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/mutexLocker.hpp"
  46 #include "runtime/objectMonitor.hpp"
  47 #include "runtime/objectMonitor.inline.hpp"
  48 #include "runtime/orderAccess.hpp"
  49 #include "runtime/osThread.hpp"
  50 #include "runtime/perfData.hpp"
  51 #include "runtime/safefetch.hpp"
  52 #include "runtime/safepointMechanism.inline.hpp"
  53 #include "runtime/sharedRuntime.hpp"
  54 #include "runtime/thread.inline.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 = NULL;                                                      \
  72   int len = 0;                                                             \
  73   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  74   Symbol* klassname = obj->klass()->name();                                \
  75   if (klassname != NULL) {                                                 \
  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 = NULL;
 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 = self->as_Java_thread();
 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(NULL),
 265   _previous_owner_tid(0),
 266   _next_om(NULL),
 267   _recursions(0),
 268   _EntryList(NULL),
 269   _cxq(NULL),
 270   _succ(NULL),
 271   _Responsible(NULL),
 272   _Spinner(0),
 273   _SpinDuration(ObjectMonitor::Knob_SpinLimit),
 274   _contentions(0),
 275   _WaitSet(NULL),
 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   if (_object.is_null()) {
 287     return NULL;
 288   }
 289   return _object.resolve();
 290 }
 291 
 292 oop ObjectMonitor::object_peek() const {
 293   if (_object.is_null()) {
 294     return NULL;
 295   }
 296   return _object.peek();
 297 }
 298 
 299 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
 300   if (current->is_suspended()) {
 301     _om->_recursions = 0;
 302     _om->_succ = NULL;
 303     // Don't need a full fence after clearing successor here because of the call to exit().
 304     _om->exit(current, false /* not_suspended */);
 305     _om_exited = true;
 306 
 307     current->set_current_pending_monitor(_om);
 308   }
 309 }
 310 
 311 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
 312   if (current->is_suspended()) {
 313     if (_om->_succ == current) {
 314       _om->_succ = NULL;
 315       OrderAccess::fence(); // always do a full fence when successor is cleared
 316     }
 317   }
 318 }
 319 
 320 // -----------------------------------------------------------------------------
 321 // Enter support
 322 
 323 bool ObjectMonitor::enter_for(JavaThread* locking_thread) {
 324   // Used by ObjectSynchronizer::enter_for to enter for another thread.
 325   // The monitor is private to or already owned by locking_thread which must be suspended.
 326   // So this code may only contend with deflation.
 327   assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
 328 
 329   // Block out deflation as soon as possible.
 330   add_to_contentions(1);
 331 
 332   bool success = false;
 333   if (!is_being_async_deflated()) {
 334     void* prev_owner = try_set_owner_from(nullptr, locking_thread);
 335 
 336     if (prev_owner == nullptr) {
 337       assert(_recursions == 0, "invariant");
 338       success = true;
 339     } else if (prev_owner == locking_thread) {
 340       _recursions++;
 341       success = true;
 342     } else if (prev_owner == DEFLATER_MARKER) {
 343       // Racing with deflation.
 344       prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread);
 345       if (prev_owner == DEFLATER_MARKER) {
 346         // Cancelled deflation. Increment contentions as part of the deflation protocol.
 347         add_to_contentions(1);
 348         success = true;
 349       } else if (prev_owner == nullptr) {
 350         // At this point we cannot race with deflation as we have both incremented
 351         // contentions, seen contention > 0 and seen a DEFLATER_MARKER.
 352         // success will only be false if this races with something other than
 353         // deflation.
 354         prev_owner = try_set_owner_from(nullptr, locking_thread);
 355         success = prev_owner == nullptr;
 356       }
 357     } else if (LockingMode == LM_LEGACY && locking_thread->is_lock_owned((address)prev_owner)) {
 358       assert(_recursions == 0, "must be");
 359       _recursions = 1;
 360       set_owner_from_BasicLock(prev_owner, locking_thread);
 361       success = true;
 362     }
 363     assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT
 364            ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT,
 365            p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner));
 366   } else {
 367     // Async deflation is in progress and our contentions increment
 368     // above lost the race to async deflation. Undo the work and
 369     // force the caller to retry.
 370     const oop l_object = object();
 371     if (l_object != nullptr) {
 372       // Attempt to restore the header/dmw to the object's header so that
 373       // we only retry once if the deflater thread happens to be slow.
 374       install_displaced_markword_in_object(l_object);
 375     }
 376   }
 377 
 378   add_to_contentions(-1);
 379 
 380   assert(!success || owner_raw() == locking_thread, "must be");
 381 
 382   return success;
 383 }
 384 
 385 bool ObjectMonitor::enter(JavaThread* current) {
 386   assert(current == JavaThread::current(), "must be");
 387   // The following code is ordered to check the most common cases first
 388   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 389 
 390   void* cur = try_set_owner_from(NULL, current);
 391   if (cur == NULL) {
 392     assert(_recursions == 0, "invariant");
 393     return true;
 394   }
 395 
 396   if (cur == current) {
 397     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 398     _recursions++;
 399     return true;
 400   }
 401 
 402   if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
 403     assert(_recursions == 0, "internal state error");
 404     _recursions = 1;
 405     set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
 406     return true;
 407   }
 408 
 409   // We've encountered genuine contention.
 410   assert(current->_Stalled == 0, "invariant");
 411   current->_Stalled = intptr_t(this);
 412 
 413   // Try one round of spinning *before* enqueueing current
 414   // and before going through the awkward and expensive state
 415   // transitions.  The following spin is strictly optional ...
 416   // Note that if we acquire the monitor from an initial spin
 417   // we forgo posting JVMTI events and firing DTRACE probes.
 418   if (TrySpin(current) > 0) {
 419     assert(owner_raw() == current, "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
 420     assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
 421     assert(object()->mark() == markWord::encode(this),
 422            "object mark must match encoded this: mark=" INTPTR_FORMAT
 423            ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
 424            markWord::encode(this).value());
 425     current->_Stalled = 0;
 426     return true;
 427   }
 428 
 429   assert(owner_raw() != current, "invariant");
 430   assert(_succ != current, "invariant");
 431   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 432   assert(current->thread_state() != _thread_blocked, "invariant");
 433 
 434   // Keep track of contention for JVM/TI and M&M queries.
 435   add_to_contentions(1);
 436   if (is_being_async_deflated()) {
 437     // Async deflation is in progress and our contentions increment
 438     // above lost the race to async deflation. Undo the work and
 439     // force the caller to retry.
 440     const oop l_object = object();
 441     if (l_object != NULL) {
 442       // Attempt to restore the header/dmw to the object's header so that
 443       // we only retry once if the deflater thread happens to be slow.
 444       install_displaced_markword_in_object(l_object);
 445     }
 446     current->_Stalled = 0;
 447     add_to_contentions(-1);
 448     return false;
 449   }
 450 
 451   JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(current);)
 452   EventJavaMonitorEnter event;
 453   if (event.is_started()) {
 454     event.set_monitorClass(object()->klass());
 455     // Set an address that is 'unique enough', such that events close in
 456     // time and with the same address are likely (but not guaranteed) to
 457     // belong to the same object.
 458     event.set_address((uintptr_t)this);
 459   }
 460 
 461   { // Change java thread status to indicate blocked on monitor enter.
 462     JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
 463 
 464     assert(current->current_pending_monitor() == NULL, "invariant");
 465     current->set_current_pending_monitor(this);
 466 
 467     DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
 468     if (JvmtiExport::should_post_monitor_contended_enter()) {
 469       JvmtiExport::post_monitor_contended_enter(current, this);
 470 
 471       // The current thread does not yet own the monitor and does not
 472       // yet appear on any queues that would get it made the successor.
 473       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 474       // handler cannot accidentally consume an unpark() meant for the
 475       // ParkEvent associated with this ObjectMonitor.
