalty          = -1;      // Penalty for failed CAS
 113 static int Knob_OXPenalty           = -1;      // Penalty for observed _owner change
 114 static int Knob_SpinSetSucc         = 1;       // spinners set the _succ field
 115 static int Knob_SpinEarly           = 1;
 116 static int Knob_SuccEnabled         = 1;       // futile wake throttling
 117 static int Knob_SuccRestrict        = 0;       // Limit successors + spinners to at-most-one
 118 static int Knob_MaxSpinners         = -1;      // Should be a function of # CPUs
 119 static int Knob_Bonus               = 100;     // spin success bonus
 120 static int Knob_BonusB              = 100;     // spin success bonus
 121 static int Knob_Penalty             = 200;     // spin failure penalty
 122 static int Knob_Poverty             = 1000;
 123 static int Knob_SpinAfterFutile     = 1;       // Spin after returning from park()
 124 static int Knob_FixedSpin           = 0;
 125 static int Knob_OState              = 3;       // Spinner checks thread state of _owner
 126 static int Knob_UsePause            = 1;
 127 static int Knob_ExitPolicy          = 0;
 128 static int Knob_PreSpin             = 10;      // 20-100 likely better
 129 static int Knob_ResetEvent          = 0;
 130 
 131 static int Knob_FastHSSEC           = 0;
 132 static int Knob_MoveNotifyee        = 2;       // notify() - disposition of notifyee
 133 static int Knob_QMode               = 0;       // EntryList-cxq policy - queue discipline
 134 static volatile int InitDone        = 0;
 135 
 136 // -----------------------------------------------------------------------------
 137 // Theory of operations -- Monitors lists, thread residency, etc:
 138 //
 139 // * A thread acquires ownership of a monitor by successfully
 140 //   CAS()ing the _owner field from null to non-null.
 141 //
 142 // * Invariant: A thread appears on at most one monitor list --
 143 //   cxq, EntryList or WaitSet -- at any one time.
 144 //
 145 // * Contending threads "push" themselves onto the cxq with CAS
 146 //   and then spin/park.
 147 //
 148 // * After a contending thread eventually acquires the lock it must
 149 //   dequeue itself from either the EntryList or the cxq.
 150 //
 151 // * The exiting thread identifies and unparks an "heir presumptive"
 152 //   tentative successor thread on the EntryList.  Critically, the
 153 //   exiting thread doesn't unlink the successor thread from the EntryList.
 154 //   After having been unparked, the wakee will recontend for ownership of
 155 //   the monitor.   The successor (wakee) will either acquire the lock or
 156 //   re-park itself.
 157 //
 158 //   Succession is provided for by a policy of competitive handoff.
 159 //   The exiting thread does _not_ grant or pass ownership to the
 160 //   successor thread.  (This is also referred to as "handoff" succession").
 161 //   Instead the exiting thread releases ownership and possibly wakes
 162 //   a successor, so the successor can (re)compete for ownership of the lock.
 163 //   If the EntryList is empty but the cxq is populated the exiting
 164 //   thread will drain the cxq into the EntryList.  It does so by
 165 //   by detaching the cxq (installing null with CAS) and folding
 166 //   the threads from the cxq into the EntryList.  The EntryList is
 167 //   doubly linked, while the cxq is singly linked because of the
 168 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 169 //
 170 // * Concurrency invariants:
 171 //
 172 //   -- only the monitor owner may access or mutate the EntryList.
 173 //      The mutex property of the monitor itself protects the EntryList
 174 //      from concurrent interference.
 175 //   -- Only the monitor owner may detach the cxq.
 176 //
 177 // * The monitor entry list operations avoid locks, but strictly speaking
 178 //   they're not lock-free.  Enter is lock-free, exit is not.
 179 //   For a description of 'Methods and apparatus providing non-blocking access
 180 //   to a resource,' see U.S. Pat. No. 7844973.
 181 //
 182 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 183 //   detaching thread.  This mechanism is immune from the ABA corruption.
 184 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 185 //
 186 // * Taken together, the cxq and the EntryList constitute or form a
 187 //   single logical queue of threads stalled trying to acquire the lock.
 188 //   We use two distinct lists to improve the odds of a constant-time
 189 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 190 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 191 //   A key desideratum is to minimize queue & monitor metadata manipulation
 192 //   that occurs while holding the monitor lock -- that is, we want to
 193 //   minimize monitor lock holds times.  Note that even a small amount of
 194 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 195 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 196 //   locks and monitor metadata.
 197 //
 198 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 199 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 200 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 201 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 202 //
 203 //   The EntryList is ordered by the prevailing queue discipline and
 204 //   can be organized in any convenient fashion, such as a doubly-linked list or
 205 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 206 //   to operate in constant-time.  If we need a priority queue then something akin
 207 //   to Solaris' sleepq would work nicely.  Viz.,
 208 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 209 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 210 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 211 //   EntryList accordingly.
 212 //
 213 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 214 //   somewhat similar to an elevator-scan.
 215 //
 216 // * The monitor synchronization subsystem avoids the use of native
 217 //   synchronization primitives except for the narrow platform-specific
 218 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 219 //   the semantics of park-unpark.  Put another way, this monitor implementation
 220 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 221 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 222 //   underlying OS manages the READY<->RUN transitions.
 223 //
 224 // * Waiting threads reside on the WaitSet list -- wait() puts
 225 //   the caller onto the WaitSet.
 226 //
 227 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 228 //   either the EntryList or cxq.  Subsequent exit() operations will
 229 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 230 //   it's likely the notifyee would simply impale itself on the lock held
 231 //   by the notifier.
 232 //
 233 // * An interesting alternative is to encode cxq as (List,LockByte) where
 234 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 235 //   variable, like _recursions, in the scheme.  The threads or Events that form
 236 //   the list would have to be aligned in 256-byte addresses.  A thread would
 237 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 238 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 239 //   Note that is is *not* word-tearing, but it does presume that full-word
 240 //   CAS operations are coherent with intermix with STB operations.  That's true
 241 //   on most common processors.
 242 //
 243 // * See also http://blogs.sun.com/dave
 244 
 245 
 246 void* ObjectMonitor::operator new (size_t size) throw() {
 247   return AllocateHeap(size, mtInternal);
 248 }
 249 void* ObjectMonitor::operator new[] (size_t size) throw() {
 250   return operator new (size);
 251 }
 252 void ObjectMonitor::operator delete(void* p) {
 253   FreeHeap(p);
 254 }
 255 void ObjectMonitor::operator delete[] (void *p) {
 256   operator delete(p);
 257 }
 258 
 259 // -----------------------------------------------------------------------------
 260 // Enter support
 261 
 262 void ObjectMonitor::enter(TRAPS) {
 263   // The following code is ordered to check the most common cases first
 264   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 265   Thread * const Self = THREAD;
 266 
 267   void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL);
 268   if (cur == NULL) {
 269     // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
 270     assert(_recursions == 0, "invariant");
 271     assert(_owner == Self, "invariant");
 272     return;
 273   }
 274 
 275   if (cur == Self) {
 276     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 277     _recursions++;
 278     return;
 279   }
 280 
 281   if (Self->is_lock_owned ((address)cur)) {
 282     assert(_recursions == 0, "internal state error");
 283     _recursions = 1;
 284     // Commute owner from a thread-specific on-stack BasicLockObject address to
 285     // a full-fledged "Thread *".
 286     _owner = Self;
 287     return;
 288   }
 289 
 290   // We've encountered genuine contention.
 291   assert(Self->_Stalled == 0, "invariant");
 292   Self->_Stalled = intptr_t(this);
 293 
 294   // Try one round of spinning *before* enqueueing Self
 295   // and before going through the awkward and expensive state
 296   // transitions.  The following spin is strictly optional ...
 297   // Note that if we acquire the monitor from an initial spin
 298   // we forgo posting JVMTI events and firing DTRACE probes.
 299   if (Knob_SpinEarly && TrySpin (Self) > 0) {
 300     assert(_owner == Self, "invariant");
 301     assert(_recursions == 0, "invariant");
 302     assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 303     Self->_Stalled = 0;
 304     return;
 305   }
 306 
 307   assert(_owner != Self, "invariant");
 308   assert(_succ != Self, "invariant");
 309   assert(Self->is_Java_thread(), "invariant");
 310   JavaThread * jt = (JavaThread *) Self;
 311   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 312   assert(jt->thread_state() != _thread_blocked, "invariant");
 313   assert(this->object() != NULL, "invariant");
 314   assert(_count >= 0, "invariant");
 315 
 316   // Prevent deflation at STW-time.  See deflate_idle_monitors() and is_busy().
 317   // Ensure the object-monitor relationship remains stable while there's contention.
 318   Atomic::inc(&_count);
 319 
 320   JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
 321   EventJavaMonitorEnter event;
 322   if (event.should_commit()) {
 323     event.set_monitorClass(((oop)this->object())->klass());
 324     event.set_address((uintptr_t)(this->object_addr()));
 325   }
 326 
 327   { // Change java thread status to indicate blocked on monitor enter.
 328     JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
 329 
 330     Self->set_current_pending_monitor(this);
 331 
 332     DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
 333     if (JvmtiExport::should_post_monitor_contended_enter()) {
 334       JvmtiExport::post_monitor_contended_enter(jt, this);
 335 
 336       // The current thread does not yet own the monitor and does not
 337       // yet appear on any queues that would get it made the successor.
 338       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 339       // handler cannot accidentally consume an unpark() meant for the
 340       // ParkEvent associated with this ObjectMonitor.
 341     }
 342 
 343     OSThreadContendState osts(Self->osthread());
 344     ThreadBlockInVM tbivm(jt);
 345 
 346     // TODO-FIXME: change the following for(;;) loop to straight-line code.
 347     for (;;) {
 348       jt->set_suspend_equivalent();
 349       // cleared by handle_special_suspend_equivalent_condition()
 350       // or java_suspend_self()
 351 
 352       EnterI(THREAD);
 353 
 354       if (!ExitSuspendEquivalent(jt)) break;
 355 
 356       // We have acquired the contended monitor, but while we were
 357       // waiting another thread suspended us. We don't want to enter
 358       // the monitor while suspended because that would surprise the
 359       // thread that suspended us.
