rev 51780 : imported patch syncknobs-00-base
rev 51781 : imported patch syncknobs-01-Knob_ReportSettings
rev 51782 : imported patch syncknobs-02-Knob_SpinBackOff
rev 51783 : imported patch syncknobs-03-BackOffMask
rev 51784 : imported patch syncknobs-04-Knob_ExitRelease
rev 51785 : imported patch syncknobs-05-Knob_InlineNotify
rev 51786 : imported patch syncknobs-06-Knob_Verbose
rev 51787 : imported patch syncknobs-07-Knob_VerifyInUse
rev 51788 : imported patch syncknobs-08-Knob_VerifyMatch
rev 51789 : imported patch syncknobs-09-Knob_SpinBase
rev 51790 : imported patch syncknobs-10-Knob_CASPenalty
rev 51791 : imported patch syncknobs-11-Knob_OXPenalty
rev 51792 : imported patch syncknobs-12-Knob_SpinSetSucc
rev 51793 : imported patch syncknobs-13-Knob_SpinEarly
rev 51794 : imported patch syncknobs-14-Knob_SuccEnabled
rev 51795 : imported patch syncknobs-15-Knob_SuccRestrict
rev 51796 : imported patch syncknobs-16-Knob_MaxSpinners
rev 51797 : imported patch syncknobs-17-Knob_SpinAfterFutile
rev 51798 : imported patch syncknobs-18-Knob_OState
rev 51799 : imported patch syncknobs-19-Knob_UsePause
rev 51800 : imported patch syncknobs-20-Knob_ExitPolicy
rev 51801 : imported patch syncknobs-21-Knob_ResetEvent
rev 51802 : imported patch syncknobs-22-Knob_FastHSSEC
rev 51803 : imported patch syncknobs-23-Knob_MoveNotifyee
rev 51804 : imported patch syncknobs-24-Knob_QMode

   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_ReportSettings      = 0;
 112 static int Knob_SpinBase            = 0;       // Floor AKA SpinMin
 113 static int Knob_SpinBackOff         = 0;       // spin-loop backoff
 114 static int Knob_CASPenalty          = -1;      // Penalty for failed CAS
 115 static int Knob_OXPenalty           = -1;      // Penalty for observed _owner change
 116 static int Knob_SpinSetSucc         = 1;       // spinners set the _succ field
 117 static int Knob_SpinEarly           = 1;
 118 static int Knob_SuccEnabled         = 1;       // futile wake throttling
 119 static int Knob_SuccRestrict        = 0;       // Limit successors + spinners to at-most-one
 120 static int Knob_MaxSpinners         = -1;      // Should be a function of # CPUs
 121 static int Knob_Bonus               = 100;     // spin success bonus
 122 static int Knob_BonusB              = 100;     // spin success bonus
 123 static int Knob_Penalty             = 200;     // spin failure penalty
 124 static int Knob_Poverty             = 1000;
 125 static int Knob_SpinAfterFutile     = 1;       // Spin after returning from park()
 126 static int Knob_FixedSpin           = 0;
 127 static int Knob_OState              = 3;       // Spinner checks thread state of _owner
 128 static int Knob_UsePause            = 1;
 129 static int Knob_ExitPolicy          = 0;
 130 static int Knob_PreSpin             = 10;      // 20-100 likely better
 131 static int Knob_ResetEvent          = 0;
 132 static int BackOffMask              = 0;
 133 
 134 static int Knob_FastHSSEC           = 0;
 135 static int Knob_MoveNotifyee        = 2;       // notify() - disposition of notifyee
 136 static int Knob_QMode               = 0;       // EntryList-cxq policy - queue discipline
 137 static volatile int InitDone        = 0;
 138 
 139 // -----------------------------------------------------------------------------
 140 // Theory of operations -- Monitors lists, thread residency, etc:
 141 //
 142 // * A thread acquires ownership of a monitor by successfully
 143 //   CAS()ing the _owner field from null to non-null.
 144 //
 145 // * Invariant: A thread appears on at most one monitor list --
 146 //   cxq, EntryList or WaitSet -- at any one time.
 147 //
 148 // * Contending threads "push" themselves onto the cxq with CAS
 149 //   and then spin/park.
 150 //
 151 // * After a contending thread eventually acquires the lock it must
 152 //   dequeue itself from either the EntryList or the cxq.
 153 //
 154 // * The exiting thread identifies and unparks an "heir presumptive"
 155 //   tentative successor thread on the EntryList.  Critically, the
 156 //   exiting thread doesn't unlink the successor thread from the EntryList.
 157 //   After having been unparked, the wakee will recontend for ownership of
 158 //   the monitor.   The successor (wakee) will either acquire the lock or
 159 //   re-park itself.
 160 //
 161 //   Succession is provided for by a policy of competitive handoff.
 162 //   The exiting thread does _not_ grant or pass ownership to the
 163 //   successor thread.  (This is also referred to as "handoff" succession").
 164 //   Instead the exiting thread releases ownership and possibly wakes
 165 //   a successor, so the successor can (re)compete for ownership of the lock.
 166 //   If the EntryList is empty but the cxq is populated the exiting
 167 //   thread will drain the cxq into the EntryList.  It does so by
 168 //   by detaching the cxq (installing null with CAS) and folding
 169 //   the threads from the cxq into the EntryList.  The EntryList is
 170 //   doubly linked, while the cxq is singly linked because of the
 171 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 172 //
 173 // * Concurrency invariants:
 174 //
 175 //   -- only the monitor owner may access or mutate the EntryList.
 176 //      The mutex property of the monitor itself protects the EntryList
 177 //      from concurrent interference.
 178 //   -- Only the monitor owner may detach the cxq.
 179 //
 180 // * The monitor entry list operations avoid locks, but strictly speaking
 181 //   they're not lock-free.  Enter is lock-free, exit is not.
 182 //   For a description of 'Methods and apparatus providing non-blocking access
 183 //   to a resource,' see U.S. Pat. No. 7844973.
 184 //
 185 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 186 //   detaching thread.  This mechanism is immune from the ABA corruption.
 187 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 188 //
 189 // * Taken together, the cxq and the EntryList constitute or form a
 190 //   single logical queue of threads stalled trying to acquire the lock.
 191 //   We use two distinct lists to improve the odds of a constant-time
 192 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 193 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 194 //   A key desideratum is to minimize queue & monitor metadata manipulation
 195 //   that occurs while holding the monitor lock -- that is, we want to
 196 //   minimize monitor lock holds times.  Note that even a small amount of
 197 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 198 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 199 //   locks and monitor metadata.
 200 //
 201 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 202 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 203 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 204 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 205 //
 206 //   The EntryList is ordered by the prevailing queue discipline and
 207 //   can be organized in any convenient fashion, such as a doubly-linked list or
 208 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 209 //   to operate in constant-time.  If we need a priority queue then something akin
 210 //   to Solaris' sleepq would work nicely.  Viz.,
 211 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 212 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 213 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 214 //   EntryList accordingly.
 215 //
 216 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 217 //   somewhat similar to an elevator-scan.
 218 //
 219 // * The monitor synchronization subsystem avoids the use of native
 220 //   synchronization primitives except for the narrow platform-specific
 221 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 222 //   the semantics of park-unpark.  Put another way, this monitor implementation
 223 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 224 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 225 //   underlying OS manages the READY<->RUN transitions.
 226 //
 227 // * Waiting threads reside on the WaitSet list -- wait() puts
 228 //   the caller onto the WaitSet.
 229 //
 230 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 231 //   either the EntryList or cxq.  Subsequent exit() operations will
 232 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 233 //   it's likely the notifyee would simply impale itself on the lock held
 234 //   by the notifier.
 235 //
 236 // * An interesting alternative is to encode cxq as (List,LockByte) where
 237 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 238 //   variable, like _recursions, in the scheme.  The threads or Events that form
 239 //   the list would have to be aligned in 256-byte addresses.  A thread would
 240 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 241 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 242 //   Note that is is *not* word-tearing, but it does presume that full-word
 243 //   CAS operations are coherent with intermix with STB operations.  That's true
 244 //   on most common processors.
 245 //
 246 // * See also http://blogs.sun.com/dave
 247 
 248 
 249 void* ObjectMonitor::operator new (size_t size) throw() {
 250   return AllocateHeap(size, mtInternal);
 251 }
 252 void* ObjectMonitor::operator new[] (size_t size) throw() {
 253   return operator new (size);
 254 }
 255 void ObjectMonitor::operator delete(void* p) {
 256   FreeHeap(p);
 257 }
 258 void ObjectMonitor::operator delete[] (void *p) {
 259   operator delete(p);
 260 }
 261 
 262 // -----------------------------------------------------------------------------
 263 // Enter support
 264 
 265 void ObjectMonitor::enter(TRAPS) {
 266   // The following code is ordered to check the most common cases first
 267   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 268   Thread * const Self = THREAD;
 269 
 270   void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL);
 271   if (cur == NULL) {
 272     // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
 273     assert(_recursions == 0, "invariant");
 274     assert(_owner == Self, "invariant");
 275     return;
 276   }
 277 
 278   if (cur == Self) {
 279     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 280     _recursions++;
 281     return;
 282   }
 283 
 284   if (Self->is_lock_owned ((address)cur)) {
 285     assert(_recursions == 0, "internal state error");
 286     _recursions = 1;
 287     // Commute owner from a thread-specific on-stack BasicLockObject address to
 288     // a full-fledged "Thread *".
 289     _owner = Self;
 290     return;
 291   }
 292 
 293   // We've encountered genuine contention.
 294   assert(Self->_Stalled == 0, "invariant");
 295   Self->_Stalled = intptr_t(this);
 296 
 297   // Try one round of spinning *before* enqueueing Self
 298   // and before going through the awkward and expensive state
 299   // transitions.  The following spin is strictly optional ...
 300   // Note that if we acquire the monitor from an initial spin
 301   // we forgo posting JVMTI events and firing DTRACE probes.
 302   if (Knob_SpinEarly && TrySpin (Self) > 0) {
 303     assert(_owner == Self, "invariant");
 304     assert(_recursions == 0, "invariant");
 305     assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 306     Self->_Stalled = 0;
 307     return;
 308   }
 309 
 310   assert(_owner != Self, "invariant");
 311   assert(_succ != Self, "invariant");
 312   assert(Self->is_Java_thread(), "invariant");
 313   JavaThread * jt = (JavaThread *) Self;
 314   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 315   assert(jt->thread_state() != _thread_blocked, "invariant");
 316   assert(this->object() != NULL, "invariant");
 317   assert(_count >= 0, "invariant");
 318 
 319   // Prevent deflation at STW-time.  See deflate_idle_monitors() and is_busy().
 320   // Ensure the object-monitor relationship remains stable while there's contention.
