1 /*
   2  * Copyright (c) 1998, 2022, 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 "gc/shared/oopStorage.hpp"
  28 #include "gc/shared/oopStorageSet.hpp"
  29 #include "jfr/jfrEvents.hpp"
  30 #include "jfr/support/jfrThreadId.hpp"
  31 #include "logging/log.hpp"
  32 #include "logging/logStream.hpp"
  33 #include "memory/allocation.inline.hpp"
  34 #include "memory/resourceArea.hpp"
  35 #include "oops/markWord.hpp"
  36 #include "oops/oop.inline.hpp"
  37 #include "oops/oopHandle.inline.hpp"
  38 #include "oops/weakHandle.inline.hpp"
  39 #include "prims/jvmtiDeferredUpdates.hpp"
  40 #include "prims/jvmtiExport.hpp"
  41 #include "runtime/atomic.hpp"
  42 #include "runtime/handles.inline.hpp"
  43 #include "runtime/interfaceSupport.inline.hpp"
  44 #include "runtime/mutexLocker.hpp"
  45 #include "runtime/objectMonitor.hpp"
  46 #include "runtime/objectMonitor.inline.hpp"
  47 #include "runtime/orderAccess.hpp"
  48 #include "runtime/osThread.hpp"
  49 #include "runtime/perfData.hpp"
  50 #include "runtime/safefetch.hpp"
  51 #include "runtime/safepointMechanism.inline.hpp"
  52 #include "runtime/sharedRuntime.hpp"
  53 #include "runtime/thread.inline.hpp"
  54 #include "services/threadService.hpp"
  55 #include "utilities/dtrace.hpp"
  56 #include "utilities/macros.hpp"
  57 #include "utilities/preserveException.hpp"
  58 #if INCLUDE_JFR
  59 #include "jfr/support/jfrFlush.hpp"
  60 #endif
  61 
  62 #ifdef DTRACE_ENABLED
  63 
  64 // Only bother with this argument setup if dtrace is available
  65 // TODO-FIXME: probes should not fire when caller is _blocked.  assert() accordingly.
  66 
  67 
  68 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread)                           \
  69   char* bytes = NULL;                                                      \
  70   int len = 0;                                                             \
  71   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  72   Symbol* klassname = obj->klass()->name();                                \
  73   if (klassname != NULL) {                                                 \
  74     bytes = (char*)klassname->bytes();                                     \
  75     len = klassname->utf8_length();                                        \
  76   }
  77 
  78 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)            \
  79   {                                                                        \
  80     if (DTraceMonitorProbes) {                                             \
  81       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  82       HOTSPOT_MONITOR_WAIT(jtid,                                           \
  83                            (monitor), bytes, len, (millis));               \
  84     }                                                                      \
  85   }
  86 
  87 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
  88 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
  89 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
  90 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
  91 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
  92 
  93 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)                  \
  94   {                                                                        \
  95     if (DTraceMonitorProbes) {                                             \
  96       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  97       HOTSPOT_MONITOR_##probe(jtid,                                        \
  98                               (uintptr_t)(monitor), bytes, len);           \
  99     }                                                                      \
 100   }
 101 
 102 #else //  ndef DTRACE_ENABLED
 103 
 104 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon)    {;}
 105 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon)          {;}
 106 
 107 #endif // ndef DTRACE_ENABLED
 108 
 109 // Tunables ...
 110 // The knob* variables are effectively final.  Once set they should
 111 // never be modified hence.  Consider using __read_mostly with GCC.
 112 
 113 int ObjectMonitor::Knob_SpinLimit    = 5000;    // derived by an external tool -
 114 
 115 static int Knob_Bonus               = 100;     // spin success bonus
 116 static int Knob_BonusB              = 100;     // spin success bonus
 117 static int Knob_Penalty             = 200;     // spin failure penalty
 118 static int Knob_Poverty             = 1000;
 119 static int Knob_FixedSpin           = 0;
 120 static int Knob_PreSpin             = 10;      // 20-100 likely better
 121 
 122 DEBUG_ONLY(static volatile bool InitDone = false;)
 123 
 124 OopStorage* ObjectMonitor::_oop_storage = NULL;
 125 
 126 // -----------------------------------------------------------------------------
 127 // Theory of operations -- Monitors lists, thread residency, etc:
 128 //
 129 // * A thread acquires ownership of a monitor by successfully
 130 //   CAS()ing the _owner field from null to non-null.
 131 //
 132 // * Invariant: A thread appears on at most one monitor list --
 133 //   cxq, EntryList or WaitSet -- at any one time.
 134 //
 135 // * Contending threads "push" themselves onto the cxq with CAS
 136 //   and then spin/park.
 137 //
 138 // * After a contending thread eventually acquires the lock it must
 139 //   dequeue itself from either the EntryList or the cxq.
 140 //
 141 // * The exiting thread identifies and unparks an "heir presumptive"
 142 //   tentative successor thread on the EntryList.  Critically, the
 143 //   exiting thread doesn't unlink the successor thread from the EntryList.
 144 //   After having been unparked, the wakee will recontend for ownership of
 145 //   the monitor.   The successor (wakee) will either acquire the lock or
 146 //   re-park itself.
 147 //
 148 //   Succession is provided for by a policy of competitive handoff.
 149 //   The exiting thread does _not_ grant or pass ownership to the
 150 //   successor thread.  (This is also referred to as "handoff" succession").
 151 //   Instead the exiting thread releases ownership and possibly wakes
 152 //   a successor, so the successor can (re)compete for ownership of the lock.
 153 //   If the EntryList is empty but the cxq is populated the exiting
 154 //   thread will drain the cxq into the EntryList.  It does so by
 155 //   by detaching the cxq (installing null with CAS) and folding
 156 //   the threads from the cxq into the EntryList.  The EntryList is
 157 //   doubly linked, while the cxq is singly linked because of the
 158 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 159 //
 160 // * Concurrency invariants:
 161 //
 162 //   -- only the monitor owner may access or mutate the EntryList.
 163 //      The mutex property of the monitor itself protects the EntryList
 164 //      from concurrent interference.
 165 //   -- Only the monitor owner may detach the cxq.
 166 //
 167 // * The monitor entry list operations avoid locks, but strictly speaking
 168 //   they're not lock-free.  Enter is lock-free, exit is not.
 169 //   For a description of 'Methods and apparatus providing non-blocking access
 170 //   to a resource,' see U.S. Pat. No. 7844973.
 171 //
 172 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 173 //   detaching thread.  This mechanism is immune from the ABA corruption.
 174 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 175 //
 176 // * Taken together, the cxq and the EntryList constitute or form a
 177 //   single logical queue of threads stalled trying to acquire the lock.
 178 //   We use two distinct lists to improve the odds of a constant-time
 179 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 180 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 181 //   A key desideratum is to minimize queue & monitor metadata manipulation
 182 //   that occurs while holding the monitor lock -- that is, we want to
 183 //   minimize monitor lock holds times.  Note that even a small amount of
 184 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 185 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 186 //   locks and monitor metadata.
 187 //
 188 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 189 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 190 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 191 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 192 //
 193 //   The EntryList is ordered by the prevailing queue discipline and
 194 //   can be organized in any convenient fashion, such as a doubly-linked list or
 195 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 196 //   to operate in constant-time.  If we need a priority queue then something akin
 197 //   to Solaris' sleepq would work nicely.  Viz.,
 198 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 199 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 200 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 201 //   EntryList accordingly.
 202 //
 203 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 204 //   somewhat similar to an elevator-scan.
 205 //
 206 // * The monitor synchronization subsystem avoids the use of native
 207 //   synchronization primitives except for the narrow platform-specific
 208 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 209 //   the semantics of park-unpark.  Put another way, this monitor implementation
 210 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 211 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 212 //   underlying OS manages the READY<->RUN transitions.
 213 //
 214 // * Waiting threads reside on the WaitSet list -- wait() puts
 215 //   the caller onto the WaitSet.
 216 //
 217 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 218 //   either the EntryList or cxq.  Subsequent exit() operations will
 219 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 220 //   it's likely the notifyee would simply impale itself on the lock held
 221 //   by the notifier.
 222 //
 223 // * An interesting alternative is to encode cxq as (List,LockByte) where
 224 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 225 //   variable, like _recursions, in the scheme.  The threads or Events that form
 226 //   the list would have to be aligned in 256-byte addresses.  A thread would
 227 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 228 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 229 //   Note that is is *not* word-tearing, but it does presume that full-word
 230 //   CAS operations are coherent with intermix with STB operations.  That's true
 231 //   on most common processors.
 232 //
 233 // * See also http://blogs.sun.com/dave
 234 
 235 
 236 // Check that object() and set_object() are called from the right context:
 237 static void check_object_context() {
 238 #ifdef ASSERT
 239   Thread* self = Thread::current();
 240   if (self->is_Java_thread()) {
 241     // Mostly called from JavaThreads so sanity check the thread state.
 242     JavaThread* jt = JavaThread::cast(self);
 243     switch (jt->thread_state()) {
 244     case _thread_in_vm:    // the usual case
 245     case _thread_in_Java:  // during deopt
 246       break;
 247     default:
 248       fatal("called from an unsafe thread state");
 249     }
 250     assert(jt->is_active_Java_thread(), "must be active JavaThread");
 251   } else {
 252     // However, ThreadService::get_current_contended_monitor()
 253     // can call here via the VMThread so sanity check it.
 254     assert(self->is_VM_thread(), "must be");
 255   }
 256 #endif // ASSERT
 257 }
 258 
 259 ObjectMonitor::ObjectMonitor(oop object) :
 260   _header(markWord::zero()),
 261   _object(_oop_storage, object),
 262   _owner(NULL),
 263   _previous_owner_tid(0),
 264   _next_om(NULL),
 265   _recursions(0),
 266   _EntryList(NULL),
 267   _cxq(NULL),
 268   _succ(NULL),
 269   _Responsible(NULL),
 270   _Spinner(0),
 271   _SpinDuration(ObjectMonitor::Knob_SpinLimit),
 272   _contentions(0),
 273   _WaitSet(NULL),
 274   _waiters(0),
 275   _WaitSetLock(0)
 276 { }
 277 
 278 ObjectMonitor::~ObjectMonitor() {
 279   _object.release(_oop_storage);
 280 }
 281 
 282 oop ObjectMonitor::object() const {
 283   check_object_context();
 284   if (_object.is_null()) {
 285     return NULL;
 286   }
 287   return _object.resolve();
 288 }
 289 
 290 oop ObjectMonitor::object_peek() const {
 291   if (_object.is_null()) {
 292     return NULL;
 293   }
 294   return _object.peek();
 295 }
 296 
 297 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
 298   if (current->is_suspended()) {
 299     _om->_recursions = 0;
 300     _om->_succ = NULL;
 301     // Don't need a full fence after clearing successor here because of the call to exit().