 360       //
 361       _recursions = 0;
 362       _succ = NULL;
 363       exit(false, Self);
 364 
 365       jt->java_suspend_self();
 366     }
 367     Self->set_current_pending_monitor(NULL);
 368 
 369     // We cleared the pending monitor info since we've just gotten past
 370     // the enter-check-for-suspend dance and we now own the monitor free
 371     // and clear, i.e., it is no longer pending. The ThreadBlockInVM
 372     // destructor can go to a safepoint at the end of this block. If we
 373     // do a thread dump during that safepoint, then this thread will show
 374     // as having "-locked" the monitor, but the OS and java.lang.Thread
 375     // states will still report that the thread is blocked trying to
 376     // acquire it.
 377   }
 378 
 379   Atomic::dec(&_count);
 380   assert(_count >= 0, "invariant");
 381   Self->_Stalled = 0;
 382 
 383   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 384   assert(_recursions == 0, "invariant");
 385   assert(_owner == Self, "invariant");
 386   assert(_succ != Self, "invariant");
 387   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 388 
 389   // The thread -- now the owner -- is back in vm mode.
 390   // Report the glorious news via TI,DTrace and jvmstat.
 391   // The probe effect is non-trivial.  All the reportage occurs
 392   // while we hold the monitor, increasing the length of the critical
 393   // section.  Amdahl's parallel speedup law comes vividly into play.
 394   //
 395   // Another option might be to aggregate the events (thread local or
 396   // per-monitor aggregation) and defer reporting until a more opportune
 397   // time -- such as next time some thread encounters contention but has
 398   // yet to acquire the lock.  While spinning that thread could
 399   // spinning we could increment JVMStat counters, etc.
 400 
 401   DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
 402   if (JvmtiExport::should_post_monitor_contended_entered()) {
 403     JvmtiExport::post_monitor_contended_entered(jt, this);
 404 
 405     // The current thread already owns the monitor and is not going to
 406     // call park() for the remainder of the monitor enter protocol. So
 407     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 408     // event handler consumed an unpark() issued by the thread that
 409     // just exited the monitor.
 410   }
 411   if (event.should_commit()) {
 412     event.set_previousOwner((uintptr_t)_previous_owner_tid);
 413     event.commit();
 414   }
 415   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 416 }
 417 
 418 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 419 // Callers must compensate as needed.
 420 
 421 int ObjectMonitor::TryLock(Thread * Self) {
 422   void * own = _owner;
 423   if (own != NULL) return 0;
 424   if (Atomic::replace_if_null(Self, &_owner)) {
 425     // Either guarantee _recursions == 0 or set _recursions = 0.
 426     assert(_recursions == 0, "invariant");
 427     assert(_owner == Self, "invariant");
 428     return 1;
 429   }
 430   // The lock had been free momentarily, but we lost the race to the lock.
 431   // Interference -- the CAS failed.
 432   // We can either return -1 or retry.
 433   // Retry doesn't make as much sense because the lock was just acquired.
 434   return -1;
 435 }
 436 
 437 #define MAX_RECHECK_INTERVAL 1000
 438 
 439 void ObjectMonitor::EnterI(TRAPS) {
 440   Thread * const Self = THREAD;
 441   assert(Self->is_Java_thread(), "invariant");
 442   assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
 443 
 444   // Try the lock - TATAS
 445   if (TryLock (Self) > 0) {
 446     assert(_succ != Self, "invariant");
 447     assert(_owner == Self, "invariant");
 448     assert(_Responsible != Self, "invariant");
 449     return;
 450   }
 451 
 452   DeferredInitialize();
 453 
 454   // We try one round of spinning *before* enqueueing Self.
 455   //
 456   // If the _owner is ready but OFFPROC we could use a YieldTo()
 457   // operation to donate the remainder of this thread's quantum
 458   // to the owner.  This has subtle but beneficial affinity
 459   // effects.
 460 
 461   if (TrySpin (Self) > 0) {
 462     assert(_owner == Self, "invariant");
 463     assert(_succ != Self, "invariant");
 464     assert(_Responsible != Self, "invariant");
 465     return;
 466   }
 467 
 468   // The Spin failed -- Enqueue and park the thread ...
 469   assert(_succ != Self, "invariant");
 470   assert(_owner != Self, "invariant");
 471   assert(_Responsible != Self, "invariant");
 472 
 473   // Enqueue "Self" on ObjectMonitor's _cxq.
 474   //
 475   // Node acts as a proxy for Self.
 476   // As an aside, if were to ever rewrite the synchronization code mostly
 477   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 478   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 479   // as well as eliminate a subset of ABA issues.
 480   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 481 
 482   ObjectWaiter node(Self);
 483   Self->_ParkEvent->reset();
 484   node._prev   = (ObjectWaiter *) 0xBAD;
 485   node.TState  = ObjectWaiter::TS_CXQ;
 486 
 487   // Push "Self" onto the front of the _cxq.
 488   // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
 489   // Note that spinning tends to reduce the rate at which threads
 490   // enqueue and dequeue on EntryList|cxq.
 491   ObjectWaiter * nxt;
 492   for (;;) {
 493     node._next = nxt = _cxq;
 494     if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break;
 495 
 496     // Interference - the CAS failed because _cxq changed.  Just retry.
 497     // As an optional optimization we retry the lock.
 498     if (TryLock (Self) > 0) {
 499       assert(_succ != Self, "invariant");
 500       assert(_owner == Self, "invariant");
 501       assert(_Responsible != Self, "invariant");
 502       return;
 503     }
 504   }
 505 
 506   // Check for cxq|EntryList edge transition to non-null.  This indicates
 507   // the onset of contention.  While contention persists exiting threads
 508   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 509   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 510   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 511   // arrange for one of the contending thread to use a timed park() operations
 512   // to detect and recover from the race.  (Stranding is form of progress failure
 513   // where the monitor is unlocked but all the contending threads remain parked).
 514   // That is, at least one of the contended threads will periodically poll _owner.
 515   // One of the contending threads will become the designated "Responsible" thread.
 516   // The Responsible thread uses a timed park instead of a normal indefinite park
 517   // operation -- it periodically wakes and checks for and recovers from potential
 518   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 519   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 520   // be responsible for a monitor.
 521   //
 522   // Currently, one of the contended threads takes on the added role of "Responsible".
 523   // A viable alternative would be to use a dedicated "stranding checker" thread
 524   // that periodically iterated over all the threads (or active monitors) and unparked
 525   // successors where there was risk of stranding.  This would help eliminate the
 526   // timer scalability issues we see on some platforms as we'd only have one thread
 527   // -- the checker -- parked on a timer.
 528 
 529   if (nxt == NULL && _EntryList == NULL) {
 530     // Try to assume the role of responsible thread for the monitor.
 531     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=Self }
 532     Atomic::replace_if_null(Self, &_Responsible);
 533   }
 534 
 535   // The lock might have been released while this thread was occupied queueing
 536   // itself onto _cxq.  To close the race and avoid "stranding" and
 537   // progress-liveness failure we must resample-retry _owner before parking.
 538   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 539   // In this case the ST-MEMBAR is accomplished with CAS().
 540   //
 541   // TODO: Defer all thread state transitions until park-time.
 542   // Since state transitions are heavy and inefficient we'd like
 543   // to defer the state transitions until absolutely necessary,
 544   // and in doing so avoid some transitions ...
 545 
 546   int nWakeups = 0;
 547   int recheckInterval = 1;
 548 
 549   for (;;) {
 550 
 551     if (TryLock(Self) > 0) break;
 552     assert(_owner != Self, "invariant");
 553 
 554     // park self
 555     if (_Responsible == Self) {
 556       Self->_ParkEvent->park((jlong) recheckInterval);
 557       // Increase the recheckInterval, but clamp the value.
 558       recheckInterval *= 8;
 559       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 560         recheckInterval = MAX_RECHECK_INTERVAL;
 561       }
 562     } else {
 563       Self->_ParkEvent->park();
 564     }
 565 
 566     if (TryLock(Self) > 0) break;
 567 
 568     // The lock is still contested.
 569     // Keep a tally of the # of futile wakeups.
 570     // Note that the counter is not protected by a lock or updated by atomics.
 571     // That is by design - we trade "lossy" counters which are exposed to
 572     // races during updates for a lower probe effect.
 573 
 574     // This PerfData object can be used in parallel with a safepoint.
 575     // See the work around in PerfDataManager::destroy().
 576     OM_PERFDATA_OP(FutileWakeups, inc());
 577     ++nWakeups;
 578 
 579     // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
 580     // We can defer clearing _succ until after the spin completes
 581     // TrySpin() must tolerate being called with _succ == Self.
 582     // Try yet another round of adaptive spinning.
 583     if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
 584 
 585     // We can find that we were unpark()ed and redesignated _succ while
 586     // we were spinning.  That's harmless.  If we iterate and call park(),
 587     // park() will consume the event and return immediately and we'll
 588     // just spin again.  This pattern can repeat, leaving _succ to simply
 589     // spin on a CPU.  Enable Knob_ResetEvent to clear pending unparks().
 590     // Alternately, we can sample fired() here, and if set, forgo spinning
 591     // in the next iteration.
 592 
 593     if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
 594       Self->_ParkEvent->reset();
 595       OrderAccess::fence();
 596     }
 597     if (_succ == Self) _succ = NULL;
 598 
 599     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 600     OrderAccess::fence();
 601   }
 602 
 603   // Egress :
 604   // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
 605   // Normally we'll find Self on the EntryList .
 606   // From the perspective of the lock owner (this thread), the
 607   // EntryList is stable and cxq is prepend-only.
 608   // The head of cxq is volatile but the interior is stable.
 609   // In addition, Self.TState is stable.
 610 
 611   assert(_owner == Self, "invariant");
 612   assert(object() != NULL, "invariant");
 613   // I'd like to write:
 614   //   guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 615   // but as we're at a safepoint that's not safe.