 321   Atomic::inc(&_count);
 322 
 323   JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
 324   EventJavaMonitorEnter event;
 325   if (event.should_commit()) {
 326     event.set_monitorClass(((oop)this->object())->klass());
 327     event.set_address((uintptr_t)(this->object_addr()));
 328   }
 329 
 330   { // Change java thread status to indicate blocked on monitor enter.
 331     JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
 332 
 333     Self->set_current_pending_monitor(this);
 334 
 335     DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
 336     if (JvmtiExport::should_post_monitor_contended_enter()) {
 337       JvmtiExport::post_monitor_contended_enter(jt, this);
 338 
 339       // The current thread does not yet own the monitor and does not
 340       // yet appear on any queues that would get it made the successor.
 341       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 342       // handler cannot accidentally consume an unpark() meant for the
 343       // ParkEvent associated with this ObjectMonitor.
 344     }
 345 
 346     OSThreadContendState osts(Self->osthread());
 347     ThreadBlockInVM tbivm(jt);
 348 
 349     // TODO-FIXME: change the following for(;;) loop to straight-line code.
 350     for (;;) {
 351       jt->set_suspend_equivalent();
 352       // cleared by handle_special_suspend_equivalent_condition()
 353       // or java_suspend_self()
 354 
 355       EnterI(THREAD);
 356 
 357       if (!ExitSuspendEquivalent(jt)) break;
 358 
 359       // We have acquired the contended monitor, but while we were
 360       // waiting another thread suspended us. We don't want to enter
 361       // the monitor while suspended because that would surprise the
 362       // thread that suspended us.
 363       //
 364       _recursions = 0;
 365       _succ = NULL;
 366       exit(false, Self);
 367 
 368       jt->java_suspend_self();
 369     }
 370     Self->set_current_pending_monitor(NULL);
 371 
 372     // We cleared the pending monitor info since we've just gotten past
 373     // the enter-check-for-suspend dance and we now own the monitor free
 374     // and clear, i.e., it is no longer pending. The ThreadBlockInVM
 375     // destructor can go to a safepoint at the end of this block. If we
 376     // do a thread dump during that safepoint, then this thread will show
 377     // as having "-locked" the monitor, but the OS and java.lang.Thread
 378     // states will still report that the thread is blocked trying to
 379     // acquire it.
 380   }
 381 
 382   Atomic::dec(&_count);
 383   assert(_count >= 0, "invariant");
 384   Self->_Stalled = 0;
 385 
 386   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 387   assert(_recursions == 0, "invariant");
 388   assert(_owner == Self, "invariant");
 389   assert(_succ != Self, "invariant");
 390   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 391 
 392   // The thread -- now the owner -- is back in vm mode.
 393   // Report the glorious news via TI,DTrace and jvmstat.
 394   // The probe effect is non-trivial.  All the reportage occurs
 395   // while we hold the monitor, increasing the length of the critical
 396   // section.  Amdahl's parallel speedup law comes vividly into play.
 397   //
 398   // Another option might be to aggregate the events (thread local or
 399   // per-monitor aggregation) and defer reporting until a more opportune
 400   // time -- such as next time some thread encounters contention but has
 401   // yet to acquire the lock.  While spinning that thread could
 402   // spinning we could increment JVMStat counters, etc.
 403 
 404   DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
 405   if (JvmtiExport::should_post_monitor_contended_entered()) {
 406     JvmtiExport::post_monitor_contended_entered(jt, this);
 407 
 408     // The current thread already owns the monitor and is not going to
 409     // call park() for the remainder of the monitor enter protocol. So
 410     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 411     // event handler consumed an unpark() issued by the thread that
 412     // just exited the monitor.
 413   }
 414   if (event.should_commit()) {
 415     event.set_previousOwner((uintptr_t)_previous_owner_tid);
 416     event.commit();
 417   }
 418   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 419 }
 420 
 421 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 422 // Callers must compensate as needed.
 423 
 424 int ObjectMonitor::TryLock(Thread * Self) {
 425   void * own = _owner;
 426   if (own != NULL) return 0;
 427   if (Atomic::replace_if_null(Self, &_owner)) {
 428     // Either guarantee _recursions == 0 or set _recursions = 0.
 429     assert(_recursions == 0, "invariant");
 430     assert(_owner == Self, "invariant");
 431     return 1;
 432   }
 433   // The lock had been free momentarily, but we lost the race to the lock.
 434   // Interference -- the CAS failed.
 435   // We can either return -1 or retry.
 436   // Retry doesn't make as much sense because the lock was just acquired.
 437   return -1;
 438 }
 439 
 440 #define MAX_RECHECK_INTERVAL 1000
 441 
 442 void ObjectMonitor::EnterI(TRAPS) {
 443   Thread * const Self = THREAD;
 444   assert(Self->is_Java_thread(), "invariant");
 445   assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
 446 
 447   // Try the lock - TATAS
 448   if (TryLock (Self) > 0) {
 449     assert(_succ != Self, "invariant");
 450     assert(_owner == Self, "invariant");
 451     assert(_Responsible != Self, "invariant");
 452     return;
 453   }
 454 
 455   DeferredInitialize();
 456 
 457   // We try one round of spinning *before* enqueueing Self.
 458   //
 459   // If the _owner is ready but OFFPROC we could use a YieldTo()
 460   // operation to donate the remainder of this thread's quantum
 461   // to the owner.  This has subtle but beneficial affinity
 462   // effects.
 463 
 464   if (TrySpin (Self) > 0) {
 465     assert(_owner == Self, "invariant");
 466     assert(_succ != Self, "invariant");
 467     assert(_Responsible != Self, "invariant");
 468     return;
 469   }
 470 
 471   // The Spin failed -- Enqueue and park the thread ...
 472   assert(_succ != Self, "invariant");
 473   assert(_owner != Self, "invariant");
 474   assert(_Responsible != Self, "invariant");
 475 
 476   // Enqueue "Self" on ObjectMonitor's _cxq.
 477   //
 478   // Node acts as a proxy for Self.
 479   // As an aside, if were to ever rewrite the synchronization code mostly
 480   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 481   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 482   // as well as eliminate a subset of ABA issues.
 483   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 484 
 485   ObjectWaiter node(Self);
 486   Self->_ParkEvent->reset();
 487   node._prev   = (ObjectWaiter *) 0xBAD;
 488   node.TState  = ObjectWaiter::TS_CXQ;
 489 
 490   // Push "Self" onto the front of the _cxq.
 491   // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
 492   // Note that spinning tends to reduce the rate at which threads
 493   // enqueue and dequeue on EntryList|cxq.
 494   ObjectWaiter * nxt;
 495   for (;;) {
 496     node._next = nxt = _cxq;
 497     if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break;
 498 
 499     // Interference - the CAS failed because _cxq changed.  Just retry.
 500     // As an optional optimization we retry the lock.
 501     if (TryLock (Self) > 0) {
 502       assert(_succ != Self, "invariant");
 503       assert(_owner == Self, "invariant");
 504       assert(_Responsible != Self, "invariant");
 505       return;
 506     }
 507   }
 508 
 509   // Check for cxq|EntryList edge transition to non-null.  This indicates
 510   // the onset of contention.  While contention persists exiting threads
 511   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 512   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 513   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 514   // arrange for one of the contending thread to use a timed park() operations
 515   // to detect and recover from the race.  (Stranding is form of progress failure
 516   // where the monitor is unlocked but all the contending threads remain parked).
 517   // That is, at least one of the contended threads will periodically poll _owner.
 518   // One of the contending threads will become the designated "Responsible" thread.
 519   // The Responsible thread uses a timed park instead of a normal indefinite park
 520   // operation -- it periodically wakes and checks for and recovers from potential
 521   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 522   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 523   // be responsible for a monitor.
 524   //
 525   // Currently, one of the contended threads takes on the added role of "Responsible".
 526   // A viable alternative would be to use a dedicated "stranding checker" thread
 527   // that periodically iterated over all the threads (or active monitors) and unparked
 528   // successors where there was risk of stranding.  This would help eliminate the
 529   // timer scalability issues we see on some platforms as we'd only have one thread
 530   // -- the checker -- parked on a timer.
 531 
 532   if (nxt == NULL && _EntryList == NULL) {
 533     // Try to assume the role of responsible thread for the monitor.
 534     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=Self }
 535     Atomic::replace_if_null(Self, &_Responsible);
 536   }
 537 
 538   // The lock might have been released while this thread was occupied queueing
 539   // itself onto _cxq.  To close the race and avoid "stranding" and
 540   // progress-liveness failure we must resample-retry _owner before parking.
 541   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 542   // In this case the ST-MEMBAR is accomplished with CAS().
 543   //
 544   // TODO: Defer all thread state transitions until park-time.
 545   // Since state transitions are heavy and inefficient we'd like
 546   // to defer the state transitions until absolutely necessary,
 547   // and in doing so avoid some transitions ...
 548 
 549   int nWakeups = 0;
 550   int recheckInterval = 1;
 551 
 552   for (;;) {
 553 
 554     if (TryLock(Self) > 0) break;
 555     assert(_owner != Self, "invariant");
 556 
 557     // park self
 558     if (_Responsible == Self) {
 559       Self->_ParkEvent->park((jlong) recheckInterval);
 560       // Increase the recheckInterval, but clamp the value.
 561       recheckInterval *= 8;
 562       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 563         recheckInterval = MAX_RECHECK_INTERVAL;
 564       }
 565     } else {
 566       Self->_ParkEvent->park();
 567     }
 568 
 569     if (TryLock(Self) > 0) break;
 570 
 571     // The lock is still contested.
 572     // Keep a tally of the # of futile wakeups.
 573     // Note that the counter is not protected by a lock or updated by atomics.
 574     // That is by design - we trade "lossy" counters which are exposed to
 575     // races during updates for a lower probe effect.
 576 
 577     // This PerfData object can be used in parallel with a safepoint.
 578     // See the work around in PerfDataManager::destroy().
 579     OM_PERFDATA_OP(FutileWakeups, inc());
 580     ++nWakeups;
 581 
 582     // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
 583     // We can defer clearing _succ until after the spin completes
 584     // TrySpin() must tolerate being called with _succ == Self.
 585     // Try yet another round of adaptive spinning.
 586     if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
 587 
 588     // We can find that we were unpark()ed and redesignated _succ while
 589     // we were spinning.  That's harmless.  If we iterate and call park(),
 590     // park() will consume the event and return immediately and we'll
 591     // just spin again.  This pattern can repeat, leaving _succ to simply
 592     // spin on a CPU.  Enable Knob_ResetEvent to clear pending unparks().
 593     // Alternately, we can sample fired() here, and if set, forgo spinning
 594     // in the next iteration.
 595 
 596     if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
 597       Self->_ParkEvent->reset();
 598       OrderAccess::fence();
 599     }
 600     if (_succ == Self) _succ = NULL;
 601 
 602     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 603     OrderAccess::fence();
 604   }
 605 
 606   // Egress :
 607   // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
 608   // Normally we'll find Self on the EntryList .