 302     _om->exit(current, false /* not_suspended */);
 303     _om_exited = true;
 304 
 305     current->set_current_pending_monitor(_om);
 306   }
 307 }
 308 
 309 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
 310   if (current->is_suspended()) {
 311     if (_om->_succ == current) {
 312       _om->_succ = NULL;
 313       OrderAccess::fence(); // always do a full fence when successor is cleared
 314     }
 315   }
 316 }
 317 
 318 // -----------------------------------------------------------------------------
 319 // Enter support
 320 
 321 bool ObjectMonitor::enter(JavaThread* current) {
 322   // The following code is ordered to check the most common cases first
 323   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 324 
 325   void* cur = try_set_owner_from(NULL, current);
 326   if (cur == NULL) {
 327     assert(_recursions == 0, "invariant");
 328     return true;
 329   }
 330 
 331   if (cur == current) {
 332     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 333     _recursions++;
 334     return true;
 335   }
 336 
 337   if (current->is_lock_owned((address)cur)) {
 338     assert(_recursions == 0, "internal state error");
 339     _recursions = 1;
 340     set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
 341     return true;
 342   }
 343 
 344   // We've encountered genuine contention.
 345   assert(current->_Stalled == 0, "invariant");
 346   current->_Stalled = intptr_t(this);
 347 
 348   // Try one round of spinning *before* enqueueing current
 349   // and before going through the awkward and expensive state
 350   // transitions.  The following spin is strictly optional ...
 351   // Note that if we acquire the monitor from an initial spin
 352   // we forgo posting JVMTI events and firing DTRACE probes.
 353   if (TrySpin(current) > 0) {
 354     assert(owner_raw() == current, "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
 355     assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
 356     assert(object()->mark() == markWord::encode(this),
 357            "object mark must match encoded this: mark=" INTPTR_FORMAT
 358            ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
 359            markWord::encode(this).value());
 360     current->_Stalled = 0;
 361     return true;
 362   }
 363 
 364   assert(owner_raw() != current, "invariant");
 365   assert(_succ != current, "invariant");
 366   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 367   assert(current->thread_state() != _thread_blocked, "invariant");
 368 
 369   // Keep track of contention for JVM/TI and M&M queries.
 370   add_to_contentions(1);
 371   if (is_being_async_deflated()) {
 372     // Async deflation is in progress and our contentions increment
 373     // above lost the race to async deflation. Undo the work and
 374     // force the caller to retry.
 375     const oop l_object = object();
 376     if (l_object != NULL) {
 377       // Attempt to restore the header/dmw to the object's header so that
 378       // we only retry once if the deflater thread happens to be slow.
 379       install_displaced_markword_in_object(l_object);
 380     }
 381     current->_Stalled = 0;
 382     add_to_contentions(-1);
 383     return false;
 384   }
 385 
 386   JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(current);)
 387   EventJavaMonitorEnter event;
 388   if (event.is_started()) {
 389     event.set_monitorClass(object()->klass());
 390     // Set an address that is 'unique enough', such that events close in
 391     // time and with the same address are likely (but not guaranteed) to
 392     // belong to the same object.
 393     event.set_address((uintptr_t)this);
 394   }
 395 
 396   { // Change java thread status to indicate blocked on monitor enter.
 397     JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
 398 
 399     assert(current->current_pending_monitor() == NULL, "invariant");
 400     current->set_current_pending_monitor(this);
 401 
 402     DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
 403     if (JvmtiExport::should_post_monitor_contended_enter()) {
 404       JvmtiExport::post_monitor_contended_enter(current, this);
 405 
 406       // The current thread does not yet own the monitor and does not
 407       // yet appear on any queues that would get it made the successor.
 408       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 409       // handler cannot accidentally consume an unpark() meant for the
 410       // ParkEvent associated with this ObjectMonitor.
 411     }
 412 
 413     OSThreadContendState osts(current->osthread());
 414 
 415     assert(current->thread_state() == _thread_in_vm, "invariant");
 416 
 417     for (;;) {
 418       ExitOnSuspend eos(this);
 419       {
 420         ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */);
 421         EnterI(current);
 422         current->set_current_pending_monitor(NULL);
 423         // We can go to a safepoint at the end of this block. If we
 424         // do a thread dump during that safepoint, then this thread will show
 425         // as having "-locked" the monitor, but the OS and java.lang.Thread
 426         // states will still report that the thread is blocked trying to
 427         // acquire it.
 428         // If there is a suspend request, ExitOnSuspend will exit the OM
 429         // and set the OM as pending.
 430       }
 431       if (!eos.exited()) {
 432         // ExitOnSuspend did not exit the OM
 433         assert(owner_raw() == current, "invariant");
 434         break;
 435       }
 436     }
 437 
 438     // We've just gotten past the enter-check-for-suspend dance and we now own
 439     // the monitor free and clear.
 440   }
 441 
 442   add_to_contentions(-1);
 443   assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
 444   current->_Stalled = 0;
 445 
 446   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 447   assert(_recursions == 0, "invariant");
 448   assert(owner_raw() == current, "invariant");
 449   assert(_succ != current, "invariant");
 450   assert(object()->mark() == markWord::encode(this), "invariant");
 451 
 452   // The thread -- now the owner -- is back in vm mode.
 453   // Report the glorious news via TI,DTrace and jvmstat.
 454   // The probe effect is non-trivial.  All the reportage occurs
 455   // while we hold the monitor, increasing the length of the critical
 456   // section.  Amdahl's parallel speedup law comes vividly into play.
 457   //
 458   // Another option might be to aggregate the events (thread local or
 459   // per-monitor aggregation) and defer reporting until a more opportune
 460   // time -- such as next time some thread encounters contention but has
 461   // yet to acquire the lock.  While spinning that thread could
 462   // spinning we could increment JVMStat counters, etc.
 463 
 464   DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
 465   if (JvmtiExport::should_post_monitor_contended_entered()) {
 466     JvmtiExport::post_monitor_contended_entered(current, this);
 467 
 468     // The current thread already owns the monitor and is not going to
 469     // call park() for the remainder of the monitor enter protocol. So
 470     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 471     // event handler consumed an unpark() issued by the thread that
 472     // just exited the monitor.
 473   }
 474   if (event.should_commit()) {
 475     event.set_previousOwner(_previous_owner_tid);
 476     event.commit();
 477   }
 478   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 479   return true;
 480 }
 481 
 482 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 483 // Callers must compensate as needed.
 484 
 485 int ObjectMonitor::TryLock(JavaThread* current) {
 486   void* own = owner_raw();
 487   if (own != NULL) return 0;
 488   if (try_set_owner_from(NULL, current) == NULL) {
 489     assert(_recursions == 0, "invariant");
 490     return 1;
 491   }
 492   // The lock had been free momentarily, but we lost the race to the lock.
 493   // Interference -- the CAS failed.
 494   // We can either return -1 or retry.
 495   // Retry doesn't make as much sense because the lock was just acquired.
 496   return -1;
 497 }
 498 
 499 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
 500 // was deflated and false otherwise.
 501 //
 502 // The async deflation protocol sets owner to DEFLATER_MARKER and
 503 // makes contentions negative as signals to contending threads that
 504 // an async deflation is in progress. There are a number of checks
 505 // as part of the protocol to make sure that the calling thread has
 506 // not lost the race to a contending thread.
 507 //
 508 // The ObjectMonitor has been successfully async deflated when:
 509 //   (contentions < 0)
 510 // Contending threads that see that condition know to retry their operation.
 511 //
 512 bool ObjectMonitor::deflate_monitor() {
 513   if (is_busy()) {
 514     // Easy checks are first - the ObjectMonitor is busy so no deflation.
 515     return false;
 516   }
 517 
 518   if (ObjectSynchronizer::is_final_audit() && owner_is_DEFLATER_MARKER()) {
 519     // The final audit can see an already deflated ObjectMonitor on the
 520     // in-use list because MonitorList::unlink_deflated() might have
 521     // blocked for the final safepoint before unlinking all the deflated
 522     // monitors.
 523     assert(contentions() < 0, "must be negative: contentions=%d", contentions());
 524     // Already returned 'true' when it was originally deflated.
 525     return false;
 526   }
 527 
 528   const oop obj = object_peek();
 529 
 530   if (obj == NULL) {
 531     // If the object died, we can recycle the monitor without racing with
 532     // Java threads. The GC already broke the association with the object.
 533     set_owner_from(NULL, DEFLATER_MARKER);
 534     assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
 535     _contentions = INT_MIN; // minimum negative int
 536   } else {
 537     // Attempt async deflation protocol.
 538 
 539     // Set a NULL owner to DEFLATER_MARKER to force any contending thread
 540     // through the slow path. This is just the first part of the async
 541     // deflation dance.
 542     if (try_set_owner_from(NULL, DEFLATER_MARKER) != NULL) {
 543       // The owner field is no longer NULL so we lost the race since the
 544       // ObjectMonitor is now busy.
 545       return false;
 546     }
 547 
 548     if (contentions() > 0 || _waiters != 0) {
 549       // Another thread has raced to enter the ObjectMonitor after
 550       // is_busy() above or has already entered and waited on
 551       // it which makes it busy so no deflation. Restore owner to
 552       // NULL if it is still DEFLATER_MARKER.
 553       if (try_set_owner_from(DEFLATER_MARKER, NULL) != DEFLATER_MARKER) {
 554         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 555         add_to_contentions(-1);
 556       }
 557       return false;
 558     }
 559 
 560     // Make a zero contentions field negative to force any contending threads
 561     // to retry. This is the second part of the async deflation dance.
 562     if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) {
 563       // Contentions was no longer 0 so we lost the race since the
 564       // ObjectMonitor is now busy. Restore owner to NULL if it is
 565       // still DEFLATER_MARKER:
 566       if (try_set_owner_from(DEFLATER_MARKER, NULL) != DEFLATER_MARKER) {
 567         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 568         add_to_contentions(-1);
 569       }
 570       return false;
 571     }
 572   }
 573 
 574   // Sanity checks for the races:
 575   guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
 576   guarantee(contentions() < 0, "must be negative: contentions=%d",
 577             contentions());
 578   guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
 579   guarantee(_cxq == NULL, "must be no contending threads: cxq="
 580             INTPTR_FORMAT, p2i(_cxq));
 581   guarantee(_EntryList == NULL,
 582             "must be no entering threads: EntryList=" INTPTR_FORMAT,
 583             p2i(_EntryList));
 584 
 585   if (obj != NULL) {
 586     if (log_is_enabled(Trace, monitorinflation)) {
 587       ResourceMark rm;
 588       log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
 589                                   ", mark=" INTPTR_FORMAT ", type='%s'",
 590                                   p2i(obj), obj->mark().value(),
 591                                   obj->klass()->external_name());
 592     }
 593 
 594     // Install the old mark word if nobody else has already done it.