 616 
 617   UnlinkAfterAcquire(Self, &node);
 618   if (_succ == Self) _succ = NULL;
 619 
 620   assert(_succ != Self, "invariant");
 621   if (_Responsible == Self) {
 622     _Responsible = NULL;
 623     OrderAccess::fence(); // Dekker pivot-point
 624 
 625     // We may leave threads on cxq|EntryList without a designated
 626     // "Responsible" thread.  This is benign.  When this thread subsequently
 627     // exits the monitor it can "see" such preexisting "old" threads --
 628     // threads that arrived on the cxq|EntryList before the fence, above --
 629     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 630     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 631     // non-null and elect a new "Responsible" timer thread.
 632     //
 633     // This thread executes:
 634     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 635     //    LD cxq|EntryList               (in subsequent exit)
 636     //
 637     // Entering threads in the slow/contended path execute:
 638     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 639     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 640     //
 641     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 642     // exit operation from floating above the ST Responsible=null.
 643   }
 644 
 645   // We've acquired ownership with CAS().
 646   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 647   // But since the CAS() this thread may have also stored into _succ,
 648   // EntryList, cxq or Responsible.  These meta-data updates must be
 649   // visible __before this thread subsequently drops the lock.
 650   // Consider what could occur if we didn't enforce this constraint --
 651   // STs to monitor meta-data and user-data could reorder with (become
 652   // visible after) the ST in exit that drops ownership of the lock.
 653   // Some other thread could then acquire the lock, but observe inconsistent
 654   // or old monitor meta-data and heap data.  That violates the JMM.
 655   // To that end, the 1-0 exit() operation must have at least STST|LDST
 656   // "release" barrier semantics.  Specifically, there must be at least a
 657   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 658   // the lock.   The barrier ensures that changes to monitor meta-data and data
 659   // protected by the lock will be visible before we release the lock, and
 660   // therefore before some other thread (CPU) has a chance to acquire the lock.
 661   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 662   //
 663   // Critically, any prior STs to _succ or EntryList must be visible before
 664   // the ST of null into _owner in the *subsequent* (following) corresponding
 665   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 666   // execute a serializing instruction.
 667 
 668   return;
 669 }
 670 
 671 // ReenterI() is a specialized inline form of the latter half of the
 672 // contended slow-path from EnterI().  We use ReenterI() only for
 673 // monitor reentry in wait().
 674 //
 675 // In the future we should reconcile EnterI() and ReenterI(), adding
 676 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
 677 // loop accordingly.
 678 
 679 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
 680   assert(Self != NULL, "invariant");
 681   assert(SelfNode != NULL, "invariant");
 682   assert(SelfNode->_thread == Self, "invariant");
 683   assert(_waiters > 0, "invariant");
 684   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 685   assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
 686   JavaThread * jt = (JavaThread *) Self;
 687 
 688   int nWakeups = 0;
 689   for (;;) {
 690     ObjectWaiter::TStates v = SelfNode->TState;
 691     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
 692     assert(_owner != Self, "invariant");
 693 
 694     if (TryLock(Self) > 0) break;
 695     if (TrySpin(Self) > 0) break;
 696 
 697     // State transition wrappers around park() ...
 698     // ReenterI() wisely defers state transitions until
 699     // it's clear we must park the thread.
 700     {
 701       OSThreadContendState osts(Self->osthread());
 702       ThreadBlockInVM tbivm(jt);
 703 
 704       // cleared by handle_special_suspend_equivalent_condition()
 705       // or java_suspend_self()
 706       jt->set_suspend_equivalent();
 707       Self->_ParkEvent->park();
 708 
 709       // were we externally suspended while we were waiting?
 710       for (;;) {
 711         if (!ExitSuspendEquivalent(jt)) break;
 712         if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
 713         jt->java_suspend_self();
 714         jt->set_suspend_equivalent();
 715       }
 716     }
 717 
 718     // Try again, but just so we distinguish between futile wakeups and
 719     // successful wakeups.  The following test isn't algorithmically
 720     // necessary, but it helps us maintain sensible statistics.
 721     if (TryLock(Self) > 0) break;
 722 
 723     // The lock is still contested.
 724     // Keep a tally of the # of futile wakeups.
 725     // Note that the counter is not protected by a lock or updated by atomics.
 726     // That is by design - we trade "lossy" counters which are exposed to
 727     // races during updates for a lower probe effect.
 728     ++nWakeups;
 729 
 730     // Assuming this is not a spurious wakeup we'll normally
 731     // find that _succ == Self.
 732     if (_succ == Self) _succ = NULL;
 733 
 734     // Invariant: after clearing _succ a contending thread
 735     // *must* retry  _owner before parking.
 736     OrderAccess::fence();
 737 
 738     // This PerfData object can be used in parallel with a safepoint.
 739     // See the work around in PerfDataManager::destroy().
 740     OM_PERFDATA_OP(FutileWakeups, inc());
 741   }
 742 
 743   // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
 744   // Normally we'll find Self on the EntryList.
 745   // Unlinking from the EntryList is constant-time and atomic-free.
 746   // From the perspective of the lock owner (this thread), the
 747   // EntryList is stable and cxq is prepend-only.
 748   // The head of cxq is volatile but the interior is stable.
 749   // In addition, Self.TState is stable.
 750 
 751   assert(_owner == Self, "invariant");
 752   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 753   UnlinkAfterAcquire(Self, SelfNode);
 754   if (_succ == Self) _succ = NULL;
 755   assert(_succ != Self, "invariant");
 756   SelfNode->TState = ObjectWaiter::TS_RUN;
 757   OrderAccess::fence();      // see comments at the end of EnterI()
 758 }
 759 
 760 // By convention we unlink a contending thread from EntryList|cxq immediately
 761 // after the thread acquires the lock in ::enter().  Equally, we could defer
 762 // unlinking the thread until ::exit()-time.
 763 
 764 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
 765   assert(_owner == Self, "invariant");
 766   assert(SelfNode->_thread == Self, "invariant");
 767 
 768   if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
 769     // Normal case: remove Self from the DLL EntryList .
 770     // This is a constant-time operation.
 771     ObjectWaiter * nxt = SelfNode->_next;
 772     ObjectWaiter * prv = SelfNode->_prev;
 773     if (nxt != NULL) nxt->_prev = prv;
 774     if (prv != NULL) prv->_next = nxt;
 775     if (SelfNode == _EntryList) _EntryList = nxt;
 776     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
 777     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
 778   } else {
 779     assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
 780     // Inopportune interleaving -- Self is still on the cxq.
 781     // This usually means the enqueue of self raced an exiting thread.
 782     // Normally we'll find Self near the front of the cxq, so
 783     // dequeueing is typically fast.  If needbe we can accelerate
 784     // this with some MCS/CHL-like bidirectional list hints and advisory
 785     // back-links so dequeueing from the interior will normally operate
 786     // in constant-time.
 787     // Dequeue Self from either the head (with CAS) or from the interior
 788     // with a linear-time scan and normal non-atomic memory operations.
 789     // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
 790     // and then unlink Self from EntryList.  We have to drain eventually,
 791     // so it might as well be now.
 792 
 793     ObjectWaiter * v = _cxq;
 794     assert(v != NULL, "invariant");
 795     if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) {
 796       // The CAS above can fail from interference IFF a "RAT" arrived.
 797       // In that case Self must be in the interior and can no longer be
 798       // at the head of cxq.
 799       if (v == SelfNode) {
 800         assert(_cxq != v, "invariant");
 801         v = _cxq;          // CAS above failed - start scan at head of list
 802       }
 803       ObjectWaiter * p;
 804       ObjectWaiter * q = NULL;
 805       for (p = v; p != NULL && p != SelfNode; p = p->_next) {
 806         q = p;
 807         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
 808       }
 809       assert(v != SelfNode, "invariant");
 810       assert(p == SelfNode, "Node not found on cxq");
 811       assert(p != _cxq, "invariant");
 812       assert(q != NULL, "invariant");
 813       assert(q->_next == p, "invariant");
 814       q->_next = p->_next;
 815     }
 816   }
 817 
 818 #ifdef ASSERT
 819   // Diagnostic hygiene ...
 820   SelfNode->_prev  = (ObjectWaiter *) 0xBAD;
 821   SelfNode->_next  = (ObjectWaiter *) 0xBAD;
 822   SelfNode->TState = ObjectWaiter::TS_RUN;
 823 #endif
 824 }
 825 
 826 // -----------------------------------------------------------------------------
 827 // Exit support
 828 //
 829 // exit()
 830 // ~~~~~~
 831 // Note that the collector can't reclaim the objectMonitor or deflate
 832 // the object out from underneath the thread calling ::exit() as the
 833 // thread calling ::exit() never transitions to a stable state.
 834 // This inhibits GC, which in turn inhibits asynchronous (and
 835 // inopportune) reclamation of "this".
 836 //
 837 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
 838 // There's one exception to the claim above, however.  EnterI() can call
 839 // exit() to drop a lock if the acquirer has been externally suspended.
 840 // In that case exit() is called with _thread_state as _thread_blocked,
 841 // but the monitor's _count field is > 0, which inhibits reclamation.
 842 //
 843 // 1-0 exit
 844 // ~~~~~~~~
 845 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
 846 // the fast-path operators have been optimized so the common ::exit()
 847 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
 848 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
 849 // greatly improves latency -- MEMBAR and CAS having considerable local
 850 // latency on modern processors -- but at the cost of "stranding".  Absent the
 851 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
 852 // ::enter() path, resulting in the entering thread being stranding
 853 // and a progress-liveness failure.   Stranding is extremely rare.
 854 // We use timers (timed park operations) & periodic polling to detect
 855 // and recover from stranding.  Potentially stranded threads periodically
 856 // wake up and poll the lock.  See the usage of the _Responsible variable.
 857 //
 858 // The CAS() in enter provides for safety and exclusion, while the CAS or
 859 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
 860 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
 861 // We detect and recover from stranding with timers.
