1 /*
   2  * Copyright (c) 1998, 2024, 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/globals.hpp"
  43 #include "runtime/handles.inline.hpp"
  44 #include "runtime/interfaceSupport.inline.hpp"
  45 #include "runtime/javaThread.inline.hpp"
  46 #include "runtime/mutexLocker.hpp"
  47 #include "runtime/objectMonitor.hpp"
  48 #include "runtime/objectMonitor.inline.hpp"
  49 #include "runtime/orderAccess.hpp"
  50 #include "runtime/osThread.hpp"
  51 #include "runtime/perfData.hpp"
  52 #include "runtime/safefetch.hpp"
  53 #include "runtime/safepointMechanism.inline.hpp"
  54 #include "runtime/sharedRuntime.hpp"
  55 #include "services/threadService.hpp"
  56 #include "utilities/dtrace.hpp"
  57 #include "utilities/globalDefinitions.hpp"
  58 #include "utilities/macros.hpp"
  59 #include "utilities/preserveException.hpp"
  60 #if INCLUDE_JFR
  61 #include "jfr/support/jfrFlush.hpp"
  62 #endif
  63 
  64 #ifdef DTRACE_ENABLED
  65 
  66 // Only bother with this argument setup if dtrace is available
  67 // TODO-FIXME: probes should not fire when caller is _blocked.  assert() accordingly.
  68 
  69 
  70 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread)                           \
  71   char* bytes = nullptr;                                                   \
  72   int len = 0;                                                             \
  73   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  74   Symbol* klassname = obj->klass()->name();                                \
  75   if (klassname != nullptr) {                                              \
  76     bytes = (char*)klassname->bytes();                                     \
  77     len = klassname->utf8_length();                                        \
  78   }
  79 
  80 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)            \
  81   {                                                                        \
  82     if (DTraceMonitorProbes) {                                             \
  83       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  84       HOTSPOT_MONITOR_WAIT(jtid,                                           \
  85                            (monitor), bytes, len, (millis));               \
  86     }                                                                      \
  87   }
  88 
  89 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
  90 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
  91 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
  92 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
  93 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
  94 
  95 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)                  \
  96   {                                                                        \
  97     if (DTraceMonitorProbes) {                                             \
  98       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  99       HOTSPOT_MONITOR_##probe(jtid,                                        \
 100                               (uintptr_t)(monitor), bytes, len);           \
 101     }                                                                      \
 102   }
 103 
 104 #else //  ndef DTRACE_ENABLED
 105 
 106 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon)    {;}
 107 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon)          {;}
 108 
 109 #endif // ndef DTRACE_ENABLED
 110 
 111 // Tunables ...
 112 // The knob* variables are effectively final.  Once set they should
 113 // never be modified hence.  Consider using __read_mostly with GCC.
 114 
 115 int ObjectMonitor::Knob_SpinLimit    = 5000;    // derived by an external tool -
 116 
 117 static int Knob_Bonus               = 100;     // spin success bonus
 118 static int Knob_BonusB              = 100;     // spin success bonus
 119 static int Knob_Penalty             = 200;     // spin failure penalty
 120 static int Knob_Poverty             = 1000;
 121 static int Knob_FixedSpin           = 0;
 122 static int Knob_PreSpin             = 10;      // 20-100 likely better
 123 
 124 DEBUG_ONLY(static volatile bool InitDone = false;)
 125 
 126 OopStorage* ObjectMonitor::_oop_storage = nullptr;
 127 
 128 // -----------------------------------------------------------------------------
 129 // Theory of operations -- Monitors lists, thread residency, etc:
 130 //
 131 // * A thread acquires ownership of a monitor by successfully
 132 //   CAS()ing the _owner field from null to non-null.
 133 //
 134 // * Invariant: A thread appears on at most one monitor list --
 135 //   cxq, EntryList or WaitSet -- at any one time.
 136 //
 137 // * Contending threads "push" themselves onto the cxq with CAS
 138 //   and then spin/park.
 139 //
 140 // * After a contending thread eventually acquires the lock it must
 141 //   dequeue itself from either the EntryList or the cxq.
 142 //
 143 // * The exiting thread identifies and unparks an "heir presumptive"
 144 //   tentative successor thread on the EntryList.  Critically, the
 145 //   exiting thread doesn't unlink the successor thread from the EntryList.
 146 //   After having been unparked, the wakee will recontend for ownership of
 147 //   the monitor.   The successor (wakee) will either acquire the lock or
 148 //   re-park itself.
 149 //
 150 //   Succession is provided for by a policy of competitive handoff.
 151 //   The exiting thread does _not_ grant or pass ownership to the
 152 //   successor thread.  (This is also referred to as "handoff" succession").
 153 //   Instead the exiting thread releases ownership and possibly wakes
 154 //   a successor, so the successor can (re)compete for ownership of the lock.
 155 //   If the EntryList is empty but the cxq is populated the exiting
 156 //   thread will drain the cxq into the EntryList.  It does so by
 157 //   by detaching the cxq (installing null with CAS) and folding
 158 //   the threads from the cxq into the EntryList.  The EntryList is
 159 //   doubly linked, while the cxq is singly linked because of the
 160 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 161 //
 162 // * Concurrency invariants:
 163 //
 164 //   -- only the monitor owner may access or mutate the EntryList.
 165 //      The mutex property of the monitor itself protects the EntryList
 166 //      from concurrent interference.
 167 //   -- Only the monitor owner may detach the cxq.
 168 //
 169 // * The monitor entry list operations avoid locks, but strictly speaking
 170 //   they're not lock-free.  Enter is lock-free, exit is not.
 171 //   For a description of 'Methods and apparatus providing non-blocking access
 172 //   to a resource,' see U.S. Pat. No. 7844973.
 173 //
 174 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 175 //   detaching thread.  This mechanism is immune from the ABA corruption.
 176 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 177 //
 178 // * Taken together, the cxq and the EntryList constitute or form a
 179 //   single logical queue of threads stalled trying to acquire the lock.
 180 //   We use two distinct lists to improve the odds of a constant-time
 181 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 182 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 183 //   A key desideratum is to minimize queue & monitor metadata manipulation
 184 //   that occurs while holding the monitor lock -- that is, we want to
 185 //   minimize monitor lock holds times.  Note that even a small amount of
 186 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 187 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 188 //   locks and monitor metadata.
 189 //
 190 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 191 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 192 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 193 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 194 //
 195 //   The EntryList is ordered by the prevailing queue discipline and
 196 //   can be organized in any convenient fashion, such as a doubly-linked list or
 197 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 198 //   to operate in constant-time.  If we need a priority queue then something akin
 199 //   to Solaris' sleepq would work nicely.  Viz.,
 200 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 201 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 202 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 203 //   EntryList accordingly.
 204 //
 205 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 206 //   somewhat similar to an elevator-scan.
 207 //
 208 // * The monitor synchronization subsystem avoids the use of native
 209 //   synchronization primitives except for the narrow platform-specific
 210 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 211 //   the semantics of park-unpark.  Put another way, this monitor implementation
 212 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 213 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 214 //   underlying OS manages the READY<->RUN transitions.
 215 //
 216 // * Waiting threads reside on the WaitSet list -- wait() puts
 217 //   the caller onto the WaitSet.
 218 //
 219 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 220 //   either the EntryList or cxq.  Subsequent exit() operations will
 221 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 222 //   it's likely the notifyee would simply impale itself on the lock held
 223 //   by the notifier.
 224 //
 225 // * An interesting alternative is to encode cxq as (List,LockByte) where
 226 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 227 //   variable, like _recursions, in the scheme.  The threads or Events that form
 228 //   the list would have to be aligned in 256-byte addresses.  A thread would
 229 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 230 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 231 //   Note that is is *not* word-tearing, but it does presume that full-word
 232 //   CAS operations are coherent with intermix with STB operations.  That's true
 233 //   on most common processors.
 234 //
 235 // * See also http://blogs.sun.com/dave
 236 
 237 
 238 // Check that object() and set_object() are called from the right context:
 239 static void check_object_context() {
 240 #ifdef ASSERT
 241   Thread* self = Thread::current();
 242   if (self->is_Java_thread()) {
 243     // Mostly called from JavaThreads so sanity check the thread state.
 244     JavaThread* jt = JavaThread::cast(self);
 245     switch (jt->thread_state()) {
 246     case _thread_in_vm:    // the usual case
 247     case _thread_in_Java:  // during deopt
 248       break;
 249     default:
 250       fatal("called from an unsafe thread state");
 251     }
 252     assert(jt->is_active_Java_thread(), "must be active JavaThread");
 253   } else {
 254     // However, ThreadService::get_current_contended_monitor()
 255     // can call here via the VMThread so sanity check it.
 256     assert(self->is_VM_thread(), "must be");
 257   }
 258 #endif // ASSERT
 259 }
 260 
 261 ObjectMonitor::ObjectMonitor(oop object) :
 262   _header(markWord::zero()),
 263   _object(_oop_storage, object),
 264   _owner(nullptr),
 265   _previous_owner_tid(0),
 266   _next_om(nullptr),
 267   _recursions(0),
 268   _EntryList(nullptr),
 269   _cxq(nullptr),
 270   _succ(nullptr),
 271   _Responsible(nullptr),
 272   _Spinner(0),
 273   _SpinDuration(ObjectMonitor::Knob_SpinLimit),
 274   _contentions(0),
 275   _WaitSet(nullptr),
 276   _waiters(0),
 277   _WaitSetLock(0)
 278 { }
 279 
 280 ObjectMonitor::~ObjectMonitor() {
 281   _object.release(_oop_storage);
 282 }
 283 
 284 oop ObjectMonitor::object() const {
 285   check_object_context();
 286   return _object.resolve();
 287 }
 288 
 289 oop ObjectMonitor::object_peek() const {
 290   return _object.peek();
 291 }
 292 
 293 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
 294   if (current->is_suspended()) {
 295     _om->_recursions = 0;
 296     _om->_succ = nullptr;
 297     // Don't need a full fence after clearing successor here because of the call to exit().
 298     _om->exit(current, false /* not_suspended */);
 299     _om_exited = true;
 300 
 301     current->set_current_pending_monitor(_om);
 302   }
 303 }
 304 
 305 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
 306   if (current->is_suspended()) {
 307     if (_om->_succ == current) {
 308       _om->_succ = nullptr;
 309       OrderAccess::fence(); // always do a full fence when successor is cleared
 310     }
 311   }
 312 }
 313 
 314 // -----------------------------------------------------------------------------
 315 // Enter support
 316 
 317 bool ObjectMonitor::enter_for(JavaThread* locking_thread) {
 318   // Used by ObjectSynchronizer::enter_for to enter for another thread.
 319   // The monitor is private to or already owned by locking_thread which must be suspended.
 320   // So this code may only contend with deflation.
 321   assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
 322 
 323   // Block out deflation as soon as possible.