 595     install_displaced_markword_in_object(obj);
 596   }
 597 
 598   // We leave owner == DEFLATER_MARKER and contentions < 0
 599   // to force any racing threads to retry.
 600   return true;  // Success, ObjectMonitor has been deflated.
 601 }
 602 
 603 // We might access the dead object headers for parsable heap walk, make sure
 604 // headers are in correct shape, e.g. monitors deflated.
 605 void ObjectMonitor::maybe_deflate_dead(oop* p) {
 606   oop obj = *p;
 607   assert(obj != NULL, "must not yet been cleared");
 608   markWord mark = obj->mark();
 609   if (mark.has_monitor()) {
 610     ObjectMonitor* monitor = mark.monitor();
 611     if (p == monitor->_object.ptr_raw()) {
 612       assert(monitor->object_peek() == obj, "lock object must match");
 613       markWord dmw = monitor->header();
 614       obj->set_mark(dmw);
 615     }
 616   }
 617 }
 618 
 619 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
 620 // into the header of the object associated with the monitor. This
 621 // idempotent method is called by a thread that is deflating a
 622 // monitor and by other threads that have detected a race with the
 623 // deflation process.
 624 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
 625   // This function must only be called when (owner == DEFLATER_MARKER
 626   // && contentions <= 0), but we can't guarantee that here because
 627   // those values could change when the ObjectMonitor gets moved from
 628   // the global free list to a per-thread free list.
 629 
 630   guarantee(obj != NULL, "must be non-NULL");
 631 
 632   // Separate loads in is_being_async_deflated(), which is almost always
 633   // called before this function, from the load of dmw/header below.
 634 
 635   // _contentions and dmw/header may get written by different threads.
 636   // Make sure to observe them in the same order when having several observers.
 637   OrderAccess::loadload_for_IRIW();
 638 
 639   const oop l_object = object_peek();
 640   if (l_object == NULL) {
 641     // ObjectMonitor's object ref has already been cleared by async
 642     // deflation or GC so we're done here.
 643     return;
 644   }
 645   assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
 646          INTPTR_FORMAT, p2i(l_object), p2i(obj));
 647 
 648   markWord dmw = header();
 649   // The dmw has to be neutral (not NULL, not locked and not marked).
 650   assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
 651 
 652   // Install displaced mark word if the object's header still points
 653   // to this ObjectMonitor. More than one racing caller to this function
 654   // can rarely reach this point, but only one can win.
 655   markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
 656   if (res != markWord::encode(this)) {
 657     // This should be rare so log at the Info level when it happens.
 658     log_info(monitorinflation)("install_displaced_markword_in_object: "
 659                                "failed cas_set_mark: new_mark=" INTPTR_FORMAT
 660                                ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
 661                                dmw.value(), markWord::encode(this).value(),
 662                                res.value());
 663   }
 664 
 665   // Note: It does not matter which thread restored the header/dmw
 666   // into the object's header. The thread deflating the monitor just
 667   // wanted the object's header restored and it is. The threads that
 668   // detected a race with the deflation process also wanted the
 669   // object's header restored before they retry their operation and
 670   // because it is restored they will only retry once.
 671 }
 672 
 673 // Convert the fields used by is_busy() to a string that can be
 674 // used for diagnostic output.
 675 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
 676   ss->print("is_busy: waiters=%d, ", _waiters);
 677   if (contentions() > 0) {
 678     ss->print("contentions=%d, ", contentions());
 679   } else {
 680     ss->print("contentions=0");
 681   }
 682   if (!owner_is_DEFLATER_MARKER()) {
 683     ss->print("owner=" INTPTR_FORMAT, p2i(owner_raw()));
 684   } else {
 685     // We report NULL instead of DEFLATER_MARKER here because is_busy()
 686     // ignores DEFLATER_MARKER values.
 687     ss->print("owner=" INTPTR_FORMAT, NULL);
 688   }
 689   ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq),
 690             p2i(_EntryList));
 691   return ss->base();
 692 }
 693 
 694 #define MAX_RECHECK_INTERVAL 1000
 695 
 696 void ObjectMonitor::EnterI(JavaThread* current) {
 697   assert(current->thread_state() == _thread_blocked, "invariant");
 698 
 699   // Try the lock - TATAS
 700   if (TryLock (current) > 0) {
 701     assert(_succ != current, "invariant");
 702     assert(owner_raw() == current, "invariant");
 703     assert(_Responsible != current, "invariant");
 704     return;
 705   }
 706 
 707   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 708     // Cancelled the in-progress async deflation by changing owner from
 709     // DEFLATER_MARKER to current. As part of the contended enter protocol,
 710     // contentions was incremented to a positive value before EnterI()
 711     // was called and that prevents the deflater thread from winning the
 712     // last part of the 2-part async deflation protocol. After EnterI()
 713     // returns to enter(), contentions is decremented because the caller
 714     // now owns the monitor. We bump contentions an extra time here to
 715     // prevent the deflater thread from winning the last part of the
 716     // 2-part async deflation protocol after the regular decrement
 717     // occurs in enter(). The deflater thread will decrement contentions
 718     // after it recognizes that the async deflation was cancelled.
 719     add_to_contentions(1);
 720     assert(_succ != current, "invariant");
 721     assert(_Responsible != current, "invariant");
 722     return;
 723   }
 724 
 725   assert(InitDone, "Unexpectedly not initialized");
 726 
 727   // We try one round of spinning *before* enqueueing current.
 728   //
 729   // If the _owner is ready but OFFPROC we could use a YieldTo()
 730   // operation to donate the remainder of this thread's quantum
 731   // to the owner.  This has subtle but beneficial affinity
 732   // effects.
 733 
 734   if (TrySpin(current) > 0) {
 735     assert(owner_raw() == current, "invariant");
 736     assert(_succ != current, "invariant");
 737     assert(_Responsible != current, "invariant");
 738     return;
 739   }
 740 
 741   // The Spin failed -- Enqueue and park the thread ...
 742   assert(_succ != current, "invariant");
 743   assert(owner_raw() != current, "invariant");
 744   assert(_Responsible != current, "invariant");
 745 
 746   // Enqueue "current" on ObjectMonitor's _cxq.
 747   //
 748   // Node acts as a proxy for current.
 749   // As an aside, if were to ever rewrite the synchronization code mostly
 750   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 751   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 752   // as well as eliminate a subset of ABA issues.
 753   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 754 
 755   ObjectWaiter node(current);
 756   current->_ParkEvent->reset();
 757   node._prev   = (ObjectWaiter*) 0xBAD;
 758   node.TState  = ObjectWaiter::TS_CXQ;
 759 
 760   // Push "current" onto the front of the _cxq.
 761   // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
 762   // Note that spinning tends to reduce the rate at which threads
 763   // enqueue and dequeue on EntryList|cxq.
 764   ObjectWaiter* nxt;
 765   for (;;) {
 766     node._next = nxt = _cxq;
 767     if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
 768 
 769     // Interference - the CAS failed because _cxq changed.  Just retry.
 770     // As an optional optimization we retry the lock.
 771     if (TryLock (current) > 0) {
 772       assert(_succ != current, "invariant");
 773       assert(owner_raw() == current, "invariant");
 774       assert(_Responsible != current, "invariant");
 775       return;
 776     }
 777   }
 778 
 779   // Check for cxq|EntryList edge transition to non-null.  This indicates
 780   // the onset of contention.  While contention persists exiting threads
 781   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 782   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 783   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 784   // arrange for one of the contending thread to use a timed park() operations
 785   // to detect and recover from the race.  (Stranding is form of progress failure
 786   // where the monitor is unlocked but all the contending threads remain parked).
 787   // That is, at least one of the contended threads will periodically poll _owner.
 788   // One of the contending threads will become the designated "Responsible" thread.
 789   // The Responsible thread uses a timed park instead of a normal indefinite park
 790   // operation -- it periodically wakes and checks for and recovers from potential
 791   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 792   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 793   // be responsible for a monitor.
 794   //
 795   // Currently, one of the contended threads takes on the added role of "Responsible".
 796   // A viable alternative would be to use a dedicated "stranding checker" thread
 797   // that periodically iterated over all the threads (or active monitors) and unparked
 798   // successors where there was risk of stranding.  This would help eliminate the
 799   // timer scalability issues we see on some platforms as we'd only have one thread
 800   // -- the checker -- parked on a timer.
 801 
 802   if (nxt == NULL && _EntryList == NULL) {
 803     // Try to assume the role of responsible thread for the monitor.
 804     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=current }
 805     Atomic::replace_if_null(&_Responsible, current);
 806   }
 807 
 808   // The lock might have been released while this thread was occupied queueing
 809   // itself onto _cxq.  To close the race and avoid "stranding" and
 810   // progress-liveness failure we must resample-retry _owner before parking.
 811   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 812   // In this case the ST-MEMBAR is accomplished with CAS().
 813   //
 814   // TODO: Defer all thread state transitions until park-time.
 815   // Since state transitions are heavy and inefficient we'd like
 816   // to defer the state transitions until absolutely necessary,
 817   // and in doing so avoid some transitions ...
 818 
 819   int nWakeups = 0;
 820   int recheckInterval = 1;
 821 
 822   for (;;) {
 823 
 824     if (TryLock(current) > 0) break;
 825     assert(owner_raw() != current, "invariant");
 826 
 827     // park self
 828     if (_Responsible == current) {
 829       current->_ParkEvent->park((jlong) recheckInterval);
 830       // Increase the recheckInterval, but clamp the value.
 831       recheckInterval *= 8;
 832       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 833         recheckInterval = MAX_RECHECK_INTERVAL;
 834       }
 835     } else {
 836       current->_ParkEvent->park();
 837     }
 838 
 839     if (TryLock(current) > 0) break;
 840 
 841     if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 842       // Cancelled the in-progress async deflation by changing owner from
 843       // DEFLATER_MARKER to current. As part of the contended enter protocol,
 844       // contentions was incremented to a positive value before EnterI()
 845       // was called and that prevents the deflater thread from winning the
 846       // last part of the 2-part async deflation protocol. After EnterI()
 847       // returns to enter(), contentions is decremented because the caller
 848       // now owns the monitor. We bump contentions an extra time here to
 849       // prevent the deflater thread from winning the last part of the
 850       // 2-part async deflation protocol after the regular decrement
 851       // occurs in enter(). The deflater thread will decrement contentions
 852       // after it recognizes that the async deflation was cancelled.
 853       add_to_contentions(1);
 854       break;
 855     }
 856 
 857     // The lock is still contested.
 858     // Keep a tally of the # of futile wakeups.