 862 //
 863 // If a thread transiently strands it'll park until (a) another
 864 // thread acquires the lock and then drops the lock, at which time the
 865 // exiting thread will notice and unpark the stranded thread, or, (b)
 866 // the timer expires.  If the lock is high traffic then the stranding latency
 867 // will be low due to (a).  If the lock is low traffic then the odds of
 868 // stranding are lower, although the worst-case stranding latency
 869 // is longer.  Critically, we don't want to put excessive load in the
 870 // platform's timer subsystem.  We want to minimize both the timer injection
 871 // rate (timers created/sec) as well as the number of timers active at
 872 // any one time.  (more precisely, we want to minimize timer-seconds, which is
 873 // the integral of the # of active timers at any instant over time).
 874 // Both impinge on OS scalability.  Given that, at most one thread parked on
 875 // a monitor will use a timer.
 876 //
 877 // There is also the risk of a futile wake-up. If we drop the lock
 878 // another thread can reacquire the lock immediately, and we can
 879 // then wake a thread unnecessarily. This is benign, and we've
 880 // structured the code so the windows are short and the frequency
 881 // of such futile wakups is low.
 882 
 883 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
 884   Thread * const Self = THREAD;
 885   if (THREAD != _owner) {
 886     if (THREAD->is_lock_owned((address) _owner)) {
 887       // Transmute _owner from a BasicLock pointer to a Thread address.
 888       // We don't need to hold _mutex for this transition.
 889       // Non-null to Non-null is safe as long as all readers can
 890       // tolerate either flavor.
 891       assert(_recursions == 0, "invariant");
 892       _owner = THREAD;
 893       _recursions = 0;
 894     } else {
 895       // Apparent unbalanced locking ...
 896       // Naively we'd like to throw IllegalMonitorStateException.
 897       // As a practical matter we can neither allocate nor throw an
 898       // exception as ::exit() can be called from leaf routines.
 899       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
 900       // Upon deeper reflection, however, in a properly run JVM the only
 901       // way we should encounter this situation is in the presence of
 902       // unbalanced JNI locking. TODO: CheckJNICalls.
 903       // See also: CR4414101
 904       assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
 905       return;
 906     }
 907   }
 908 
 909   if (_recursions != 0) {
 910     _recursions--;        // this is simple recursive enter
 911     return;
 912   }
 913 
 914   // Invariant: after setting Responsible=null an thread must execute
 915   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
 916   _Responsible = NULL;
 917 
 918 #if INCLUDE_JFR
 919   // get the owner's thread id for the MonitorEnter event
 920   // if it is enabled and the thread isn't suspended
 921   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
 922     _previous_owner_tid = JFR_THREAD_ID(Self);
 923   }
 924 #endif
 925 
 926   for (;;) {
 927     assert(THREAD == _owner, "invariant");
 928 
 929     if (Knob_ExitPolicy == 0) {
 930       // release semantics: prior loads and stores from within the critical section
 931       // must not float (reorder) past the following store that drops the lock.
 932       // On SPARC that requires MEMBAR #loadstore|#storestore.
 933       // But of course in TSO #loadstore|#storestore is not required.
 934       // I'd like to write one of the following:
 935       // A.  OrderAccess::release() ; _owner = NULL
 936       // B.  OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
 937       // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
 938       // store into a _dummy variable.  That store is not needed, but can result
 939       // in massive wasteful coherency traffic on classic SMP systems.
 940       // Instead, I use release_store(), which is implemented as just a simple
 941       // ST on x64, x86 and SPARC.
 942       OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 943       OrderAccess::storeload();                        // See if we need to wake a successor
 944       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 945         return;
 946       }
 947       // Other threads are blocked trying to acquire the lock.
 948 
 949       // Normally the exiting thread is responsible for ensuring succession,
 950       // but if other successors are ready or other entering threads are spinning
 951       // then this thread can simply store NULL into _owner and exit without
 952       // waking a successor.  The existence of spinners or ready successors
 953       // guarantees proper succession (liveness).  Responsibility passes to the
 954       // ready or running successors.  The exiting thread delegates the duty.
 955       // More precisely, if a successor already exists this thread is absolved
 956       // of the responsibility of waking (unparking) one.
 957       //
 958       // The _succ variable is critical to reducing futile wakeup frequency.
 959       // _succ identifies the "heir presumptive" thread that has been made
 960       // ready (unparked) but that has not yet run.  We need only one such
 961       // successor thread to guarantee progress.
 962       // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 963       // section 3.3 "Futile Wakeup Throttling" for details.
 964       //
 965       // Note that spinners in Enter() also set _succ non-null.
 966       // In the current implementation spinners opportunistically set
 967       // _succ so that exiting threads might avoid waking a successor.
 968       // Another less appealing alternative would be for the exiting thread
 969       // to drop the lock and then spin briefly to see if a spinner managed
 970       // to acquire the lock.  If so, the exiting thread could exit
 971       // immediately without waking a successor, otherwise the exiting
 972       // thread would need to dequeue and wake a successor.
 973       // (Note that we'd need to make the post-drop spin short, but no
 974       // shorter than the worst-case round-trip cache-line migration time.
 975       // The dropped lock needs to become visible to the spinner, and then
 976       // the acquisition of the lock by the spinner must become visible to
 977       // the exiting thread).
 978 
 979       // It appears that an heir-presumptive (successor) must be made ready.
 980       // Only the current lock owner can manipulate the EntryList or
 981       // drain _cxq, so we need to reacquire the lock.  If we fail
 982       // to reacquire the lock the responsibility for ensuring succession
 983       // falls to the new owner.
 984       //
 985       if (!Atomic::replace_if_null(THREAD, &_owner)) {
 986         return;
 987       }
 988     } else {
 989       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 990         OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 991         OrderAccess::storeload();
 992         // Ratify the previously observed values.
 993         if (_cxq == NULL || _succ != NULL) {
 994           return;
 995         }
 996 
 997         // inopportune interleaving -- the exiting thread (this thread)
 998         // in the fast-exit path raced an entering thread in the slow-enter
 999         // path.
1000         // We have two choices:
1001         // A.  Try to reacquire the lock.
1002         //     If the CAS() fails return immediately, otherwise
1003         //     we either restart/rerun the exit operation, or simply
1004         //     fall-through into the code below which wakes a successor.
1005         // B.  If the elements forming the EntryList|cxq are TSM
1006         //     we could simply unpark() the lead thread and return
1007         //     without having set _succ.
1008         if (!Atomic::replace_if_null(THREAD, &_owner)) {
1009           return;
1010         }
1011       }
1012     }
1013 
1014     guarantee(_owner == THREAD, "invariant");
1015 
1016     ObjectWaiter * w = NULL;
1017     int QMode = Knob_QMode;
1018 
1019     if (QMode == 2 && _cxq != NULL) {
1020       // QMode == 2 : cxq has precedence over EntryList.
1021       // Try to directly wake a successor from the cxq.
1022       // If successful, the successor will need to unlink itself from cxq.
1023       w = _cxq;
1024       assert(w != NULL, "invariant");
1025       assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1026       ExitEpilog(Self, w);
1027       return;
1028     }
1029 
1030     if (QMode == 3 && _cxq != NULL) {
1031       // Aggressively drain cxq into EntryList at the first opportunity.
1032       // This policy ensure that recently-run threads live at the head of EntryList.
1033       // Drain _cxq into EntryList - bulk transfer.
1034       // First, detach _cxq.
1035       // The following loop is tantamount to: w = swap(&cxq, NULL)
1036       w = _cxq;
1037       for (;;) {
1038         assert(w != NULL, "Invariant");
1039         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1040         if (u == w) break;
1041         w = u;
1042       }
1043       assert(w != NULL, "invariant");
1044 
1045       ObjectWaiter * q = NULL;
1046       ObjectWaiter * p;
1047       for (p = w; p != NULL; p = p->_next) {
1048         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1049         p->TState = ObjectWaiter::TS_ENTER;
1050         p->_prev = q;
1051         q = p;
1052       }
1053 
1054       // Append the RATs to the EntryList
1055       // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1056       ObjectWaiter * Tail;
1057       for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1058            Tail = Tail->_next)
1059         /* empty */;
1060       if (Tail == NULL) {
1061         _EntryList = w;
1062       } else {
1063         Tail->_next = w;
1064         w->_prev = Tail;
1065       }
1066 
1067       // Fall thru into code that tries to wake a successor from EntryList
1068     }
1069 
1070     if (QMode == 4 && _cxq != NULL) {
1071       // Aggressively drain cxq into EntryList at the first opportunity.
1072       // This policy ensure that recently-run threads live at the head of EntryList.
1073 
1074       // Drain _cxq into EntryList - bulk transfer.
1075       // First, detach _cxq.
1076       // The following loop is tantamount to: w = swap(&cxq, NULL)
1077       w = _cxq;
1078       for (;;) {
1079         assert(w != NULL, "Invariant");
1080         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1081         if (u == w) break;
1082         w = u;
1083       }
1084       assert(w != NULL, "invariant");
1085 
1086       ObjectWaiter * q = NULL;
1087       ObjectWaiter * p;
1088       for (p = w; p != NULL; p = p->_next) {
1089         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1090         p->TState = ObjectWaiter::TS_ENTER;
1091         p->_prev = q;
1092         q = p;
1093       }
1094 
1095       // Prepend the RATs to the EntryList
1096       if (_EntryList != NULL) {
1097         q->_next = _EntryList;
1098         _EntryList->_prev = q;
1099       }
1100       _EntryList = w;
1101 
1102       // Fall thru into code that tries to wake a successor from EntryList
1103     }
1104 
1105     w = _EntryList;
1106     if (w != NULL) {
1107       // I'd like to write: guarantee (w->_thread != Self).
1108       // But in practice an exiting thread may find itself on the EntryList.
1109       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1110       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1111       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1112       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1113       // release the lock "O".  T2 resumes immediately after the ST of null into
1114       // _owner, above.  T2 notices that the EntryList is populated, so it
1115       // reacquires the lock and then finds itself on the EntryList.