 324   add_to_contentions(1);
 325 
 326   bool success = false;
 327   if (!is_being_async_deflated()) {
 328     void* prev_owner = try_set_owner_from(nullptr, locking_thread);
 329 
 330     if (prev_owner == nullptr) {
 331       assert(_recursions == 0, "invariant");
 332       success = true;
 333     } else if (prev_owner == locking_thread) {
 334       _recursions++;
 335       success = true;
 336     } else if (prev_owner == DEFLATER_MARKER) {
 337       // Racing with deflation.
 338       prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread);
 339       if (prev_owner == DEFLATER_MARKER) {
 340         // Cancelled deflation. Increment contentions as part of the deflation protocol.
 341         add_to_contentions(1);
 342         success = true;
 343       } else if (prev_owner == nullptr) {
 344         // At this point we cannot race with deflation as we have both incremented
 345         // contentions, seen contention > 0 and seen a DEFLATER_MARKER.
 346         // success will only be false if this races with something other than
 347         // deflation.
 348         prev_owner = try_set_owner_from(nullptr, locking_thread);
 349         success = prev_owner == nullptr;
 350       }
 351     } else if (LockingMode == LM_LEGACY && locking_thread->is_lock_owned((address)prev_owner)) {
 352       assert(_recursions == 0, "must be");
 353       _recursions = 1;
 354       set_owner_from_BasicLock(prev_owner, locking_thread);
 355       success = true;
 356     }
 357     assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT
 358            ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT,
 359            p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner));
 360   } else {
 361     // Async deflation is in progress and our contentions increment
 362     // above lost the race to async deflation. Undo the work and
 363     // force the caller to retry.
 364     const oop l_object = object();
 365     if (l_object != nullptr) {
 366       // Attempt to restore the header/dmw to the object's header so that
 367       // we only retry once if the deflater thread happens to be slow.
 368       install_displaced_markword_in_object(l_object);
 369     }
 370   }
 371 
 372   add_to_contentions(-1);
 373 
 374   assert(!success || owner_raw() == locking_thread, "must be");
 375 
 376   return success;
 377 }
 378 
 379 bool ObjectMonitor::enter(JavaThread* current) {
 380   assert(current == JavaThread::current(), "must be");
 381   // The following code is ordered to check the most common cases first
 382   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 383 
 384   void* cur = try_set_owner_from(nullptr, current);
 385   if (cur == nullptr) {
 386     assert(_recursions == 0, "invariant");
 387     return true;
 388   }
 389 
 390   if (cur == current) {
 391     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 392     _recursions++;
 393     return true;
 394   }
 395 
 396   if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
 397     assert(_recursions == 0, "internal state error");
 398     _recursions = 1;
 399     set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
 400     return true;
 401   }
 402 
 403   // We've encountered genuine contention.
 404   assert(current->_Stalled == 0, "invariant");
 405   current->_Stalled = intptr_t(this);
 406 
 407   // Try one round of spinning *before* enqueueing current
 408   // and before going through the awkward and expensive state
 409   // transitions.  The following spin is strictly optional ...
 410   // Note that if we acquire the monitor from an initial spin
 411   // we forgo posting JVMTI events and firing DTRACE probes.
 412   if (TrySpin(current) > 0) {
 413     assert(owner_raw() == current, "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
 414     assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
 415     assert(object()->mark() == markWord::encode(this),
 416            "object mark must match encoded this: mark=" INTPTR_FORMAT
 417            ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
 418            markWord::encode(this).value());
 419     current->_Stalled = 0;
 420     return true;
 421   }
 422 
 423   assert(owner_raw() != current, "invariant");
 424   assert(_succ != current, "invariant");
 425   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 426   assert(current->thread_state() != _thread_blocked, "invariant");
 427 
 428   // Keep track of contention for JVM/TI and M&M queries.
 429   add_to_contentions(1);
 430   if (is_being_async_deflated()) {
 431     // Async deflation is in progress and our contentions increment
 432     // above lost the race to async deflation. Undo the work and
 433     // force the caller to retry.
 434     const oop l_object = object();
 435     if (l_object != nullptr) {
 436       // Attempt to restore the header/dmw to the object's header so that
 437       // we only retry once if the deflater thread happens to be slow.
 438       install_displaced_markword_in_object(l_object);
 439     }
 440     current->_Stalled = 0;
 441     add_to_contentions(-1);
 442     return false;
 443   }
 444 
 445   JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);)
 446   EventJavaMonitorEnter event;
 447   if (event.is_started()) {
 448     event.set_monitorClass(object()->klass());
 449     // Set an address that is 'unique enough', such that events close in
 450     // time and with the same address are likely (but not guaranteed) to
 451     // belong to the same object.
 452     event.set_address((uintptr_t)this);
 453   }
 454 
 455   { // Change java thread status to indicate blocked on monitor enter.
 456     JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
 457 
 458     assert(current->current_pending_monitor() == nullptr, "invariant");
 459     current->set_current_pending_monitor(this);
 460 
 461     DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
 462     if (JvmtiExport::should_post_monitor_contended_enter()) {
 463       JvmtiExport::post_monitor_contended_enter(current, this);
 464 
 465       // The current thread does not yet own the monitor and does not
 466       // yet appear on any queues that would get it made the successor.
 467       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 468       // handler cannot accidentally consume an unpark() meant for the
 469       // ParkEvent associated with this ObjectMonitor.
 470     }
 471 
 472     OSThreadContendState osts(current->osthread());
 473 
 474     assert(current->thread_state() == _thread_in_vm, "invariant");
 475 
 476     for (;;) {
 477       ExitOnSuspend eos(this);
 478       {
 479         ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */);
 480         EnterI(current);
 481         current->set_current_pending_monitor(nullptr);
 482         // We can go to a safepoint at the end of this block. If we
 483         // do a thread dump during that safepoint, then this thread will show
 484         // as having "-locked" the monitor, but the OS and java.lang.Thread
 485         // states will still report that the thread is blocked trying to
 486         // acquire it.
 487         // If there is a suspend request, ExitOnSuspend will exit the OM
 488         // and set the OM as pending.
 489       }
 490       if (!eos.exited()) {
 491         // ExitOnSuspend did not exit the OM
 492         assert(owner_raw() == current, "invariant");
 493         break;
 494       }
 495     }
 496 
 497     // We've just gotten past the enter-check-for-suspend dance and we now own
 498     // the monitor free and clear.
 499   }
 500 
 501   add_to_contentions(-1);
 502   assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
 503   current->_Stalled = 0;
 504 
 505   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 506   assert(_recursions == 0, "invariant");
 507   assert(owner_raw() == current, "invariant");
 508   assert(_succ != current, "invariant");
 509   assert(object()->mark() == markWord::encode(this), "invariant");
 510 
 511   // The thread -- now the owner -- is back in vm mode.
 512   // Report the glorious news via TI,DTrace and jvmstat.
 513   // The probe effect is non-trivial.  All the reportage occurs
 514   // while we hold the monitor, increasing the length of the critical
 515   // section.  Amdahl's parallel speedup law comes vividly into play.
 516   //
 517   // Another option might be to aggregate the events (thread local or
 518   // per-monitor aggregation) and defer reporting until a more opportune
 519   // time -- such as next time some thread encounters contention but has
 520   // yet to acquire the lock.  While spinning that thread could
 521   // spinning we could increment JVMStat counters, etc.
 522 
 523   DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
 524   if (JvmtiExport::should_post_monitor_contended_entered()) {
 525     JvmtiExport::post_monitor_contended_entered(current, this);
 526 
 527     // The current thread already owns the monitor and is not going to
 528     // call park() for the remainder of the monitor enter protocol. So
 529     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 530     // event handler consumed an unpark() issued by the thread that
 531     // just exited the monitor.
 532   }
 533   if (event.should_commit()) {
 534     event.set_previousOwner(_previous_owner_tid);
 535     event.commit();
 536   }
 537   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 538   return true;
 539 }
 540 
 541 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 542 // Callers must compensate as needed.
 543 
 544 int ObjectMonitor::TryLock(JavaThread* current) {
 545   void* own = owner_raw();
 546   if (own != nullptr) return 0;
 547   if (try_set_owner_from(nullptr, current) == nullptr) {
 548     assert(_recursions == 0, "invariant");
 549     return 1;
 550   }
 551   // The lock had been free momentarily, but we lost the race to the lock.
 552   // Interference -- the CAS failed.
 553   // We can either return -1 or retry.
 554   // Retry doesn't make as much sense because the lock was just acquired.
 555   return -1;
 556 }
 557 
 558 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
 559 // was deflated and false otherwise.
 560 //
 561 // The async deflation protocol sets owner to DEFLATER_MARKER and
 562 // makes contentions negative as signals to contending threads that
 563 // an async deflation is in progress. There are a number of checks
 564 // as part of the protocol to make sure that the calling thread has
 565 // not lost the race to a contending thread.
 566 //
 567 // The ObjectMonitor has been successfully async deflated when:
 568 //   (contentions < 0)
 569 // Contending threads that see that condition know to retry their operation.
 570 //
 571 bool ObjectMonitor::deflate_monitor() {
 572   if (is_busy()) {
 573     // Easy checks are first - the ObjectMonitor is busy so no deflation.
 574     return false;
 575   }
 576 
 577   const oop obj = object_peek();
 578 
 579   if (obj == nullptr) {
 580     // If the object died, we can recycle the monitor without racing with
 581     // Java threads. The GC already broke the association with the object.
 582     set_owner_from(nullptr, DEFLATER_MARKER);
 583     assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
 584     _contentions = INT_MIN; // minimum negative int
 585   } else {
 586     // Attempt async deflation protocol.
 587 
 588     // Set a null owner to DEFLATER_MARKER to force any contending thread
 589     // through the slow path. This is just the first part of the async
 590     // deflation dance.
 591     if (try_set_owner_from(nullptr, DEFLATER_MARKER) != nullptr) {
 592       // The owner field is no longer null so we lost the race since the
 593       // ObjectMonitor is now busy.
 594       return false;
 595     }
 596 
 597     if (contentions() > 0 || _waiters != 0) {
 598       // Another thread has raced to enter the ObjectMonitor after
 599       // is_busy() above or has already entered and waited on
 600       // it which makes it busy so no deflation. Restore owner to
 601       // null if it is still DEFLATER_MARKER.
 602       if (try_set_owner_from(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
 603         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 604         add_to_contentions(-1);
 605       }
 606       return false;
 607     }
 608 
 609     // Make a zero contentions field negative to force any contending threads
 610     // to retry. This is the second part of the async deflation dance.