 859     // Note that the counter is not protected by a lock or updated by atomics.
 860     // That is by design - we trade "lossy" counters which are exposed to
 861     // races during updates for a lower probe effect.
 862 
 863     // This PerfData object can be used in parallel with a safepoint.
 864     // See the work around in PerfDataManager::destroy().
 865     OM_PERFDATA_OP(FutileWakeups, inc());
 866     ++nWakeups;
 867 
 868     // Assuming this is not a spurious wakeup we'll normally find _succ == current.
 869     // We can defer clearing _succ until after the spin completes
 870     // TrySpin() must tolerate being called with _succ == current.
 871     // Try yet another round of adaptive spinning.
 872     if (TrySpin(current) > 0) break;
 873 
 874     // We can find that we were unpark()ed and redesignated _succ while
 875     // we were spinning.  That's harmless.  If we iterate and call park(),
 876     // park() will consume the event and return immediately and we'll
 877     // just spin again.  This pattern can repeat, leaving _succ to simply
 878     // spin on a CPU.
 879 
 880     if (_succ == current) _succ = NULL;
 881 
 882     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 883     OrderAccess::fence();
 884   }
 885 
 886   // Egress :
 887   // current has acquired the lock -- Unlink current from the cxq or EntryList.
 888   // Normally we'll find current on the EntryList .
 889   // From the perspective of the lock owner (this thread), the
 890   // EntryList is stable and cxq is prepend-only.
 891   // The head of cxq is volatile but the interior is stable.
 892   // In addition, current.TState is stable.
 893 
 894   assert(owner_raw() == current, "invariant");
 895 
 896   UnlinkAfterAcquire(current, &node);
 897   if (_succ == current) _succ = NULL;
 898 
 899   assert(_succ != current, "invariant");
 900   if (_Responsible == current) {
 901     _Responsible = NULL;
 902     OrderAccess::fence(); // Dekker pivot-point
 903 
 904     // We may leave threads on cxq|EntryList without a designated
 905     // "Responsible" thread.  This is benign.  When this thread subsequently
 906     // exits the monitor it can "see" such preexisting "old" threads --
 907     // threads that arrived on the cxq|EntryList before the fence, above --
 908     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 909     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 910     // non-null and elect a new "Responsible" timer thread.
 911     //
 912     // This thread executes:
 913     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 914     //    LD cxq|EntryList               (in subsequent exit)
 915     //
 916     // Entering threads in the slow/contended path execute:
 917     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 918     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 919     //
 920     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 921     // exit operation from floating above the ST Responsible=null.
 922   }
 923 
 924   // We've acquired ownership with CAS().
 925   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 926   // But since the CAS() this thread may have also stored into _succ,
 927   // EntryList, cxq or Responsible.  These meta-data updates must be
 928   // visible __before this thread subsequently drops the lock.
 929   // Consider what could occur if we didn't enforce this constraint --
 930   // STs to monitor meta-data and user-data could reorder with (become
 931   // visible after) the ST in exit that drops ownership of the lock.
 932   // Some other thread could then acquire the lock, but observe inconsistent
 933   // or old monitor meta-data and heap data.  That violates the JMM.
 934   // To that end, the 1-0 exit() operation must have at least STST|LDST
 935   // "release" barrier semantics.  Specifically, there must be at least a
 936   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 937   // the lock.   The barrier ensures that changes to monitor meta-data and data
 938   // protected by the lock will be visible before we release the lock, and
 939   // therefore before some other thread (CPU) has a chance to acquire the lock.
 940   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 941   //
 942   // Critically, any prior STs to _succ or EntryList must be visible before
 943   // the ST of null into _owner in the *subsequent* (following) corresponding
 944   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 945   // execute a serializing instruction.
 946 
 947   return;
 948 }
 949 
 950 // ReenterI() is a specialized inline form of the latter half of the
 951 // contended slow-path from EnterI().  We use ReenterI() only for
 952 // monitor reentry in wait().
 953 //
 954 // In the future we should reconcile EnterI() and ReenterI().
 955 
 956 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
 957   assert(current != NULL, "invariant");
 958   assert(currentNode != NULL, "invariant");
 959   assert(currentNode->_thread == current, "invariant");
 960   assert(_waiters > 0, "invariant");
 961   assert(object()->mark() == markWord::encode(this), "invariant");
 962 
 963   assert(current->thread_state() != _thread_blocked, "invariant");
 964 
 965   int nWakeups = 0;
 966   for (;;) {
 967     ObjectWaiter::TStates v = currentNode->TState;
 968     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
 969     assert(owner_raw() != current, "invariant");
 970 
 971     if (TryLock(current) > 0) break;
 972     if (TrySpin(current) > 0) break;
 973 
 974     {
 975       OSThreadContendState osts(current->osthread());
 976 
 977       assert(current->thread_state() == _thread_in_vm, "invariant");
 978 
 979       {
 980         ClearSuccOnSuspend csos(this);
 981         ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
 982         current->_ParkEvent->park();
 983       }
 984     }
 985 
 986     // Try again, but just so we distinguish between futile wakeups and
 987     // successful wakeups.  The following test isn't algorithmically
 988     // necessary, but it helps us maintain sensible statistics.
 989     if (TryLock(current) > 0) break;
 990 
 991     // The lock is still contested.
 992     // Keep a tally of the # of futile wakeups.
 993     // Note that the counter is not protected by a lock or updated by atomics.
 994     // That is by design - we trade "lossy" counters which are exposed to
 995     // races during updates for a lower probe effect.
 996     ++nWakeups;
 997 
 998     // Assuming this is not a spurious wakeup we'll normally
 999     // find that _succ == current.
1000     if (_succ == current) _succ = NULL;
1001 
1002     // Invariant: after clearing _succ a contending thread
1003     // *must* retry  _owner before parking.
1004     OrderAccess::fence();
1005 
1006     // This PerfData object can be used in parallel with a safepoint.
1007     // See the work around in PerfDataManager::destroy().
1008     OM_PERFDATA_OP(FutileWakeups, inc());
1009   }
1010 
1011   // current has acquired the lock -- Unlink current from the cxq or EntryList .
1012   // Normally we'll find current on the EntryList.
1013   // Unlinking from the EntryList is constant-time and atomic-free.
1014   // From the perspective of the lock owner (this thread), the
1015   // EntryList is stable and cxq is prepend-only.
1016   // The head of cxq is volatile but the interior is stable.
1017   // In addition, current.TState is stable.
1018 
1019   assert(owner_raw() == current, "invariant");
1020   assert(object()->mark() == markWord::encode(this), "invariant");
1021   UnlinkAfterAcquire(current, currentNode);
1022   if (_succ == current) _succ = NULL;
1023   assert(_succ != current, "invariant");
1024   currentNode->TState = ObjectWaiter::TS_RUN;
1025   OrderAccess::fence();      // see comments at the end of EnterI()
1026 }
1027 
1028 // By convention we unlink a contending thread from EntryList|cxq immediately
1029 // after the thread acquires the lock in ::enter().  Equally, we could defer
1030 // unlinking the thread until ::exit()-time.
1031 
1032 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode) {
1033   assert(owner_raw() == current, "invariant");
1034   assert(currentNode->_thread == current, "invariant");
1035 
1036   if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1037     // Normal case: remove current from the DLL EntryList .
1038     // This is a constant-time operation.
1039     ObjectWaiter* nxt = currentNode->_next;
1040     ObjectWaiter* prv = currentNode->_prev;
1041     if (nxt != NULL) nxt->_prev = prv;
1042     if (prv != NULL) prv->_next = nxt;
1043     if (currentNode == _EntryList) _EntryList = nxt;
1044     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1045     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1046   } else {
1047     assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1048     // Inopportune interleaving -- current is still on the cxq.
1049     // This usually means the enqueue of self raced an exiting thread.
1050     // Normally we'll find current near the front of the cxq, so
1051     // dequeueing is typically fast.  If needbe we can accelerate
1052     // this with some MCS/CHL-like bidirectional list hints and advisory
1053     // back-links so dequeueing from the interior will normally operate
1054     // in constant-time.
1055     // Dequeue current from either the head (with CAS) or from the interior
1056     // with a linear-time scan and normal non-atomic memory operations.
1057     // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1058     // and then unlink current from EntryList.  We have to drain eventually,
1059     // so it might as well be now.
1060 
1061     ObjectWaiter* v = _cxq;
1062     assert(v != NULL, "invariant");
1063     if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1064       // The CAS above can fail from interference IFF a "RAT" arrived.
1065       // In that case current must be in the interior and can no longer be
1066       // at the head of cxq.
1067       if (v == currentNode) {
1068         assert(_cxq != v, "invariant");
1069         v = _cxq;          // CAS above failed - start scan at head of list
1070       }
1071       ObjectWaiter* p;
1072       ObjectWaiter* q = NULL;
1073       for (p = v; p != NULL && p != currentNode; p = p->_next) {
1074         q = p;
1075         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1076       }
1077       assert(v != currentNode, "invariant");
1078       assert(p == currentNode, "Node not found on cxq");
1079       assert(p != _cxq, "invariant");
1080       assert(q != NULL, "invariant");
1081       assert(q->_next == p, "invariant");
1082       q->_next = p->_next;
1083     }
1084   }
1085 
1086 #ifdef ASSERT
1087   // Diagnostic hygiene ...
1088   currentNode->_prev  = (ObjectWaiter*) 0xBAD;
1089   currentNode->_next  = (ObjectWaiter*) 0xBAD;
1090   currentNode->TState = ObjectWaiter::TS_RUN;
1091 #endif
1092 }
1093 
1094 // -----------------------------------------------------------------------------
1095 // Exit support
1096 //
1097 // exit()
1098 // ~~~~~~
1099 // Note that the collector can't reclaim the objectMonitor or deflate
1100 // the object out from underneath the thread calling ::exit() as the
1101 // thread calling ::exit() never transitions to a stable state.
1102 // This inhibits GC, which in turn inhibits asynchronous (and
1103 // inopportune) reclamation of "this".
1104 //
1105 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1106 // There's one exception to the claim above, however.  EnterI() can call
1107 // exit() to drop a lock if the acquirer has been externally suspended.
1108 // In that case exit() is called with _thread_state == _thread_blocked,
1109 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1110 //
1111 // 1-0 exit
1112 // ~~~~~~~~
1113 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1114 // the fast-path operators have been optimized so the common ::exit()
1115 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1116 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
1117 // greatly improves latency -- MEMBAR and CAS having considerable local
1118 // latency on modern processors -- but at the cost of "stranding".  Absent the
1119 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1120 // ::enter() path, resulting in the entering thread being stranding
1121 // and a progress-liveness failure.   Stranding is extremely rare.