1116       // Given all that, we have to tolerate the circumstance where "w" is
1117       // associated with Self.
1118       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1119       ExitEpilog(Self, w);
1120       return;
1121     }
1122 
1123     // If we find that both _cxq and EntryList are null then just
1124     // re-run the exit protocol from the top.
1125     w = _cxq;
1126     if (w == NULL) continue;
1127 
1128     // Drain _cxq into EntryList - bulk transfer.
1129     // First, detach _cxq.
1130     // The following loop is tantamount to: w = swap(&cxq, NULL)
1131     for (;;) {
1132       assert(w != NULL, "Invariant");
1133       ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1134       if (u == w) break;
1135       w = u;
1136     }
1137 
1138     assert(w != NULL, "invariant");
1139     assert(_EntryList == NULL, "invariant");
1140 
1141     // Convert the LIFO SLL anchored by _cxq into a DLL.
1142     // The list reorganization step operates in O(LENGTH(w)) time.
1143     // It's critical that this step operate quickly as
1144     // "Self" still holds the outer-lock, restricting parallelism
1145     // and effectively lengthening the critical section.
1146     // Invariant: s chases t chases u.
1147     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1148     // we have faster access to the tail.
1149 
1150     if (QMode == 1) {
1151       // QMode == 1 : drain cxq to EntryList, reversing order
1152       // We also reverse the order of the list.
1153       ObjectWaiter * s = NULL;
1154       ObjectWaiter * t = w;
1155       ObjectWaiter * u = NULL;
1156       while (t != NULL) {
1157         guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1158         t->TState = ObjectWaiter::TS_ENTER;
1159         u = t->_next;
1160         t->_prev = u;
1161         t->_next = s;
1162         s = t;
1163         t = u;
1164       }
1165       _EntryList  = s;
1166       assert(s != NULL, "invariant");
1167     } else {
1168       // QMode == 0 or QMode == 2
1169       _EntryList = w;
1170       ObjectWaiter * q = NULL;
1171       ObjectWaiter * p;
1172       for (p = w; p != NULL; p = p->_next) {
1173         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1174         p->TState = ObjectWaiter::TS_ENTER;
1175         p->_prev = q;
1176         q = p;
1177       }
1178     }
1179 
1180     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1181     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1182 
1183     // See if we can abdicate to a spinner instead of waking a thread.
1184     // A primary goal of the implementation is to reduce the
1185     // context-switch rate.
1186     if (_succ != NULL) continue;
1187 
1188     w = _EntryList;
1189     if (w != NULL) {
1190       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1191       ExitEpilog(Self, w);
1192       return;
1193     }
1194   }
1195 }
1196 
1197 // ExitSuspendEquivalent:
1198 // A faster alternate to handle_special_suspend_equivalent_condition()
1199 //
1200 // handle_special_suspend_equivalent_condition() unconditionally
1201 // acquires the SR_lock.  On some platforms uncontended MutexLocker()
1202 // operations have high latency.  Note that in ::enter() we call HSSEC
1203 // while holding the monitor, so we effectively lengthen the critical sections.
1204 //
1205 // There are a number of possible solutions:
1206 //
1207 // A.  To ameliorate the problem we might also defer state transitions
1208 //     to as late as possible -- just prior to parking.
1209 //     Given that, we'd call HSSEC after having returned from park(),
1210 //     but before attempting to acquire the monitor.  This is only a
1211 //     partial solution.  It avoids calling HSSEC while holding the
1212 //     monitor (good), but it still increases successor reacquisition latency --
1213 //     the interval between unparking a successor and the time the successor
1214 //     resumes and retries the lock.  See ReenterI(), which defers state transitions.
1215 //     If we use this technique we can also avoid EnterI()-exit() loop
1216 //     in ::enter() where we iteratively drop the lock and then attempt
1217 //     to reacquire it after suspending.
1218 //
1219 // B.  In the future we might fold all the suspend bits into a
1220 //     composite per-thread suspend flag and then update it with CAS().
1221 //     Alternately, a Dekker-like mechanism with multiple variables
1222 //     would suffice:
1223 //       ST Self->_suspend_equivalent = false
1224 //       MEMBAR
1225 //       LD Self_>_suspend_flags
1226 //
1227 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1228 // with a more efficient implementation, the need to use "FastHSSEC" has
1229 // decreased. - Dave
1230 
1231 
1232 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1233   const int Mode = Knob_FastHSSEC;
1234   if (Mode && !jSelf->is_external_suspend()) {
1235     assert(jSelf->is_suspend_equivalent(), "invariant");
1236     jSelf->clear_suspend_equivalent();
1237     if (2 == Mode) OrderAccess::storeload();
1238     if (!jSelf->is_external_suspend()) return false;
1239     // We raced a suspension -- fall thru into the slow path
1240     jSelf->set_suspend_equivalent();
1241   }
1242   return jSelf->handle_special_suspend_equivalent_condition();
1243 }
1244 
1245 
1246 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1247   assert(_owner == Self, "invariant");
1248 
1249   // Exit protocol:
1250   // 1. ST _succ = wakee
1251   // 2. membar #loadstore|#storestore;
1252   // 2. ST _owner = NULL
1253   // 3. unpark(wakee)
1254 
1255   _succ = Knob_SuccEnabled ? Wakee->_thread : NULL;
1256   ParkEvent * Trigger = Wakee->_event;
1257 
1258   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1259   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1260   // out-of-scope (non-extant).
1261   Wakee  = NULL;
1262 
1263   // Drop the lock
1264   OrderAccess::release_store(&_owner, (void*)NULL);
1265   OrderAccess::fence();                               // ST _owner vs LD in unpark()
1266 
1267   DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1268   Trigger->unpark();
1269 
1270   // Maintain stats and report events to JVMTI
1271   OM_PERFDATA_OP(Parks, inc());
1272 }
1273 
1274 
1275 // -----------------------------------------------------------------------------
1276 // Class Loader deadlock handling.
1277 //
1278 // complete_exit exits a lock returning recursion count
1279 // complete_exit/reenter operate as a wait without waiting
1280 // complete_exit requires an inflated monitor
1281 // The _owner field is not always the Thread addr even with an
1282 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1283 // thread due to contention.
1284 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1285   Thread * const Self = THREAD;
1286   assert(Self->is_Java_thread(), "Must be Java thread!");
1287   JavaThread *jt = (JavaThread *)THREAD;
1288 
1289   DeferredInitialize();
1290 
1291   if (THREAD != _owner) {
1292     if (THREAD->is_lock_owned ((address)_owner)) {
1293       assert(_recursions == 0, "internal state error");
1294       _owner = THREAD;   // Convert from basiclock addr to Thread addr
1295       _recursions = 0;
1296     }
1297   }
1298 
1299   guarantee(Self == _owner, "complete_exit not owner");
1300   intptr_t save = _recursions; // record the old recursion count
1301   _recursions = 0;        // set the recursion level to be 0
1302   exit(true, Self);           // exit the monitor
1303   guarantee(_owner != Self, "invariant");
1304   return save;
1305 }
1306 
1307 // reenter() enters a lock and sets recursion count
1308 // complete_exit/reenter operate as a wait without waiting
1309 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1310   Thread * const Self = THREAD;
1311   assert(Self->is_Java_thread(), "Must be Java thread!");
1312   JavaThread *jt = (JavaThread *)THREAD;
1313 
1314   guarantee(_owner != Self, "reenter already owner");
1315   enter(THREAD);       // enter the monitor
1316   guarantee(_recursions == 0, "reenter recursion");
1317   _recursions = recursions;
1318   return;
1319 }
1320 
1321 
1322 // -----------------------------------------------------------------------------
1323 // A macro is used below because there may already be a pending
1324 // exception which should not abort the execution of the routines
1325 // which use this (which is why we don't put this into check_slow and
1326 // call it with a CHECK argument).
1327 
1328 #define CHECK_OWNER()                                                       \
1329   do {                                                                      \
1330     if (THREAD != _owner) {                                                 \
1331       if (THREAD->is_lock_owned((address) _owner)) {                        \
1332         _owner = THREAD;  /* Convert from basiclock addr to Thread addr */  \
1333         _recursions = 0;                                                    \
1334       } else {                                                              \
1335         THROW(vmSymbols::java_lang_IllegalMonitorStateException());         \
1336       }                                                                     \
1337     }                                                                       \
1338   } while (false)
1339 
1340 // check_slow() is a misnomer.  It's called to simply to throw an IMSX exception.
1341 // TODO-FIXME: remove check_slow() -- it's likely dead.
1342 
1343 void ObjectMonitor::check_slow(TRAPS) {
1344   assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1345   THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1346 }
1347 
1348 static int Adjust(volatile int * adr, int dx) {
1349   int v;
1350   for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1351   return v;
1352 }
1353 
1354 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1355                                     ObjectMonitor* monitor,
1356                                     jlong notifier_tid,
1357                                     jlong timeout,
1358                                     bool timedout) {
1359   assert(event != NULL, "invariant");
1360   assert(monitor != NULL, "invariant");
1361   event->set_monitorClass(((oop)monitor->object())->klass());
1362   event->set_timeout(timeout);
1363   event->set_address((uintptr_t)monitor->object_addr());
1364   event->set_notifier(notifier_tid);
1365   event->set_timedOut(timedout);
1366   event->commit();
1367 }
1368 
1369 // -----------------------------------------------------------------------------
1370 // Wait/Notify/NotifyAll
1371 //
1372 // Note: a subset of changes to ObjectMonitor::wait()
1373 // will need to be replicated in complete_exit
1374 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1375   Thread * const Self = THREAD;
1376   assert(Self->is_Java_thread(), "Must be Java thread!");
1377   JavaThread *jt = (JavaThread *)THREAD;
1378 
1379   DeferredInitialize();
1380 
1381   // Throw IMSX or IEX.
1382   CHECK_OWNER();
1383 
1384   EventJavaMonitorWait event;
1385 
1386   // check for a pending interrupt
1387   if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1388     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1389     if (JvmtiExport::should_post_monitor_waited()) {
1390       // Note: 'false' parameter is passed here because the
1391       // wait was not timed out due to thread interrupt.