 609   // From the perspective of the lock owner (this thread), the
 610   // EntryList is stable and cxq is prepend-only.
 611   // The head of cxq is volatile but the interior is stable.
 612   // In addition, Self.TState is stable.
 613 
 614   assert(_owner == Self, "invariant");
 615   assert(object() != NULL, "invariant");
 616   // I'd like to write:
 617   //   guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 618   // but as we're at a safepoint that's not safe.
 619 
 620   UnlinkAfterAcquire(Self, &node);
 621   if (_succ == Self) _succ = NULL;
 622 
 623   assert(_succ != Self, "invariant");
 624   if (_Responsible == Self) {
 625     _Responsible = NULL;
 626     OrderAccess::fence(); // Dekker pivot-point
 627 
 628     // We may leave threads on cxq|EntryList without a designated
 629     // "Responsible" thread.  This is benign.  When this thread subsequently
 630     // exits the monitor it can "see" such preexisting "old" threads --
 631     // threads that arrived on the cxq|EntryList before the fence, above --
 632     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 633     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 634     // non-null and elect a new "Responsible" timer thread.
 635     //
 636     // This thread executes:
 637     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 638     //    LD cxq|EntryList               (in subsequent exit)
 639     //
 640     // Entering threads in the slow/contended path execute:
 641     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 642     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 643     //
 644     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 645     // exit operation from floating above the ST Responsible=null.
 646   }
 647 
 648   // We've acquired ownership with CAS().
 649   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 650   // But since the CAS() this thread may have also stored into _succ,
 651   // EntryList, cxq or Responsible.  These meta-data updates must be
 652   // visible __before this thread subsequently drops the lock.
 653   // Consider what could occur if we didn't enforce this constraint --
 654   // STs to monitor meta-data and user-data could reorder with (become
 655   // visible after) the ST in exit that drops ownership of the lock.
 656   // Some other thread could then acquire the lock, but observe inconsistent
 657   // or old monitor meta-data and heap data.  That violates the JMM.
 658   // To that end, the 1-0 exit() operation must have at least STST|LDST
 659   // "release" barrier semantics.  Specifically, there must be at least a
 660   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 661   // the lock.   The barrier ensures that changes to monitor meta-data and data
 662   // protected by the lock will be visible before we release the lock, and
 663   // therefore before some other thread (CPU) has a chance to acquire the lock.
 664   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 665   //
 666   // Critically, any prior STs to _succ or EntryList must be visible before
 667   // the ST of null into _owner in the *subsequent* (following) corresponding
 668   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 669   // execute a serializing instruction.
 670 
 671   return;
 672 }
 673 
 674 // ReenterI() is a specialized inline form of the latter half of the
 675 // contended slow-path from EnterI().  We use ReenterI() only for
 676 // monitor reentry in wait().
 677 //
 678 // In the future we should reconcile EnterI() and ReenterI(), adding
 679 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
 680 // loop accordingly.
 681 
 682 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
 683   assert(Self != NULL, "invariant");
 684   assert(SelfNode != NULL, "invariant");
 685   assert(SelfNode->_thread == Self, "invariant");
 686   assert(_waiters > 0, "invariant");
 687   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 688   assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
 689   JavaThread * jt = (JavaThread *) Self;
 690 
 691   int nWakeups = 0;
 692   for (;;) {
 693     ObjectWaiter::TStates v = SelfNode->TState;
 694     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
 695     assert(_owner != Self, "invariant");
 696 
 697     if (TryLock(Self) > 0) break;
 698     if (TrySpin(Self) > 0) break;
 699 
 700     // State transition wrappers around park() ...
 701     // ReenterI() wisely defers state transitions until
 702     // it's clear we must park the thread.
 703     {
 704       OSThreadContendState osts(Self->osthread());
 705       ThreadBlockInVM tbivm(jt);
 706 
 707       // cleared by handle_special_suspend_equivalent_condition()
 708       // or java_suspend_self()
 709       jt->set_suspend_equivalent();
 710       Self->_ParkEvent->park();
 711 
 712       // were we externally suspended while we were waiting?
 713       for (;;) {
 714         if (!ExitSuspendEquivalent(jt)) break;
 715         if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
 716         jt->java_suspend_self();
 717         jt->set_suspend_equivalent();
 718       }
 719     }
 720 
 721     // Try again, but just so we distinguish between futile wakeups and
 722     // successful wakeups.  The following test isn't algorithmically
 723     // necessary, but it helps us maintain sensible statistics.
 724     if (TryLock(Self) > 0) break;
 725 
 726     // The lock is still contested.
 727     // Keep a tally of the # of futile wakeups.
 728     // Note that the counter is not protected by a lock or updated by atomics.
 729     // That is by design - we trade "lossy" counters which are exposed to
 730     // races during updates for a lower probe effect.
 731     ++nWakeups;
 732 
 733     // Assuming this is not a spurious wakeup we'll normally
 734     // find that _succ == Self.
 735     if (_succ == Self) _succ = NULL;
 736 
 737     // Invariant: after clearing _succ a contending thread
 738     // *must* retry  _owner before parking.
 739     OrderAccess::fence();
 740 
 741     // This PerfData object can be used in parallel with a safepoint.
 742     // See the work around in PerfDataManager::destroy().
 743     OM_PERFDATA_OP(FutileWakeups, inc());
 744   }
 745 
 746   // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
 747   // Normally we'll find Self on the EntryList.
 748   // Unlinking from the EntryList is constant-time and atomic-free.
 749   // From the perspective of the lock owner (this thread), the
 750   // EntryList is stable and cxq is prepend-only.
 751   // The head of cxq is volatile but the interior is stable.
 752   // In addition, Self.TState is stable.
 753 
 754   assert(_owner == Self, "invariant");
 755   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 756   UnlinkAfterAcquire(Self, SelfNode);
 757   if (_succ == Self) _succ = NULL;
 758   assert(_succ != Self, "invariant");
 759   SelfNode->TState = ObjectWaiter::TS_RUN;
 760   OrderAccess::fence();      // see comments at the end of EnterI()
 761 }
 762 
 763 // By convention we unlink a contending thread from EntryList|cxq immediately
 764 // after the thread acquires the lock in ::enter().  Equally, we could defer
 765 // unlinking the thread until ::exit()-time.
 766 
 767 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
 768   assert(_owner == Self, "invariant");
 769   assert(SelfNode->_thread == Self, "invariant");
 770 
 771   if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
 772     // Normal case: remove Self from the DLL EntryList .
 773     // This is a constant-time operation.
 774     ObjectWaiter * nxt = SelfNode->_next;
 775     ObjectWaiter * prv = SelfNode->_prev;
 776     if (nxt != NULL) nxt->_prev = prv;
 777     if (prv != NULL) prv->_next = nxt;
 778     if (SelfNode == _EntryList) _EntryList = nxt;
 779     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
 780     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
 781   } else {
 782     assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
 783     // Inopportune interleaving -- Self is still on the cxq.
 784     // This usually means the enqueue of self raced an exiting thread.
 785     // Normally we'll find Self near the front of the cxq, so
 786     // dequeueing is typically fast.  If needbe we can accelerate
 787     // this with some MCS/CHL-like bidirectional list hints and advisory
 788     // back-links so dequeueing from the interior will normally operate
 789     // in constant-time.
 790     // Dequeue Self from either the head (with CAS) or from the interior
 791     // with a linear-time scan and normal non-atomic memory operations.
 792     // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
 793     // and then unlink Self from EntryList.  We have to drain eventually,
 794     // so it might as well be now.
 795 
 796     ObjectWaiter * v = _cxq;
 797     assert(v != NULL, "invariant");
 798     if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) {
 799       // The CAS above can fail from interference IFF a "RAT" arrived.
 800       // In that case Self must be in the interior and can no longer be
 801       // at the head of cxq.
 802       if (v == SelfNode) {
 803         assert(_cxq != v, "invariant");
 804         v = _cxq;          // CAS above failed - start scan at head of list
 805       }
 806       ObjectWaiter * p;
 807       ObjectWaiter * q = NULL;
 808       for (p = v; p != NULL && p != SelfNode; p = p->_next) {
 809         q = p;
 810         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
 811       }
 812       assert(v != SelfNode, "invariant");
 813       assert(p == SelfNode, "Node not found on cxq");
 814       assert(p != _cxq, "invariant");
 815       assert(q != NULL, "invariant");
 816       assert(q->_next == p, "invariant");
 817       q->_next = p->_next;
 818     }
 819   }
 820 
 821 #ifdef ASSERT
 822   // Diagnostic hygiene ...
 823   SelfNode->_prev  = (ObjectWaiter *) 0xBAD;
 824   SelfNode->_next  = (ObjectWaiter *) 0xBAD;
 825   SelfNode->TState = ObjectWaiter::TS_RUN;
 826 #endif
 827 }
 828 
 829 // -----------------------------------------------------------------------------
 830 // Exit support
 831 //
 832 // exit()
 833 // ~~~~~~
 834 // Note that the collector can't reclaim the objectMonitor or deflate
 835 // the object out from underneath the thread calling ::exit() as the
 836 // thread calling ::exit() never transitions to a stable state.
 837 // This inhibits GC, which in turn inhibits asynchronous (and
 838 // inopportune) reclamation of "this".
 839 //
 840 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
 841 // There's one exception to the claim above, however.  EnterI() can call
 842 // exit() to drop a lock if the acquirer has been externally suspended.
 843 // In that case exit() is called with _thread_state as _thread_blocked,
 844 // but the monitor's _count field is > 0, which inhibits reclamation.
 845 //
 846 // 1-0 exit
 847 // ~~~~~~~~
 848 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
 849 // the fast-path operators have been optimized so the common ::exit()
 850 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
 851 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
 852 // greatly improves latency -- MEMBAR and CAS having considerable local
 853 // latency on modern processors -- but at the cost of "stranding".  Absent the
 854 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
 855 // ::enter() path, resulting in the entering thread being stranding
 856 // and a progress-liveness failure.   Stranding is extremely rare.
 857 // We use timers (timed park operations) & periodic polling to detect
 858 // and recover from stranding.  Potentially stranded threads periodically
 859 // wake up and poll the lock.  See the usage of the _Responsible variable.
 860 //
 861 // The CAS() in enter provides for safety and exclusion, while the CAS or
 862 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
 863 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
 864 // We detect and recover from stranding with timers.