 611     if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) {
 612       // Contentions was no longer 0 so we lost the race since the
 613       // ObjectMonitor is now busy. Restore owner to null if it is
 614       // still DEFLATER_MARKER:
 615       if (try_set_owner_from(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
 616         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 617         add_to_contentions(-1);
 618       }
 619       return false;
 620     }
 621   }
 622 
 623   // Sanity checks for the races:
 624   guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
 625   guarantee(contentions() < 0, "must be negative: contentions=%d",
 626             contentions());
 627   guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
 628   guarantee(_cxq == nullptr, "must be no contending threads: cxq="
 629             INTPTR_FORMAT, p2i(_cxq));
 630   guarantee(_EntryList == nullptr,
 631             "must be no entering threads: EntryList=" INTPTR_FORMAT,
 632             p2i(_EntryList));
 633 
 634   if (obj != nullptr) {
 635     if (log_is_enabled(Trace, monitorinflation)) {
 636       ResourceMark rm;
 637       log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
 638                                   ", mark=" INTPTR_FORMAT ", type='%s'",
 639                                   p2i(obj), obj->mark().value(),
 640                                   obj->klass()->external_name());
 641     }
 642 
 643     // Install the old mark word if nobody else has already done it.
 644     install_displaced_markword_in_object(obj);
 645   }
 646 
 647   // We leave owner == DEFLATER_MARKER and contentions < 0
 648   // to force any racing threads to retry.
 649   return true;  // Success, ObjectMonitor has been deflated.
 650 }
 651 
 652 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
 653 // into the header of the object associated with the monitor. This
 654 // idempotent method is called by a thread that is deflating a
 655 // monitor and by other threads that have detected a race with the
 656 // deflation process.
 657 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
 658   // This function must only be called when (owner == DEFLATER_MARKER
 659   // && contentions <= 0), but we can't guarantee that here because
 660   // those values could change when the ObjectMonitor gets moved from
 661   // the global free list to a per-thread free list.
 662 
 663   guarantee(obj != nullptr, "must be non-null");
 664 
 665   // Separate loads in is_being_async_deflated(), which is almost always
 666   // called before this function, from the load of dmw/header below.
 667 
 668   // _contentions and dmw/header may get written by different threads.
 669   // Make sure to observe them in the same order when having several observers.
 670   OrderAccess::loadload_for_IRIW();
 671 
 672   const oop l_object = object_peek();
 673   if (l_object == nullptr) {
 674     // ObjectMonitor's object ref has already been cleared by async
 675     // deflation or GC so we're done here.
 676     return;
 677   }
 678   assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
 679          INTPTR_FORMAT, p2i(l_object), p2i(obj));
 680 
 681   markWord dmw = header();
 682   // The dmw has to be neutral (not null, not locked and not marked).
 683   assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
 684 
 685   // Install displaced mark word if the object's header still points
 686   // to this ObjectMonitor. More than one racing caller to this function
 687   // can rarely reach this point, but only one can win.
 688   markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
 689   if (res != markWord::encode(this)) {
 690     // This should be rare so log at the Info level when it happens.
 691     log_info(monitorinflation)("install_displaced_markword_in_object: "
 692                                "failed cas_set_mark: new_mark=" INTPTR_FORMAT
 693                                ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
 694                                dmw.value(), markWord::encode(this).value(),
 695                                res.value());
 696   }
 697 
 698   // Note: It does not matter which thread restored the header/dmw
 699   // into the object's header. The thread deflating the monitor just
 700   // wanted the object's header restored and it is. The threads that
 701   // detected a race with the deflation process also wanted the
 702   // object's header restored before they retry their operation and
 703   // because it is restored they will only retry once.
 704 }
 705 
 706 // Convert the fields used by is_busy() to a string that can be
 707 // used for diagnostic output.
 708 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
 709   ss->print("is_busy: waiters=%d, ", _waiters);
 710   if (contentions() > 0) {
 711     ss->print("contentions=%d, ", contentions());
 712   } else {
 713     ss->print("contentions=0");
 714   }
 715   if (!owner_is_DEFLATER_MARKER()) {
 716     ss->print("owner=" INTPTR_FORMAT, p2i(owner_raw()));
 717   } else {
 718     // We report null instead of DEFLATER_MARKER here because is_busy()
 719     // ignores DEFLATER_MARKER values.
 720     ss->print("owner=" INTPTR_FORMAT, NULL_WORD);
 721   }
 722   ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq),
 723             p2i(_EntryList));
 724   return ss->base();
 725 }
 726 
 727 #define MAX_RECHECK_INTERVAL 1000
 728 
 729 void ObjectMonitor::EnterI(JavaThread* current) {
 730   assert(current->thread_state() == _thread_blocked, "invariant");
 731 
 732   // Try the lock - TATAS
 733   if (TryLock (current) > 0) {
 734     assert(_succ != current, "invariant");
 735     assert(owner_raw() == current, "invariant");
 736     assert(_Responsible != current, "invariant");
 737     return;
 738   }
 739 
 740   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 741     // Cancelled the in-progress async deflation by changing owner from
 742     // DEFLATER_MARKER to current. As part of the contended enter protocol,
 743     // contentions was incremented to a positive value before EnterI()
 744     // was called and that prevents the deflater thread from winning the
 745     // last part of the 2-part async deflation protocol. After EnterI()
 746     // returns to enter(), contentions is decremented because the caller
 747     // now owns the monitor. We bump contentions an extra time here to
 748     // prevent the deflater thread from winning the last part of the
 749     // 2-part async deflation protocol after the regular decrement
 750     // occurs in enter(). The deflater thread will decrement contentions
 751     // after it recognizes that the async deflation was cancelled.
 752     add_to_contentions(1);
 753     assert(_succ != current, "invariant");
 754     assert(_Responsible != current, "invariant");
 755     return;
 756   }
 757 
 758   assert(InitDone, "Unexpectedly not initialized");
 759 
 760   // We try one round of spinning *before* enqueueing current.
 761   //
 762   // If the _owner is ready but OFFPROC we could use a YieldTo()
 763   // operation to donate the remainder of this thread's quantum
 764   // to the owner.  This has subtle but beneficial affinity
 765   // effects.
 766 
 767   if (TrySpin(current) > 0) {
 768     assert(owner_raw() == current, "invariant");
 769     assert(_succ != current, "invariant");
 770     assert(_Responsible != current, "invariant");
 771     return;
 772   }
 773 
 774   // The Spin failed -- Enqueue and park the thread ...
 775   assert(_succ != current, "invariant");
 776   assert(owner_raw() != current, "invariant");
 777   assert(_Responsible != current, "invariant");
 778 
 779   // Enqueue "current" on ObjectMonitor's _cxq.
 780   //
 781   // Node acts as a proxy for current.
 782   // As an aside, if were to ever rewrite the synchronization code mostly
 783   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 784   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 785   // as well as eliminate a subset of ABA issues.
 786   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 787 
 788   ObjectWaiter node(current);
 789   current->_ParkEvent->reset();
 790   node._prev   = (ObjectWaiter*) 0xBAD;
 791   node.TState  = ObjectWaiter::TS_CXQ;
 792 
 793   // Push "current" onto the front of the _cxq.
 794   // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
 795   // Note that spinning tends to reduce the rate at which threads
 796   // enqueue and dequeue on EntryList|cxq.
 797   ObjectWaiter* nxt;
 798   for (;;) {
 799     node._next = nxt = _cxq;
 800     if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
 801 
 802     // Interference - the CAS failed because _cxq changed.  Just retry.
 803     // As an optional optimization we retry the lock.
 804     if (TryLock (current) > 0) {
 805       assert(_succ != current, "invariant");
 806       assert(owner_raw() == current, "invariant");
 807       assert(_Responsible != current, "invariant");
 808       return;
 809     }
 810   }
 811 
 812   // Check for cxq|EntryList edge transition to non-null.  This indicates
 813   // the onset of contention.  While contention persists exiting threads
 814   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 815   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 816   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 817   // arrange for one of the contending thread to use a timed park() operations
 818   // to detect and recover from the race.  (Stranding is form of progress failure
 819   // where the monitor is unlocked but all the contending threads remain parked).
 820   // That is, at least one of the contended threads will periodically poll _owner.
 821   // One of the contending threads will become the designated "Responsible" thread.
 822   // The Responsible thread uses a timed park instead of a normal indefinite park
 823   // operation -- it periodically wakes and checks for and recovers from potential
 824   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 825   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 826   // be responsible for a monitor.
 827   //
 828   // Currently, one of the contended threads takes on the added role of "Responsible".
 829   // A viable alternative would be to use a dedicated "stranding checker" thread
 830   // that periodically iterated over all the threads (or active monitors) and unparked
 831   // successors where there was risk of stranding.  This would help eliminate the
 832   // timer scalability issues we see on some platforms as we'd only have one thread
 833   // -- the checker -- parked on a timer.
 834 
 835   if (nxt == nullptr && _EntryList == nullptr) {
 836     // Try to assume the role of responsible thread for the monitor.
 837     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=current }
 838     Atomic::replace_if_null(&_Responsible, current);
 839   }
 840 
 841   // The lock might have been released while this thread was occupied queueing
 842   // itself onto _cxq.  To close the race and avoid "stranding" and
 843   // progress-liveness failure we must resample-retry _owner before parking.
 844   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 845   // In this case the ST-MEMBAR is accomplished with CAS().
 846   //
 847   // TODO: Defer all thread state transitions until park-time.
 848   // Since state transitions are heavy and inefficient we'd like
 849   // to defer the state transitions until absolutely necessary,
 850   // and in doing so avoid some transitions ...
 851 
 852   int nWakeups = 0;
 853   int recheckInterval = 1;
 854 
 855   for (;;) {
 856 
 857     if (TryLock(current) > 0) break;
 858     assert(owner_raw() != current, "invariant");
 859 
 860     // park self
 861     if (_Responsible == current) {
 862       current->_ParkEvent->park((jlong) recheckInterval);
 863       // Increase the recheckInterval, but clamp the value.
 864       recheckInterval *= 8;
 865       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 866         recheckInterval = MAX_RECHECK_INTERVAL;
 867       }
 868     } else {
 869       current->_ParkEvent->park();
 870     }
 871 
 872     if (TryLock(current) > 0) break;
 873 
 874     if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 875       // Cancelled the in-progress async deflation by changing owner from
 876       // DEFLATER_MARKER to current. As part of the contended enter protocol,
 877       // contentions was incremented to a positive value before EnterI()
 878       // was called and that prevents the deflater thread from winning the
 879       // last part of the 2-part async deflation protocol. After EnterI()
 880       // returns to enter(), contentions is decremented because the caller
 881       // now owns the monitor. We bump contentions an extra time here to
 882       // prevent the deflater thread from winning the last part of the
 883       // 2-part async deflation protocol after the regular decrement
 884       // occurs in enter(). The deflater thread will decrement contentions
 885       // after it recognizes that the async deflation was cancelled.
 886       add_to_contentions(1);
 887       break;
 888     }
 889 
 890     // The lock is still contested.