1392       JvmtiExport::post_monitor_waited(jt, this, false);
1393 
1394       // In this short circuit of the monitor wait protocol, the
1395       // current thread never drops ownership of the monitor and
1396       // never gets added to the wait queue so the current thread
1397       // cannot be made the successor. This means that the
1398       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1399       // consume an unpark() meant for the ParkEvent associated with
1400       // this ObjectMonitor.
1401     }
1402     if (event.should_commit()) {
1403       post_monitor_wait_event(&event, this, 0, millis, false);
1404     }
1405     THROW(vmSymbols::java_lang_InterruptedException());
1406     return;
1407   }
1408 
1409   assert(Self->_Stalled == 0, "invariant");
1410   Self->_Stalled = intptr_t(this);
1411   jt->set_current_waiting_monitor(this);
1412 
1413   // create a node to be put into the queue
1414   // Critically, after we reset() the event but prior to park(), we must check
1415   // for a pending interrupt.
1416   ObjectWaiter node(Self);
1417   node.TState = ObjectWaiter::TS_WAIT;
1418   Self->_ParkEvent->reset();
1419   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1420 
1421   // Enter the waiting queue, which is a circular doubly linked list in this case
1422   // but it could be a priority queue or any data structure.
1423   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1424   // by the the owner of the monitor *except* in the case where park()
1425   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1426   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1427 
1428   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1429   AddWaiter(&node);
1430   Thread::SpinRelease(&_WaitSetLock);
1431 
1432   _Responsible = NULL;
1433 
1434   intptr_t save = _recursions; // record the old recursion count
1435   _waiters++;                  // increment the number of waiters
1436   _recursions = 0;             // set the recursion level to be 1
1437   exit(true, Self);                    // exit the monitor
1438   guarantee(_owner != Self, "invariant");
1439 
1440   // The thread is on the WaitSet list - now park() it.
1441   // On MP systems it's conceivable that a brief spin before we park
1442   // could be profitable.
1443   //
1444   // TODO-FIXME: change the following logic to a loop of the form
1445   //   while (!timeout && !interrupted && _notified == 0) park()
1446 
1447   int ret = OS_OK;
1448   int WasNotified = 0;
1449   { // State transition wrappers
1450     OSThread* osthread = Self->osthread();
1451     OSThreadWaitState osts(osthread, true);
1452     {
1453       ThreadBlockInVM tbivm(jt);
1454       // Thread is in thread_blocked state and oop access is unsafe.
1455       jt->set_suspend_equivalent();
1456 
1457       if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1458         // Intentionally empty
1459       } else if (node._notified == 0) {
1460         if (millis <= 0) {
1461           Self->_ParkEvent->park();
1462         } else {
1463           ret = Self->_ParkEvent->park(millis);
1464         }
1465       }
1466 
1467       // were we externally suspended while we were waiting?
1468       if (ExitSuspendEquivalent (jt)) {
1469         // TODO-FIXME: add -- if succ == Self then succ = null.
1470         jt->java_suspend_self();
1471       }
1472 
1473     } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1474 
1475     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1476     // from the WaitSet to the EntryList.
1477     // See if we need to remove Node from the WaitSet.
1478     // We use double-checked locking to avoid grabbing _WaitSetLock
1479     // if the thread is not on the wait queue.
1480     //
1481     // Note that we don't need a fence before the fetch of TState.
1482     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1483     // written by the is thread. (perhaps the fetch might even be satisfied
1484     // by a look-aside into the processor's own store buffer, although given
1485     // the length of the code path between the prior ST and this load that's
1486     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1487     // then we'll acquire the lock and then re-fetch a fresh TState value.
1488     // That is, we fail toward safety.
1489 
1490     if (node.TState == ObjectWaiter::TS_WAIT) {
1491       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1492       if (node.TState == ObjectWaiter::TS_WAIT) {
1493         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1494         assert(node._notified == 0, "invariant");
1495         node.TState = ObjectWaiter::TS_RUN;
1496       }
1497       Thread::SpinRelease(&_WaitSetLock);
1498     }
1499 
1500     // The thread is now either on off-list (TS_RUN),
1501     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1502     // The Node's TState variable is stable from the perspective of this thread.
1503     // No other threads will asynchronously modify TState.
1504     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1505     OrderAccess::loadload();
1506     if (_succ == Self) _succ = NULL;
1507     WasNotified = node._notified;
1508 
1509     // Reentry phase -- reacquire the monitor.
1510     // re-enter contended monitor after object.wait().
1511     // retain OBJECT_WAIT state until re-enter successfully completes
1512     // Thread state is thread_in_vm and oop access is again safe,
1513     // although the raw address of the object may have changed.
1514     // (Don't cache naked oops over safepoints, of course).
1515 
1516     // post monitor waited event. Note that this is past-tense, we are done waiting.
1517     if (JvmtiExport::should_post_monitor_waited()) {
1518       JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1519 
1520       if (node._notified != 0 && _succ == Self) {
1521         // In this part of the monitor wait-notify-reenter protocol it
1522         // is possible (and normal) for another thread to do a fastpath
1523         // monitor enter-exit while this thread is still trying to get
1524         // to the reenter portion of the protocol.
1525         //
1526         // The ObjectMonitor was notified and the current thread is
1527         // the successor which also means that an unpark() has already
1528         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1529         // consume the unpark() that was done when the successor was
1530         // set because the same ParkEvent is shared between Java
1531         // monitors and JVM/TI RawMonitors (for now).
1532         //
1533         // We redo the unpark() to ensure forward progress, i.e., we
1534         // don't want all pending threads hanging (parked) with none
1535         // entering the unlocked monitor.
1536         node._event->unpark();
1537       }
1538     }
1539 
1540     if (event.should_commit()) {
1541       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1542     }
1543 
1544     OrderAccess::fence();
1545 
1546     assert(Self->_Stalled != 0, "invariant");
1547     Self->_Stalled = 0;
1548 
1549     assert(_owner != Self, "invariant");
1550     ObjectWaiter::TStates v = node.TState;
1551     if (v == ObjectWaiter::TS_RUN) {
1552       enter(Self);
1553     } else {
1554       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1555       ReenterI(Self, &node);
1556       node.wait_reenter_end(this);
1557     }
1558 
1559     // Self has reacquired the lock.
1560     // Lifecycle - the node representing Self must not appear on any queues.
1561     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1562     // want residual elements associated with this thread left on any lists.
1563     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1564     assert(_owner == Self, "invariant");
1565     assert(_succ != Self, "invariant");
1566   } // OSThreadWaitState()
1567 
1568   jt->set_current_waiting_monitor(NULL);
1569 
1570   guarantee(_recursions == 0, "invariant");
1571   _recursions = save;     // restore the old recursion count
1572   _waiters--;             // decrement the number of waiters
1573 
1574   // Verify a few postconditions
1575   assert(_owner == Self, "invariant");
1576   assert(_succ != Self, "invariant");
1577   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1578 
1579   // check if the notification happened
1580   if (!WasNotified) {
1581     // no, it could be timeout or Thread.interrupt() or both
1582     // check for interrupt event, otherwise it is timeout
1583     if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1584       THROW(vmSymbols::java_lang_InterruptedException());
1585     }
1586   }
1587 
1588   // NOTE: Spurious wake up will be consider as timeout.
1589   // Monitor notify has precedence over thread interrupt.
1590 }
1591 
1592 
1593 // Consider:
1594 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1595 // then instead of transferring a thread from the WaitSet to the EntryList
1596 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1597 
1598 void ObjectMonitor::INotify(Thread * Self) {
1599   const int policy = Knob_MoveNotifyee;
1600 
1601   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1602   ObjectWaiter * iterator = DequeueWaiter();
1603   if (iterator != NULL) {
1604     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1605     guarantee(iterator->_notified == 0, "invariant");
1606     // Disposition - what might we do with iterator ?
1607     // a.  add it directly to the EntryList - either tail (policy == 1)
1608     //     or head (policy == 0).
1609     // b.  push it onto the front of the _cxq (policy == 2).
1610     // For now we use (b).
1611     if (policy != 4) {
1612       iterator->TState = ObjectWaiter::TS_ENTER;
1613     }
1614     iterator->_notified = 1;
1615     iterator->_notifier_tid = JFR_THREAD_ID(Self);
1616 
1617     ObjectWaiter * list = _EntryList;
1618     if (list != NULL) {
1619       assert(list->_prev == NULL, "invariant");
1620       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1621       assert(list != iterator, "invariant");
1622     }
1623 
1624     if (policy == 0) {       // prepend to EntryList
1625       if (list == NULL) {
1626         iterator->_next = iterator->_prev = NULL;
1627         _EntryList = iterator;
1628       } else {
1629         list->_prev = iterator;
1630         iterator->_next = list;
1631         iterator->_prev = NULL;
1632         _EntryList = iterator;
1633       }
1634     } else if (policy == 1) {      // append to EntryList
1635       if (list == NULL) {
1636         iterator->_next = iterator->_prev = NULL;
1637         _EntryList = iterator;
1638       } else {
1639         // CONSIDER:  finding the tail currently requires a linear-time walk of
1640         // the EntryList.  We can make tail access constant-time by converting to
1641         // a CDLL instead of using our current DLL.