 865 //
 866 // If a thread transiently strands it'll park until (a) another
 867 // thread acquires the lock and then drops the lock, at which time the
 868 // exiting thread will notice and unpark the stranded thread, or, (b)
 869 // the timer expires.  If the lock is high traffic then the stranding latency
 870 // will be low due to (a).  If the lock is low traffic then the odds of
 871 // stranding are lower, although the worst-case stranding latency
 872 // is longer.  Critically, we don't want to put excessive load in the
 873 // platform's timer subsystem.  We want to minimize both the timer injection
 874 // rate (timers created/sec) as well as the number of timers active at
 875 // any one time.  (more precisely, we want to minimize timer-seconds, which is
 876 // the integral of the # of active timers at any instant over time).
 877 // Both impinge on OS scalability.  Given that, at most one thread parked on
 878 // a monitor will use a timer.
 879 //
 880 // There is also the risk of a futile wake-up. If we drop the lock
 881 // another thread can reacquire the lock immediately, and we can
 882 // then wake a thread unnecessarily. This is benign, and we've
 883 // structured the code so the windows are short and the frequency
 884 // of such futile wakups is low.
 885 
 886 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
 887   Thread * const Self = THREAD;
 888   if (THREAD != _owner) {
 889     if (THREAD->is_lock_owned((address) _owner)) {
 890       // Transmute _owner from a BasicLock pointer to a Thread address.
 891       // We don't need to hold _mutex for this transition.
 892       // Non-null to Non-null is safe as long as all readers can
 893       // tolerate either flavor.
 894       assert(_recursions == 0, "invariant");
 895       _owner = THREAD;
 896       _recursions = 0;
 897     } else {
 898       // Apparent unbalanced locking ...
 899       // Naively we'd like to throw IllegalMonitorStateException.
 900       // As a practical matter we can neither allocate nor throw an
 901       // exception as ::exit() can be called from leaf routines.
 902       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
 903       // Upon deeper reflection, however, in a properly run JVM the only
 904       // way we should encounter this situation is in the presence of
 905       // unbalanced JNI locking. TODO: CheckJNICalls.
 906       // See also: CR4414101
 907       assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
 908       return;
 909     }
 910   }
 911 
 912   if (_recursions != 0) {
 913     _recursions--;        // this is simple recursive enter
 914     return;
 915   }
 916 
 917   // Invariant: after setting Responsible=null an thread must execute
 918   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
 919   _Responsible = NULL;
 920 
 921 #if INCLUDE_JFR
 922   // get the owner's thread id for the MonitorEnter event
 923   // if it is enabled and the thread isn't suspended
 924   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
 925     _previous_owner_tid = JFR_THREAD_ID(Self);
 926   }
 927 #endif
 928 
 929   for (;;) {
 930     assert(THREAD == _owner, "invariant");
 931 
 932     if (Knob_ExitPolicy == 0) {
 933       // release semantics: prior loads and stores from within the critical section
 934       // must not float (reorder) past the following store that drops the lock.
 935       // On SPARC that requires MEMBAR #loadstore|#storestore.
 936       // But of course in TSO #loadstore|#storestore is not required.
 937       // I'd like to write one of the following:
 938       // A.  OrderAccess::release() ; _owner = NULL
 939       // B.  OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
 940       // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
 941       // store into a _dummy variable.  That store is not needed, but can result
 942       // in massive wasteful coherency traffic on classic SMP systems.
 943       // Instead, I use release_store(), which is implemented as just a simple
 944       // ST on x64, x86 and SPARC.
 945       OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 946       OrderAccess::storeload();                        // See if we need to wake a successor
 947       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 948         return;
 949       }
 950       // Other threads are blocked trying to acquire the lock.
 951 
 952       // Normally the exiting thread is responsible for ensuring succession,
 953       // but if other successors are ready or other entering threads are spinning
 954       // then this thread can simply store NULL into _owner and exit without
 955       // waking a successor.  The existence of spinners or ready successors
 956       // guarantees proper succession (liveness).  Responsibility passes to the
 957       // ready or running successors.  The exiting thread delegates the duty.
 958       // More precisely, if a successor already exists this thread is absolved
 959       // of the responsibility of waking (unparking) one.
 960       //
 961       // The _succ variable is critical to reducing futile wakeup frequency.
 962       // _succ identifies the "heir presumptive" thread that has been made
 963       // ready (unparked) but that has not yet run.  We need only one such
 964       // successor thread to guarantee progress.
 965       // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 966       // section 3.3 "Futile Wakeup Throttling" for details.
 967       //
 968       // Note that spinners in Enter() also set _succ non-null.
 969       // In the current implementation spinners opportunistically set
 970       // _succ so that exiting threads might avoid waking a successor.
 971       // Another less appealing alternative would be for the exiting thread
 972       // to drop the lock and then spin briefly to see if a spinner managed
 973       // to acquire the lock.  If so, the exiting thread could exit
 974       // immediately without waking a successor, otherwise the exiting
 975       // thread would need to dequeue and wake a successor.
 976       // (Note that we'd need to make the post-drop spin short, but no
 977       // shorter than the worst-case round-trip cache-line migration time.
 978       // The dropped lock needs to become visible to the spinner, and then
 979       // the acquisition of the lock by the spinner must become visible to
 980       // the exiting thread).
 981 
 982       // It appears that an heir-presumptive (successor) must be made ready.
 983       // Only the current lock owner can manipulate the EntryList or
 984       // drain _cxq, so we need to reacquire the lock.  If we fail
 985       // to reacquire the lock the responsibility for ensuring succession
 986       // falls to the new owner.
 987       //
 988       if (!Atomic::replace_if_null(THREAD, &_owner)) {
 989         return;
 990       }
 991     } else {
 992       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 993         OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 994         OrderAccess::storeload();
 995         // Ratify the previously observed values.
 996         if (_cxq == NULL || _succ != NULL) {
 997           return;
 998         }
 999 
1000         // inopportune interleaving -- the exiting thread (this thread)
1001         // in the fast-exit path raced an entering thread in the slow-enter
1002         // path.
1003         // We have two choices:
1004         // A.  Try to reacquire the lock.
1005         //     If the CAS() fails return immediately, otherwise
1006         //     we either restart/rerun the exit operation, or simply
1007         //     fall-through into the code below which wakes a successor.
1008         // B.  If the elements forming the EntryList|cxq are TSM
1009         //     we could simply unpark() the lead thread and return
1010         //     without having set _succ.
1011         if (!Atomic::replace_if_null(THREAD, &_owner)) {
1012           return;
1013         }
1014       }
1015     }
1016 
1017     guarantee(_owner == THREAD, "invariant");
1018 
1019     ObjectWaiter * w = NULL;
1020     int QMode = Knob_QMode;
1021 
1022     if (QMode == 2 && _cxq != NULL) {
1023       // QMode == 2 : cxq has precedence over EntryList.
1024       // Try to directly wake a successor from the cxq.
1025       // If successful, the successor will need to unlink itself from cxq.
1026       w = _cxq;
1027       assert(w != NULL, "invariant");
1028       assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1029       ExitEpilog(Self, w);
1030       return;
1031     }
1032 
1033     if (QMode == 3 && _cxq != NULL) {
1034       // Aggressively drain cxq into EntryList at the first opportunity.
1035       // This policy ensure that recently-run threads live at the head of EntryList.
1036       // Drain _cxq into EntryList - bulk transfer.
1037       // First, detach _cxq.
1038       // The following loop is tantamount to: w = swap(&cxq, NULL)
1039       w = _cxq;
1040       for (;;) {
1041         assert(w != NULL, "Invariant");
1042         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1043         if (u == w) break;
1044         w = u;
1045       }
1046       assert(w != NULL, "invariant");
1047 
1048       ObjectWaiter * q = NULL;
1049       ObjectWaiter * p;
1050       for (p = w; p != NULL; p = p->_next) {
1051         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1052         p->TState = ObjectWaiter::TS_ENTER;
1053         p->_prev = q;
1054         q = p;
1055       }
1056 
1057       // Append the RATs to the EntryList
1058       // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1059       ObjectWaiter * Tail;
1060       for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1061            Tail = Tail->_next)
1062         /* empty */;
1063       if (Tail == NULL) {
1064         _EntryList = w;
1065       } else {
1066         Tail->_next = w;
1067         w->_prev = Tail;
1068       }
1069 
1070       // Fall thru into code that tries to wake a successor from EntryList
1071     }
1072 
1073     if (QMode == 4 && _cxq != NULL) {
1074       // Aggressively drain cxq into EntryList at the first opportunity.
1075       // This policy ensure that recently-run threads live at the head of EntryList.
1076 
1077       // Drain _cxq into EntryList - bulk transfer.
1078       // First, detach _cxq.
1079       // The following loop is tantamount to: w = swap(&cxq, NULL)
1080       w = _cxq;
1081       for (;;) {
1082         assert(w != NULL, "Invariant");
1083         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1084         if (u == w) break;
1085         w = u;
1086       }
1087       assert(w != NULL, "invariant");
1088 
1089       ObjectWaiter * q = NULL;
1090       ObjectWaiter * p;
1091       for (p = w; p != NULL; p = p->_next) {
1092         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1093         p->TState = ObjectWaiter::TS_ENTER;
1094         p->_prev = q;
1095         q = p;
1096       }
1097 
1098       // Prepend the RATs to the EntryList
1099       if (_EntryList != NULL) {
1100         q->_next = _EntryList;
1101         _EntryList->_prev = q;
1102       }
1103       _EntryList = w;
1104 
1105       // Fall thru into code that tries to wake a successor from EntryList
1106     }
1107 
1108     w = _EntryList;
1109     if (w != NULL) {
1110       // I'd like to write: guarantee (w->_thread != Self).
1111       // But in practice an exiting thread may find itself on the EntryList.
1112       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1113       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1114       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1115       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1116       // release the lock "O".  T2 resumes immediately after the ST of null into
1117       // _owner, above.  T2 notices that the EntryList is populated, so it
1118       // reacquires the lock and then finds itself on the EntryList.
1119       // Given all that, we have to tolerate the circumstance where "w" is
1120       // associated with Self.
1121       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1122       ExitEpilog(Self, w);
1123       return;
1124     }
1125 
1126     // If we find that both _cxq and EntryList are null then just
1127     // re-run the exit protocol from the top.
1128     w = _cxq;
1129     if (w == NULL) continue;
1130 
1131     // Drain _cxq into EntryList - bulk transfer.
1132     // First, detach _cxq.
1133     // The following loop is tantamount to: w = swap(&cxq, NULL)
1134     for (;;) {
1135       assert(w != NULL, "Invariant");
1136       ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1137       if (u == w) break;
1138       w = u;
1139     }
1140 
1141     assert(w != NULL, "invariant");
1142     assert(_EntryList == NULL, "invariant");
1143 
1144     // Convert the LIFO SLL anchored by _cxq into a DLL.
1145     // The list reorganization step operates in O(LENGTH(w)) time.
1146     // It's critical that this step operate quickly as
1147     // "Self" still holds the outer-lock, restricting parallelism
1148     // and effectively lengthening the critical section.
1149     // Invariant: s chases t chases u.