 476     }
 477 
 478     OSThreadContendState osts(current->osthread());
 479 
 480     assert(current->thread_state() == _thread_in_vm, "invariant");
 481 
 482     for (;;) {
 483       ExitOnSuspend eos(this);
 484       {
 485         ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos);
 486         EnterI(current);
 487         current->set_current_pending_monitor(NULL);
 488         // We can go to a safepoint at the end of this block. If we
 489         // do a thread dump during that safepoint, then this thread will show
 490         // as having "-locked" the monitor, but the OS and java.lang.Thread
 491         // states will still report that the thread is blocked trying to
 492         // acquire it.
 493         // If there is a suspend request, ExitOnSuspend will exit the OM
 494         // and set the OM as pending.
 495       }
 496       if (!eos.exited()) {
 497         // ExitOnSuspend did not exit the OM
 498         assert(owner_raw() == current, "invariant");
 499         break;
 500       }
 501     }
 502 
 503     // We've just gotten past the enter-check-for-suspend dance and we now own
 504     // the monitor free and clear.
 505   }
 506 
 507   add_to_contentions(-1);
 508   assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
 509   current->_Stalled = 0;
 510 
 511   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 512   assert(_recursions == 0, "invariant");
 513   assert(owner_raw() == current, "invariant");
 514   assert(_succ != current, "invariant");
 515   assert(object()->mark() == markWord::encode(this), "invariant");
 516 
 517   // The thread -- now the owner -- is back in vm mode.
 518   // Report the glorious news via TI,DTrace and jvmstat.
 519   // The probe effect is non-trivial.  All the reportage occurs
 520   // while we hold the monitor, increasing the length of the critical
 521   // section.  Amdahl's parallel speedup law comes vividly into play.
 522   //
 523   // Another option might be to aggregate the events (thread local or
 524   // per-monitor aggregation) and defer reporting until a more opportune
 525   // time -- such as next time some thread encounters contention but has
 526   // yet to acquire the lock.  While spinning that thread could
 527   // spinning we could increment JVMStat counters, etc.
 528 
 529   DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
 530   if (JvmtiExport::should_post_monitor_contended_entered()) {
 531     JvmtiExport::post_monitor_contended_entered(current, this);
 532 
 533     // The current thread already owns the monitor and is not going to
 534     // call park() for the remainder of the monitor enter protocol. So
 535     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 536     // event handler consumed an unpark() issued by the thread that
 537     // just exited the monitor.
 538   }
 539   if (event.should_commit()) {
 540     event.set_previousOwner((uintptr_t)_previous_owner_tid);
 541     event.commit();
 542   }
 543   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 544   return true;
 545 }
 546 
 547 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 548 // Callers must compensate as needed.
 549 
 550 int ObjectMonitor::TryLock(JavaThread* current) {
 551   void* own = owner_raw();
 552   if (own != NULL) return 0;
 553   if (try_set_owner_from(NULL, current) == NULL) {
 554     assert(_recursions == 0, "invariant");
 555     return 1;
 556   }
 557   // The lock had been free momentarily, but we lost the race to the lock.
 558   // Interference -- the CAS failed.
 559   // We can either return -1 or retry.
 560   // Retry doesn't make as much sense because the lock was just acquired.
 561   return -1;
 562 }
 563 
 564 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
 565 // was deflated and false otherwise.
 566 //
 567 // The async deflation protocol sets owner to DEFLATER_MARKER and
 568 // makes contentions negative as signals to contending threads that
 569 // an async deflation is in progress. There are a number of checks
 570 // as part of the protocol to make sure that the calling thread has
 571 // not lost the race to a contending thread.
 572 //
 573 // The ObjectMonitor has been successfully async deflated when:
 574 //   (contentions < 0)
 575 // Contending threads that see that condition know to retry their operation.
 576 //
 577 bool ObjectMonitor::deflate_monitor() {
 578   if (is_busy()) {
 579     // Easy checks are first - the ObjectMonitor is busy so no deflation.
 580     return false;
 581   }
 582 
 583   if (ObjectSynchronizer::is_final_audit() && owner_is_DEFLATER_MARKER()) {
 584     // The final audit can see an already deflated ObjectMonitor on the
 585     // in-use list because MonitorList::unlink_deflated() might have
 586     // blocked for the final safepoint before unlinking all the deflated
 587     // monitors.
 588     assert(contentions() < 0, "must be negative: contentions=%d", contentions());
 589     // Already returned 'true' when it was originally deflated.
 590     return false;
 591   }
 592 
 593   const oop obj = object_peek();
 594 
 595   if (obj == NULL) {
 596     // If the object died, we can recycle the monitor without racing with
 597     // Java threads. The GC already broke the association with the object.
 598     set_owner_from(NULL, DEFLATER_MARKER);
 599     assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
 600     _contentions = -max_jint;
 601   } else {
 602     // Attempt async deflation protocol.
 603 
 604     // Set a NULL owner to DEFLATER_MARKER to force any contending thread
 605     // through the slow path. This is just the first part of the async
 606     // deflation dance.
 607     if (try_set_owner_from(NULL, DEFLATER_MARKER) != NULL) {
 608       // The owner field is no longer NULL so we lost the race since the
 609       // ObjectMonitor is now busy.
 610       return false;
 611     }
 612 
 613     if (contentions() > 0 || _waiters != 0) {
 614       // Another thread has raced to enter the ObjectMonitor after
 615       // is_busy() above or has already entered and waited on
 616       // it which makes it busy so no deflation. Restore owner to
 617       // NULL if it is still DEFLATER_MARKER.
 618       if (try_set_owner_from(DEFLATER_MARKER, NULL) != DEFLATER_MARKER) {
 619         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 620         add_to_contentions(-1);
 621       }
 622       return false;
 623     }
 624 
 625     // Make a zero contentions field negative to force any contending threads
 626     // to retry. This is the second part of the async deflation dance.
 627     if (Atomic::cmpxchg(&_contentions, (jint)0, -max_jint) != 0) {
 628       // Contentions was no longer 0 so we lost the race since the
 629       // ObjectMonitor is now busy. Restore owner to NULL if it is
 630       // still DEFLATER_MARKER:
 631       if (try_set_owner_from(DEFLATER_MARKER, NULL) != DEFLATER_MARKER) {
 632         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 633         add_to_contentions(-1);
 634       }
 635       return false;
 636     }
 637   }
 638 
 639   // Sanity checks for the races:
 640   guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
 641   guarantee(contentions() < 0, "must be negative: contentions=%d",
 642             contentions());
 643   guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
 644   guarantee(_cxq == NULL, "must be no contending threads: cxq="
 645             INTPTR_FORMAT, p2i(_cxq));
 646   guarantee(_EntryList == NULL,
 647             "must be no entering threads: EntryList=" INTPTR_FORMAT,
 648             p2i(_EntryList));
 649 
 650   if (obj != NULL) {
 651     if (log_is_enabled(Trace, monitorinflation)) {
 652       ResourceMark rm;
 653       log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
 654                                   ", mark=" INTPTR_FORMAT ", type='%s'",
 655                                   p2i(obj), obj->mark().value(),
 656                                   obj->klass()->external_name());
 657     }
 658 
 659     // Install the old mark word if nobody else has already done it.
 660     install_displaced_markword_in_object(obj);
 661   }
 662 
 663   // We leave owner == DEFLATER_MARKER and contentions < 0
 664   // to force any racing threads to retry.
 665   return true;  // Success, ObjectMonitor has been deflated.
 666 }
 667 
 668 // We might access the dead object headers for parsable heap walk, make sure
 669 // headers are in correct shape, e.g. monitors deflated.