1122 // We use timers (timed park operations) & periodic polling to detect
1123 // and recover from stranding.  Potentially stranded threads periodically
1124 // wake up and poll the lock.  See the usage of the _Responsible variable.
1125 //
1126 // The CAS() in enter provides for safety and exclusion, while the CAS or
1127 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
1128 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1129 // We detect and recover from stranding with timers.
1130 //
1131 // If a thread transiently strands it'll park until (a) another
1132 // thread acquires the lock and then drops the lock, at which time the
1133 // exiting thread will notice and unpark the stranded thread, or, (b)
1134 // the timer expires.  If the lock is high traffic then the stranding latency
1135 // will be low due to (a).  If the lock is low traffic then the odds of
1136 // stranding are lower, although the worst-case stranding latency
1137 // is longer.  Critically, we don't want to put excessive load in the
1138 // platform's timer subsystem.  We want to minimize both the timer injection
1139 // rate (timers created/sec) as well as the number of timers active at
1140 // any one time.  (more precisely, we want to minimize timer-seconds, which is
1141 // the integral of the # of active timers at any instant over time).
1142 // Both impinge on OS scalability.  Given that, at most one thread parked on
1143 // a monitor will use a timer.
1144 //
1145 // There is also the risk of a futile wake-up. If we drop the lock
1146 // another thread can reacquire the lock immediately, and we can
1147 // then wake a thread unnecessarily. This is benign, and we've
1148 // structured the code so the windows are short and the frequency
1149 // of such futile wakups is low.
1150 
1151 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1152   void* cur = owner_raw();
1153   if (current != cur) {
1154     if (current->is_lock_owned((address)cur)) {
1155       assert(_recursions == 0, "invariant");
1156       set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1157       _recursions = 0;
1158     } else {
1159       // Apparent unbalanced locking ...
1160       // Naively we'd like to throw IllegalMonitorStateException.
1161       // As a practical matter we can neither allocate nor throw an
1162       // exception as ::exit() can be called from leaf routines.
1163       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1164       // Upon deeper reflection, however, in a properly run JVM the only
1165       // way we should encounter this situation is in the presence of
1166       // unbalanced JNI locking. TODO: CheckJNICalls.
1167       // See also: CR4414101
1168 #ifdef ASSERT
1169       LogStreamHandle(Error, monitorinflation) lsh;
1170       lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1171                     " is exiting an ObjectMonitor it does not own.", p2i(current));
1172       lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1173       print_debug_style_on(&lsh);
1174       assert(false, "Non-balanced monitor enter/exit!");
1175 #endif
1176       return;
1177     }
1178   }
1179 
1180   if (_recursions != 0) {
1181     _recursions--;        // this is simple recursive enter
1182     return;
1183   }
1184 
1185   // Invariant: after setting Responsible=null an thread must execute
1186   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1187   _Responsible = NULL;
1188 
1189 #if INCLUDE_JFR
1190   // get the owner's thread id for the MonitorEnter event
1191   // if it is enabled and the thread isn't suspended
1192   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1193     _previous_owner_tid = JFR_THREAD_ID(current);
1194   }
1195 #endif
1196 
1197   for (;;) {
1198     assert(current == owner_raw(), "invariant");
1199 
1200     // Drop the lock.
1201     // release semantics: prior loads and stores from within the critical section
1202     // must not float (reorder) past the following store that drops the lock.
1203     // Uses a storeload to separate release_store(owner) from the
1204     // successor check. The try_set_owner() below uses cmpxchg() so
1205     // we get the fence down there.
1206     release_clear_owner(current);
1207     OrderAccess::storeload();
1208 
1209     if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1210       return;
1211     }
1212     // Other threads are blocked trying to acquire the lock.
1213 
1214     // Normally the exiting thread is responsible for ensuring succession,
1215     // but if other successors are ready or other entering threads are spinning
1216     // then this thread can simply store NULL into _owner and exit without
1217     // waking a successor.  The existence of spinners or ready successors
1218     // guarantees proper succession (liveness).  Responsibility passes to the
1219     // ready or running successors.  The exiting thread delegates the duty.
1220     // More precisely, if a successor already exists this thread is absolved
1221     // of the responsibility of waking (unparking) one.
1222     //
1223     // The _succ variable is critical to reducing futile wakeup frequency.
1224     // _succ identifies the "heir presumptive" thread that has been made
1225     // ready (unparked) but that has not yet run.  We need only one such
1226     // successor thread to guarantee progress.
1227     // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1228     // section 3.3 "Futile Wakeup Throttling" for details.
1229     //
1230     // Note that spinners in Enter() also set _succ non-null.
1231     // In the current implementation spinners opportunistically set
1232     // _succ so that exiting threads might avoid waking a successor.
1233     // Another less appealing alternative would be for the exiting thread
1234     // to drop the lock and then spin briefly to see if a spinner managed
1235     // to acquire the lock.  If so, the exiting thread could exit
1236     // immediately without waking a successor, otherwise the exiting
1237     // thread would need to dequeue and wake a successor.
1238     // (Note that we'd need to make the post-drop spin short, but no
1239     // shorter than the worst-case round-trip cache-line migration time.
1240     // The dropped lock needs to become visible to the spinner, and then
1241     // the acquisition of the lock by the spinner must become visible to
1242     // the exiting thread).
1243 
1244     // It appears that an heir-presumptive (successor) must be made ready.
1245     // Only the current lock owner can manipulate the EntryList or
1246     // drain _cxq, so we need to reacquire the lock.  If we fail
1247     // to reacquire the lock the responsibility for ensuring succession
1248     // falls to the new owner.
1249     //
1250     if (try_set_owner_from(NULL, current) != NULL) {
1251       return;
1252     }
1253 
1254     guarantee(owner_raw() == current, "invariant");
1255 
1256     ObjectWaiter* w = NULL;
1257 
1258     w = _EntryList;
1259     if (w != NULL) {
1260       // I'd like to write: guarantee (w->_thread != current).
1261       // But in practice an exiting thread may find itself on the EntryList.
1262       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1263       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1264       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1265       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1266       // release the lock "O".  T2 resumes immediately after the ST of null into
1267       // _owner, above.  T2 notices that the EntryList is populated, so it
1268       // reacquires the lock and then finds itself on the EntryList.
1269       // Given all that, we have to tolerate the circumstance where "w" is
1270       // associated with current.
1271       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1272       ExitEpilog(current, w);
1273       return;
1274     }
1275 
1276     // If we find that both _cxq and EntryList are null then just
1277     // re-run the exit protocol from the top.
1278     w = _cxq;
1279     if (w == NULL) continue;
1280 
1281     // Drain _cxq into EntryList - bulk transfer.
1282     // First, detach _cxq.
1283     // The following loop is tantamount to: w = swap(&cxq, NULL)
1284     for (;;) {
1285       assert(w != NULL, "Invariant");
1286       ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL);
1287       if (u == w) break;
1288       w = u;
1289     }
1290 
1291     assert(w != NULL, "invariant");
1292     assert(_EntryList == NULL, "invariant");
1293 
1294     // Convert the LIFO SLL anchored by _cxq into a DLL.
1295     // The list reorganization step operates in O(LENGTH(w)) time.
1296     // It's critical that this step operate quickly as
1297     // "current" still holds the outer-lock, restricting parallelism
1298     // and effectively lengthening the critical section.
1299     // Invariant: s chases t chases u.
1300     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1301     // we have faster access to the tail.
1302 
1303     _EntryList = w;
1304     ObjectWaiter* q = NULL;
1305     ObjectWaiter* p;
1306     for (p = w; p != NULL; p = p->_next) {
1307       guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1308       p->TState = ObjectWaiter::TS_ENTER;
1309       p->_prev = q;
1310       q = p;
1311     }
1312 
1313     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1314     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1315 
1316     // See if we can abdicate to a spinner instead of waking a thread.
1317     // A primary goal of the implementation is to reduce the
1318     // context-switch rate.
1319     if (_succ != NULL) continue;
1320 
1321     w = _EntryList;
1322     if (w != NULL) {
1323       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1324       ExitEpilog(current, w);
1325       return;
1326     }
1327   }
1328 }
1329 
1330 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1331   assert(owner_raw() == current, "invariant");
1332 
1333   // Exit protocol:
1334   // 1. ST _succ = wakee
1335   // 2. membar #loadstore|#storestore;
1336   // 2. ST _owner = NULL
1337   // 3. unpark(wakee)
1338 
1339   _succ = Wakee->_thread;
1340   ParkEvent * Trigger = Wakee->_event;
1341 
1342   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1343   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1344   // out-of-scope (non-extant).
1345   Wakee  = NULL;
1346 
1347   // Drop the lock.
1348   // Uses a fence to separate release_store(owner) from the LD in unpark().
1349   release_clear_owner(current);
1350   OrderAccess::fence();
1351 
1352   DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1353   Trigger->unpark();
1354 
1355   // Maintain stats and report events to JVMTI
1356   OM_PERFDATA_OP(Parks, inc());
1357 }
1358 
1359 
1360 // -----------------------------------------------------------------------------
1361 // Class Loader deadlock handling.
1362 //
1363 // complete_exit exits a lock returning recursion count
1364 // complete_exit/reenter operate as a wait without waiting
1365 // complete_exit requires an inflated monitor
1366 // The _owner field is not always the Thread addr even with an
1367 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1368 // thread due to contention.
1369 intx ObjectMonitor::complete_exit(JavaThread* current) {
1370   assert(InitDone, "Unexpectedly not initialized");
1371 
1372   void* cur = owner_raw();
1373   if (current != cur) {
1374     if (current->is_lock_owned((address)cur)) {
1375       assert(_recursions == 0, "internal state error");
1376       set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1377       _recursions = 0;
1378     }
1379   }
1380 
1381   guarantee(current == owner_raw(), "complete_exit not owner");
1382   intx save = _recursions; // record the old recursion count
1383   _recursions = 0;         // set the recursion level to be 0
1384   exit(current);           // exit the monitor
1385   guarantee(owner_raw() != current, "invariant");
1386   return save;
1387 }
1388 
1389 // reenter() enters a lock and sets recursion count
1390 // complete_exit/reenter operate as a wait without waiting
1391 bool ObjectMonitor::reenter(intx recursions, JavaThread* current) {
1392 
1393   guarantee(owner_raw() != current, "reenter already owner");
1394   if (!enter(current)) {
1395     return false;
1396   }
1397   // Entered the monitor.
1398   guarantee(_recursions == 0, "reenter recursion");
1399   _recursions = recursions;
1400   return true;
1401 }
1402 
1403 // Checks that the current THREAD owns this monitor and causes an
1404 // immediate return if it doesn't. We don't use the CHECK macro
1405 // because we want the IMSE to be the only exception that is thrown
1406 // from the call site when false is returned. Any other pending
1407 // exception is ignored.