1150     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1151     // we have faster access to the tail.
1152 
1153     if (QMode == 1) {
1154       // QMode == 1 : drain cxq to EntryList, reversing order
1155       // We also reverse the order of the list.
1156       ObjectWaiter * s = NULL;
1157       ObjectWaiter * t = w;
1158       ObjectWaiter * u = NULL;
1159       while (t != NULL) {
1160         guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1161         t->TState = ObjectWaiter::TS_ENTER;
1162         u = t->_next;
1163         t->_prev = u;
1164         t->_next = s;
1165         s = t;
1166         t = u;
1167       }
1168       _EntryList  = s;
1169       assert(s != NULL, "invariant");
1170     } else {
1171       // QMode == 0 or QMode == 2
1172       _EntryList = w;
1173       ObjectWaiter * q = NULL;
1174       ObjectWaiter * p;
1175       for (p = w; p != NULL; p = p->_next) {
1176         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1177         p->TState = ObjectWaiter::TS_ENTER;
1178         p->_prev = q;
1179         q = p;
1180       }
1181     }
1182 
1183     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1184     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1185 
1186     // See if we can abdicate to a spinner instead of waking a thread.
1187     // A primary goal of the implementation is to reduce the
1188     // context-switch rate.
1189     if (_succ != NULL) continue;
1190 
1191     w = _EntryList;
1192     if (w != NULL) {
1193       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1194       ExitEpilog(Self, w);
1195       return;
1196     }
1197   }
1198 }
1199 
1200 // ExitSuspendEquivalent:
1201 // A faster alternate to handle_special_suspend_equivalent_condition()
1202 //
1203 // handle_special_suspend_equivalent_condition() unconditionally
1204 // acquires the SR_lock.  On some platforms uncontended MutexLocker()
1205 // operations have high latency.  Note that in ::enter() we call HSSEC
1206 // while holding the monitor, so we effectively lengthen the critical sections.
1207 //
1208 // There are a number of possible solutions:
1209 //
1210 // A.  To ameliorate the problem we might also defer state transitions
1211 //     to as late as possible -- just prior to parking.
1212 //     Given that, we'd call HSSEC after having returned from park(),
1213 //     but before attempting to acquire the monitor.  This is only a
1214 //     partial solution.  It avoids calling HSSEC while holding the
1215 //     monitor (good), but it still increases successor reacquisition latency --
1216 //     the interval between unparking a successor and the time the successor
1217 //     resumes and retries the lock.  See ReenterI(), which defers state transitions.
1218 //     If we use this technique we can also avoid EnterI()-exit() loop
1219 //     in ::enter() where we iteratively drop the lock and then attempt
1220 //     to reacquire it after suspending.
1221 //
1222 // B.  In the future we might fold all the suspend bits into a
1223 //     composite per-thread suspend flag and then update it with CAS().
1224 //     Alternately, a Dekker-like mechanism with multiple variables
1225 //     would suffice:
1226 //       ST Self->_suspend_equivalent = false
1227 //       MEMBAR
1228 //       LD Self_>_suspend_flags
1229 //
1230 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1231 // with a more efficient implementation, the need to use "FastHSSEC" has
1232 // decreased. - Dave
1233 
1234 
1235 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1236   const int Mode = Knob_FastHSSEC;
1237   if (Mode && !jSelf->is_external_suspend()) {
1238     assert(jSelf->is_suspend_equivalent(), "invariant");
1239     jSelf->clear_suspend_equivalent();
1240     if (2 == Mode) OrderAccess::storeload();
1241     if (!jSelf->is_external_suspend()) return false;
1242     // We raced a suspension -- fall thru into the slow path
1243     jSelf->set_suspend_equivalent();
1244   }
1245   return jSelf->handle_special_suspend_equivalent_condition();
1246 }
1247 
1248 
1249 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1250   assert(_owner == Self, "invariant");
1251 
1252   // Exit protocol:
1253   // 1. ST _succ = wakee
1254   // 2. membar #loadstore|#storestore;
1255   // 2. ST _owner = NULL
1256   // 3. unpark(wakee)
1257 
1258   _succ = Knob_SuccEnabled ? Wakee->_thread : NULL;
1259   ParkEvent * Trigger = Wakee->_event;
1260 
1261   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1262   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1263   // out-of-scope (non-extant).
1264   Wakee  = NULL;
1265 
1266   // Drop the lock
1267   OrderAccess::release_store(&_owner, (void*)NULL);
1268   OrderAccess::fence();                               // ST _owner vs LD in unpark()
1269 
1270   DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1271   Trigger->unpark();
1272 
1273   // Maintain stats and report events to JVMTI
1274   OM_PERFDATA_OP(Parks, inc());
1275 }
1276 
1277 
1278 // -----------------------------------------------------------------------------
1279 // Class Loader deadlock handling.
1280 //
1281 // complete_exit exits a lock returning recursion count
1282 // complete_exit/reenter operate as a wait without waiting
1283 // complete_exit requires an inflated monitor
1284 // The _owner field is not always the Thread addr even with an
1285 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1286 // thread due to contention.
1287 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1288   Thread * const Self = THREAD;
1289   assert(Self->is_Java_thread(), "Must be Java thread!");
1290   JavaThread *jt = (JavaThread *)THREAD;
1291 
1292   DeferredInitialize();
1293 
1294   if (THREAD != _owner) {
1295     if (THREAD->is_lock_owned ((address)_owner)) {
1296       assert(_recursions == 0, "internal state error");
1297       _owner = THREAD;   // Convert from basiclock addr to Thread addr
1298       _recursions = 0;
1299     }
1300   }
1301 
1302   guarantee(Self == _owner, "complete_exit not owner");
1303   intptr_t save = _recursions; // record the old recursion count
1304   _recursions = 0;        // set the recursion level to be 0
1305   exit(true, Self);           // exit the monitor
1306   guarantee(_owner != Self, "invariant");
1307   return save;
1308 }
1309 
1310 // reenter() enters a lock and sets recursion count
1311 // complete_exit/reenter operate as a wait without waiting
1312 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1313   Thread * const Self = THREAD;
1314   assert(Self->is_Java_thread(), "Must be Java thread!");
1315   JavaThread *jt = (JavaThread *)THREAD;
1316 
1317   guarantee(_owner != Self, "reenter already owner");
1318   enter(THREAD);       // enter the monitor
1319   guarantee(_recursions == 0, "reenter recursion");
1320   _recursions = recursions;
1321   return;
1322 }
1323 
1324 
1325 // -----------------------------------------------------------------------------
1326 // A macro is used below because there may already be a pending
1327 // exception which should not abort the execution of the routines
1328 // which use this (which is why we don't put this into check_slow and
1329 // call it with a CHECK argument).
1330 
1331 #define CHECK_OWNER()                                                       \
1332   do {                                                                      \
1333     if (THREAD != _owner) {                                                 \
1334       if (THREAD->is_lock_owned((address) _owner)) {                        \
1335         _owner = THREAD;  /* Convert from basiclock addr to Thread addr */  \
1336         _recursions = 0;                                                    \
1337       } else {                                                              \
1338         THROW(vmSymbols::java_lang_IllegalMonitorStateException());         \
1339       }                                                                     \
1340     }                                                                       \
1341   } while (false)
1342 
1343 // check_slow() is a misnomer.  It's called to simply to throw an IMSX exception.
1344 // TODO-FIXME: remove check_slow() -- it's likely dead.
1345 
1346 void ObjectMonitor::check_slow(TRAPS) {
1347   assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1348   THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1349 }
1350 
1351 static int Adjust(volatile int * adr, int dx) {
1352   int v;
1353   for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1354   return v;
1355 }
1356 
1357 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1358                                     ObjectMonitor* monitor,
1359                                     jlong notifier_tid,
1360                                     jlong timeout,
1361                                     bool timedout) {
1362   assert(event != NULL, "invariant");
1363   assert(monitor != NULL, "invariant");
1364   event->set_monitorClass(((oop)monitor->object())->klass());
1365   event->set_timeout(timeout);
1366   event->set_address((uintptr_t)monitor->object_addr());
1367   event->set_notifier(notifier_tid);
1368   event->set_timedOut(timedout);
1369   event->commit();
1370 }
1371 
1372 // -----------------------------------------------------------------------------
1373 // Wait/Notify/NotifyAll
1374 //
1375 // Note: a subset of changes to ObjectMonitor::wait()
1376 // will need to be replicated in complete_exit
1377 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1378   Thread * const Self = THREAD;
1379   assert(Self->is_Java_thread(), "Must be Java thread!");
1380   JavaThread *jt = (JavaThread *)THREAD;
1381 
1382   DeferredInitialize();
1383 
1384   // Throw IMSX or IEX.
1385   CHECK_OWNER();
1386 
1387   EventJavaMonitorWait event;
1388 
1389   // check for a pending interrupt
1390   if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1391     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1392     if (JvmtiExport::should_post_monitor_waited()) {
1393       // Note: 'false' parameter is passed here because the
1394       // wait was not timed out due to thread interrupt.
1395       JvmtiExport::post_monitor_waited(jt, this, false);
1396 
1397       // In this short circuit of the monitor wait protocol, the
1398       // current thread never drops ownership of the monitor and
1399       // never gets added to the wait queue so the current thread
1400       // cannot be made the successor. This means that the
1401       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1402       // consume an unpark() meant for the ParkEvent associated with
1403       // this ObjectMonitor.
1404     }
1405     if (event.should_commit()) {
1406       post_monitor_wait_event(&event, this, 0, millis, false);
1407     }
1408     THROW(vmSymbols::java_lang_InterruptedException());
1409     return;
1410   }
1411 
1412   assert(Self->_Stalled == 0, "invariant");
1413   Self->_Stalled = intptr_t(this);
1414   jt->set_current_waiting_monitor(this);
1415 
1416   // create a node to be put into the queue
1417   // Critically, after we reset() the event but prior to park(), we must check
1418   // for a pending interrupt.
1419   ObjectWaiter node(Self);
1420   node.TState = ObjectWaiter::TS_WAIT;
1421   Self->_ParkEvent->reset();
1422   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1423 
1424   // Enter the waiting queue, which is a circular doubly linked list in this case
1425   // but it could be a priority queue or any data structure.
1426   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1427   // by the the owner of the monitor *except* in the case where park()
1428   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1429   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1430 
1431   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1432   AddWaiter(&node);
1433   Thread::SpinRelease(&_WaitSetLock);
1434 
1435   _Responsible = NULL;
1436 
1437   intptr_t save = _recursions; // record the old recursion count
1438   _waiters++;                  // increment the number of waiters
1439   _recursions = 0;             // set the recursion level to be 1
1440   exit(true, Self);                    // exit the monitor
1441   guarantee(_owner != Self, "invariant");
1442 
1443   // The thread is on the WaitSet list - now park() it.