 670 void ObjectMonitor::maybe_deflate_dead(oop* p) {
 671   oop obj = *p;
 672   assert(obj != NULL, "must not yet been cleared");
 673   markWord mark = obj->mark();
 674   if (mark.has_monitor()) {
 675     ObjectMonitor* monitor = mark.monitor();
 676     if (p == monitor->_object.ptr_raw()) {
 677       assert(monitor->object_peek() == obj, "lock object must match");
 678       markWord dmw = monitor->header();
 679       obj->set_mark(dmw);
 680     }
 681   }
 682 }
 683 
 684 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
 685 // into the header of the object associated with the monitor. This
 686 // idempotent method is called by a thread that is deflating a
 687 // monitor and by other threads that have detected a race with the
 688 // deflation process.
 689 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
 690   // This function must only be called when (owner == DEFLATER_MARKER
 691   // && contentions <= 0), but we can't guarantee that here because
 692   // those values could change when the ObjectMonitor gets moved from
 693   // the global free list to a per-thread free list.
 694 
 695   guarantee(obj != NULL, "must be non-NULL");
 696 
 697   // Separate loads in is_being_async_deflated(), which is almost always
 698   // called before this function, from the load of dmw/header below.
 699 
 700   // _contentions and dmw/header may get written by different threads.
 701   // Make sure to observe them in the same order when having several observers.
 702   OrderAccess::loadload_for_IRIW();
 703 
 704   const oop l_object = object_peek();
 705   if (l_object == NULL) {
 706     // ObjectMonitor's object ref has already been cleared by async
 707     // deflation or GC so we're done here.
 708     return;
 709   }
 710   assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
 711          INTPTR_FORMAT, p2i(l_object), p2i(obj));
 712 
 713   markWord dmw = header();
 714   // The dmw has to be neutral (not NULL, not locked and not marked).
 715   assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
 716 
 717   // Install displaced mark word if the object's header still points
 718   // to this ObjectMonitor. More than one racing caller to this function
 719   // can rarely reach this point, but only one can win.
 720   markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
 721   if (res != markWord::encode(this)) {
 722     // This should be rare so log at the Info level when it happens.
 723     log_info(monitorinflation)("install_displaced_markword_in_object: "
 724                                "failed cas_set_mark: new_mark=" INTPTR_FORMAT
 725                                ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
 726                                dmw.value(), markWord::encode(this).value(),
 727                                res.value());
 728   }
 729 
 730   // Note: It does not matter which thread restored the header/dmw
 731   // into the object's header. The thread deflating the monitor just
 732   // wanted the object's header restored and it is. The threads that
 733   // detected a race with the deflation process also wanted the
 734   // object's header restored before they retry their operation and
 735   // because it is restored they will only retry once.
 736 }
 737 
 738 // Convert the fields used by is_busy() to a string that can be
 739 // used for diagnostic output.
 740 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
 741   ss->print("is_busy: waiters=%d, ", _waiters);
 742   if (contentions() > 0) {
 743     ss->print("contentions=%d, ", contentions());
 744   } else {
 745     ss->print("contentions=0");
 746   }
 747   if (!owner_is_DEFLATER_MARKER()) {
 748     ss->print("owner=" INTPTR_FORMAT, p2i(owner_raw()));
 749   } else {
 750     // We report NULL instead of DEFLATER_MARKER here because is_busy()
 751     // ignores DEFLATER_MARKER values.
 752     ss->print("owner=" INTPTR_FORMAT, NULL_WORD);
 753   }
 754   ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq),
 755             p2i(_EntryList));
 756   return ss->base();
 757 }
 758 
 759 #define MAX_RECHECK_INTERVAL 1000
 760 
 761 void ObjectMonitor::EnterI(JavaThread* current) {
 762   assert(current->thread_state() == _thread_blocked, "invariant");
 763 
 764   // Try the lock - TATAS
 765   if (TryLock (current) > 0) {
 766     assert(_succ != current, "invariant");
 767     assert(owner_raw() == current, "invariant");
 768     assert(_Responsible != current, "invariant");
 769     return;
 770   }
 771 
 772   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 773     // Cancelled the in-progress async deflation by changing owner from
 774     // DEFLATER_MARKER to current. As part of the contended enter protocol,
 775     // contentions was incremented to a positive value before EnterI()
 776     // was called and that prevents the deflater thread from winning the
 777     // last part of the 2-part async deflation protocol. After EnterI()
 778     // returns to enter(), contentions is decremented because the caller
 779     // now owns the monitor. We bump contentions an extra time here to
 780     // prevent the deflater thread from winning the last part of the
 781     // 2-part async deflation protocol after the regular decrement
 782     // occurs in enter(). The deflater thread will decrement contentions
 783     // after it recognizes that the async deflation was cancelled.
 784     add_to_contentions(1);
 785     assert(_succ != current, "invariant");
 786     assert(_Responsible != current, "invariant");
 787     return;
 788   }
 789 
 790   assert(InitDone, "Unexpectedly not initialized");
 791 
 792   // We try one round of spinning *before* enqueueing current.
 793   //
 794   // If the _owner is ready but OFFPROC we could use a YieldTo()
 795   // operation to donate the remainder of this thread's quantum
 796   // to the owner.  This has subtle but beneficial affinity
 797   // effects.
 798 
 799   if (TrySpin(current) > 0) {
 800     assert(owner_raw() == current, "invariant");
 801     assert(_succ != current, "invariant");
 802     assert(_Responsible != current, "invariant");
 803     return;
 804   }
 805 
 806   // The Spin failed -- Enqueue and park the thread ...
 807   assert(_succ != current, "invariant");
 808   assert(owner_raw() != current, "invariant");
 809   assert(_Responsible != current, "invariant");
 810 
 811   // Enqueue "current" on ObjectMonitor's _cxq.
 812   //
 813   // Node acts as a proxy for current.
 814   // As an aside, if were to ever rewrite the synchronization code mostly
 815   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 816   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 817   // as well as eliminate a subset of ABA issues.
 818   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 819 
 820   ObjectWaiter node(current);
 821   current->_ParkEvent->reset();
 822   node._prev   = (ObjectWaiter*) 0xBAD;
 823   node.TState  = ObjectWaiter::TS_CXQ;
 824 
 825   // Push "current" onto the front of the _cxq.
 826   // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
 827   // Note that spinning tends to reduce the rate at which threads
 828   // enqueue and dequeue on EntryList|cxq.
 829   ObjectWaiter* nxt;
 830   for (;;) {
 831     node._next = nxt = _cxq;
 832     if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
 833 
 834     // Interference - the CAS failed because _cxq changed.  Just retry.
 835     // As an optional optimization we retry the lock.
 836     if (TryLock (current) > 0) {
 837       assert(_succ != current, "invariant");
 838       assert(owner_raw() == current, "invariant");
 839       assert(_Responsible != current, "invariant");
 840       return;
 841     }
 842   }
 843 
 844   // Check for cxq|EntryList edge transition to non-null.  This indicates
 845   // the onset of contention.  While contention persists exiting threads
 846   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 847   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 848   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 849   // arrange for one of the contending thread to use a timed park() operations
 850   // to detect and recover from the race.  (Stranding is form of progress failure
 851   // where the monitor is unlocked but all the contending threads remain parked).
 852   // That is, at least one of the contended threads will periodically poll _owner.
 853   // One of the contending threads will become the designated "Responsible" thread.
 854   // The Responsible thread uses a timed park instead of a normal indefinite park
 855   // operation -- it periodically wakes and checks for and recovers from potential
 856   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 857   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 858   // be responsible for a monitor.
 859   //
 860   // Currently, one of the contended threads takes on the added role of "Responsible".
 861   // A viable alternative would be to use a dedicated "stranding checker" thread
 862   // that periodically iterated over all the threads (or active monitors) and unparked
 863   // successors where there was risk of stranding.  This would help eliminate the
 864   // timer scalability issues we see on some platforms as we'd only have one thread
 865   // -- the checker -- parked on a timer.
 866 
 867   if (nxt == NULL && _EntryList == NULL) {
 868     // Try to assume the role of responsible thread for the monitor.