 891     // Keep a tally of the # of futile wakeups.
 892     // Note that the counter is not protected by a lock or updated by atomics.
 893     // That is by design - we trade "lossy" counters which are exposed to
 894     // races during updates for a lower probe effect.
 895 
 896     // This PerfData object can be used in parallel with a safepoint.
 897     // See the work around in PerfDataManager::destroy().
 898     OM_PERFDATA_OP(FutileWakeups, inc());
 899     ++nWakeups;
 900 
 901     // Assuming this is not a spurious wakeup we'll normally find _succ == current.
 902     // We can defer clearing _succ until after the spin completes
 903     // TrySpin() must tolerate being called with _succ == current.
 904     // Try yet another round of adaptive spinning.
 905     if (TrySpin(current) > 0) break;
 906 
 907     // We can find that we were unpark()ed and redesignated _succ while
 908     // we were spinning.  That's harmless.  If we iterate and call park(),
 909     // park() will consume the event and return immediately and we'll
 910     // just spin again.  This pattern can repeat, leaving _succ to simply
 911     // spin on a CPU.
 912 
 913     if (_succ == current) _succ = nullptr;
 914 
 915     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 916     OrderAccess::fence();
 917   }
 918 
 919   // Egress :
 920   // current has acquired the lock -- Unlink current from the cxq or EntryList.
 921   // Normally we'll find current on the EntryList .
 922   // From the perspective of the lock owner (this thread), the
 923   // EntryList is stable and cxq is prepend-only.
 924   // The head of cxq is volatile but the interior is stable.
 925   // In addition, current.TState is stable.
 926 
 927   assert(owner_raw() == current, "invariant");
 928 
 929   UnlinkAfterAcquire(current, &node);
 930   if (_succ == current) _succ = nullptr;
 931 
 932   assert(_succ != current, "invariant");
 933   if (_Responsible == current) {
 934     _Responsible = nullptr;
 935     OrderAccess::fence(); // Dekker pivot-point
 936 
 937     // We may leave threads on cxq|EntryList without a designated
 938     // "Responsible" thread.  This is benign.  When this thread subsequently
 939     // exits the monitor it can "see" such preexisting "old" threads --
 940     // threads that arrived on the cxq|EntryList before the fence, above --
 941     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 942     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 943     // non-null and elect a new "Responsible" timer thread.
 944     //
 945     // This thread executes:
 946     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 947     //    LD cxq|EntryList               (in subsequent exit)
 948     //
 949     // Entering threads in the slow/contended path execute:
 950     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 951     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 952     //
 953     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 954     // exit operation from floating above the ST Responsible=null.
 955   }
 956 
 957   // We've acquired ownership with CAS().
 958   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 959   // But since the CAS() this thread may have also stored into _succ,
 960   // EntryList, cxq or Responsible.  These meta-data updates must be
 961   // visible __before this thread subsequently drops the lock.
 962   // Consider what could occur if we didn't enforce this constraint --
 963   // STs to monitor meta-data and user-data could reorder with (become
 964   // visible after) the ST in exit that drops ownership of the lock.
 965   // Some other thread could then acquire the lock, but observe inconsistent
 966   // or old monitor meta-data and heap data.  That violates the JMM.
 967   // To that end, the 1-0 exit() operation must have at least STST|LDST
 968   // "release" barrier semantics.  Specifically, there must be at least a
 969   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 970   // the lock.   The barrier ensures that changes to monitor meta-data and data
 971   // protected by the lock will be visible before we release the lock, and
 972   // therefore before some other thread (CPU) has a chance to acquire the lock.
 973   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 974   //
 975   // Critically, any prior STs to _succ or EntryList must be visible before
 976   // the ST of null into _owner in the *subsequent* (following) corresponding
 977   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 978   // execute a serializing instruction.
 979 
 980   return;
 981 }
 982 
 983 // ReenterI() is a specialized inline form of the latter half of the
 984 // contended slow-path from EnterI().  We use ReenterI() only for
 985 // monitor reentry in wait().
 986 //
 987 // In the future we should reconcile EnterI() and ReenterI().
 988 
 989 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
 990   assert(current != nullptr, "invariant");
 991   assert(currentNode != nullptr, "invariant");
 992   assert(currentNode->_thread == current, "invariant");
 993   assert(_waiters > 0, "invariant");
 994   assert(object()->mark() == markWord::encode(this), "invariant");
 995 
 996   assert(current->thread_state() != _thread_blocked, "invariant");
 997 
 998   int nWakeups = 0;
 999   for (;;) {
1000     ObjectWaiter::TStates v = currentNode->TState;
1001     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1002     assert(owner_raw() != current, "invariant");
1003 
1004     if (TryLock(current) > 0) break;
1005     if (TrySpin(current) > 0) break;
1006 
1007     {
1008       OSThreadContendState osts(current->osthread());
1009 
1010       assert(current->thread_state() == _thread_in_vm, "invariant");
1011 
1012       {
1013         ClearSuccOnSuspend csos(this);
1014         ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1015         current->_ParkEvent->park();
1016       }
1017     }
1018 
1019     // Try again, but just so we distinguish between futile wakeups and
1020     // successful wakeups.  The following test isn't algorithmically
1021     // necessary, but it helps us maintain sensible statistics.
1022     if (TryLock(current) > 0) break;
1023 
1024     // The lock is still contested.
1025     // Keep a tally of the # of futile wakeups.
1026     // Note that the counter is not protected by a lock or updated by atomics.
1027     // That is by design - we trade "lossy" counters which are exposed to
1028     // races during updates for a lower probe effect.
1029     ++nWakeups;
1030 
1031     // Assuming this is not a spurious wakeup we'll normally
1032     // find that _succ == current.
1033     if (_succ == current) _succ = nullptr;
1034 
1035     // Invariant: after clearing _succ a contending thread
1036     // *must* retry  _owner before parking.
1037     OrderAccess::fence();
1038 
1039     // This PerfData object can be used in parallel with a safepoint.
1040     // See the work around in PerfDataManager::destroy().
1041     OM_PERFDATA_OP(FutileWakeups, inc());
1042   }
1043 
1044   // current has acquired the lock -- Unlink current from the cxq or EntryList .
1045   // Normally we'll find current on the EntryList.
1046   // Unlinking from the EntryList is constant-time and atomic-free.
1047   // From the perspective of the lock owner (this thread), the
1048   // EntryList is stable and cxq is prepend-only.
1049   // The head of cxq is volatile but the interior is stable.
1050   // In addition, current.TState is stable.
1051 
1052   assert(owner_raw() == current, "invariant");
1053   assert(object()->mark() == markWord::encode(this), "invariant");
1054   UnlinkAfterAcquire(current, currentNode);
1055   if (_succ == current) _succ = nullptr;
1056   assert(_succ != current, "invariant");
1057   currentNode->TState = ObjectWaiter::TS_RUN;
1058   OrderAccess::fence();      // see comments at the end of EnterI()
1059 }
1060 
1061 // By convention we unlink a contending thread from EntryList|cxq immediately
1062 // after the thread acquires the lock in ::enter().  Equally, we could defer
1063 // unlinking the thread until ::exit()-time.
1064 
1065 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode) {
1066   assert(owner_raw() == current, "invariant");
1067   assert(currentNode->_thread == current, "invariant");
1068 
1069   if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1070     // Normal case: remove current from the DLL EntryList .
1071     // This is a constant-time operation.
1072     ObjectWaiter* nxt = currentNode->_next;
1073     ObjectWaiter* prv = currentNode->_prev;
1074     if (nxt != nullptr) nxt->_prev = prv;
1075     if (prv != nullptr) prv->_next = nxt;
1076     if (currentNode == _EntryList) _EntryList = nxt;
1077     assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1078     assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1079   } else {
1080     assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1081     // Inopportune interleaving -- current is still on the cxq.
1082     // This usually means the enqueue of self raced an exiting thread.
1083     // Normally we'll find current near the front of the cxq, so
1084     // dequeueing is typically fast.  If needbe we can accelerate
1085     // this with some MCS/CHL-like bidirectional list hints and advisory
1086     // back-links so dequeueing from the interior will normally operate
1087     // in constant-time.
1088     // Dequeue current from either the head (with CAS) or from the interior
1089     // with a linear-time scan and normal non-atomic memory operations.
1090     // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1091     // and then unlink current from EntryList.  We have to drain eventually,
1092     // so it might as well be now.
1093 
1094     ObjectWaiter* v = _cxq;
1095     assert(v != nullptr, "invariant");
1096     if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1097       // The CAS above can fail from interference IFF a "RAT" arrived.
1098       // In that case current must be in the interior and can no longer be
1099       // at the head of cxq.
1100       if (v == currentNode) {
1101         assert(_cxq != v, "invariant");
1102         v = _cxq;          // CAS above failed - start scan at head of list
1103       }
1104       ObjectWaiter* p;
1105       ObjectWaiter* q = nullptr;
1106       for (p = v; p != nullptr && p != currentNode; p = p->_next) {
1107         q = p;
1108         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1109       }
1110       assert(v != currentNode, "invariant");
1111       assert(p == currentNode, "Node not found on cxq");
1112       assert(p != _cxq, "invariant");
1113       assert(q != nullptr, "invariant");
1114       assert(q->_next == p, "invariant");
1115       q->_next = p->_next;
1116     }
1117   }
1118 
1119 #ifdef ASSERT
1120   // Diagnostic hygiene ...
1121   currentNode->_prev  = (ObjectWaiter*) 0xBAD;
1122   currentNode->_next  = (ObjectWaiter*) 0xBAD;
1123   currentNode->TState = ObjectWaiter::TS_RUN;
1124 #endif
1125 }
1126 
1127 // -----------------------------------------------------------------------------
1128 // Exit support
1129 //
1130 // exit()
1131 // ~~~~~~
1132 // Note that the collector can't reclaim the objectMonitor or deflate
1133 // the object out from underneath the thread calling ::exit() as the
1134 // thread calling ::exit() never transitions to a stable state.
1135 // This inhibits GC, which in turn inhibits asynchronous (and
1136 // inopportune) reclamation of "this".
1137 //
1138 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1139 // There's one exception to the claim above, however.  EnterI() can call
1140 // exit() to drop a lock if the acquirer has been externally suspended.
1141 // In that case exit() is called with _thread_state == _thread_blocked,
1142 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1143 //
1144 // 1-0 exit
1145 // ~~~~~~~~
1146 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1147 // the fast-path operators have been optimized so the common ::exit()
1148 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1149 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
1150 // greatly improves latency -- MEMBAR and CAS having considerable local
1151 // latency on modern processors -- but at the cost of "stranding".  Absent the
1152 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1153 // ::enter() path, resulting in the entering thread being stranding
1154 // and a progress-liveness failure.   Stranding is extremely rare.