1408 #define CHECK_OWNER()                                                  \
1409   do {                                                                 \
1410     if (!check_owner(THREAD)) {                                        \
1411        assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1412        return;                                                         \
1413      }                                                                 \
1414   } while (false)
1415 
1416 // Returns true if the specified thread owns the ObjectMonitor.
1417 // Otherwise returns false and throws IllegalMonitorStateException
1418 // (IMSE). If there is a pending exception and the specified thread
1419 // is not the owner, that exception will be replaced by the IMSE.
1420 bool ObjectMonitor::check_owner(TRAPS) {
1421   JavaThread* current = THREAD;
1422   void* cur = owner_raw();
1423   if (cur == current) {
1424     return true;
1425   }
1426   if (current->is_lock_owned((address)cur)) {
1427     set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1428     _recursions = 0;
1429     return true;
1430   }
1431   THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1432              "current thread is not owner", false);
1433 }
1434 
1435 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1436                                     ObjectMonitor* monitor,
1437                                     uint64_t notifier_tid,
1438                                     jlong timeout,
1439                                     bool timedout) {
1440   assert(event != NULL, "invariant");
1441   assert(monitor != NULL, "invariant");
1442   event->set_monitorClass(monitor->object()->klass());
1443   event->set_timeout(timeout);
1444   // Set an address that is 'unique enough', such that events close in
1445   // time and with the same address are likely (but not guaranteed) to
1446   // belong to the same object.
1447   event->set_address((uintptr_t)monitor);
1448   event->set_notifier(notifier_tid);
1449   event->set_timedOut(timedout);
1450   event->commit();
1451 }
1452 
1453 // -----------------------------------------------------------------------------
1454 // Wait/Notify/NotifyAll
1455 //
1456 // Note: a subset of changes to ObjectMonitor::wait()
1457 // will need to be replicated in complete_exit
1458 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1459   JavaThread* current = THREAD;
1460 
1461   assert(InitDone, "Unexpectedly not initialized");
1462 
1463   CHECK_OWNER();  // Throws IMSE if not owner.
1464 
1465   EventJavaMonitorWait event;
1466 
1467   // check for a pending interrupt
1468   if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1469     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1470     if (JvmtiExport::should_post_monitor_waited()) {
1471       // Note: 'false' parameter is passed here because the
1472       // wait was not timed out due to thread interrupt.
1473       JvmtiExport::post_monitor_waited(current, this, false);
1474 
1475       // In this short circuit of the monitor wait protocol, the
1476       // current thread never drops ownership of the monitor and
1477       // never gets added to the wait queue so the current thread
1478       // cannot be made the successor. This means that the
1479       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1480       // consume an unpark() meant for the ParkEvent associated with
1481       // this ObjectMonitor.
1482     }
1483     if (event.should_commit()) {
1484       post_monitor_wait_event(&event, this, 0, millis, false);
1485     }
1486     THROW(vmSymbols::java_lang_InterruptedException());
1487     return;
1488   }
1489 
1490   assert(current->_Stalled == 0, "invariant");
1491   current->_Stalled = intptr_t(this);
1492   current->set_current_waiting_monitor(this);
1493 
1494   // create a node to be put into the queue
1495   // Critically, after we reset() the event but prior to park(), we must check
1496   // for a pending interrupt.
1497   ObjectWaiter node(current);
1498   node.TState = ObjectWaiter::TS_WAIT;
1499   current->_ParkEvent->reset();
1500   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1501 
1502   // Enter the waiting queue, which is a circular doubly linked list in this case
1503   // but it could be a priority queue or any data structure.
1504   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1505   // by the the owner of the monitor *except* in the case where park()
1506   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1507   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1508 
1509   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1510   AddWaiter(&node);
1511   Thread::SpinRelease(&_WaitSetLock);
1512 
1513   _Responsible = NULL;
1514 
1515   intx save = _recursions;     // record the old recursion count
1516   _waiters++;                  // increment the number of waiters
1517   _recursions = 0;             // set the recursion level to be 1
1518   exit(current);               // exit the monitor
1519   guarantee(owner_raw() != current, "invariant");
1520 
1521   // The thread is on the WaitSet list - now park() it.
1522   // On MP systems it's conceivable that a brief spin before we park
1523   // could be profitable.
1524   //
1525   // TODO-FIXME: change the following logic to a loop of the form
1526   //   while (!timeout && !interrupted && _notified == 0) park()
1527 
1528   int ret = OS_OK;
1529   int WasNotified = 0;
1530 
1531   // Need to check interrupt state whilst still _thread_in_vm
1532   bool interrupted = interruptible && current->is_interrupted(false);
1533 
1534   { // State transition wrappers
1535     OSThread* osthread = current->osthread();
1536     OSThreadWaitState osts(osthread, true);
1537 
1538     assert(current->thread_state() == _thread_in_vm, "invariant");
1539 
1540     {
1541       ClearSuccOnSuspend csos(this);
1542       ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1543       if (interrupted || HAS_PENDING_EXCEPTION) {
1544         // Intentionally empty
1545       } else if (node._notified == 0) {
1546         if (millis <= 0) {
1547           current->_ParkEvent->park();
1548         } else {
1549           ret = current->_ParkEvent->park(millis);
1550         }
1551       }
1552     }
1553 
1554     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1555     // from the WaitSet to the EntryList.
1556     // See if we need to remove Node from the WaitSet.
1557     // We use double-checked locking to avoid grabbing _WaitSetLock
1558     // if the thread is not on the wait queue.
1559     //
1560     // Note that we don't need a fence before the fetch of TState.
1561     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1562     // written by the is thread. (perhaps the fetch might even be satisfied
1563     // by a look-aside into the processor's own store buffer, although given
1564     // the length of the code path between the prior ST and this load that's
1565     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1566     // then we'll acquire the lock and then re-fetch a fresh TState value.
1567     // That is, we fail toward safety.
1568 
1569     if (node.TState == ObjectWaiter::TS_WAIT) {
1570       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1571       if (node.TState == ObjectWaiter::TS_WAIT) {
1572         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1573         assert(node._notified == 0, "invariant");
1574         node.TState = ObjectWaiter::TS_RUN;
1575       }
1576       Thread::SpinRelease(&_WaitSetLock);
1577     }
1578 
1579     // The thread is now either on off-list (TS_RUN),
1580     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1581     // The Node's TState variable is stable from the perspective of this thread.
1582     // No other threads will asynchronously modify TState.
1583     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1584     OrderAccess::loadload();
1585     if (_succ == current) _succ = NULL;
1586     WasNotified = node._notified;
1587 
1588     // Reentry phase -- reacquire the monitor.
1589     // re-enter contended monitor after object.wait().
1590     // retain OBJECT_WAIT state until re-enter successfully completes
1591     // Thread state is thread_in_vm and oop access is again safe,
1592     // although the raw address of the object may have changed.
1593     // (Don't cache naked oops over safepoints, of course).
1594 
1595     // post monitor waited event. Note that this is past-tense, we are done waiting.
1596     if (JvmtiExport::should_post_monitor_waited()) {
1597       JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1598 
1599       if (node._notified != 0 && _succ == current) {
1600         // In this part of the monitor wait-notify-reenter protocol it
1601         // is possible (and normal) for another thread to do a fastpath
1602         // monitor enter-exit while this thread is still trying to get
1603         // to the reenter portion of the protocol.
1604         //
1605         // The ObjectMonitor was notified and the current thread is
1606         // the successor which also means that an unpark() has already
1607         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1608         // consume the unpark() that was done when the successor was
1609         // set because the same ParkEvent is shared between Java
1610         // monitors and JVM/TI RawMonitors (for now).
1611         //
1612         // We redo the unpark() to ensure forward progress, i.e., we
1613         // don't want all pending threads hanging (parked) with none
1614         // entering the unlocked monitor.
1615         node._event->unpark();
1616       }
1617     }
1618 
1619     if (event.should_commit()) {
1620       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1621     }
1622 
1623     OrderAccess::fence();
1624 
1625     assert(current->_Stalled != 0, "invariant");
1626     current->_Stalled = 0;
1627 
1628     assert(owner_raw() != current, "invariant");
1629     ObjectWaiter::TStates v = node.TState;
1630     if (v == ObjectWaiter::TS_RUN) {
1631       enter(current);
1632     } else {
1633       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1634       ReenterI(current, &node);
1635       node.wait_reenter_end(this);
1636     }
1637 
1638     // current has reacquired the lock.
1639     // Lifecycle - the node representing current must not appear on any queues.
1640     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1641     // want residual elements associated with this thread left on any lists.
1642     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1643     assert(owner_raw() == current, "invariant");
1644     assert(_succ != current, "invariant");
1645   } // OSThreadWaitState()
1646 
1647   current->set_current_waiting_monitor(NULL);
1648 
1649   guarantee(_recursions == 0, "invariant");
1650   _recursions = save      // restore the old recursion count
1651                 + JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current); //  increased by the deferred relock count
1652   _waiters--;             // decrement the number of waiters
1653 
1654   // Verify a few postconditions
1655   assert(owner_raw() == current, "invariant");
1656   assert(_succ != current, "invariant");
1657   assert(object()->mark() == markWord::encode(this), "invariant");
1658 
1659   // check if the notification happened
1660   if (!WasNotified) {
1661     // no, it could be timeout or Thread.interrupt() or both
1662     // check for interrupt event, otherwise it is timeout
1663     if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1664       THROW(vmSymbols::java_lang_InterruptedException());
1665     }
1666   }
1667 
1668   // NOTE: Spurious wake up will be consider as timeout.
1669   // Monitor notify has precedence over thread interrupt.
1670 }
1671 
1672 
1673 // Consider:
1674 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1675 // then instead of transferring a thread from the WaitSet to the EntryList
1676 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1677 
1678 void ObjectMonitor::INotify(JavaThread* current) {
1679   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1680   ObjectWaiter* iterator = DequeueWaiter();
1681   if (iterator != NULL) {
1682     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1683     guarantee(iterator->_notified == 0, "invariant");
1684     // Disposition - what might we do with iterator ?
1685     // a.  add it directly to the EntryList - either tail (policy == 1)
1686     //     or head (policy == 0).
1687     // b.  push it onto the front of the _cxq (policy == 2).
1688     // For now we use (b).