1444   // On MP systems it's conceivable that a brief spin before we park
1445   // could be profitable.
1446   //
1447   // TODO-FIXME: change the following logic to a loop of the form
1448   //   while (!timeout && !interrupted && _notified == 0) park()
1449 
1450   int ret = OS_OK;
1451   int WasNotified = 0;
1452   { // State transition wrappers
1453     OSThread* osthread = Self->osthread();
1454     OSThreadWaitState osts(osthread, true);
1455     {
1456       ThreadBlockInVM tbivm(jt);
1457       // Thread is in thread_blocked state and oop access is unsafe.
1458       jt->set_suspend_equivalent();
1459 
1460       if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1461         // Intentionally empty
1462       } else if (node._notified == 0) {
1463         if (millis <= 0) {
1464           Self->_ParkEvent->park();
1465         } else {
1466           ret = Self->_ParkEvent->park(millis);
1467         }
1468       }
1469 
1470       // were we externally suspended while we were waiting?
1471       if (ExitSuspendEquivalent (jt)) {
1472         // TODO-FIXME: add -- if succ == Self then succ = null.
1473         jt->java_suspend_self();
1474       }
1475 
1476     } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1477 
1478     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1479     // from the WaitSet to the EntryList.
1480     // See if we need to remove Node from the WaitSet.
1481     // We use double-checked locking to avoid grabbing _WaitSetLock
1482     // if the thread is not on the wait queue.
1483     //
1484     // Note that we don't need a fence before the fetch of TState.
1485     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1486     // written by the is thread. (perhaps the fetch might even be satisfied
1487     // by a look-aside into the processor's own store buffer, although given
1488     // the length of the code path between the prior ST and this load that's
1489     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1490     // then we'll acquire the lock and then re-fetch a fresh TState value.
1491     // That is, we fail toward safety.
1492 
1493     if (node.TState == ObjectWaiter::TS_WAIT) {
1494       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1495       if (node.TState == ObjectWaiter::TS_WAIT) {
1496         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1497         assert(node._notified == 0, "invariant");
1498         node.TState = ObjectWaiter::TS_RUN;
1499       }
1500       Thread::SpinRelease(&_WaitSetLock);
1501     }
1502 
1503     // The thread is now either on off-list (TS_RUN),
1504     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1505     // The Node's TState variable is stable from the perspective of this thread.
1506     // No other threads will asynchronously modify TState.
1507     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1508     OrderAccess::loadload();
1509     if (_succ == Self) _succ = NULL;
1510     WasNotified = node._notified;
1511 
1512     // Reentry phase -- reacquire the monitor.
1513     // re-enter contended monitor after object.wait().
1514     // retain OBJECT_WAIT state until re-enter successfully completes
1515     // Thread state is thread_in_vm and oop access is again safe,
1516     // although the raw address of the object may have changed.
1517     // (Don't cache naked oops over safepoints, of course).
1518 
1519     // post monitor waited event. Note that this is past-tense, we are done waiting.
1520     if (JvmtiExport::should_post_monitor_waited()) {
1521       JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1522 
1523       if (node._notified != 0 && _succ == Self) {
1524         // In this part of the monitor wait-notify-reenter protocol it
1525         // is possible (and normal) for another thread to do a fastpath
1526         // monitor enter-exit while this thread is still trying to get
1527         // to the reenter portion of the protocol.
1528         //
1529         // The ObjectMonitor was notified and the current thread is
1530         // the successor which also means that an unpark() has already
1531         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1532         // consume the unpark() that was done when the successor was
1533         // set because the same ParkEvent is shared between Java
1534         // monitors and JVM/TI RawMonitors (for now).
1535         //
1536         // We redo the unpark() to ensure forward progress, i.e., we
1537         // don't want all pending threads hanging (parked) with none
1538         // entering the unlocked monitor.
1539         node._event->unpark();
1540       }
1541     }
1542 
1543     if (event.should_commit()) {
1544       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1545     }
1546 
1547     OrderAccess::fence();
1548 
1549     assert(Self->_Stalled != 0, "invariant");
1550     Self->_Stalled = 0;
1551 
1552     assert(_owner != Self, "invariant");
1553     ObjectWaiter::TStates v = node.TState;
1554     if (v == ObjectWaiter::TS_RUN) {
1555       enter(Self);
1556     } else {
1557       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1558       ReenterI(Self, &node);
1559       node.wait_reenter_end(this);
1560     }
1561 
1562     // Self has reacquired the lock.
1563     // Lifecycle - the node representing Self must not appear on any queues.
1564     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1565     // want residual elements associated with this thread left on any lists.
1566     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1567     assert(_owner == Self, "invariant");
1568     assert(_succ != Self, "invariant");
1569   } // OSThreadWaitState()
1570 
1571   jt->set_current_waiting_monitor(NULL);
1572 
1573   guarantee(_recursions == 0, "invariant");
1574   _recursions = save;     // restore the old recursion count
1575   _waiters--;             // decrement the number of waiters
1576 
1577   // Verify a few postconditions
1578   assert(_owner == Self, "invariant");
1579   assert(_succ != Self, "invariant");
1580   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1581 
1582   // check if the notification happened
1583   if (!WasNotified) {
1584     // no, it could be timeout or Thread.interrupt() or both
1585     // check for interrupt event, otherwise it is timeout
1586     if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1587       THROW(vmSymbols::java_lang_InterruptedException());
1588     }
1589   }
1590 
1591   // NOTE: Spurious wake up will be consider as timeout.
1592   // Monitor notify has precedence over thread interrupt.
1593 }
1594 
1595 
1596 // Consider:
1597 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1598 // then instead of transferring a thread from the WaitSet to the EntryList
1599 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1600 
1601 void ObjectMonitor::INotify(Thread * Self) {
1602   const int policy = Knob_MoveNotifyee;
1603 
1604   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1605   ObjectWaiter * iterator = DequeueWaiter();
1606   if (iterator != NULL) {
1607     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1608     guarantee(iterator->_notified == 0, "invariant");
1609     // Disposition - what might we do with iterator ?
1610     // a.  add it directly to the EntryList - either tail (policy == 1)
1611     //     or head (policy == 0).
1612     // b.  push it onto the front of the _cxq (policy == 2).
1613     // For now we use (b).
1614     if (policy != 4) {
1615       iterator->TState = ObjectWaiter::TS_ENTER;
1616     }
1617     iterator->_notified = 1;
1618     iterator->_notifier_tid = JFR_THREAD_ID(Self);
1619 
1620     ObjectWaiter * list = _EntryList;
1621     if (list != NULL) {
1622       assert(list->_prev == NULL, "invariant");
1623       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1624       assert(list != iterator, "invariant");
1625     }
1626 
1627     if (policy == 0) {       // prepend to EntryList
1628       if (list == NULL) {
1629         iterator->_next = iterator->_prev = NULL;
1630         _EntryList = iterator;
1631       } else {
1632         list->_prev = iterator;
1633         iterator->_next = list;
1634         iterator->_prev = NULL;
1635         _EntryList = iterator;
1636       }
1637     } else if (policy == 1) {      // append to EntryList
1638       if (list == NULL) {
1639         iterator->_next = iterator->_prev = NULL;
1640         _EntryList = iterator;
1641       } else {
1642         // CONSIDER:  finding the tail currently requires a linear-time walk of
1643         // the EntryList.  We can make tail access constant-time by converting to
1644         // a CDLL instead of using our current DLL.
1645         ObjectWaiter * tail;
1646         for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1647         assert(tail != NULL && tail->_next == NULL, "invariant");
1648         tail->_next = iterator;
1649         iterator->_prev = tail;
1650         iterator->_next = NULL;
1651       }
1652     } else if (policy == 2) {      // prepend to cxq
1653       if (list == NULL) {
1654         iterator->_next = iterator->_prev = NULL;
1655         _EntryList = iterator;
1656       } else {
1657         iterator->TState = ObjectWaiter::TS_CXQ;
1658         for (;;) {
1659           ObjectWaiter * front = _cxq;
1660           iterator->_next = front;
1661           if (Atomic::cmpxchg(iterator, &_cxq, front) == front) {
1662             break;
1663           }
1664         }
1665       }
1666     } else if (policy == 3) {      // append to cxq
1667       iterator->TState = ObjectWaiter::TS_CXQ;
1668       for (;;) {
1669         ObjectWaiter * tail = _cxq;
1670         if (tail == NULL) {
1671           iterator->_next = NULL;
1672           if (Atomic::replace_if_null(iterator, &_cxq)) {
1673             break;
1674           }
1675         } else {
1676           while (tail->_next != NULL) tail = tail->_next;
1677           tail->_next = iterator;
1678           iterator->_prev = tail;
1679           iterator->_next = NULL;
1680           break;
1681         }
1682       }
1683     } else {
1684       ParkEvent * ev = iterator->_event;
1685       iterator->TState = ObjectWaiter::TS_RUN;
1686       OrderAccess::fence();
1687       ev->unpark();
1688     }
1689 
1690     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1691     // move the add-to-EntryList operation, above, outside the critical section
1692     // protected by _WaitSetLock.  In practice that's not useful.  With the
1693     // exception of  wait() timeouts and interrupts the monitor owner
1694     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1695     // on _WaitSetLock so it's not profitable to reduce the length of the
1696     // critical section.
1697 
1698     if (policy < 4) {
1699       iterator->wait_reenter_begin(this);
1700     }
1701   }
1702   Thread::SpinRelease(&_WaitSetLock);
1703 }
1704 
1705 // Consider: a not-uncommon synchronization bug is to use notify() when
1706 // notifyAll() is more appropriate, potentially resulting in stranded
1707 // threads; this is one example of a lost wakeup. A useful diagnostic
1708 // option is to force all notify() operations to behave as notifyAll().
1709 //
1710 // Note: We can also detect many such problems with a "minimum wait".
1711 // When the "minimum wait" is set to a small non-zero timeout value
1712 // and the program does not hang whereas it did absent "minimum wait",
1713 // that suggests a lost wakeup bug.
1714 
1715 void ObjectMonitor::notify(TRAPS) {
1716   CHECK_OWNER();
1717   if (_WaitSet == NULL) {
1718     return;
1719   }
1720   DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1721   INotify(THREAD);
1722   OM_PERFDATA_OP(Notifications, inc(1));
1723 }
1724 
1725 
1726 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1727 // from the waitset to the EntryList. This could be done more efficiently with a
1728 // single bulk transfer but in practice it's not time-critical. Beware too,
1729 // that in prepend-mode we invert the order of the waiters. Let's say that the
1730 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1731 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1732 
1733 void ObjectMonitor::notifyAll(TRAPS) {
1734   CHECK_OWNER();
1735   if (_WaitSet == NULL) {
1736     return;
1737   }
1738 
1739   DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1740   int tally = 0;
1741   while (_WaitSet != NULL) {
1742     tally++;
1743     INotify(THREAD);
1744   }
1745 
1746   OM_PERFDATA_OP(Notifications, inc(tally));
1747 }
1748 
1749 // -----------------------------------------------------------------------------
1750 // Adaptive Spinning Support
1751 //
1752 // Adaptive spin-then-block - rational spinning
1753 //
1754 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1755 // algorithm.  On high order SMP systems it would be better to start with
1756 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1757 // a contending thread could enqueue itself on the cxq and then spin locally
1758 // on a thread-specific variable such as its ParkEvent._Event flag.