1155 // We use timers (timed park operations) & periodic polling to detect
1156 // and recover from stranding.  Potentially stranded threads periodically
1157 // wake up and poll the lock.  See the usage of the _Responsible variable.
1158 //
1159 // The CAS() in enter provides for safety and exclusion, while the CAS or
1160 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
1161 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1162 // We detect and recover from stranding with timers.
1163 //
1164 // If a thread transiently strands it'll park until (a) another
1165 // thread acquires the lock and then drops the lock, at which time the
1166 // exiting thread will notice and unpark the stranded thread, or, (b)
1167 // the timer expires.  If the lock is high traffic then the stranding latency
1168 // will be low due to (a).  If the lock is low traffic then the odds of
1169 // stranding are lower, although the worst-case stranding latency
1170 // is longer.  Critically, we don't want to put excessive load in the
1171 // platform's timer subsystem.  We want to minimize both the timer injection
1172 // rate (timers created/sec) as well as the number of timers active at
1173 // any one time.  (more precisely, we want to minimize timer-seconds, which is
1174 // the integral of the # of active timers at any instant over time).
1175 // Both impinge on OS scalability.  Given that, at most one thread parked on
1176 // a monitor will use a timer.
1177 //
1178 // There is also the risk of a futile wake-up. If we drop the lock
1179 // another thread can reacquire the lock immediately, and we can
1180 // then wake a thread unnecessarily. This is benign, and we've
1181 // structured the code so the windows are short and the frequency
1182 // of such futile wakups is low.
1183 
1184 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1185   void* cur = owner_raw();
1186   if (current != cur) {
1187     if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1188       assert(_recursions == 0, "invariant");
1189       set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1190       _recursions = 0;
1191     } else {
1192       // Apparent unbalanced locking ...
1193       // Naively we'd like to throw IllegalMonitorStateException.
1194       // As a practical matter we can neither allocate nor throw an
1195       // exception as ::exit() can be called from leaf routines.
1196       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1197       // Upon deeper reflection, however, in a properly run JVM the only
1198       // way we should encounter this situation is in the presence of
1199       // unbalanced JNI locking. TODO: CheckJNICalls.
1200       // See also: CR4414101
1201 #ifdef ASSERT
1202       LogStreamHandle(Error, monitorinflation) lsh;
1203       lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1204                     " is exiting an ObjectMonitor it does not own.", p2i(current));
1205       lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1206       print_debug_style_on(&lsh);
1207       assert(false, "Non-balanced monitor enter/exit!");
1208 #endif
1209       return;
1210     }
1211   }
1212 
1213   if (_recursions != 0) {
1214     _recursions--;        // this is simple recursive enter
1215     return;
1216   }
1217 
1218   // Invariant: after setting Responsible=null an thread must execute
1219   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1220   _Responsible = nullptr;
1221 
1222 #if INCLUDE_JFR
1223   // get the owner's thread id for the MonitorEnter event
1224   // if it is enabled and the thread isn't suspended
1225   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1226     _previous_owner_tid = JFR_THREAD_ID(current);
1227   }
1228 #endif
1229 
1230   for (;;) {
1231     assert(current == owner_raw(), "invariant");
1232 
1233     // Drop the lock.
1234     // release semantics: prior loads and stores from within the critical section
1235     // must not float (reorder) past the following store that drops the lock.
1236     // Uses a storeload to separate release_store(owner) from the
1237     // successor check. The try_set_owner() below uses cmpxchg() so
1238     // we get the fence down there.
1239     release_clear_owner(current);
1240     OrderAccess::storeload();
1241 
1242     if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) {
1243       return;
1244     }
1245     // Other threads are blocked trying to acquire the lock.
1246 
1247     // Normally the exiting thread is responsible for ensuring succession,
1248     // but if other successors are ready or other entering threads are spinning
1249     // then this thread can simply store null into _owner and exit without
1250     // waking a successor.  The existence of spinners or ready successors
1251     // guarantees proper succession (liveness).  Responsibility passes to the
1252     // ready or running successors.  The exiting thread delegates the duty.
1253     // More precisely, if a successor already exists this thread is absolved
1254     // of the responsibility of waking (unparking) one.
1255     //
1256     // The _succ variable is critical to reducing futile wakeup frequency.
1257     // _succ identifies the "heir presumptive" thread that has been made
1258     // ready (unparked) but that has not yet run.  We need only one such
1259     // successor thread to guarantee progress.
1260     // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1261     // section 3.3 "Futile Wakeup Throttling" for details.
1262     //
1263     // Note that spinners in Enter() also set _succ non-null.
1264     // In the current implementation spinners opportunistically set
1265     // _succ so that exiting threads might avoid waking a successor.
1266     // Another less appealing alternative would be for the exiting thread
1267     // to drop the lock and then spin briefly to see if a spinner managed
1268     // to acquire the lock.  If so, the exiting thread could exit
1269     // immediately without waking a successor, otherwise the exiting
1270     // thread would need to dequeue and wake a successor.
1271     // (Note that we'd need to make the post-drop spin short, but no
1272     // shorter than the worst-case round-trip cache-line migration time.
1273     // The dropped lock needs to become visible to the spinner, and then
1274     // the acquisition of the lock by the spinner must become visible to
1275     // the exiting thread).
1276 
1277     // It appears that an heir-presumptive (successor) must be made ready.
1278     // Only the current lock owner can manipulate the EntryList or
1279     // drain _cxq, so we need to reacquire the lock.  If we fail
1280     // to reacquire the lock the responsibility for ensuring succession
1281     // falls to the new owner.
1282     //
1283     if (try_set_owner_from(nullptr, current) != nullptr) {
1284       return;
1285     }
1286 
1287     guarantee(owner_raw() == current, "invariant");
1288 
1289     ObjectWaiter* w = nullptr;
1290 
1291     w = _EntryList;
1292     if (w != nullptr) {
1293       // I'd like to write: guarantee (w->_thread != current).
1294       // But in practice an exiting thread may find itself on the EntryList.
1295       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1296       // then calls exit().  Exit release the lock by setting O._owner to null.
1297       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1298       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1299       // release the lock "O".  T2 resumes immediately after the ST of null into
1300       // _owner, above.  T2 notices that the EntryList is populated, so it
1301       // reacquires the lock and then finds itself on the EntryList.
1302       // Given all that, we have to tolerate the circumstance where "w" is
1303       // associated with current.
1304       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1305       ExitEpilog(current, w);
1306       return;
1307     }
1308 
1309     // If we find that both _cxq and EntryList are null then just
1310     // re-run the exit protocol from the top.
1311     w = _cxq;
1312     if (w == nullptr) continue;
1313 
1314     // Drain _cxq into EntryList - bulk transfer.
1315     // First, detach _cxq.
1316     // The following loop is tantamount to: w = swap(&cxq, nullptr)
1317     for (;;) {
1318       assert(w != nullptr, "Invariant");
1319       ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr);
1320       if (u == w) break;
1321       w = u;
1322     }
1323 
1324     assert(w != nullptr, "invariant");
1325     assert(_EntryList == nullptr, "invariant");
1326 
1327     // Convert the LIFO SLL anchored by _cxq into a DLL.
1328     // The list reorganization step operates in O(LENGTH(w)) time.
1329     // It's critical that this step operate quickly as
1330     // "current" still holds the outer-lock, restricting parallelism
1331     // and effectively lengthening the critical section.
1332     // Invariant: s chases t chases u.
1333     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1334     // we have faster access to the tail.
1335 
1336     _EntryList = w;
1337     ObjectWaiter* q = nullptr;
1338     ObjectWaiter* p;
1339     for (p = w; p != nullptr; p = p->_next) {
1340       guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1341       p->TState = ObjectWaiter::TS_ENTER;
1342       p->_prev = q;
1343       q = p;
1344     }
1345 
1346     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr
1347     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1348 
1349     // See if we can abdicate to a spinner instead of waking a thread.
1350     // A primary goal of the implementation is to reduce the
1351     // context-switch rate.
1352     if (_succ != nullptr) continue;
1353 
1354     w = _EntryList;
1355     if (w != nullptr) {
1356       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1357       ExitEpilog(current, w);
1358       return;
1359     }
1360   }
1361 }
1362 
1363 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1364   assert(owner_raw() == current, "invariant");
1365 
1366   // Exit protocol:
1367   // 1. ST _succ = wakee
1368   // 2. membar #loadstore|#storestore;
1369   // 2. ST _owner = nullptr
1370   // 3. unpark(wakee)
1371 
1372   _succ = Wakee->_thread;
1373   ParkEvent * Trigger = Wakee->_event;
1374 
1375   // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again.
1376   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1377   // out-of-scope (non-extant).
1378   Wakee  = nullptr;
1379 
1380   // Drop the lock.
1381   // Uses a fence to separate release_store(owner) from the LD in unpark().
1382   release_clear_owner(current);
1383   OrderAccess::fence();
1384 
1385   DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1386   Trigger->unpark();
1387 
1388   // Maintain stats and report events to JVMTI
1389   OM_PERFDATA_OP(Parks, inc());
1390 }
1391 
1392 // complete_exit exits a lock returning recursion count
1393 // complete_exit requires an inflated monitor
1394 // The _owner field is not always the Thread addr even with an
1395 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1396 // thread due to contention.
1397 intx ObjectMonitor::complete_exit(JavaThread* current) {
1398   assert(InitDone, "Unexpectedly not initialized");
1399 
1400   void* cur = owner_raw();
1401   if (current != cur) {
1402     if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1403       assert(_recursions == 0, "internal state error");
1404       set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1405       _recursions = 0;
1406     }
1407   }
1408 
1409   guarantee(current == owner_raw(), "complete_exit not owner");
1410   intx save = _recursions; // record the old recursion count
1411   _recursions = 0;         // set the recursion level to be 0
1412   exit(current);           // exit the monitor
1413   guarantee(owner_raw() != current, "invariant");
1414   return save;
1415 }
1416 
1417 // Checks that the current THREAD owns this monitor and causes an
1418 // immediate return if it doesn't. We don't use the CHECK macro
1419 // because we want the IMSE to be the only exception that is thrown
1420 // from the call site when false is returned. Any other pending
1421 // exception is ignored.
1422 #define CHECK_OWNER()                                                  \
1423   do {                                                                 \
1424     if (!check_owner(THREAD)) {                                        \
1425        assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1426        return;                                                         \
1427      }                                                                 \
1428   } while (false)
1429 
1430 // Returns true if the specified thread owns the ObjectMonitor.
1431 // Otherwise returns false and throws IllegalMonitorStateException
1432 // (IMSE). If there is a pending exception and the specified thread
1433 // is not the owner, that exception will be replaced by the IMSE.