1689 
1690     iterator->TState = ObjectWaiter::TS_ENTER;
1691 
1692     iterator->_notified = 1;
1693     iterator->_notifier_tid = JFR_THREAD_ID(current);
1694 
1695     ObjectWaiter* list = _EntryList;
1696     if (list != NULL) {
1697       assert(list->_prev == NULL, "invariant");
1698       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1699       assert(list != iterator, "invariant");
1700     }
1701 
1702     // prepend to cxq
1703     if (list == NULL) {
1704       iterator->_next = iterator->_prev = NULL;
1705       _EntryList = iterator;
1706     } else {
1707       iterator->TState = ObjectWaiter::TS_CXQ;
1708       for (;;) {
1709         ObjectWaiter* front = _cxq;
1710         iterator->_next = front;
1711         if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1712           break;
1713         }
1714       }
1715     }
1716 
1717     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1718     // move the add-to-EntryList operation, above, outside the critical section
1719     // protected by _WaitSetLock.  In practice that's not useful.  With the
1720     // exception of  wait() timeouts and interrupts the monitor owner
1721     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1722     // on _WaitSetLock so it's not profitable to reduce the length of the
1723     // critical section.
1724 
1725     iterator->wait_reenter_begin(this);
1726   }
1727   Thread::SpinRelease(&_WaitSetLock);
1728 }
1729 
1730 // Consider: a not-uncommon synchronization bug is to use notify() when
1731 // notifyAll() is more appropriate, potentially resulting in stranded
1732 // threads; this is one example of a lost wakeup. A useful diagnostic
1733 // option is to force all notify() operations to behave as notifyAll().
1734 //
1735 // Note: We can also detect many such problems with a "minimum wait".
1736 // When the "minimum wait" is set to a small non-zero timeout value
1737 // and the program does not hang whereas it did absent "minimum wait",
1738 // that suggests a lost wakeup bug.
1739 
1740 void ObjectMonitor::notify(TRAPS) {
1741   JavaThread* current = THREAD;
1742   CHECK_OWNER();  // Throws IMSE if not owner.
1743   if (_WaitSet == NULL) {
1744     return;
1745   }
1746   DTRACE_MONITOR_PROBE(notify, this, object(), current);
1747   INotify(current);
1748   OM_PERFDATA_OP(Notifications, inc(1));
1749 }
1750 
1751 
1752 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1753 // from the waitset to the EntryList. This could be done more efficiently with a
1754 // single bulk transfer but in practice it's not time-critical. Beware too,
1755 // that in prepend-mode we invert the order of the waiters. Let's say that the
1756 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1757 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1758 
1759 void ObjectMonitor::notifyAll(TRAPS) {
1760   JavaThread* current = THREAD;
1761   CHECK_OWNER();  // Throws IMSE if not owner.
1762   if (_WaitSet == NULL) {
1763     return;
1764   }
1765 
1766   DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1767   int tally = 0;
1768   while (_WaitSet != NULL) {
1769     tally++;
1770     INotify(current);
1771   }
1772 
1773   OM_PERFDATA_OP(Notifications, inc(tally));
1774 }
1775 
1776 // -----------------------------------------------------------------------------
1777 // Adaptive Spinning Support
1778 //
1779 // Adaptive spin-then-block - rational spinning
1780 //
1781 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1782 // algorithm.  On high order SMP systems it would be better to start with
1783 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1784 // a contending thread could enqueue itself on the cxq and then spin locally
1785 // on a thread-specific variable such as its ParkEvent._Event flag.
1786 // That's left as an exercise for the reader.  Note that global spinning is
1787 // not problematic on Niagara, as the L2 cache serves the interconnect and
1788 // has both low latency and massive bandwidth.
1789 //
1790 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1791 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1792 // (duration) or we can fix the count at approximately the duration of
1793 // a context switch and vary the frequency.   Of course we could also
1794 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1795 // For a description of 'Adaptive spin-then-block mutual exclusion in
1796 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1797 //
1798 // This implementation varies the duration "D", where D varies with
1799 // the success rate of recent spin attempts. (D is capped at approximately
1800 // length of a round-trip context switch).  The success rate for recent
1801 // spin attempts is a good predictor of the success rate of future spin
1802 // attempts.  The mechanism adapts automatically to varying critical
1803 // section length (lock modality), system load and degree of parallelism.
1804 // D is maintained per-monitor in _SpinDuration and is initialized
1805 // optimistically.  Spin frequency is fixed at 100%.
1806 //
1807 // Note that _SpinDuration is volatile, but we update it without locks
1808 // or atomics.  The code is designed so that _SpinDuration stays within
1809 // a reasonable range even in the presence of races.  The arithmetic
1810 // operations on _SpinDuration are closed over the domain of legal values,
1811 // so at worst a race will install and older but still legal value.
1812 // At the very worst this introduces some apparent non-determinism.
1813 // We might spin when we shouldn't or vice-versa, but since the spin
1814 // count are relatively short, even in the worst case, the effect is harmless.
1815 //
1816 // Care must be taken that a low "D" value does not become an
1817 // an absorbing state.  Transient spinning failures -- when spinning
1818 // is overall profitable -- should not cause the system to converge
1819 // on low "D" values.  We want spinning to be stable and predictable
1820 // and fairly responsive to change and at the same time we don't want
1821 // it to oscillate, become metastable, be "too" non-deterministic,
1822 // or converge on or enter undesirable stable absorbing states.
1823 //
1824 // We implement a feedback-based control system -- using past behavior
1825 // to predict future behavior.  We face two issues: (a) if the
1826 // input signal is random then the spin predictor won't provide optimal
1827 // results, and (b) if the signal frequency is too high then the control
1828 // system, which has some natural response lag, will "chase" the signal.
1829 // (b) can arise from multimodal lock hold times.  Transient preemption
1830 // can also result in apparent bimodal lock hold times.
1831 // Although sub-optimal, neither condition is particularly harmful, as
1832 // in the worst-case we'll spin when we shouldn't or vice-versa.
1833 // The maximum spin duration is rather short so the failure modes aren't bad.
1834 // To be conservative, I've tuned the gain in system to bias toward
1835 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1836 // "rings" or oscillates between spinning and not spinning.  This happens
1837 // when spinning is just on the cusp of profitability, however, so the
1838 // situation is not dire.  The state is benign -- there's no need to add
1839 // hysteresis control to damp the transition rate between spinning and
1840 // not spinning.
1841 
1842 // Spinning: Fixed frequency (100%), vary duration
1843 int ObjectMonitor::TrySpin(JavaThread* current) {
1844   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1845   int ctr = Knob_FixedSpin;
1846   if (ctr != 0) {
1847     while (--ctr >= 0) {
1848       if (TryLock(current) > 0) return 1;
1849       SpinPause();
1850     }
1851     return 0;
1852   }
1853 
1854   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1855     if (TryLock(current) > 0) {
1856       // Increase _SpinDuration ...
1857       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1858       // Raising _SpurDuration to the poverty line is key.
1859       int x = _SpinDuration;
1860       if (x < Knob_SpinLimit) {
1861         if (x < Knob_Poverty) x = Knob_Poverty;
1862         _SpinDuration = x + Knob_BonusB;
1863       }
1864       return 1;
1865     }
1866     SpinPause();
1867   }
1868 
1869   // Admission control - verify preconditions for spinning
1870   //
1871   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1872   // becoming an absorbing state.  Put another way, we spin briefly to
1873   // sample, just in case the system load, parallelism, contention, or lock
1874   // modality changed.
1875   //
1876   // Consider the following alternative:
1877   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1878   // spin attempt.  "Periodically" might mean after a tally of
1879   // the # of failed spin attempts (or iterations) reaches some threshold.
1880   // This takes us into the realm of 1-out-of-N spinning, where we
1881   // hold the duration constant but vary the frequency.
1882 
1883   ctr = _SpinDuration;
1884   if (ctr <= 0) return 0;
1885 
1886   if (NotRunnable(current, (JavaThread*) owner_raw())) {
1887     return 0;
1888   }
1889 
1890   // We're good to spin ... spin ingress.
1891   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1892   // when preparing to LD...CAS _owner, etc and the CAS is likely
1893   // to succeed.
1894   if (_succ == NULL) {
1895     _succ = current;
1896   }
1897   Thread* prv = NULL;
1898 
1899   // There are three ways to exit the following loop:
1900   // 1.  A successful spin where this thread has acquired the lock.
1901   // 2.  Spin failure with prejudice
1902   // 3.  Spin failure without prejudice
1903 
1904   while (--ctr >= 0) {
1905 
1906     // Periodic polling -- Check for pending GC
1907     // Threads may spin while they're unsafe.
1908     // We don't want spinning threads to delay the JVM from reaching
1909     // a stop-the-world safepoint or to steal cycles from GC.
1910     // If we detect a pending safepoint we abort in order that
1911     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1912     // this thread, if safe, doesn't steal cycles from GC.
1913     // This is in keeping with the "no loitering in runtime" rule.
1914     // We periodically check to see if there's a safepoint pending.
1915     if ((ctr & 0xFF) == 0) {
1916       // Can't call SafepointMechanism::should_process() since that
1917       // might update the poll values and we could be in a thread_blocked
1918       // state here which is not allowed so just check the poll.
1919       if (SafepointMechanism::local_poll_armed(current)) {
1920         goto Abort;           // abrupt spin egress
1921       }
1922       SpinPause();
1923     }
1924 
1925     // Probe _owner with TATAS
1926     // If this thread observes the monitor transition or flicker
1927     // from locked to unlocked to locked, then the odds that this
1928     // thread will acquire the lock in this spin attempt go down
1929     // considerably.  The same argument applies if the CAS fails
1930     // or if we observe _owner change from one non-null value to
1931     // another non-null value.   In such cases we might abort
1932     // the spin without prejudice or apply a "penalty" to the
1933     // spin count-down variable "ctr", reducing it by 100, say.
1934 
1935     JavaThread* ox = (JavaThread*) owner_raw();
1936     if (ox == NULL) {
1937       ox = (JavaThread*)try_set_owner_from(NULL, current);
1938       if (ox == NULL) {
1939         // The CAS succeeded -- this thread acquired ownership
1940         // Take care of some bookkeeping to exit spin state.
1941         if (_succ == current) {
1942           _succ = NULL;
1943         }
1944 
1945         // Increase _SpinDuration :
1946         // The spin was successful (profitable) so we tend toward
1947         // longer spin attempts in the future.
1948         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1949         // If we acquired the lock early in the spin cycle it
1950         // makes sense to increase _SpinDuration proportionally.
1951         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1952         int x = _SpinDuration;
1953         if (x < Knob_SpinLimit) {
1954           if (x < Knob_Poverty) x = Knob_Poverty;
1955           _SpinDuration = x + Knob_Bonus;
1956         }
1957         return 1;
1958       }
1959 
1960       // The CAS failed ... we can take any of the following actions:
1961       // * penalize: ctr -= CASPenalty
1962       // * exit spin with prejudice -- goto Abort;
1963       // * exit spin without prejudice.
1964       // * Since CAS is high-latency, retry again immediately.
1965       prv = ox;
1966       goto Abort;
1967     }
1968 
1969     // Did lock ownership change hands ?