 869     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=current }
 870     Atomic::replace_if_null(&_Responsible, current);
 871   }
 872 
 873   // The lock might have been released while this thread was occupied queueing
 874   // itself onto _cxq.  To close the race and avoid "stranding" and
 875   // progress-liveness failure we must resample-retry _owner before parking.
 876   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 877   // In this case the ST-MEMBAR is accomplished with CAS().
 878   //
 879   // TODO: Defer all thread state transitions until park-time.
 880   // Since state transitions are heavy and inefficient we'd like
 881   // to defer the state transitions until absolutely necessary,
 882   // and in doing so avoid some transitions ...
 883 
 884   int nWakeups = 0;
 885   int recheckInterval = 1;
 886 
 887   for (;;) {
 888 
 889     if (TryLock(current) > 0) break;
 890     assert(owner_raw() != current, "invariant");
 891 
 892     // park self
 893     if (_Responsible == current) {
 894       current->_ParkEvent->park((jlong) recheckInterval);
 895       // Increase the recheckInterval, but clamp the value.
 896       recheckInterval *= 8;
 897       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 898         recheckInterval = MAX_RECHECK_INTERVAL;
 899       }
 900     } else {
 901       current->_ParkEvent->park();
 902     }
 903 
 904     if (TryLock(current) > 0) break;
 905 
 906     if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 907       // Cancelled the in-progress async deflation by changing owner from
 908       // DEFLATER_MARKER to current. As part of the contended enter protocol,
 909       // contentions was incremented to a positive value before EnterI()
 910       // was called and that prevents the deflater thread from winning the
 911       // last part of the 2-part async deflation protocol. After EnterI()
 912       // returns to enter(), contentions is decremented because the caller
 913       // now owns the monitor. We bump contentions an extra time here to
 914       // prevent the deflater thread from winning the last part of the
 915       // 2-part async deflation protocol after the regular decrement
 916       // occurs in enter(). The deflater thread will decrement contentions
 917       // after it recognizes that the async deflation was cancelled.
 918       add_to_contentions(1);
 919       break;
 920     }
 921 
 922     // The lock is still contested.
 923     // Keep a tally of the # of futile wakeups.
 924     // Note that the counter is not protected by a lock or updated by atomics.
 925     // That is by design - we trade "lossy" counters which are exposed to
 926     // races during updates for a lower probe effect.
 927 
 928     // This PerfData object can be used in parallel with a safepoint.
 929     // See the work around in PerfDataManager::destroy().
 930     OM_PERFDATA_OP(FutileWakeups, inc());
 931     ++nWakeups;
 932 
 933     // Assuming this is not a spurious wakeup we'll normally find _succ == current.
 934     // We can defer clearing _succ until after the spin completes
 935     // TrySpin() must tolerate being called with _succ == current.
 936     // Try yet another round of adaptive spinning.
 937     if (TrySpin(current) > 0) break;
 938 
 939     // We can find that we were unpark()ed and redesignated _succ while
 940     // we were spinning.  That's harmless.  If we iterate and call park(),
 941     // park() will consume the event and return immediately and we'll
 942     // just spin again.  This pattern can repeat, leaving _succ to simply
 943     // spin on a CPU.
 944 
 945     if (_succ == current) _succ = NULL;
 946 
 947     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 948     OrderAccess::fence();
 949   }
 950 
 951   // Egress :
 952   // current has acquired the lock -- Unlink current from the cxq or EntryList.
 953   // Normally we'll find current on the EntryList .
 954   // From the perspective of the lock owner (this thread), the
 955   // EntryList is stable and cxq is prepend-only.
 956   // The head of cxq is volatile but the interior is stable.
 957   // In addition, current.TState is stable.
 958 
 959   assert(owner_raw() == current, "invariant");
 960 
 961   UnlinkAfterAcquire(current, &node);
 962   if (_succ == current) _succ = NULL;
 963 
 964   assert(_succ != current, "invariant");
 965   if (_Responsible == current) {
 966     _Responsible = NULL;
 967     OrderAccess::fence(); // Dekker pivot-point
 968 
 969     // We may leave threads on cxq|EntryList without a designated
 970     // "Responsible" thread.  This is benign.  When this thread subsequently
 971     // exits the monitor it can "see" such preexisting "old" threads --
 972     // threads that arrived on the cxq|EntryList before the fence, above --
 973     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 974     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 975     // non-null and elect a new "Responsible" timer thread.
 976     //
 977     // This thread executes:
 978     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 979     //    LD cxq|EntryList               (in subsequent exit)
 980     //
 981     // Entering threads in the slow/contended path execute:
 982     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 983     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 984     //
 985     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 986     // exit operation from floating above the ST Responsible=null.
 987   }
 988 
 989   // We've acquired ownership with CAS().
 990   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 991   // But since the CAS() this thread may have also stored into _succ,
 992   // EntryList, cxq or Responsible.  These meta-data updates must be
 993   // visible __before this thread subsequently drops the lock.
 994   // Consider what could occur if we didn't enforce this constraint --
 995   // STs to monitor meta-data and user-data could reorder with (become
 996   // visible after) the ST in exit that drops ownership of the lock.
 997   // Some other thread could then acquire the lock, but observe inconsistent
 998   // or old monitor meta-data and heap data.  That violates the JMM.
 999   // To that end, the 1-0 exit() operation must have at least STST|LDST
1000   // "release" barrier semantics.  Specifically, there must be at least a
1001   // STST|LDST barrier in exit() before the ST of null into _owner that drops
1002   // the lock.   The barrier ensures that changes to monitor meta-data and data
1003   // protected by the lock will be visible before we release the lock, and
1004   // therefore before some other thread (CPU) has a chance to acquire the lock.
1005   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
1006   //
1007   // Critically, any prior STs to _succ or EntryList must be visible before
1008   // the ST of null into _owner in the *subsequent* (following) corresponding
1009   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
1010   // execute a serializing instruction.
1011 
1012   return;
1013 }
1014 
1015 // ReenterI() is a specialized inline form of the latter half of the
1016 // contended slow-path from EnterI().  We use ReenterI() only for
1017 // monitor reentry in wait().
1018 //
1019 // In the future we should reconcile EnterI() and ReenterI().
1020 
1021 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
1022   assert(current != NULL, "invariant");
1023   assert(currentNode != NULL, "invariant");
1024   assert(currentNode->_thread == current, "invariant");
1025   assert(_waiters > 0, "invariant");
1026   assert(object()->mark() == markWord::encode(this), "invariant");
1027 
1028   assert(current->thread_state() != _thread_blocked, "invariant");
1029 
1030   int nWakeups = 0;
1031   for (;;) {
1032     ObjectWaiter::TStates v = currentNode->TState;
1033     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1034     assert(owner_raw() != current, "invariant");
1035 
1036     if (TryLock(current) > 0) break;
1037     if (TrySpin(current) > 0) break;
1038 
1039     {
1040       OSThreadContendState osts(current->osthread());
1041 
1042       assert(current->thread_state() == _thread_in_vm, "invariant");
1043 
1044       {
1045         ClearSuccOnSuspend csos(this);
1046         ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos);
1047         current->_ParkEvent->park();
1048       }
1049     }
1050 
1051     // Try again, but just so we distinguish between futile wakeups and
1052     // successful wakeups.  The following test isn't algorithmically
1053     // necessary, but it helps us maintain sensible statistics.
1054     if (TryLock(current) > 0) break;
1055 
1056     // The lock is still contested.
1057     // Keep a tally of the # of futile wakeups.
1058     // Note that the counter is not protected by a lock or updated by atomics.
1059     // That is by design - we trade "lossy" counters which are exposed to
1060     // races during updates for a lower probe effect.
1061     ++nWakeups;
1062 
1063     // Assuming this is not a spurious wakeup we'll normally
1064     // find that _succ == current.