1434 bool ObjectMonitor::check_owner(TRAPS) {
1435   JavaThread* current = THREAD;
1436   void* cur = owner_raw();
1437   assert(cur != anon_owner_ptr(), "no anon owner here");
1438   if (cur == current) {
1439     return true;
1440   }
1441   if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1442     set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1443     _recursions = 0;
1444     return true;
1445   }
1446   THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1447              "current thread is not owner", false);
1448 }
1449 
1450 static inline bool is_excluded(const Klass* monitor_klass) {
1451   assert(monitor_klass != nullptr, "invariant");
1452   NOT_JFR_RETURN_(false);
1453   JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name());
1454 }
1455 
1456 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1457                                     ObjectMonitor* monitor,
1458                                     uint64_t notifier_tid,
1459                                     jlong timeout,
1460                                     bool timedout) {
1461   assert(event != nullptr, "invariant");
1462   assert(monitor != nullptr, "invariant");
1463   const Klass* monitor_klass = monitor->object()->klass();
1464   if (is_excluded(monitor_klass)) {
1465     return;
1466   }
1467   event->set_monitorClass(monitor_klass);
1468   event->set_timeout(timeout);
1469   // Set an address that is 'unique enough', such that events close in
1470   // time and with the same address are likely (but not guaranteed) to
1471   // belong to the same object.
1472   event->set_address((uintptr_t)monitor);
1473   event->set_notifier(notifier_tid);
1474   event->set_timedOut(timedout);
1475   event->commit();
1476 }
1477 
1478 // -----------------------------------------------------------------------------
1479 // Wait/Notify/NotifyAll
1480 //
1481 // Note: a subset of changes to ObjectMonitor::wait()
1482 // will need to be replicated in complete_exit
1483 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1484   JavaThread* current = THREAD;
1485 
1486   assert(InitDone, "Unexpectedly not initialized");
1487 
1488   CHECK_OWNER();  // Throws IMSE if not owner.
1489 
1490   EventJavaMonitorWait event;
1491 
1492   // check for a pending interrupt
1493   if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1494     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1495     if (JvmtiExport::should_post_monitor_waited()) {
1496       // Note: 'false' parameter is passed here because the
1497       // wait was not timed out due to thread interrupt.
1498       JvmtiExport::post_monitor_waited(current, this, false);
1499 
1500       // In this short circuit of the monitor wait protocol, the
1501       // current thread never drops ownership of the monitor and
1502       // never gets added to the wait queue so the current thread
1503       // cannot be made the successor. This means that the
1504       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1505       // consume an unpark() meant for the ParkEvent associated with
1506       // this ObjectMonitor.
1507     }
1508     if (event.should_commit()) {
1509       post_monitor_wait_event(&event, this, 0, millis, false);
1510     }
1511     THROW(vmSymbols::java_lang_InterruptedException());
1512     return;
1513   }
1514 
1515   assert(current->_Stalled == 0, "invariant");
1516   current->_Stalled = intptr_t(this);
1517   current->set_current_waiting_monitor(this);
1518 
1519   // create a node to be put into the queue
1520   // Critically, after we reset() the event but prior to park(), we must check
1521   // for a pending interrupt.
1522   ObjectWaiter node(current);
1523   node.TState = ObjectWaiter::TS_WAIT;
1524   current->_ParkEvent->reset();
1525   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1526 
1527   // Enter the waiting queue, which is a circular doubly linked list in this case
1528   // but it could be a priority queue or any data structure.
1529   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1530   // by the owner of the monitor *except* in the case where park()
1531   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1532   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1533 
1534   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1535   AddWaiter(&node);
1536   Thread::SpinRelease(&_WaitSetLock);
1537 
1538   _Responsible = nullptr;
1539 
1540   intx save = _recursions;     // record the old recursion count
1541   _waiters++;                  // increment the number of waiters
1542   _recursions = 0;             // set the recursion level to be 1
1543   exit(current);               // exit the monitor
1544   guarantee(owner_raw() != current, "invariant");
1545 
1546   // The thread is on the WaitSet list - now park() it.
1547   // On MP systems it's conceivable that a brief spin before we park
1548   // could be profitable.
1549   //
1550   // TODO-FIXME: change the following logic to a loop of the form
1551   //   while (!timeout && !interrupted && _notified == 0) park()
1552 
1553   int ret = OS_OK;
1554   int WasNotified = 0;
1555 
1556   // Need to check interrupt state whilst still _thread_in_vm
1557   bool interrupted = interruptible && current->is_interrupted(false);
1558 
1559   { // State transition wrappers
1560     OSThread* osthread = current->osthread();
1561     OSThreadWaitState osts(osthread, true);
1562 
1563     assert(current->thread_state() == _thread_in_vm, "invariant");
1564 
1565     {
1566       ClearSuccOnSuspend csos(this);
1567       ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1568       if (interrupted || HAS_PENDING_EXCEPTION) {
1569         // Intentionally empty
1570       } else if (node._notified == 0) {
1571         if (millis <= 0) {
1572           current->_ParkEvent->park();
1573         } else {
1574           ret = current->_ParkEvent->park(millis);
1575         }
1576       }
1577     }
1578 
1579     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1580     // from the WaitSet to the EntryList.
1581     // See if we need to remove Node from the WaitSet.
1582     // We use double-checked locking to avoid grabbing _WaitSetLock
1583     // if the thread is not on the wait queue.
1584     //
1585     // Note that we don't need a fence before the fetch of TState.
1586     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1587     // written by the is thread. (perhaps the fetch might even be satisfied
1588     // by a look-aside into the processor's own store buffer, although given
1589     // the length of the code path between the prior ST and this load that's
1590     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1591     // then we'll acquire the lock and then re-fetch a fresh TState value.
1592     // That is, we fail toward safety.
1593 
1594     if (node.TState == ObjectWaiter::TS_WAIT) {
1595       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1596       if (node.TState == ObjectWaiter::TS_WAIT) {
1597         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1598         assert(node._notified == 0, "invariant");
1599         node.TState = ObjectWaiter::TS_RUN;
1600       }
1601       Thread::SpinRelease(&_WaitSetLock);
1602     }
1603 
1604     // The thread is now either on off-list (TS_RUN),
1605     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1606     // The Node's TState variable is stable from the perspective of this thread.
1607     // No other threads will asynchronously modify TState.
1608     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1609     OrderAccess::loadload();
1610     if (_succ == current) _succ = nullptr;
1611     WasNotified = node._notified;
1612 
1613     // Reentry phase -- reacquire the monitor.
1614     // re-enter contended monitor after object.wait().
1615     // retain OBJECT_WAIT state until re-enter successfully completes
1616     // Thread state is thread_in_vm and oop access is again safe,
1617     // although the raw address of the object may have changed.
1618     // (Don't cache naked oops over safepoints, of course).
1619 
1620     // post monitor waited event. Note that this is past-tense, we are done waiting.
1621     if (JvmtiExport::should_post_monitor_waited()) {
1622       JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1623 
1624       if (node._notified != 0 && _succ == current) {
1625         // In this part of the monitor wait-notify-reenter protocol it
1626         // is possible (and normal) for another thread to do a fastpath
1627         // monitor enter-exit while this thread is still trying to get
1628         // to the reenter portion of the protocol.
1629         //
1630         // The ObjectMonitor was notified and the current thread is
1631         // the successor which also means that an unpark() has already
1632         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1633         // consume the unpark() that was done when the successor was
1634         // set because the same ParkEvent is shared between Java
1635         // monitors and JVM/TI RawMonitors (for now).
1636         //
1637         // We redo the unpark() to ensure forward progress, i.e., we
1638         // don't want all pending threads hanging (parked) with none
1639         // entering the unlocked monitor.
1640         node._event->unpark();
1641       }
1642     }
1643 
1644     if (event.should_commit()) {
1645       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1646     }
1647 
1648     OrderAccess::fence();
1649 
1650     assert(current->_Stalled != 0, "invariant");
1651     current->_Stalled = 0;
1652 
1653     assert(owner_raw() != current, "invariant");
1654     ObjectWaiter::TStates v = node.TState;
1655     if (v == ObjectWaiter::TS_RUN) {
1656       enter(current);
1657     } else {
1658       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1659       ReenterI(current, &node);
1660       node.wait_reenter_end(this);
1661     }
1662 
1663     // current has reacquired the lock.
1664     // Lifecycle - the node representing current must not appear on any queues.
1665     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1666     // want residual elements associated with this thread left on any lists.
1667     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1668     assert(owner_raw() == current, "invariant");
1669     assert(_succ != current, "invariant");
1670   } // OSThreadWaitState()
1671 
1672   current->set_current_waiting_monitor(nullptr);
1673 
1674   guarantee(_recursions == 0, "invariant");
1675   int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current);
1676   _recursions =   save          // restore the old recursion count
1677                 + relock_count; //  increased by the deferred relock count
1678   current->inc_held_monitor_count(relock_count); // Deopt never entered these counts.
1679   _waiters--;             // decrement the number of waiters
1680 
1681   // Verify a few postconditions
1682   assert(owner_raw() == current, "invariant");
1683   assert(_succ != current, "invariant");
1684   assert(object()->mark() == markWord::encode(this), "invariant");
1685 
1686   // check if the notification happened
1687   if (!WasNotified) {
1688     // no, it could be timeout or Thread.interrupt() or both
1689     // check for interrupt event, otherwise it is timeout
1690     if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1691       THROW(vmSymbols::java_lang_InterruptedException());
1692     }
1693   }
1694 
1695   // NOTE: Spurious wake up will be consider as timeout.
1696   // Monitor notify has precedence over thread interrupt.
1697 }
1698 
1699 
1700 // Consider:
1701 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1702 // then instead of transferring a thread from the WaitSet to the EntryList
1703 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1704 
1705 void ObjectMonitor::INotify(JavaThread* current) {
1706   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1707   ObjectWaiter* iterator = DequeueWaiter();
1708   if (iterator != nullptr) {
1709     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1710     guarantee(iterator->_notified == 0, "invariant");
1711     // Disposition - what might we do with iterator ?
1712     // a.  add it directly to the EntryList - either tail (policy == 1)
1713     //     or head (policy == 0).
1714     // b.  push it onto the front of the _cxq (policy == 2).
1715     // For now we use (b).
1716 
1717     iterator->TState = ObjectWaiter::TS_ENTER;
1718 
1719     iterator->_notified = 1;
1720     iterator->_notifier_tid = JFR_THREAD_ID(current);
1721 
1722     ObjectWaiter* list = _EntryList;
1723     if (list != nullptr) {
1724       assert(list->_prev == nullptr, "invariant");
1725       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1726       assert(list != iterator, "invariant");
1727     }
1728 
1729     // prepend to cxq
1730     if (list == nullptr) {
1731       iterator->_next = iterator->_prev = nullptr;
1732       _EntryList = iterator;
1733     } else {
1734       iterator->TState = ObjectWaiter::TS_CXQ;
1735       for (;;) {
1736         ObjectWaiter* front = _cxq;
1737         iterator->_next = front;
1738         if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1739           break;
1740         }
1741       }
1742     }
1743 
1744     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1745     // move the add-to-EntryList operation, above, outside the critical section
1746     // protected by _WaitSetLock.  In practice that's not useful.  With the
1747     // exception of  wait() timeouts and interrupts the monitor owner
1748     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1749     // on _WaitSetLock so it's not profitable to reduce the length of the
1750     // critical section.