1759 // That's left as an exercise for the reader.  Note that global spinning is
1760 // not problematic on Niagara, as the L2 cache serves the interconnect and
1761 // has both low latency and massive bandwidth.
1762 //
1763 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1764 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1765 // (duration) or we can fix the count at approximately the duration of
1766 // a context switch and vary the frequency.   Of course we could also
1767 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1768 // For a description of 'Adaptive spin-then-block mutual exclusion in
1769 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1770 //
1771 // This implementation varies the duration "D", where D varies with
1772 // the success rate of recent spin attempts. (D is capped at approximately
1773 // length of a round-trip context switch).  The success rate for recent
1774 // spin attempts is a good predictor of the success rate of future spin
1775 // attempts.  The mechanism adapts automatically to varying critical
1776 // section length (lock modality), system load and degree of parallelism.
1777 // D is maintained per-monitor in _SpinDuration and is initialized
1778 // optimistically.  Spin frequency is fixed at 100%.
1779 //
1780 // Note that _SpinDuration is volatile, but we update it without locks
1781 // or atomics.  The code is designed so that _SpinDuration stays within
1782 // a reasonable range even in the presence of races.  The arithmetic
1783 // operations on _SpinDuration are closed over the domain of legal values,
1784 // so at worst a race will install and older but still legal value.
1785 // At the very worst this introduces some apparent non-determinism.
1786 // We might spin when we shouldn't or vice-versa, but since the spin
1787 // count are relatively short, even in the worst case, the effect is harmless.
1788 //
1789 // Care must be taken that a low "D" value does not become an
1790 // an absorbing state.  Transient spinning failures -- when spinning
1791 // is overall profitable -- should not cause the system to converge
1792 // on low "D" values.  We want spinning to be stable and predictable
1793 // and fairly responsive to change and at the same time we don't want
1794 // it to oscillate, become metastable, be "too" non-deterministic,
1795 // or converge on or enter undesirable stable absorbing states.
1796 //
1797 // We implement a feedback-based control system -- using past behavior
1798 // to predict future behavior.  We face two issues: (a) if the
1799 // input signal is random then the spin predictor won't provide optimal
1800 // results, and (b) if the signal frequency is too high then the control
1801 // system, which has some natural response lag, will "chase" the signal.
1802 // (b) can arise from multimodal lock hold times.  Transient preemption
1803 // can also result in apparent bimodal lock hold times.
1804 // Although sub-optimal, neither condition is particularly harmful, as
1805 // in the worst-case we'll spin when we shouldn't or vice-versa.
1806 // The maximum spin duration is rather short so the failure modes aren't bad.
1807 // To be conservative, I've tuned the gain in system to bias toward
1808 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1809 // "rings" or oscillates between spinning and not spinning.  This happens
1810 // when spinning is just on the cusp of profitability, however, so the
1811 // situation is not dire.  The state is benign -- there's no need to add
1812 // hysteresis control to damp the transition rate between spinning and
1813 // not spinning.
1814 
1815 // Spinning: Fixed frequency (100%), vary duration
1816 int ObjectMonitor::TrySpin(Thread * Self) {
1817   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1818   int ctr = Knob_FixedSpin;
1819   if (ctr != 0) {
1820     while (--ctr >= 0) {
1821       if (TryLock(Self) > 0) return 1;
1822       SpinPause();
1823     }
1824     return 0;
1825   }
1826 
1827   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1828     if (TryLock(Self) > 0) {
1829       // Increase _SpinDuration ...
1830       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1831       // Raising _SpurDuration to the poverty line is key.
1832       int x = _SpinDuration;
1833       if (x < Knob_SpinLimit) {
1834         if (x < Knob_Poverty) x = Knob_Poverty;
1835         _SpinDuration = x + Knob_BonusB;
1836       }
1837       return 1;
1838     }
1839     SpinPause();
1840   }
1841 
1842   // Admission control - verify preconditions for spinning
1843   //
1844   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1845   // becoming an absorbing state.  Put another way, we spin briefly to
1846   // sample, just in case the system load, parallelism, contention, or lock
1847   // modality changed.
1848   //
1849   // Consider the following alternative:
1850   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1851   // spin attempt.  "Periodically" might mean after a tally of
1852   // the # of failed spin attempts (or iterations) reaches some threshold.
1853   // This takes us into the realm of 1-out-of-N spinning, where we
1854   // hold the duration constant but vary the frequency.
1855 
1856   ctr = _SpinDuration;
1857   if (ctr < Knob_SpinBase) ctr = Knob_SpinBase;
1858   if (ctr <= 0) return 0;
1859 
1860   if (Knob_SuccRestrict && _succ != NULL) return 0;
1861   if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1862     return 0;
1863   }
1864 
1865   int MaxSpin = Knob_MaxSpinners;
1866   if (MaxSpin >= 0) {
1867     if (_Spinner > MaxSpin) {
1868       return 0;
1869     }
1870     // Slightly racy, but benign ...
1871     Adjust(&_Spinner, 1);
1872   }
1873 
1874   // We're good to spin ... spin ingress.
1875   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1876   // when preparing to LD...CAS _owner, etc and the CAS is likely
1877   // to succeed.
1878   int hits    = 0;
1879   int msk     = 0;
1880   int caspty  = Knob_CASPenalty;
1881   int oxpty   = Knob_OXPenalty;
1882   int sss     = Knob_SpinSetSucc;
1883   if (sss && _succ == NULL) _succ = Self;
1884   Thread * prv = NULL;
1885 
1886   // There are three ways to exit the following loop:
1887   // 1.  A successful spin where this thread has acquired the lock.
1888   // 2.  Spin failure with prejudice
1889   // 3.  Spin failure without prejudice
1890 
1891   while (--ctr >= 0) {
1892 
1893     // Periodic polling -- Check for pending GC
1894     // Threads may spin while they're unsafe.
1895     // We don't want spinning threads to delay the JVM from reaching
1896     // a stop-the-world safepoint or to steal cycles from GC.
1897     // If we detect a pending safepoint we abort in order that
1898     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1899     // this thread, if safe, doesn't steal cycles from GC.
1900     // This is in keeping with the "no loitering in runtime" rule.
1901     // We periodically check to see if there's a safepoint pending.
1902     if ((ctr & 0xFF) == 0) {
1903       if (SafepointMechanism::poll(Self)) {
1904         goto Abort;           // abrupt spin egress
1905       }
1906       if (Knob_UsePause & 1) SpinPause();
1907     }
1908 
1909     if (Knob_UsePause & 2) SpinPause();
1910 
1911     // Exponential back-off ...  Stay off the bus to reduce coherency traffic.
1912     // This is useful on classic SMP systems, but is of less utility on
1913     // N1-style CMT platforms.
1914     //
1915     // Trade-off: lock acquisition latency vs coherency bandwidth.
1916     // Lock hold times are typically short.  A histogram
1917     // of successful spin attempts shows that we usually acquire
1918     // the lock early in the spin.  That suggests we want to
1919     // sample _owner frequently in the early phase of the spin,
1920     // but then back-off and sample less frequently as the spin
1921     // progresses.  The back-off makes a good citizen on SMP big
1922     // SMP systems.  Oversampling _owner can consume excessive
1923     // coherency bandwidth.  Relatedly, if we _oversample _owner we
1924     // can inadvertently interfere with the the ST m->owner=null.
1925     // executed by the lock owner.
1926     if (ctr & msk) continue;
1927     ++hits;
1928     if ((hits & 0xF) == 0) {
1929       // The 0xF, above, corresponds to the exponent.
1930       // Consider: (msk+1)|msk
1931       msk = ((msk << 2)|3) & BackOffMask;
1932     }
1933 
1934     // Probe _owner with TATAS
1935     // If this thread observes the monitor transition or flicker
1936     // from locked to unlocked to locked, then the odds that this
1937     // thread will acquire the lock in this spin attempt go down
1938     // considerably.  The same argument applies if the CAS fails
1939     // or if we observe _owner change from one non-null value to
1940     // another non-null value.   In such cases we might abort
1941     // the spin without prejudice or apply a "penalty" to the
1942     // spin count-down variable "ctr", reducing it by 100, say.
1943 
1944     Thread * ox = (Thread *) _owner;
1945     if (ox == NULL) {
1946       ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL);
1947       if (ox == NULL) {
1948         // The CAS succeeded -- this thread acquired ownership
1949         // Take care of some bookkeeping to exit spin state.
1950         if (sss && _succ == Self) {
1951           _succ = NULL;
1952         }
1953         if (MaxSpin > 0) Adjust(&_Spinner, -1);
1954 
1955         // Increase _SpinDuration :
1956         // The spin was successful (profitable) so we tend toward
1957         // longer spin attempts in the future.
1958         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1959         // If we acquired the lock early in the spin cycle it
1960         // makes sense to increase _SpinDuration proportionally.
1961         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1962         int x = _SpinDuration;
1963         if (x < Knob_SpinLimit) {
1964           if (x < Knob_Poverty) x = Knob_Poverty;
1965           _SpinDuration = x + Knob_Bonus;
1966         }
1967         return 1;
1968       }
1969 
1970       // The CAS failed ... we can take any of the following actions:
1971       // * penalize: ctr -= Knob_CASPenalty
1972       // * exit spin with prejudice -- goto Abort;
1973       // * exit spin without prejudice.
1974       // * Since CAS is high-latency, retry again immediately.
1975       prv = ox;
1976       if (caspty == -2) break;
1977       if (caspty == -1) goto Abort;
1978       ctr -= caspty;
1979       continue;
1980     }
1981 
1982     // Did lock ownership change hands ?
1983     if (ox != prv && prv != NULL) {
1984       if (oxpty == -2) break;
1985       if (oxpty == -1) goto Abort;
1986       ctr -= oxpty;
1987     }
1988     prv = ox;
1989 
1990     // Abort the spin if the owner is not executing.
1991     // The owner must be executing in order to drop the lock.
1992     // Spinning while the owner is OFFPROC is idiocy.
1993     // Consider: ctr -= RunnablePenalty ;
1994     if (Knob_OState && NotRunnable (Self, ox)) {
1995       goto Abort;
1996     }
1997     if (sss && _succ == NULL) _succ = Self;


1998   }
1999 
2000   // Spin failed with prejudice -- reduce _SpinDuration.