1065     if (_succ == current) _succ = NULL;
1066 
1067     // Invariant: after clearing _succ a contending thread
1068     // *must* retry  _owner before parking.
1069     OrderAccess::fence();
1070 
1071     // This PerfData object can be used in parallel with a safepoint.
1072     // See the work around in PerfDataManager::destroy().
1073     OM_PERFDATA_OP(FutileWakeups, inc());
1074   }
1075 
1076   // current has acquired the lock -- Unlink current from the cxq or EntryList .
1077   // Normally we'll find current on the EntryList.
1078   // Unlinking from the EntryList is constant-time and atomic-free.
1079   // From the perspective of the lock owner (this thread), the
1080   // EntryList is stable and cxq is prepend-only.
1081   // The head of cxq is volatile but the interior is stable.
1082   // In addition, current.TState is stable.
1083 
1084   assert(owner_raw() == current, "invariant");
1085   assert(object()->mark() == markWord::encode(this), "invariant");
1086   UnlinkAfterAcquire(current, currentNode);
1087   if (_succ == current) _succ = NULL;
1088   assert(_succ != current, "invariant");
1089   currentNode->TState = ObjectWaiter::TS_RUN;
1090   OrderAccess::fence();      // see comments at the end of EnterI()
1091 }
1092 
1093 // By convention we unlink a contending thread from EntryList|cxq immediately
1094 // after the thread acquires the lock in ::enter().  Equally, we could defer
1095 // unlinking the thread until ::exit()-time.
1096 
1097 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode) {
1098   assert(owner_raw() == current, "invariant");
1099   assert(currentNode->_thread == current, "invariant");
1100 
1101   if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1102     // Normal case: remove current from the DLL EntryList .
1103     // This is a constant-time operation.
1104     ObjectWaiter* nxt = currentNode->_next;
1105     ObjectWaiter* prv = currentNode->_prev;
1106     if (nxt != NULL) nxt->_prev = prv;
1107     if (prv != NULL) prv->_next = nxt;
1108     if (currentNode == _EntryList) _EntryList = nxt;
1109     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1110     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1111   } else {
1112     assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1113     // Inopportune interleaving -- current is still on the cxq.
1114     // This usually means the enqueue of self raced an exiting thread.
1115     // Normally we'll find current near the front of the cxq, so
1116     // dequeueing is typically fast.  If needbe we can accelerate
1117     // this with some MCS/CHL-like bidirectional list hints and advisory
1118     // back-links so dequeueing from the interior will normally operate
1119     // in constant-time.
1120     // Dequeue current from either the head (with CAS) or from the interior
1121     // with a linear-time scan and normal non-atomic memory operations.
1122     // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1123     // and then unlink current from EntryList.  We have to drain eventually,
1124     // so it might as well be now.
1125 
1126     ObjectWaiter* v = _cxq;
1127     assert(v != NULL, "invariant");
1128     if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1129       // The CAS above can fail from interference IFF a "RAT" arrived.
1130       // In that case current must be in the interior and can no longer be
1131       // at the head of cxq.
1132       if (v == currentNode) {
1133         assert(_cxq != v, "invariant");
1134         v = _cxq;          // CAS above failed - start scan at head of list
1135       }
1136       ObjectWaiter* p;
1137       ObjectWaiter* q = NULL;
1138       for (p = v; p != NULL && p != currentNode; p = p->_next) {
1139         q = p;
1140         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1141       }
1142       assert(v != currentNode, "invariant");
1143       assert(p == currentNode, "Node not found on cxq");
1144       assert(p != _cxq, "invariant");
1145       assert(q != NULL, "invariant");
1146       assert(q->_next == p, "invariant");
1147       q->_next = p->_next;
1148     }
1149   }
1150 
1151 #ifdef ASSERT
1152   // Diagnostic hygiene ...
1153   currentNode->_prev  = (ObjectWaiter*) 0xBAD;
1154   currentNode->_next  = (ObjectWaiter*) 0xBAD;
1155   currentNode->TState = ObjectWaiter::TS_RUN;
1156 #endif
1157 }
1158 
1159 // -----------------------------------------------------------------------------
1160 // Exit support
1161 //
1162 // exit()
1163 // ~~~~~~
1164 // Note that the collector can't reclaim the objectMonitor or deflate
1165 // the object out from underneath the thread calling ::exit() as the
1166 // thread calling ::exit() never transitions to a stable state.
1167 // This inhibits GC, which in turn inhibits asynchronous (and
1168 // inopportune) reclamation of "this".
1169 //
1170 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1171 // There's one exception to the claim above, however.  EnterI() can call
1172 // exit() to drop a lock if the acquirer has been externally suspended.
1173 // In that case exit() is called with _thread_state == _thread_blocked,
1174 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1175 //
1176 // 1-0 exit
1177 // ~~~~~~~~
1178 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1179 // the fast-path operators have been optimized so the common ::exit()
1180 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1181 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
1182 // greatly improves latency -- MEMBAR and CAS having considerable local
1183 // latency on modern processors -- but at the cost of "stranding".  Absent the
1184 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1185 // ::enter() path, resulting in the entering thread being stranding
1186 // and a progress-liveness failure.   Stranding is extremely rare.
1187 // We use timers (timed park operations) & periodic polling to detect
1188 // and recover from stranding.  Potentially stranded threads periodically
1189 // wake up and poll the lock.  See the usage of the _Responsible variable.
1190 //
1191 // The CAS() in enter provides for safety and exclusion, while the CAS or
1192 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
1193 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1194 // We detect and recover from stranding with timers.
1195 //
1196 // If a thread transiently strands it'll park until (a) another
1197 // thread acquires the lock and then drops the lock, at which time the
1198 // exiting thread will notice and unpark the stranded thread, or, (b)
1199 // the timer expires.  If the lock is high traffic then the stranding latency
1200 // will be low due to (a).  If the lock is low traffic then the odds of
1201 // stranding are lower, although the worst-case stranding latency
1202 // is longer.  Critically, we don't want to put excessive load in the
1203 // platform's timer subsystem.  We want to minimize both the timer injection
1204 // rate (timers created/sec) as well as the number of timers active at
1205 // any one time.  (more precisely, we want to minimize timer-seconds, which is
1206 // the integral of the # of active timers at any instant over time).
1207 // Both impinge on OS scalability.  Given that, at most one thread parked on
1208 // a monitor will use a timer.
1209 //
1210 // There is also the risk of a futile wake-up. If we drop the lock
1211 // another thread can reacquire the lock immediately, and we can
1212 // then wake a thread unnecessarily. This is benign, and we've
1213 // structured the code so the windows are short and the frequency
1214 // of such futile wakups is low.
1215 
1216 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1217   void* cur = owner_raw();
1218   if (current != cur) {
1219     if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1220       assert(_recursions == 0, "invariant");
1221       set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1222       _recursions = 0;
1223     } else {
1224       // Apparent unbalanced locking ...
1225       // Naively we'd like to throw IllegalMonitorStateException.
1226       // As a practical matter we can neither allocate nor throw an
1227       // exception as ::exit() can be called from leaf routines.
1228       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1229       // Upon deeper reflection, however, in a properly run JVM the only
1230       // way we should encounter this situation is in the presence of
1231       // unbalanced JNI locking. TODO: CheckJNICalls.
1232       // See also: CR4414101
1233 #ifdef ASSERT
1234       LogStreamHandle(Error, monitorinflation) lsh;
1235       lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1236                     " is exiting an ObjectMonitor it does not own.", p2i(current));
1237       lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1238       print_debug_style_on(&lsh);
1239       assert(false, "Non-balanced monitor enter/exit!");
1240 #endif
1241       return;
1242     }
1243   }
1244 
1245   if (_recursions != 0) {
1246     _recursions--;        // this is simple recursive enter
1247     return;
1248   }
1249 
1250   // Invariant: after setting Responsible=null an thread must execute
1251   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1252   _Responsible = NULL;
1253 
1254 #if INCLUDE_JFR
1255   // get the owner's thread id for the MonitorEnter event
1256   // if it is enabled and the thread isn't suspended
1257   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1258     _previous_owner_tid = JFR_THREAD_ID(current);
1259   }
1260 #endif
1261 
1262   for (;;) {
1263     assert(current == owner_raw(), "invariant");
1264 
1265     // Drop the lock.