1751 
1752     iterator->wait_reenter_begin(this);
1753   }
1754   Thread::SpinRelease(&_WaitSetLock);
1755 }
1756 
1757 // Consider: a not-uncommon synchronization bug is to use notify() when
1758 // notifyAll() is more appropriate, potentially resulting in stranded
1759 // threads; this is one example of a lost wakeup. A useful diagnostic
1760 // option is to force all notify() operations to behave as notifyAll().
1761 //
1762 // Note: We can also detect many such problems with a "minimum wait".
1763 // When the "minimum wait" is set to a small non-zero timeout value
1764 // and the program does not hang whereas it did absent "minimum wait",
1765 // that suggests a lost wakeup bug.
1766 
1767 void ObjectMonitor::notify(TRAPS) {
1768   JavaThread* current = THREAD;
1769   CHECK_OWNER();  // Throws IMSE if not owner.
1770   if (_WaitSet == nullptr) {
1771     return;
1772   }
1773   DTRACE_MONITOR_PROBE(notify, this, object(), current);
1774   INotify(current);
1775   OM_PERFDATA_OP(Notifications, inc(1));
1776 }
1777 
1778 
1779 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1780 // from the waitset to the EntryList. This could be done more efficiently with a
1781 // single bulk transfer but in practice it's not time-critical. Beware too,
1782 // that in prepend-mode we invert the order of the waiters. Let's say that the
1783 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1784 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1785 
1786 void ObjectMonitor::notifyAll(TRAPS) {
1787   JavaThread* current = THREAD;
1788   CHECK_OWNER();  // Throws IMSE if not owner.
1789   if (_WaitSet == nullptr) {
1790     return;
1791   }
1792 
1793   DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1794   int tally = 0;
1795   while (_WaitSet != nullptr) {
1796     tally++;
1797     INotify(current);
1798   }
1799 
1800   OM_PERFDATA_OP(Notifications, inc(tally));
1801 }
1802 
1803 // -----------------------------------------------------------------------------
1804 // Adaptive Spinning Support
1805 //
1806 // Adaptive spin-then-block - rational spinning
1807 //
1808 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1809 // algorithm.  On high order SMP systems it would be better to start with
1810 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1811 // a contending thread could enqueue itself on the cxq and then spin locally
1812 // on a thread-specific variable such as its ParkEvent._Event flag.
1813 // That's left as an exercise for the reader.  Note that global spinning is
1814 // not problematic on Niagara, as the L2 cache serves the interconnect and
1815 // has both low latency and massive bandwidth.
1816 //
1817 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1818 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1819 // (duration) or we can fix the count at approximately the duration of
1820 // a context switch and vary the frequency.   Of course we could also
1821 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1822 // For a description of 'Adaptive spin-then-block mutual exclusion in
1823 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1824 //
1825 // This implementation varies the duration "D", where D varies with
1826 // the success rate of recent spin attempts. (D is capped at approximately
1827 // length of a round-trip context switch).  The success rate for recent
1828 // spin attempts is a good predictor of the success rate of future spin
1829 // attempts.  The mechanism adapts automatically to varying critical
1830 // section length (lock modality), system load and degree of parallelism.
1831 // D is maintained per-monitor in _SpinDuration and is initialized
1832 // optimistically.  Spin frequency is fixed at 100%.
1833 //
1834 // Note that _SpinDuration is volatile, but we update it without locks
1835 // or atomics.  The code is designed so that _SpinDuration stays within
1836 // a reasonable range even in the presence of races.  The arithmetic
1837 // operations on _SpinDuration are closed over the domain of legal values,
1838 // so at worst a race will install and older but still legal value.
1839 // At the very worst this introduces some apparent non-determinism.
1840 // We might spin when we shouldn't or vice-versa, but since the spin
1841 // count are relatively short, even in the worst case, the effect is harmless.
1842 //
1843 // Care must be taken that a low "D" value does not become an
1844 // an absorbing state.  Transient spinning failures -- when spinning
1845 // is overall profitable -- should not cause the system to converge
1846 // on low "D" values.  We want spinning to be stable and predictable
1847 // and fairly responsive to change and at the same time we don't want
1848 // it to oscillate, become metastable, be "too" non-deterministic,
1849 // or converge on or enter undesirable stable absorbing states.
1850 //
1851 // We implement a feedback-based control system -- using past behavior
1852 // to predict future behavior.  We face two issues: (a) if the
1853 // input signal is random then the spin predictor won't provide optimal
1854 // results, and (b) if the signal frequency is too high then the control
1855 // system, which has some natural response lag, will "chase" the signal.
1856 // (b) can arise from multimodal lock hold times.  Transient preemption
1857 // can also result in apparent bimodal lock hold times.
1858 // Although sub-optimal, neither condition is particularly harmful, as
1859 // in the worst-case we'll spin when we shouldn't or vice-versa.
1860 // The maximum spin duration is rather short so the failure modes aren't bad.
1861 // To be conservative, I've tuned the gain in system to bias toward
1862 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1863 // "rings" or oscillates between spinning and not spinning.  This happens
1864 // when spinning is just on the cusp of profitability, however, so the
1865 // situation is not dire.  The state is benign -- there's no need to add
1866 // hysteresis control to damp the transition rate between spinning and
1867 // not spinning.
1868 
1869 // Spinning: Fixed frequency (100%), vary duration
1870 int ObjectMonitor::TrySpin(JavaThread* current) {
1871   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1872   int ctr = Knob_FixedSpin;
1873   if (ctr != 0) {
1874     while (--ctr >= 0) {
1875       if (TryLock(current) > 0) return 1;
1876       SpinPause();
1877     }
1878     return 0;
1879   }
1880 
1881   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1882     if (TryLock(current) > 0) {
1883       // Increase _SpinDuration ...
1884       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1885       // Raising _SpurDuration to the poverty line is key.
1886       int x = _SpinDuration;
1887       if (x < Knob_SpinLimit) {
1888         if (x < Knob_Poverty) x = Knob_Poverty;
1889         _SpinDuration = x + Knob_BonusB;
1890       }
1891       return 1;
1892     }
1893     SpinPause();
1894   }
1895 
1896   // Admission control - verify preconditions for spinning
1897   //
1898   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1899   // becoming an absorbing state.  Put another way, we spin briefly to
1900   // sample, just in case the system load, parallelism, contention, or lock
1901   // modality changed.
1902   //
1903   // Consider the following alternative:
1904   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1905   // spin attempt.  "Periodically" might mean after a tally of
1906   // the # of failed spin attempts (or iterations) reaches some threshold.
1907   // This takes us into the realm of 1-out-of-N spinning, where we
1908   // hold the duration constant but vary the frequency.
1909 
1910   ctr = _SpinDuration;
1911   if (ctr <= 0) return 0;
1912 
1913   if (NotRunnable(current, static_cast<JavaThread*>(owner_raw()))) {
1914     return 0;
1915   }
1916 
1917   // We're good to spin ... spin ingress.
1918   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1919   // when preparing to LD...CAS _owner, etc and the CAS is likely
1920   // to succeed.
1921   if (_succ == nullptr) {
1922     _succ = current;
1923   }
1924   Thread* prv = nullptr;
1925 
1926   // There are three ways to exit the following loop:
1927   // 1.  A successful spin where this thread has acquired the lock.
1928   // 2.  Spin failure with prejudice
1929   // 3.  Spin failure without prejudice
1930 
1931   while (--ctr >= 0) {
1932 
1933     // Periodic polling -- Check for pending GC
1934     // Threads may spin while they're unsafe.
1935     // We don't want spinning threads to delay the JVM from reaching
1936     // a stop-the-world safepoint or to steal cycles from GC.
1937     // If we detect a pending safepoint we abort in order that
1938     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1939     // this thread, if safe, doesn't steal cycles from GC.
1940     // This is in keeping with the "no loitering in runtime" rule.
1941     // We periodically check to see if there's a safepoint pending.
1942     if ((ctr & 0xFF) == 0) {
1943       // Can't call SafepointMechanism::should_process() since that
1944       // might update the poll values and we could be in a thread_blocked
1945       // state here which is not allowed so just check the poll.
1946       if (SafepointMechanism::local_poll_armed(current)) {
1947         goto Abort;           // abrupt spin egress
1948       }
1949       SpinPause();
1950     }
1951 
1952     // Probe _owner with TATAS
1953     // If this thread observes the monitor transition or flicker
1954     // from locked to unlocked to locked, then the odds that this
1955     // thread will acquire the lock in this spin attempt go down
1956     // considerably.  The same argument applies if the CAS fails
1957     // or if we observe _owner change from one non-null value to
1958     // another non-null value.   In such cases we might abort
1959     // the spin without prejudice or apply a "penalty" to the
1960     // spin count-down variable "ctr", reducing it by 100, say.
1961 
1962     JavaThread* ox = static_cast<JavaThread*>(owner_raw());
1963     if (ox == nullptr) {
1964       ox = static_cast<JavaThread*>(try_set_owner_from(nullptr, current));
1965       if (ox == nullptr) {
1966         // The CAS succeeded -- this thread acquired ownership
1967         // Take care of some bookkeeping to exit spin state.
1968         if (_succ == current) {
1969           _succ = nullptr;
1970         }
1971 
1972         // Increase _SpinDuration :
1973         // The spin was successful (profitable) so we tend toward
1974         // longer spin attempts in the future.
1975         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1976         // If we acquired the lock early in the spin cycle it
1977         // makes sense to increase _SpinDuration proportionally.
1978         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1979         int x = _SpinDuration;
1980         if (x < Knob_SpinLimit) {
1981           if (x < Knob_Poverty) x = Knob_Poverty;
1982           _SpinDuration = x + Knob_Bonus;
1983         }
1984         return 1;
1985       }
1986 
1987       // The CAS failed ... we can take any of the following actions:
1988       // * penalize: ctr -= CASPenalty
1989       // * exit spin with prejudice -- goto Abort;
1990       // * exit spin without prejudice.
1991       // * Since CAS is high-latency, retry again immediately.
1992       prv = ox;
1993       goto Abort;
1994     }
1995 
1996     // Did lock ownership change hands ?
1997     if (ox != prv && prv != nullptr) {
1998       goto Abort;
1999     }
2000     prv = ox;
2001 
2002     // Abort the spin if the owner is not executing.
2003     // The owner must be executing in order to drop the lock.
2004     // Spinning while the owner is OFFPROC is idiocy.