1642         ObjectWaiter * tail;
1643         for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1644         assert(tail != NULL && tail->_next == NULL, "invariant");
1645         tail->_next = iterator;
1646         iterator->_prev = tail;
1647         iterator->_next = NULL;
1648       }
1649     } else if (policy == 2) {      // prepend to cxq
1650       if (list == NULL) {
1651         iterator->_next = iterator->_prev = NULL;
1652         _EntryList = iterator;
1653       } else {
1654         iterator->TState = ObjectWaiter::TS_CXQ;
1655         for (;;) {
1656           ObjectWaiter * front = _cxq;
1657           iterator->_next = front;
1658           if (Atomic::cmpxchg(iterator, &_cxq, front) == front) {
1659             break;
1660           }
1661         }
1662       }
1663     } else if (policy == 3) {      // append to cxq
1664       iterator->TState = ObjectWaiter::TS_CXQ;
1665       for (;;) {
1666         ObjectWaiter * tail = _cxq;
1667         if (tail == NULL) {
1668           iterator->_next = NULL;
1669           if (Atomic::replace_if_null(iterator, &_cxq)) {
1670             break;
1671           }
1672         } else {
1673           while (tail->_next != NULL) tail = tail->_next;
1674           tail->_next = iterator;
1675           iterator->_prev = tail;
1676           iterator->_next = NULL;
1677           break;
1678         }
1679       }
1680     } else {
1681       ParkEvent * ev = iterator->_event;
1682       iterator->TState = ObjectWaiter::TS_RUN;
1683       OrderAccess::fence();
1684       ev->unpark();
1685     }
1686 
1687     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1688     // move the add-to-EntryList operation, above, outside the critical section
1689     // protected by _WaitSetLock.  In practice that's not useful.  With the
1690     // exception of  wait() timeouts and interrupts the monitor owner
1691     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1692     // on _WaitSetLock so it's not profitable to reduce the length of the
1693     // critical section.
1694 
1695     if (policy < 4) {
1696       iterator->wait_reenter_begin(this);
1697     }
1698   }
1699   Thread::SpinRelease(&_WaitSetLock);
1700 }
1701 
1702 // Consider: a not-uncommon synchronization bug is to use notify() when
1703 // notifyAll() is more appropriate, potentially resulting in stranded
1704 // threads; this is one example of a lost wakeup. A useful diagnostic
1705 // option is to force all notify() operations to behave as notifyAll().
1706 //
1707 // Note: We can also detect many such problems with a "minimum wait".
1708 // When the "minimum wait" is set to a small non-zero timeout value
1709 // and the program does not hang whereas it did absent "minimum wait",
1710 // that suggests a lost wakeup bug.
1711 
1712 void ObjectMonitor::notify(TRAPS) {
1713   CHECK_OWNER();
1714   if (_WaitSet == NULL) {
1715     return;
1716   }
1717   DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1718   INotify(THREAD);
1719   OM_PERFDATA_OP(Notifications, inc(1));
1720 }
1721 
1722 
1723 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1724 // from the waitset to the EntryList. This could be done more efficiently with a
1725 // single bulk transfer but in practice it's not time-critical. Beware too,
1726 // that in prepend-mode we invert the order of the waiters. Let's say that the
1727 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1728 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1729 
1730 void ObjectMonitor::notifyAll(TRAPS) {
1731   CHECK_OWNER();
1732   if (_WaitSet == NULL) {
1733     return;
1734   }
1735 
1736   DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1737   int tally = 0;
1738   while (_WaitSet != NULL) {
1739     tally++;
1740     INotify(THREAD);
1741   }
1742 
1743   OM_PERFDATA_OP(Notifications, inc(tally));
1744 }
1745 
1746 // -----------------------------------------------------------------------------
1747 // Adaptive Spinning Support
1748 //
1749 // Adaptive spin-then-block - rational spinning
1750 //
1751 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1752 // algorithm.  On high order SMP systems it would be better to start with
1753 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1754 // a contending thread could enqueue itself on the cxq and then spin locally
1755 // on a thread-specific variable such as its ParkEvent._Event flag.
1756 // That's left as an exercise for the reader.  Note that global spinning is
1757 // not problematic on Niagara, as the L2 cache serves the interconnect and
1758 // has both low latency and massive bandwidth.
1759 //
1760 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1761 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1762 // (duration) or we can fix the count at approximately the duration of
1763 // a context switch and vary the frequency.   Of course we could also
1764 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1765 // For a description of 'Adaptive spin-then-block mutual exclusion in
1766 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1767 //
1768 // This implementation varies the duration "D", where D varies with
1769 // the success rate of recent spin attempts. (D is capped at approximately
1770 // length of a round-trip context switch).  The success rate for recent
1771 // spin attempts is a good predictor of the success rate of future spin
1772 // attempts.  The mechanism adapts automatically to varying critical
1773 // section length (lock modality), system load and degree of parallelism.
1774 // D is maintained per-monitor in _SpinDuration and is initialized
1775 // optimistically.  Spin frequency is fixed at 100%.
1776 //
1777 // Note that _SpinDuration is volatile, but we update it without locks
1778 // or atomics.  The code is designed so that _SpinDuration stays within
1779 // a reasonable range even in the presence of races.  The arithmetic
1780 // operations on _SpinDuration are closed over the domain of legal values,
1781 // so at worst a race will install and older but still legal value.
1782 // At the very worst this introduces some apparent non-determinism.
1783 // We might spin when we shouldn't or vice-versa, but since the spin
1784 // count are relatively short, even in the worst case, the effect is harmless.
1785 //
1786 // Care must be taken that a low "D" value does not become an
1787 // an absorbing state.  Transient spinning failures -- when spinning
1788 // is overall profitable -- should not cause the system to converge
1789 // on low "D" values.  We want spinning to be stable and predictable
1790 // and fairly responsive to change and at the same time we don't want
1791 // it to oscillate, become metastable, be "too" non-deterministic,
1792 // or converge on or enter undesirable stable absorbing states.
1793 //
1794 // We implement a feedback-based control system -- using past behavior
1795 // to predict future behavior.  We face two issues: (a) if the
1796 // input signal is random then the spin predictor won't provide optimal
1797 // results, and (b) if the signal frequency is too high then the control
1798 // system, which has some natural response lag, will "chase" the signal.
1799 // (b) can arise from multimodal lock hold times.  Transient preemption
1800 // can also result in apparent bimodal lock hold times.
1801 // Although sub-optimal, neither condition is particularly harmful, as
1802 // in the worst-case we'll spin when we shouldn't or vice-versa.
1803 // The maximum spin duration is rather short so the failure modes aren't bad.
1804 // To be conservative, I've tuned the gain in system to bias toward
1805 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1806 // "rings" or oscillates between spinning and not spinning.  This happens
1807 // when spinning is just on the cusp of profitability, however, so the
1808 // situation is not dire.  The state is benign -- there's no need to add
1809 // hysteresis control to damp the transition rate between spinning and
1810 // not spinning.
1811 
1812 // Spinning: Fixed frequency (100%), vary duration
1813 int ObjectMonitor::TrySpin(Thread * Self) {
1814   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1815   int ctr = Knob_FixedSpin;
1816   if (ctr != 0) {
1817     while (--ctr >= 0) {
1818       if (TryLock(Self) > 0) return 1;
1819       SpinPause();
1820     }
1821     return 0;
1822   }
1823 
1824   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1825     if (TryLock(Self) > 0) {
1826       // Increase _SpinDuration ...
1827       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1828       // Raising _SpurDuration to the poverty line is key.
1829       int x = _SpinDuration;
1830       if (x < Knob_SpinLimit) {
1831         if (x < Knob_Poverty) x = Knob_Poverty;
1832         _SpinDuration = x + Knob_BonusB;
1833       }
1834       return 1;
1835     }
1836     SpinPause();
1837   }
1838 
1839   // Admission control - verify preconditions for spinning
1840   //
1841   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1842   // becoming an absorbing state.  Put another way, we spin briefly to
1843   // sample, just in case the system load, parallelism, contention, or lock
1844   // modality changed.
1845   //
1846   // Consider the following alternative:
1847   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1848   // spin attempt.  "Periodically" might mean after a tally of
1849   // the # of failed spin attempts (or iterations) reaches some threshold.
1850   // This takes us into the realm of 1-out-of-N spinning, where we
1851   // hold the duration constant but vary the frequency.
1852 
1853   ctr = _SpinDuration;
1854   if (ctr < Knob_SpinBase) ctr = Knob_SpinBase;
1855   if (ctr <= 0) return 0;
1856 
1857   if (Knob_SuccRestrict && _succ != NULL) return 0;
1858   if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1859     return 0;
1860   }
1861 
1862   int MaxSpin = Knob_MaxSpinners;
1863   if (MaxSpin >= 0) {
1864     if (_Spinner > MaxSpin) {
1865       return 0;
1866     }
1867     // Slightly racy, but benign ...
1868     Adjust(&_Spinner, 1);
1869   }
1870 
1871   // We're good to spin ... spin ingress.
1872   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1873   // when preparing to LD...CAS _owner, etc and the CAS is likely
1874   // to succeed.
1875   int caspty  = Knob_CASPenalty;
1876   int oxpty   = Knob_OXPenalty;
1877   int sss     = Knob_SpinSetSucc;
1878   if (sss && _succ == NULL) _succ = Self;
1879   Thread * prv = NULL;
1880 
1881   // There are three ways to exit the following loop:
1882   // 1.  A successful spin where this thread has acquired the lock.
1883   // 2.  Spin failure with prejudice
1884   // 3.  Spin failure without prejudice
1885 
1886   while (--ctr >= 0) {
1887 
1888     // Periodic polling -- Check for pending GC
1889     // Threads may spin while they're unsafe.
1890     // We don't want spinning threads to delay the JVM from reaching
1891     // a stop-the-world safepoint or to steal cycles from GC.
1892     // If we detect a pending safepoint we abort in order that
1893     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1894     // this thread, if safe, doesn't steal cycles from GC.
1895     // This is in keeping with the "no loitering in runtime" rule.
1896     // We periodically check to see if there's a safepoint pending.
1897     if ((ctr & 0xFF) == 0) {
1898       if (SafepointMechanism::poll(Self)) {
1899         goto Abort;           // abrupt spin egress
1900       }
1901       if (Knob_UsePause & 1) SpinPause();
1902     }
1903 
1904     if (Knob_UsePause & 2) SpinPause();
1905 
1906     // Probe _owner with TATAS
1907     // If this thread observes the monitor transition or flicker
1908     // from locked to unlocked to locked, then the odds that this
1909     // thread will acquire the lock in this spin attempt go down
1910     // considerably.  The same argument applies if the CAS fails
1911     // or if we observe _owner change from one non-null value to
1912     // another non-null value.   In such cases we might abort
1913     // the spin without prejudice or apply a "penalty" to the
1914     // spin count-down variable "ctr", reducing it by 100, say.