1266     // release semantics: prior loads and stores from within the critical section
1267     // must not float (reorder) past the following store that drops the lock.
1268     // Uses a storeload to separate release_store(owner) from the
1269     // successor check. The try_set_owner() below uses cmpxchg() so
1270     // we get the fence down there.
1271     release_clear_owner(current);
1272     OrderAccess::storeload();
1273 
1274     if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1275       return;
1276     }
1277     // Other threads are blocked trying to acquire the lock.
1278 
1279     // Normally the exiting thread is responsible for ensuring succession,
1280     // but if other successors are ready or other entering threads are spinning
1281     // then this thread can simply store NULL into _owner and exit without
1282     // waking a successor.  The existence of spinners or ready successors
1283     // guarantees proper succession (liveness).  Responsibility passes to the
1284     // ready or running successors.  The exiting thread delegates the duty.
1285     // More precisely, if a successor already exists this thread is absolved
1286     // of the responsibility of waking (unparking) one.
1287     //
1288     // The _succ variable is critical to reducing futile wakeup frequency.
1289     // _succ identifies the "heir presumptive" thread that has been made
1290     // ready (unparked) but that has not yet run.  We need only one such
1291     // successor thread to guarantee progress.
1292     // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1293     // section 3.3 "Futile Wakeup Throttling" for details.
1294     //
1295     // Note that spinners in Enter() also set _succ non-null.
1296     // In the current implementation spinners opportunistically set
1297     // _succ so that exiting threads might avoid waking a successor.
1298     // Another less appealing alternative would be for the exiting thread
1299     // to drop the lock and then spin briefly to see if a spinner managed
1300     // to acquire the lock.  If so, the exiting thread could exit
1301     // immediately without waking a successor, otherwise the exiting
1302     // thread would need to dequeue and wake a successor.
1303     // (Note that we'd need to make the post-drop spin short, but no
1304     // shorter than the worst-case round-trip cache-line migration time.
1305     // The dropped lock needs to become visible to the spinner, and then
1306     // the acquisition of the lock by the spinner must become visible to
1307     // the exiting thread).
1308 
1309     // It appears that an heir-presumptive (successor) must be made ready.
1310     // Only the current lock owner can manipulate the EntryList or
1311     // drain _cxq, so we need to reacquire the lock.  If we fail
1312     // to reacquire the lock the responsibility for ensuring succession
1313     // falls to the new owner.
1314     //
1315     if (try_set_owner_from(NULL, current) != NULL) {
1316       return;
1317     }
1318 
1319     guarantee(owner_raw() == current, "invariant");
1320 
1321     ObjectWaiter* w = NULL;
1322 
1323     w = _EntryList;
1324     if (w != NULL) {
1325       // I'd like to write: guarantee (w->_thread != current).
1326       // But in practice an exiting thread may find itself on the EntryList.
1327       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1328       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1329       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1330       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1331       // release the lock "O".  T2 resumes immediately after the ST of null into
1332       // _owner, above.  T2 notices that the EntryList is populated, so it
1333       // reacquires the lock and then finds itself on the EntryList.
1334       // Given all that, we have to tolerate the circumstance where "w" is
1335       // associated with current.
1336       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1337       ExitEpilog(current, w);
1338       return;
1339     }
1340 
1341     // If we find that both _cxq and EntryList are null then just
1342     // re-run the exit protocol from the top.
1343     w = _cxq;
1344     if (w == NULL) continue;
1345 
1346     // Drain _cxq into EntryList - bulk transfer.
1347     // First, detach _cxq.
1348     // The following loop is tantamount to: w = swap(&cxq, NULL)
1349     for (;;) {
1350       assert(w != NULL, "Invariant");
1351       ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL);
1352       if (u == w) break;
1353       w = u;
1354     }
1355 
1356     assert(w != NULL, "invariant");
1357     assert(_EntryList == NULL, "invariant");
1358 
1359     // Convert the LIFO SLL anchored by _cxq into a DLL.
1360     // The list reorganization step operates in O(LENGTH(w)) time.
1361     // It's critical that this step operate quickly as
1362     // "current" still holds the outer-lock, restricting parallelism
1363     // and effectively lengthening the critical section.
1364     // Invariant: s chases t chases u.
1365     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1366     // we have faster access to the tail.
1367 
1368     _EntryList = w;
1369     ObjectWaiter* q = NULL;
1370     ObjectWaiter* p;
1371     for (p = w; p != NULL; p = p->_next) {
1372       guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1373       p->TState = ObjectWaiter::TS_ENTER;
1374       p->_prev = q;
1375       q = p;
1376     }
1377 
1378     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1379     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1380 
1381     // See if we can abdicate to a spinner instead of waking a thread.
1382     // A primary goal of the implementation is to reduce the
1383     // context-switch rate.
1384     if (_succ != NULL) continue;
1385 
1386     w = _EntryList;
1387     if (w != NULL) {
1388       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1389       ExitEpilog(current, w);
1390       return;
1391     }
1392   }
1393 }
1394 
1395 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1396   assert(owner_raw() == current, "invariant");
1397 
1398   // Exit protocol:
1399   // 1. ST _succ = wakee
1400   // 2. membar #loadstore|#storestore;
1401   // 2. ST _owner = NULL
1402   // 3. unpark(wakee)
1403 
1404   _succ = Wakee->_thread;
1405   ParkEvent * Trigger = Wakee->_event;
1406 
1407   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1408   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1409   // out-of-scope (non-extant).
1410   Wakee  = NULL;
1411 
1412   // Drop the lock.
1413   // Uses a fence to separate release_store(owner) from the LD in unpark().
1414   release_clear_owner(current);
1415   OrderAccess::fence();
1416 
1417   DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1418   Trigger->unpark();
1419 
1420   // Maintain stats and report events to JVMTI
1421   OM_PERFDATA_OP(Parks, inc());
1422 }
1423 
1424 
1425 // -----------------------------------------------------------------------------
1426 // Class Loader deadlock handling.
1427 //
1428 // complete_exit exits a lock returning recursion count
1429 // complete_exit/reenter operate as a wait without waiting
1430 // complete_exit requires an inflated monitor
1431 // The _owner field is not always the Thread addr even with an
1432 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1433 // thread due to contention.
1434 intx ObjectMonitor::complete_exit(JavaThread* current) {
1435   assert(InitDone, "Unexpectedly not initialized");
1436 
1437   void* cur = owner_raw();
1438   if (current != cur) {
1439     if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1440       assert(_recursions == 0, "internal state error");
1441       set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1442       _recursions = 0;
1443     }
1444   }
1445 
1446   guarantee(current == owner_raw(), "complete_exit not owner");
1447   intx save = _recursions; // record the old recursion count
1448   _recursions = 0;         // set the recursion level to be 0
1449   exit(current);           // exit the monitor
1450   guarantee(owner_raw() != current, "invariant");
1451   return save;
1452 }
1453 
1454 // reenter() enters a lock and sets recursion count
1455 // complete_exit/reenter operate as a wait without waiting
1456 bool ObjectMonitor::reenter(intx recursions, JavaThread* current) {
1457 
1458   guarantee(owner_raw() != current, "reenter already owner");
1459   if (!enter(current)) {
1460     return false;
1461   }
1462   // Entered the monitor.
1463   guarantee(_recursions == 0, "reenter recursion");
1464   _recursions = recursions;
1465   return true;
1466 }
1467 
1468 // Checks that the current THREAD owns this monitor and causes an
1469 // immediate return if it doesn't. We don't use the CHECK macro
1470 // because we want the IMSE to be the only exception that is thrown
1471 // from the call site when false is returned. Any other pending
1472 // exception is ignored.