2005     // Consider: ctr -= RunnablePenalty ;
2006     if (NotRunnable(current, ox)) {
2007       goto Abort;
2008     }
2009     if (_succ == nullptr) {
2010       _succ = current;
2011     }
2012   }
2013 
2014   // Spin failed with prejudice -- reduce _SpinDuration.
2015   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2016   // AIMD is globally stable.
2017   {
2018     int x = _SpinDuration;
2019     if (x > 0) {
2020       // Consider an AIMD scheme like: x -= (x >> 3) + 100
2021       // This is globally sample and tends to damp the response.
2022       x -= Knob_Penalty;
2023       if (x < 0) x = 0;
2024       _SpinDuration = x;
2025     }
2026   }
2027 
2028  Abort:
2029   if (_succ == current) {
2030     _succ = nullptr;
2031     // Invariant: after setting succ=null a contending thread
2032     // must recheck-retry _owner before parking.  This usually happens
2033     // in the normal usage of TrySpin(), but it's safest
2034     // to make TrySpin() as foolproof as possible.
2035     OrderAccess::fence();
2036     if (TryLock(current) > 0) return 1;
2037   }
2038   return 0;
2039 }
2040 
2041 // NotRunnable() -- informed spinning
2042 //
2043 // Don't bother spinning if the owner is not eligible to drop the lock.
2044 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2045 // The thread must be runnable in order to drop the lock in timely fashion.
2046 // If the _owner is not runnable then spinning will not likely be
2047 // successful (profitable).
2048 //
2049 // Beware -- the thread referenced by _owner could have died
2050 // so a simply fetch from _owner->_thread_state might trap.
2051 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2052 // Because of the lifecycle issues, the _thread_state values
2053 // observed by NotRunnable() might be garbage.  NotRunnable must
2054 // tolerate this and consider the observed _thread_state value
2055 // as advisory.
2056 //
2057 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2058 // a thread pointer.
2059 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2060 // with the LSB of _owner.  Another option would be to probabilistically probe
2061 // the putative _owner->TypeTag value.
2062 //
2063 // Checking _thread_state isn't perfect.  Even if the thread is
2064 // in_java it might be blocked on a page-fault or have been preempted
2065 // and sitting on a ready/dispatch queue.
2066 //
2067 // The return value from NotRunnable() is *advisory* -- the
2068 // result is based on sampling and is not necessarily coherent.
2069 // The caller must tolerate false-negative and false-positive errors.
2070 // Spinning, in general, is probabilistic anyway.
2071 
2072 
2073 int ObjectMonitor::NotRunnable(JavaThread* current, JavaThread* ox) {
2074   // Check ox->TypeTag == 2BAD.
2075   if (ox == nullptr) return 0;
2076 
2077   // Avoid transitive spinning ...
2078   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2079   // Immediately after T1 acquires L it's possible that T2, also
2080   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2081   // This occurs transiently after T1 acquired L but before
2082   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2083   // its spin in this circumstance.
2084   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2085 
2086   if (BlockedOn == 1) return 1;
2087   if (BlockedOn != 0) {
2088     return BlockedOn != intptr_t(this) && owner_raw() == ox;
2089   }
2090 
2091   assert(sizeof(ox->_thread_state == sizeof(int)), "invariant");
2092   int jst = SafeFetch32((int *) &ox->_thread_state, -1);;
2093   // consider also: jst != _thread_in_Java -- but that's overspecific.
2094   return jst == _thread_blocked || jst == _thread_in_native;
2095 }
2096 
2097 
2098 // -----------------------------------------------------------------------------
2099 // WaitSet management ...
2100 
2101 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2102   _next     = nullptr;
2103   _prev     = nullptr;
2104   _notified = 0;
2105   _notifier_tid = 0;
2106   TState    = TS_RUN;
2107   _thread   = current;
2108   _event    = _thread->_ParkEvent;
2109   _active   = false;
2110   assert(_event != nullptr, "invariant");
2111 }
2112 
2113 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2114   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2115 }
2116 
2117 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2118   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2119 }
2120 
2121 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2122   assert(node != nullptr, "should not add null node");
2123   assert(node->_prev == nullptr, "node already in list");
2124   assert(node->_next == nullptr, "node already in list");
2125   // put node at end of queue (circular doubly linked list)
2126   if (_WaitSet == nullptr) {
2127     _WaitSet = node;
2128     node->_prev = node;
2129     node->_next = node;
2130   } else {
2131     ObjectWaiter* head = _WaitSet;
2132     ObjectWaiter* tail = head->_prev;
2133     assert(tail->_next == head, "invariant check");
2134     tail->_next = node;
2135     head->_prev = node;
2136     node->_next = head;
2137     node->_prev = tail;
2138   }
2139 }
2140 
2141 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2142   // dequeue the very first waiter
2143   ObjectWaiter* waiter = _WaitSet;
2144   if (waiter) {
2145     DequeueSpecificWaiter(waiter);
2146   }
2147   return waiter;
2148 }
2149 
2150 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2151   assert(node != nullptr, "should not dequeue nullptr node");
2152   assert(node->_prev != nullptr, "node already removed from list");
2153   assert(node->_next != nullptr, "node already removed from list");
2154   // when the waiter has woken up because of interrupt,
2155   // timeout or other spurious wake-up, dequeue the
2156   // waiter from waiting list
2157   ObjectWaiter* next = node->_next;
2158   if (next == node) {
2159     assert(node->_prev == node, "invariant check");
2160     _WaitSet = nullptr;
2161   } else {
2162     ObjectWaiter* prev = node->_prev;
2163     assert(prev->_next == node, "invariant check");
2164     assert(next->_prev == node, "invariant check");
2165     next->_prev = prev;
2166     prev->_next = next;
2167     if (_WaitSet == node) {
2168       _WaitSet = next;
2169     }
2170   }
2171   node->_next = nullptr;
2172   node->_prev = nullptr;
2173 }
2174 
2175 // -----------------------------------------------------------------------------
2176 // PerfData support
2177 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = nullptr;
2178 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = nullptr;
2179 PerfCounter * ObjectMonitor::_sync_Parks                       = nullptr;
2180 PerfCounter * ObjectMonitor::_sync_Notifications               = nullptr;
2181 PerfCounter * ObjectMonitor::_sync_Inflations                  = nullptr;
2182 PerfCounter * ObjectMonitor::_sync_Deflations                  = nullptr;
2183 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = nullptr;
2184 
2185 // One-shot global initialization for the sync subsystem.
2186 // We could also defer initialization and initialize on-demand
2187 // the first time we call ObjectSynchronizer::inflate().
2188 // Initialization would be protected - like so many things - by
2189 // the MonitorCache_lock.
2190 
2191 void ObjectMonitor::Initialize() {
2192   assert(!InitDone, "invariant");
2193 
2194   if (!os::is_MP()) {
2195     Knob_SpinLimit = 0;
2196     Knob_PreSpin   = 0;
2197     Knob_FixedSpin = -1;
2198   }
2199 
2200   if (UsePerfData) {
2201     EXCEPTION_MARK;
2202 #define NEWPERFCOUNTER(n)                                                \
2203   {                                                                      \
2204     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2205                                         CHECK);                          \
2206   }
2207 #define NEWPERFVARIABLE(n)                                                \
2208   {                                                                       \
2209     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2210                                          CHECK);                          \
2211   }
2212     NEWPERFCOUNTER(_sync_Inflations);
2213     NEWPERFCOUNTER(_sync_Deflations);
2214     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2215     NEWPERFCOUNTER(_sync_FutileWakeups);
2216     NEWPERFCOUNTER(_sync_Parks);
2217     NEWPERFCOUNTER(_sync_Notifications);
2218     NEWPERFVARIABLE(_sync_MonExtant);
2219 #undef NEWPERFCOUNTER
2220 #undef NEWPERFVARIABLE
2221   }
2222 
2223   _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2224 
2225   DEBUG_ONLY(InitDone = true;)
2226 }
2227 
2228 void ObjectMonitor::print_on(outputStream* st) const {
2229   // The minimal things to print for markWord printing, more can be added for debugging and logging.
2230   st->print("{contentions=0x%08x,waiters=0x%08x"
2231             ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2232             contentions(), waiters(), recursions(),
2233             p2i(owner()));
2234 }
2235 void ObjectMonitor::print() const { print_on(tty); }
2236 
2237 #ifdef ASSERT
2238 // Print the ObjectMonitor like a debugger would:
2239 //
2240 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2241 //   _header = 0x0000000000000001
2242 //   _object = 0x000000070ff45fd0
2243 //   _pad_buf0 = {
2244 //     [0] = '\0'
2245 //     ...
2246 //     [43] = '\0'
2247 //   }
2248 //   _owner = 0x0000000000000000
2249 //   _previous_owner_tid = 0
2250 //   _pad_buf1 = {
2251 //     [0] = '\0'
2252 //     ...
2253 //     [47] = '\0'
2254 //   }
2255 //   _next_om = 0x0000000000000000
2256 //   _recursions = 0
2257 //   _EntryList = 0x0000000000000000
2258 //   _cxq = 0x0000000000000000
2259 //   _succ = 0x0000000000000000
2260 //   _Responsible = 0x0000000000000000
2261 //   _Spinner = 0
2262 //   _SpinDuration = 5000
2263 //   _contentions = 0
2264 //   _WaitSet = 0x0000700009756248
2265 //   _waiters = 1
2266 //   _WaitSetLock = 0
2267 // }
2268 //
2269 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2270   st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2271   st->print_cr("  _header = " INTPTR_FORMAT, header().value());
2272   st->print_cr("  _object = " INTPTR_FORMAT, p2i(object_peek()));
2273   st->print_cr("  _pad_buf0 = {");
2274   st->print_cr("    [0] = '\\0'");
2275   st->print_cr("    ...");
2276   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2277   st->print_cr("  }");
2278   st->print_cr("  _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2279   st->print_cr("  _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid);
2280   st->print_cr("  _pad_buf1 = {");
2281   st->print_cr("    [0] = '\\0'");
2282   st->print_cr("    ...");
2283   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2284   st->print_cr("  }");
2285   st->print_cr("  _next_om = " INTPTR_FORMAT, p2i(next_om()));
2286   st->print_cr("  _recursions = " INTX_FORMAT, _recursions);
2287   st->print_cr("  _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2288   st->print_cr("  _cxq = " INTPTR_FORMAT, p2i(_cxq));
2289   st->print_cr("  _succ = " INTPTR_FORMAT, p2i(_succ));
2290   st->print_cr("  _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2291   st->print_cr("  _Spinner = %d", _Spinner);
2292   st->print_cr("  _SpinDuration = %d", _SpinDuration);
2293   st->print_cr("  _contentions = %d", contentions());
2294   st->print_cr("  _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2295   st->print_cr("  _waiters = %d", _waiters);
2296   st->print_cr("  _WaitSetLock = %d", _WaitSetLock);
2297   st->print_cr("}");
2298 }
2299 #endif