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