1473 #define CHECK_OWNER()                                                  \
1474   do {                                                                 \
1475     if (!check_owner(THREAD)) {                                        \
1476        assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1477        return;                                                         \
1478      }                                                                 \
1479   } while (false)
1480 
1481 // Returns true if the specified thread owns the ObjectMonitor.
1482 // Otherwise returns false and throws IllegalMonitorStateException
1483 // (IMSE). If there is a pending exception and the specified thread
1484 // is not the owner, that exception will be replaced by the IMSE.
1485 bool ObjectMonitor::check_owner(TRAPS) {
1486   JavaThread* current = THREAD;
1487   void* cur = owner_raw();
1488   assert(cur != anon_owner_ptr(), "no anon owner here");
1489   if (cur == current) {
1490     return true;
1491   }
1492   if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1493     set_owner_from_BasicLock(cur, current);  // Convert from BasicLock* to Thread*.
1494     _recursions = 0;
1495     return true;
1496   }
1497   THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1498              "current thread is not owner", false);
1499 }
1500 
1501 static inline bool is_excluded(const Klass* monitor_klass) {
1502   assert(monitor_klass != nullptr, "invariant");
1503   NOT_JFR_RETURN_(false);
1504   JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name());
1505 }
1506 
1507 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1508                                     ObjectMonitor* monitor,
1509                                     jlong notifier_tid,
1510                                     jlong timeout,
1511                                     bool timedout) {
1512   assert(event != NULL, "invariant");
1513   assert(monitor != NULL, "invariant");
1514   const Klass* monitor_klass = monitor->object()->klass();
1515   if (is_excluded(monitor_klass)) {
1516     return;
1517   }
1518   event->set_monitorClass(monitor_klass);
1519   event->set_timeout(timeout);
1520   // Set an address that is 'unique enough', such that events close in
1521   // time and with the same address are likely (but not guaranteed) to
1522   // belong to the same object.
1523   event->set_address((uintptr_t)monitor);
1524   event->set_notifier(notifier_tid);
1525   event->set_timedOut(timedout);
1526   event->commit();
1527 }
1528 
1529 // -----------------------------------------------------------------------------
1530 // Wait/Notify/NotifyAll
1531 //
1532 // Note: a subset of changes to ObjectMonitor::wait()
1533 // will need to be replicated in complete_exit
1534 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1535   JavaThread* current = THREAD;
1536 
1537   assert(InitDone, "Unexpectedly not initialized");
1538 
1539   CHECK_OWNER();  // Throws IMSE if not owner.
1540 
1541   EventJavaMonitorWait event;
1542 
1543   // check for a pending interrupt
1544   if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1545     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1546     if (JvmtiExport::should_post_monitor_waited()) {
1547       // Note: 'false' parameter is passed here because the
1548       // wait was not timed out due to thread interrupt.
1549       JvmtiExport::post_monitor_waited(current, this, false);
1550 
1551       // In this short circuit of the monitor wait protocol, the
1552       // current thread never drops ownership of the monitor and
1553       // never gets added to the wait queue so the current thread
1554       // cannot be made the successor. This means that the
1555       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1556       // consume an unpark() meant for the ParkEvent associated with
1557       // this ObjectMonitor.
1558     }
1559     if (event.should_commit()) {
1560       post_monitor_wait_event(&event, this, 0, millis, false);
1561     }
1562     THROW(vmSymbols::java_lang_InterruptedException());
1563     return;
1564   }
1565 
1566   assert(current->_Stalled == 0, "invariant");
1567   current->_Stalled = intptr_t(this);
1568   current->set_current_waiting_monitor(this);
1569 
1570   // create a node to be put into the queue
1571   // Critically, after we reset() the event but prior to park(), we must check
1572   // for a pending interrupt.
1573   ObjectWaiter node(current);
1574   node.TState = ObjectWaiter::TS_WAIT;
1575   current->_ParkEvent->reset();
1576   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1577 
1578   // Enter the waiting queue, which is a circular doubly linked list in this case
1579   // but it could be a priority queue or any data structure.
1580   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1581   // by the the owner of the monitor *except* in the case where park()
1582   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1583   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1584 
1585   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1586   AddWaiter(&node);
1587   Thread::SpinRelease(&_WaitSetLock);
1588 
1589   _Responsible = NULL;
1590 
1591   intx save = _recursions;     // record the old recursion count
1592   _waiters++;                  // increment the number of waiters
1593   _recursions = 0;             // set the recursion level to be 1
1594   exit(current);               // exit the monitor
1595   guarantee(owner_raw() != current, "invariant");
1596 
1597   // The thread is on the WaitSet list - now park() it.
1598   // On MP systems it's conceivable that a brief spin before we park
1599   // could be profitable.
1600   //
1601   // TODO-FIXME: change the following logic to a loop of the form
1602   //   while (!timeout && !interrupted && _notified == 0) park()
1603 
1604   int ret = OS_OK;
1605   int WasNotified = 0;
1606 
1607   // Need to check interrupt state whilst still _thread_in_vm
1608   bool interrupted = interruptible && current->is_interrupted(false);
1609 
1610   { // State transition wrappers
1611     OSThread* osthread = current->osthread();
1612     OSThreadWaitState osts(osthread, true);
1613 
1614     assert(current->thread_state() == _thread_in_vm, "invariant");
1615 
1616     {
1617       ClearSuccOnSuspend csos(this);
1618       ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos);
1619       if (interrupted || HAS_PENDING_EXCEPTION) {
1620         // Intentionally empty
1621       } else if (node._notified == 0) {
1622         if (millis <= 0) {
1623           current->_ParkEvent->park();
1624         } else {
1625           ret = current->_ParkEvent->park(millis);
1626         }
1627       }
1628     }
1629 
1630     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1631     // from the WaitSet to the EntryList.
1632     // See if we need to remove Node from the WaitSet.
1633     // We use double-checked locking to avoid grabbing _WaitSetLock
1634     // if the thread is not on the wait queue.
1635     //
1636     // Note that we don't need a fence before the fetch of TState.
1637     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1638     // written by the is thread. (perhaps the fetch might even be satisfied
1639     // by a look-aside into the processor's own store buffer, although given
1640     // the length of the code path between the prior ST and this load that's
1641     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1642     // then we'll acquire the lock and then re-fetch a fresh TState value.
1643     // That is, we fail toward safety.
1644 
1645     if (node.TState == ObjectWaiter::TS_WAIT) {
1646       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1647       if (node.TState == ObjectWaiter::TS_WAIT) {
1648         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1649         assert(node._notified == 0, "invariant");
1650         node.TState = ObjectWaiter::TS_RUN;
1651       }
1652       Thread::SpinRelease(&_WaitSetLock);
1653     }
1654 
1655     // The thread is now either on off-list (TS_RUN),
1656     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1657     // The Node's TState variable is stable from the perspective of this thread.
1658     // No other threads will asynchronously modify TState.
1659     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1660     OrderAccess::loadload();
1661     if (_succ == current) _succ = NULL;
1662     WasNotified = node._notified;
1663 
1664     // Reentry phase -- reacquire the monitor.
1665     // re-enter contended monitor after object.wait().
1666     // retain OBJECT_WAIT state until re-enter successfully completes
1667     // Thread state is thread_in_vm and oop access is again safe,
1668     // although the raw address of the object may have changed.
1669     // (Don't cache naked oops over safepoints, of course).
1670 
1671     // post monitor waited event. Note that this is past-tense, we are done waiting.
1672     if (JvmtiExport::should_post_monitor_waited()) {
1673       JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1674 
1675       if (node._notified != 0 && _succ == current) {
1676         // In this part of the monitor wait-notify-reenter protocol it
1677         // is possible (and normal) for another thread to do a fastpath
1678         // monitor enter-exit while this thread is still trying to get
1679         // to the reenter portion of the protocol.