1970     if (ox != prv && prv != NULL) {
1971       goto Abort;
1972     }
1973     prv = ox;
1974 
1975     // Abort the spin if the owner is not executing.
1976     // The owner must be executing in order to drop the lock.
1977     // Spinning while the owner is OFFPROC is idiocy.
1978     // Consider: ctr -= RunnablePenalty ;
1979     if (NotRunnable(current, ox)) {
1980       goto Abort;
1981     }
1982     if (_succ == NULL) {
1983       _succ = current;
1984     }
1985   }
1986 
1987   // Spin failed with prejudice -- reduce _SpinDuration.
1988   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1989   // AIMD is globally stable.
1990   {
1991     int x = _SpinDuration;
1992     if (x > 0) {
1993       // Consider an AIMD scheme like: x -= (x >> 3) + 100
1994       // This is globally sample and tends to damp the response.
1995       x -= Knob_Penalty;
1996       if (x < 0) x = 0;
1997       _SpinDuration = x;
1998     }
1999   }
2000 
2001  Abort:
2002   if (_succ == current) {
2003     _succ = NULL;
2004     // Invariant: after setting succ=null a contending thread
2005     // must recheck-retry _owner before parking.  This usually happens
2006     // in the normal usage of TrySpin(), but it's safest
2007     // to make TrySpin() as foolproof as possible.
2008     OrderAccess::fence();
2009     if (TryLock(current) > 0) return 1;
2010   }
2011   return 0;
2012 }
2013 
2014 // NotRunnable() -- informed spinning
2015 //
2016 // Don't bother spinning if the owner is not eligible to drop the lock.
2017 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2018 // The thread must be runnable in order to drop the lock in timely fashion.
2019 // If the _owner is not runnable then spinning will not likely be
2020 // successful (profitable).
2021 //
2022 // Beware -- the thread referenced by _owner could have died
2023 // so a simply fetch from _owner->_thread_state might trap.
2024 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2025 // Because of the lifecycle issues, the _thread_state values
2026 // observed by NotRunnable() might be garbage.  NotRunnable must
2027 // tolerate this and consider the observed _thread_state value
2028 // as advisory.
2029 //
2030 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2031 // a thread pointer.
2032 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2033 // with the LSB of _owner.  Another option would be to probabilistically probe
2034 // the putative _owner->TypeTag value.
2035 //
2036 // Checking _thread_state isn't perfect.  Even if the thread is
2037 // in_java it might be blocked on a page-fault or have been preempted
2038 // and sitting on a ready/dispatch queue.
2039 //
2040 // The return value from NotRunnable() is *advisory* -- the
2041 // result is based on sampling and is not necessarily coherent.
2042 // The caller must tolerate false-negative and false-positive errors.
2043 // Spinning, in general, is probabilistic anyway.
2044 
2045 
2046 int ObjectMonitor::NotRunnable(JavaThread* current, JavaThread* ox) {
2047   // Check ox->TypeTag == 2BAD.
2048   if (ox == NULL) return 0;
2049 
2050   // Avoid transitive spinning ...
2051   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2052   // Immediately after T1 acquires L it's possible that T2, also
2053   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2054   // This occurs transiently after T1 acquired L but before
2055   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2056   // its spin in this circumstance.
2057   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2058 
2059   if (BlockedOn == 1) return 1;
2060   if (BlockedOn != 0) {
2061     return BlockedOn != intptr_t(this) && owner_raw() == ox;
2062   }
2063 
2064   assert(sizeof(ox->_thread_state == sizeof(int)), "invariant");
2065   int jst = SafeFetch32((int *) &ox->_thread_state, -1);;
2066   // consider also: jst != _thread_in_Java -- but that's overspecific.
2067   return jst == _thread_blocked || jst == _thread_in_native;
2068 }
2069 
2070 
2071 // -----------------------------------------------------------------------------
2072 // WaitSet management ...
2073 
2074 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2075   _next     = NULL;
2076   _prev     = NULL;
2077   _notified = 0;
2078   _notifier_tid = 0;
2079   TState    = TS_RUN;
2080   _thread   = current;
2081   _event    = _thread->_ParkEvent;
2082   _active   = false;
2083   assert(_event != NULL, "invariant");
2084 }
2085 
2086 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2087   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2088 }
2089 
2090 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2091   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2092 }
2093 
2094 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2095   assert(node != NULL, "should not add NULL node");
2096   assert(node->_prev == NULL, "node already in list");
2097   assert(node->_next == NULL, "node already in list");
2098   // put node at end of queue (circular doubly linked list)
2099   if (_WaitSet == NULL) {
2100     _WaitSet = node;
2101     node->_prev = node;
2102     node->_next = node;
2103   } else {
2104     ObjectWaiter* head = _WaitSet;
2105     ObjectWaiter* tail = head->_prev;
2106     assert(tail->_next == head, "invariant check");
2107     tail->_next = node;
2108     head->_prev = node;
2109     node->_next = head;
2110     node->_prev = tail;
2111   }
2112 }
2113 
2114 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2115   // dequeue the very first waiter
2116   ObjectWaiter* waiter = _WaitSet;
2117   if (waiter) {
2118     DequeueSpecificWaiter(waiter);
2119   }
2120   return waiter;
2121 }
2122 
2123 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2124   assert(node != NULL, "should not dequeue NULL node");
2125   assert(node->_prev != NULL, "node already removed from list");
2126   assert(node->_next != NULL, "node already removed from list");
2127   // when the waiter has woken up because of interrupt,
2128   // timeout or other spurious wake-up, dequeue the
2129   // waiter from waiting list
2130   ObjectWaiter* next = node->_next;
2131   if (next == node) {
2132     assert(node->_prev == node, "invariant check");
2133     _WaitSet = NULL;
2134   } else {
2135     ObjectWaiter* prev = node->_prev;
2136     assert(prev->_next == node, "invariant check");
2137     assert(next->_prev == node, "invariant check");
2138     next->_prev = prev;
2139     prev->_next = next;
2140     if (_WaitSet == node) {
2141       _WaitSet = next;
2142     }
2143   }
2144   node->_next = NULL;
2145   node->_prev = NULL;
2146 }
2147 
2148 // -----------------------------------------------------------------------------
2149 // PerfData support
2150 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2151 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2152 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2153 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2154 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2155 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2156 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2157 
2158 // One-shot global initialization for the sync subsystem.
2159 // We could also defer initialization and initialize on-demand
2160 // the first time we call ObjectSynchronizer::inflate().
2161 // Initialization would be protected - like so many things - by
2162 // the MonitorCache_lock.
2163 
2164 void ObjectMonitor::Initialize() {
2165   assert(!InitDone, "invariant");
2166 
2167   if (!os::is_MP()) {
2168     Knob_SpinLimit = 0;
2169     Knob_PreSpin   = 0;
2170     Knob_FixedSpin = -1;
2171   }
2172 
2173   if (UsePerfData) {
2174     EXCEPTION_MARK;
2175 #define NEWPERFCOUNTER(n)                                                \
2176   {                                                                      \
2177     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2178                                         CHECK);                          \
2179   }
2180 #define NEWPERFVARIABLE(n)                                                \
2181   {                                                                       \
2182     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2183                                          CHECK);                          \
2184   }
2185     NEWPERFCOUNTER(_sync_Inflations);
2186     NEWPERFCOUNTER(_sync_Deflations);
2187     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2188     NEWPERFCOUNTER(_sync_FutileWakeups);
2189     NEWPERFCOUNTER(_sync_Parks);
2190     NEWPERFCOUNTER(_sync_Notifications);
2191     NEWPERFVARIABLE(_sync_MonExtant);
2192 #undef NEWPERFCOUNTER
2193 #undef NEWPERFVARIABLE
2194   }
2195 
2196   _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2197 
2198   DEBUG_ONLY(InitDone = true;)
2199 }
2200 
2201 void ObjectMonitor::print_on(outputStream* st) const {
2202   // The minimal things to print for markWord printing, more can be added for debugging and logging.
2203   st->print("{contentions=0x%08x,waiters=0x%08x"
2204             ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2205             contentions(), waiters(), recursions(),
2206             p2i(owner()));
2207 }
2208 void ObjectMonitor::print() const { print_on(tty); }
2209 
2210 #ifdef ASSERT
2211 // Print the ObjectMonitor like a debugger would:
2212 //
2213 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2214 //   _header = 0x0000000000000001
2215 //   _object = 0x000000070ff45fd0
2216 //   _pad_buf0 = {
2217 //     [0] = '\0'
2218 //     ...
2219 //     [43] = '\0'
2220 //   }
2221 //   _owner = 0x0000000000000000
2222 //   _previous_owner_tid = 0
2223 //   _pad_buf1 = {
2224 //     [0] = '\0'
2225 //     ...
2226 //     [47] = '\0'
2227 //   }
2228 //   _next_om = 0x0000000000000000
2229 //   _recursions = 0
2230 //   _EntryList = 0x0000000000000000
2231 //   _cxq = 0x0000000000000000
2232 //   _succ = 0x0000000000000000
2233 //   _Responsible = 0x0000000000000000
2234 //   _Spinner = 0
2235 //   _SpinDuration = 5000
2236 //   _contentions = 0
2237 //   _WaitSet = 0x0000700009756248
2238 //   _waiters = 1
2239 //   _WaitSetLock = 0
2240 // }
2241 //
2242 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2243   st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2244   st->print_cr("  _header = " INTPTR_FORMAT, header().value());
2245   st->print_cr("  _object = " INTPTR_FORMAT, p2i(object_peek()));
2246   st->print_cr("  _pad_buf0 = {");
2247   st->print_cr("    [0] = '\\0'");
2248   st->print_cr("    ...");
2249   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2250   st->print_cr("  }");
2251   st->print_cr("  _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2252   st->print_cr("  _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid);
2253   st->print_cr("  _pad_buf1 = {");
2254   st->print_cr("    [0] = '\\0'");
2255   st->print_cr("    ...");
2256   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2257   st->print_cr("  }");
2258   st->print_cr("  _next_om = " INTPTR_FORMAT, p2i(next_om()));
2259   st->print_cr("  _recursions = " INTX_FORMAT, _recursions);
2260   st->print_cr("  _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2261   st->print_cr("  _cxq = " INTPTR_FORMAT, p2i(_cxq));
2262   st->print_cr("  _succ = " INTPTR_FORMAT, p2i(_succ));
2263   st->print_cr("  _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2264   st->print_cr("  _Spinner = %d", _Spinner);
2265   st->print_cr("  _SpinDuration = %d", _SpinDuration);
2266   st->print_cr("  _contentions = %d", contentions());
2267   st->print_cr("  _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2268   st->print_cr("  _waiters = %d", _waiters);
2269   st->print_cr("  _WaitSetLock = %d", _WaitSetLock);
2270   st->print_cr("}");
2271 }
2272 #endif