1915 
1916     Thread * ox = (Thread *) _owner;
1917     if (ox == NULL) {
1918       ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL);
1919       if (ox == NULL) {
1920         // The CAS succeeded -- this thread acquired ownership
1921         // Take care of some bookkeeping to exit spin state.
1922         if (sss && _succ == Self) {
1923           _succ = NULL;
1924         }
1925         if (MaxSpin > 0) Adjust(&_Spinner, -1);
1926 
1927         // Increase _SpinDuration :
1928         // The spin was successful (profitable) so we tend toward
1929         // longer spin attempts in the future.
1930         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1931         // If we acquired the lock early in the spin cycle it
1932         // makes sense to increase _SpinDuration proportionally.
1933         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1934         int x = _SpinDuration;
1935         if (x < Knob_SpinLimit) {
1936           if (x < Knob_Poverty) x = Knob_Poverty;
1937           _SpinDuration = x + Knob_Bonus;
1938         }
1939         return 1;
1940       }
1941 
1942       // The CAS failed ... we can take any of the following actions:
1943       // * penalize: ctr -= Knob_CASPenalty
1944       // * exit spin with prejudice -- goto Abort;
1945       // * exit spin without prejudice.
1946       // * Since CAS is high-latency, retry again immediately.
1947       prv = ox;
1948       if (caspty == -2) break;
1949       if (caspty == -1) goto Abort;
1950       ctr -= caspty;
1951       continue;
1952     }
1953 
1954     // Did lock ownership change hands ?
1955     if (ox != prv && prv != NULL) {
1956       if (oxpty == -2) break;
1957       if (oxpty == -1) goto Abort;
1958       ctr -= oxpty;
1959     }
1960     prv = ox;
1961 
1962     // Abort the spin if the owner is not executing.
1963     // The owner must be executing in order to drop the lock.
1964     // Spinning while the owner is OFFPROC is idiocy.
1965     // Consider: ctr -= RunnablePenalty ;
1966     if (Knob_OState && NotRunnable (Self, ox)) {
1967       goto Abort;
1968     }
1969     if (sss && _succ == NULL) _succ = Self;
1970   }
1971 
1972   // Spin failed with prejudice -- reduce _SpinDuration.
1973   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1974   // AIMD is globally stable.
1975   {
1976     int x = _SpinDuration;
1977     if (x > 0) {
1978       // Consider an AIMD scheme like: x -= (x >> 3) + 100
1979       // This is globally sample and tends to damp the response.
1980       x -= Knob_Penalty;
1981       if (x < 0) x = 0;
1982       _SpinDuration = x;
1983     }
1984   }
1985 
1986  Abort:
1987   if (MaxSpin >= 0) Adjust(&_Spinner, -1);
1988   if (sss && _succ == Self) {
1989     _succ = NULL;
1990     // Invariant: after setting succ=null a contending thread
1991     // must recheck-retry _owner before parking.  This usually happens
1992     // in the normal usage of TrySpin(), but it's safest
1993     // to make TrySpin() as foolproof as possible.
1994     OrderAccess::fence();
1995     if (TryLock(Self) > 0) return 1;
1996   }
1997   return 0;
1998 }
1999 
2000 // NotRunnable() -- informed spinning
2001 //
2002 // Don't bother spinning if the owner is not eligible to drop the lock.
2003 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2004 // The thread must be runnable in order to drop the lock in timely fashion.
2005 // If the _owner is not runnable then spinning will not likely be
2006 // successful (profitable).
2007 //
2008 // Beware -- the thread referenced by _owner could have died
2009 // so a simply fetch from _owner->_thread_state might trap.
2010 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2011 // Because of the lifecycle issues, the _thread_state values
2012 // observed by NotRunnable() might be garbage.  NotRunnable must
2013 // tolerate this and consider the observed _thread_state value
2014 // as advisory.
2015 //
2016 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2017 // a thread pointer.
2018 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2019 // with the LSB of _owner.  Another option would be to probabilistically probe
2020 // the putative _owner->TypeTag value.
2021 //
2022 // Checking _thread_state isn't perfect.  Even if the thread is
2023 // in_java it might be blocked on a page-fault or have been preempted
2024 // and sitting on a ready/dispatch queue.
2025 //
2026 // The return value from NotRunnable() is *advisory* -- the
2027 // result is based on sampling and is not necessarily coherent.
2028 // The caller must tolerate false-negative and false-positive errors.
2029 // Spinning, in general, is probabilistic anyway.
2030 
2031 
2032 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2033   // Check ox->TypeTag == 2BAD.
2034   if (ox == NULL) return 0;
2035 
2036   // Avoid transitive spinning ...
2037   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2038   // Immediately after T1 acquires L it's possible that T2, also
2039   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2040   // This occurs transiently after T1 acquired L but before
2041   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2042   // its spin in this circumstance.
2043   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2044 
2045   if (BlockedOn == 1) return 1;
2046   if (BlockedOn != 0) {
2047     return BlockedOn != intptr_t(this) && _owner == ox;
2048   }
2049 
2050   assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2051   int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2052   // consider also: jst != _thread_in_Java -- but that's overspecific.
2053   return jst == _thread_blocked || jst == _thread_in_native;
2054 }
2055 
2056 
2057 // -----------------------------------------------------------------------------
2058 // WaitSet management ...
2059 
2060 ObjectWaiter::ObjectWaiter(Thread* thread) {
2061   _next     = NULL;
2062   _prev     = NULL;
2063   _notified = 0;
2064   _notifier_tid = 0;
2065   TState    = TS_RUN;
2066   _thread   = thread;
2067   _event    = thread->_ParkEvent;
2068   _active   = false;
2069   assert(_event != NULL, "invariant");
2070 }
2071 
2072 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2073   JavaThread *jt = (JavaThread *)this->_thread;
2074   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2075 }
2076 
2077 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2078   JavaThread *jt = (JavaThread *)this->_thread;
2079   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2080 }
2081 
2082 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2083   assert(node != NULL, "should not add NULL node");
2084   assert(node->_prev == NULL, "node already in list");
2085   assert(node->_next == NULL, "node already in list");
2086   // put node at end of queue (circular doubly linked list)
2087   if (_WaitSet == NULL) {
2088     _WaitSet = node;
2089     node->_prev = node;
2090     node->_next = node;
2091   } else {
2092     ObjectWaiter* head = _WaitSet;
2093     ObjectWaiter* tail = head->_prev;
2094     assert(tail->_next == head, "invariant check");
2095     tail->_next = node;
2096     head->_prev = node;
2097     node->_next = head;
2098     node->_prev = tail;
2099   }
2100 }
2101 
2102 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2103   // dequeue the very first waiter
2104   ObjectWaiter* waiter = _WaitSet;
2105   if (waiter) {
2106     DequeueSpecificWaiter(waiter);
2107   }
2108   return waiter;
2109 }
2110 
2111 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2112   assert(node != NULL, "should not dequeue NULL node");
2113   assert(node->_prev != NULL, "node already removed from list");
2114   assert(node->_next != NULL, "node already removed from list");
2115   // when the waiter has woken up because of interrupt,
2116   // timeout or other spurious wake-up, dequeue the
2117   // waiter from waiting list
2118   ObjectWaiter* next = node->_next;
2119   if (next == node) {
2120     assert(node->_prev == node, "invariant check");
2121     _WaitSet = NULL;
2122   } else {
2123     ObjectWaiter* prev = node->_prev;
2124     assert(prev->_next == node, "invariant check");
2125     assert(next->_prev == node, "invariant check");
2126     next->_prev = prev;
2127     prev->_next = next;
2128     if (_WaitSet == node) {
2129       _WaitSet = next;
2130     }
2131   }
2132   node->_next = NULL;
2133   node->_prev = NULL;
2134 }
2135 
2136 // -----------------------------------------------------------------------------
2137 // PerfData support
2138 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2139 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2140 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2141 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2142 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2143 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2144 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2145 
2146 // One-shot global initialization for the sync subsystem.
2147 // We could also defer initialization and initialize on-demand
2148 // the first time we call inflate().  Initialization would
2149 // be protected - like so many things - by the MonitorCache_lock.
2150 
2151 void ObjectMonitor::Initialize() {
2152   static int InitializationCompleted = 0;
2153   assert(InitializationCompleted == 0, "invariant");
2154   InitializationCompleted = 1;
2155   if (UsePerfData) {
2156     EXCEPTION_MARK;
2157 #define NEWPERFCOUNTER(n)                                                \
2158   {                                                                      \
2159     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2160                                         CHECK);                          \
2161   }
2162 #define NEWPERFVARIABLE(n)                                                \
2163   {                                                                       \
2164     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2165                                          CHECK);                          \
2166   }
2167     NEWPERFCOUNTER(_sync_Inflations);
2168     NEWPERFCOUNTER(_sync_Deflations);
2169     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2170     NEWPERFCOUNTER(_sync_FutileWakeups);
2171     NEWPERFCOUNTER(_sync_Parks);
2172     NEWPERFCOUNTER(_sync_Notifications);
2173     NEWPERFVARIABLE(_sync_MonExtant);
2174 #undef NEWPERFCOUNTER
2175 #undef NEWPERFVARIABLE
2176   }
2177 }
2178 
2179 void ObjectMonitor::DeferredInitialize() {
2180   if (InitDone > 0) return;
2181   if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2182     while (InitDone != 1) /* empty */;
2183     return;
2184   }
2185 
2186   // One-shot global initialization ...
2187   // The initialization is idempotent, so we don't need locks.
2188   // In the future consider doing this via os::init_2().
2189 
2190   if (!os::is_MP()) {
2191     Knob_SpinLimit = 0;
2192     Knob_SpinBase  = 0;
2193     Knob_PreSpin   = 0;
2194     Knob_FixedSpin = -1;
2195   }
2196 
2197   OrderAccess::fence();
2198   InitDone = 1;
2199 }
2200