2001   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2002   // AIMD is globally stable.
2003   {
2004     int x = _SpinDuration;
2005     if (x > 0) {
2006       // Consider an AIMD scheme like: x -= (x >> 3) + 100
2007       // This is globally sample and tends to damp the response.
2008       x -= Knob_Penalty;
2009       if (x < 0) x = 0;
2010       _SpinDuration = x;
2011     }
2012   }
2013 
2014  Abort:
2015   if (MaxSpin >= 0) Adjust(&_Spinner, -1);
2016   if (sss && _succ == Self) {
2017     _succ = NULL;
2018     // Invariant: after setting succ=null a contending thread
2019     // must recheck-retry _owner before parking.  This usually happens
2020     // in the normal usage of TrySpin(), but it's safest
2021     // to make TrySpin() as foolproof as possible.
2022     OrderAccess::fence();
2023     if (TryLock(Self) > 0) return 1;
2024   }
2025   return 0;
2026 }
2027 
2028 // NotRunnable() -- informed spinning
2029 //
2030 // Don't bother spinning if the owner is not eligible to drop the lock.
2031 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2032 // The thread must be runnable in order to drop the lock in timely fashion.
2033 // If the _owner is not runnable then spinning will not likely be
2034 // successful (profitable).
2035 //
2036 // Beware -- the thread referenced by _owner could have died
2037 // so a simply fetch from _owner->_thread_state might trap.
2038 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2039 // Because of the lifecycle issues, the _thread_state values
2040 // observed by NotRunnable() might be garbage.  NotRunnable must
2041 // tolerate this and consider the observed _thread_state value
2042 // as advisory.
2043 //
2044 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2045 // a thread pointer.
2046 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2047 // with the LSB of _owner.  Another option would be to probabilistically probe
2048 // the putative _owner->TypeTag value.
2049 //
2050 // Checking _thread_state isn't perfect.  Even if the thread is
2051 // in_java it might be blocked on a page-fault or have been preempted
2052 // and sitting on a ready/dispatch queue.
2053 //
2054 // The return value from NotRunnable() is *advisory* -- the
2055 // result is based on sampling and is not necessarily coherent.
2056 // The caller must tolerate false-negative and false-positive errors.
2057 // Spinning, in general, is probabilistic anyway.
2058 
2059 
2060 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2061   // Check ox->TypeTag == 2BAD.
2062   if (ox == NULL) return 0;
2063 
2064   // Avoid transitive spinning ...
2065   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2066   // Immediately after T1 acquires L it's possible that T2, also
2067   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2068   // This occurs transiently after T1 acquired L but before
2069   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2070   // its spin in this circumstance.
2071   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2072 
2073   if (BlockedOn == 1) return 1;
2074   if (BlockedOn != 0) {
2075     return BlockedOn != intptr_t(this) && _owner == ox;
2076   }
2077 
2078   assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2079   int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2080   // consider also: jst != _thread_in_Java -- but that's overspecific.
2081   return jst == _thread_blocked || jst == _thread_in_native;
2082 }
2083 
2084 
2085 // -----------------------------------------------------------------------------
2086 // WaitSet management ...
2087 
2088 ObjectWaiter::ObjectWaiter(Thread* thread) {
2089   _next     = NULL;
2090   _prev     = NULL;
2091   _notified = 0;
2092   _notifier_tid = 0;
2093   TState    = TS_RUN;
2094   _thread   = thread;
2095   _event    = thread->_ParkEvent;
2096   _active   = false;
2097   assert(_event != NULL, "invariant");
2098 }
2099 
2100 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2101   JavaThread *jt = (JavaThread *)this->_thread;
2102   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2103 }
2104 
2105 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2106   JavaThread *jt = (JavaThread *)this->_thread;
2107   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2108 }
2109 
2110 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2111   assert(node != NULL, "should not add NULL node");
2112   assert(node->_prev == NULL, "node already in list");
2113   assert(node->_next == NULL, "node already in list");
2114   // put node at end of queue (circular doubly linked list)
2115   if (_WaitSet == NULL) {
2116     _WaitSet = node;
2117     node->_prev = node;
2118     node->_next = node;
2119   } else {
2120     ObjectWaiter* head = _WaitSet;
2121     ObjectWaiter* tail = head->_prev;
2122     assert(tail->_next == head, "invariant check");
2123     tail->_next = node;
2124     head->_prev = node;
2125     node->_next = head;
2126     node->_prev = tail;
2127   }
2128 }
2129 
2130 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2131   // dequeue the very first waiter
2132   ObjectWaiter* waiter = _WaitSet;
2133   if (waiter) {
2134     DequeueSpecificWaiter(waiter);
2135   }
2136   return waiter;
2137 }
2138 
2139 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2140   assert(node != NULL, "should not dequeue NULL node");
2141   assert(node->_prev != NULL, "node already removed from list");
2142   assert(node->_next != NULL, "node already removed from list");
2143   // when the waiter has woken up because of interrupt,
2144   // timeout or other spurious wake-up, dequeue the
2145   // waiter from waiting list
2146   ObjectWaiter* next = node->_next;
2147   if (next == node) {
2148     assert(node->_prev == node, "invariant check");
2149     _WaitSet = NULL;
2150   } else {
2151     ObjectWaiter* prev = node->_prev;
2152     assert(prev->_next == node, "invariant check");
2153     assert(next->_prev == node, "invariant check");
2154     next->_prev = prev;
2155     prev->_next = next;
2156     if (_WaitSet == node) {
2157       _WaitSet = next;
2158     }
2159   }
2160   node->_next = NULL;
2161   node->_prev = NULL;
2162 }
2163 
2164 // -----------------------------------------------------------------------------
2165 // PerfData support
2166 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2167 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2168 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2169 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2170 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2171 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2172 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2173 
2174 // One-shot global initialization for the sync subsystem.
2175 // We could also defer initialization and initialize on-demand
2176 // the first time we call inflate().  Initialization would
2177 // be protected - like so many things - by the MonitorCache_lock.
2178 
2179 void ObjectMonitor::Initialize() {
2180   static int InitializationCompleted = 0;
2181   assert(InitializationCompleted == 0, "invariant");
2182   InitializationCompleted = 1;
2183   if (UsePerfData) {
2184     EXCEPTION_MARK;
2185 #define NEWPERFCOUNTER(n)                                                \
2186   {                                                                      \
2187     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2188                                         CHECK);                          \
2189   }
2190 #define NEWPERFVARIABLE(n)                                                \
2191   {                                                                       \
2192     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2193                                          CHECK);                          \
2194   }
2195     NEWPERFCOUNTER(_sync_Inflations);
2196     NEWPERFCOUNTER(_sync_Deflations);
2197     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2198     NEWPERFCOUNTER(_sync_FutileWakeups);
2199     NEWPERFCOUNTER(_sync_Parks);
2200     NEWPERFCOUNTER(_sync_Notifications);
2201     NEWPERFVARIABLE(_sync_MonExtant);
2202 #undef NEWPERFCOUNTER
2203 #undef NEWPERFVARIABLE
2204   }
2205 }
2206 
2207 static char * kvGet(char * kvList, const char * Key) {
2208   if (kvList == NULL) return NULL;
2209   size_t n = strlen(Key);
2210   char * Search;
2211   for (Search = kvList; *Search; Search += strlen(Search) + 1) {
2212     if (strncmp (Search, Key, n) == 0) {
2213       if (Search[n] == '=') return Search + n + 1;
2214       if (Search[n] == 0)   return(char *) "1";
2215     }
2216   }
2217   return NULL;
2218 }
2219 
2220 static int kvGetInt(char * kvList, const char * Key, int Default) {
2221   char * v = kvGet(kvList, Key);
2222   int rslt = v ? ::strtol(v, NULL, 0) : Default;
2223   if (Knob_ReportSettings && v != NULL) {
2224     tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ;
2225     tty->flush();
2226   }
2227   return rslt;
2228 }
2229 
2230 void ObjectMonitor::DeferredInitialize() {
2231   if (InitDone > 0) return;
2232   if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2233     while (InitDone != 1) /* empty */;
2234     return;
2235   }
2236 
2237   // One-shot global initialization ...
2238   // The initialization is idempotent, so we don't need locks.
2239   // In the future consider doing this via os::init_2().
2240   // SyncKnobs consist of <Key>=<Value> pairs in the style
2241   // of environment variables.  Start by converting ':' to NUL.
2242 
2243   if (SyncKnobs == NULL) SyncKnobs = "";
2244 
2245   size_t sz = strlen(SyncKnobs);
2246   char * knobs = (char *) os::malloc(sz + 2, mtInternal);
2247   if (knobs == NULL) {
2248     vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs");
2249     guarantee(0, "invariant");
2250   }
2251   strcpy(knobs, SyncKnobs);
2252   knobs[sz+1] = 0;
2253   for (char * p = knobs; *p; p++) {
2254     if (*p == ':') *p = 0;
2255   }
2256 
2257   #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); }
2258   SETKNOB(ReportSettings);
2259   SETKNOB(ExitRelease);
2260   SETKNOB(InlineNotify);
2261   SETKNOB(Verbose);
2262   SETKNOB(VerifyInUse);
2263   SETKNOB(VerifyMatch);
2264   SETKNOB(FixedSpin);
2265   SETKNOB(SpinLimit);
2266   SETKNOB(SpinBase);
2267   SETKNOB(SpinBackOff);
2268   SETKNOB(CASPenalty);
2269   SETKNOB(OXPenalty);
2270   SETKNOB(SpinSetSucc);
2271   SETKNOB(SuccEnabled);
2272   SETKNOB(SuccRestrict);
2273   SETKNOB(Penalty);
2274   SETKNOB(Bonus);
2275   SETKNOB(BonusB);
2276   SETKNOB(Poverty);
2277   SETKNOB(SpinAfterFutile);
2278   SETKNOB(UsePause);
2279   SETKNOB(SpinEarly);
2280   SETKNOB(OState);
2281   SETKNOB(MaxSpinners);
2282   SETKNOB(PreSpin);
2283   SETKNOB(ExitPolicy);
2284   SETKNOB(QMode);
2285   SETKNOB(ResetEvent);
2286   SETKNOB(MoveNotifyee);
2287   SETKNOB(FastHSSEC);
2288   #undef SETKNOB
2289 
2290   if (os::is_MP()) {
2291     BackOffMask = (1 << Knob_SpinBackOff) - 1;
2292     if (Knob_ReportSettings) {
2293       tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask);
2294     }
2295     // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2296   } else {
2297     Knob_SpinLimit = 0;
2298     Knob_SpinBase  = 0;
2299     Knob_PreSpin   = 0;
2300     Knob_FixedSpin = -1;
2301   }
2302 
2303   os::free(knobs);
2304   OrderAccess::fence();
2305   InitDone = 1;
2306 }
2307 
--- EOF ---