1680         //
1681         // The ObjectMonitor was notified and the current thread is
1682         // the successor which also means that an unpark() has already
1683         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1684         // consume the unpark() that was done when the successor was
1685         // set because the same ParkEvent is shared between Java
1686         // monitors and JVM/TI RawMonitors (for now).
1687         //
1688         // We redo the unpark() to ensure forward progress, i.e., we
1689         // don't want all pending threads hanging (parked) with none
1690         // entering the unlocked monitor.
1691         node._event->unpark();
1692       }
1693     }
1694 
1695     if (event.should_commit()) {
1696       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1697     }
1698 
1699     OrderAccess::fence();
1700 
1701     assert(current->_Stalled != 0, "invariant");
1702     current->_Stalled = 0;
1703 
1704     assert(owner_raw() != current, "invariant");
1705     ObjectWaiter::TStates v = node.TState;
1706     if (v == ObjectWaiter::TS_RUN) {
1707       enter(current);
1708     } else {
1709       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1710       ReenterI(current, &node);
1711       node.wait_reenter_end(this);
1712     }
1713 
1714     // current has reacquired the lock.
1715     // Lifecycle - the node representing current must not appear on any queues.
1716     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1717     // want residual elements associated with this thread left on any lists.
1718     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1719     assert(owner_raw() == current, "invariant");
1720     assert(_succ != current, "invariant");
1721   } // OSThreadWaitState()
1722 
1723   current->set_current_waiting_monitor(NULL);
1724 
1725   guarantee(_recursions == 0, "invariant");
1726   _recursions = save      // restore the old recursion count
1727                 + JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current); //  increased by the deferred relock count
1728   _waiters--;             // decrement the number of waiters
1729 
1730   // Verify a few postconditions
1731   assert(owner_raw() == current, "invariant");
1732   assert(_succ != current, "invariant");
1733   assert(object()->mark() == markWord::encode(this), "invariant");
1734 
1735   // check if the notification happened
1736   if (!WasNotified) {
1737     // no, it could be timeout or Thread.interrupt() or both
1738     // check for interrupt event, otherwise it is timeout
1739     if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1740       THROW(vmSymbols::java_lang_InterruptedException());
1741     }
1742   }
1743 
1744   // NOTE: Spurious wake up will be consider as timeout.
1745   // Monitor notify has precedence over thread interrupt.
1746 }
1747 
1748 
1749 // Consider:
1750 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1751 // then instead of transferring a thread from the WaitSet to the EntryList
1752 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1753 
1754 void ObjectMonitor::INotify(JavaThread* current) {
1755   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1756   ObjectWaiter* iterator = DequeueWaiter();
1757   if (iterator != NULL) {
1758     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1759     guarantee(iterator->_notified == 0, "invariant");
1760     // Disposition - what might we do with iterator ?
1761     // a.  add it directly to the EntryList - either tail (policy == 1)
1762     //     or head (policy == 0).
1763     // b.  push it onto the front of the _cxq (policy == 2).
1764     // For now we use (b).
1765 
1766     iterator->TState = ObjectWaiter::TS_ENTER;
1767 
1768     iterator->_notified = 1;
1769     iterator->_notifier_tid = JFR_THREAD_ID(current);
1770 
1771     ObjectWaiter* list = _EntryList;
1772     if (list != NULL) {
1773       assert(list->_prev == NULL, "invariant");
1774       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1775       assert(list != iterator, "invariant");
1776     }
1777 
1778     // prepend to cxq
1779     if (list == NULL) {
1780       iterator->_next = iterator->_prev = NULL;
1781       _EntryList = iterator;
1782     } else {
1783       iterator->TState = ObjectWaiter::TS_CXQ;
1784       for (;;) {
1785         ObjectWaiter* front = _cxq;
1786         iterator->_next = front;
1787         if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1788           break;
1789         }
1790       }
1791     }
1792 
1793     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1794     // move the add-to-EntryList operation, above, outside the critical section
1795     // protected by _WaitSetLock.  In practice that's not useful.  With the
1796     // exception of  wait() timeouts and interrupts the monitor owner
1797     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1798     // on _WaitSetLock so it's not profitable to reduce the length of the
1799     // critical section.
1800 
1801     iterator->wait_reenter_begin(this);
1802   }
1803   Thread::SpinRelease(&_WaitSetLock);
1804 }
1805 
1806 // Consider: a not-uncommon synchronization bug is to use notify() when
1807 // notifyAll() is more appropriate, potentially resulting in stranded
1808 // threads; this is one example of a lost wakeup. A useful diagnostic
1809 // option is to force all notify() operations to behave as notifyAll().
1810 //
1811 // Note: We can also detect many such problems with a "minimum wait".
1812 // When the "minimum wait" is set to a small non-zero timeout value
1813 // and the program does not hang whereas it did absent "minimum wait",
1814 // that suggests a lost wakeup bug.
1815 
1816 void ObjectMonitor::notify(TRAPS) {
1817   JavaThread* current = THREAD;
1818   CHECK_OWNER();  // Throws IMSE if not owner.
1819   if (_WaitSet == NULL) {
1820     return;
1821   }
1822   DTRACE_MONITOR_PROBE(notify, this, object(), current);
1823   INotify(current);
1824   OM_PERFDATA_OP(Notifications, inc(1));
1825 }
1826 
1827 
1828 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1829 // from the waitset to the EntryList. This could be done more efficiently with a
1830 // single bulk transfer but in practice it's not time-critical. Beware too,
1831 // that in prepend-mode we invert the order of the waiters. Let's say that the
1832 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1833 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1834 
1835 void ObjectMonitor::notifyAll(TRAPS) {
1836   JavaThread* current = THREAD;
1837   CHECK_OWNER();  // Throws IMSE if not owner.
1838   if (_WaitSet == NULL) {
1839     return;
1840   }
1841 
1842   DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1843   int tally = 0;
1844   while (_WaitSet != NULL) {
1845     tally++;
1846     INotify(current);
1847   }
1848 
1849   OM_PERFDATA_OP(Notifications, inc(tally));
1850 }
1851 
1852 // -----------------------------------------------------------------------------
1853 // Adaptive Spinning Support
1854 //
1855 // Adaptive spin-then-block - rational spinning
1856 //
1857 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1858 // algorithm.  On high order SMP systems it would be better to start with
1859 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1860 // a contending thread could enqueue itself on the cxq and then spin locally
1861 // on a thread-specific variable such as its ParkEvent._Event flag.
1862 // That's left as an exercise for the reader.  Note that global spinning is
1863 // not problematic on Niagara, as the L2 cache serves the interconnect and
1864 // has both low latency and massive bandwidth.
1865 //
1866 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1867 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1868 // (duration) or we can fix the count at approximately the duration of
1869 // a context switch and vary the frequency.   Of course we could also
1870 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1871 // For a description of 'Adaptive spin-then-block mutual exclusion in
1872 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1873 //
1874 // This implementation varies the duration "D", where D varies with
1875 // the success rate of recent spin attempts. (D is capped at approximately
1876 // length of a round-trip context switch).  The success rate for recent
1877 // spin attempts is a good predictor of the success rate of future spin
1878 // attempts.  The mechanism adapts automatically to varying critical
1879 // section length (lock modality), system load and degree of parallelism.
1880 // D is maintained per-monitor in _SpinDuration and is initialized
1881 // optimistically.  Spin frequency is fixed at 100%.
1882 //
1883 // Note that _SpinDuration is volatile, but we update it without locks
1884 // or atomics.  The code is designed so that _SpinDuration stays within
1885 // a reasonable range even in the presence of races.  The arithmetic
1886 // operations on _SpinDuration are closed over the domain of legal values,
1887 // so at worst a race will install and older but still legal value.
1888 // At the very worst this introduces some apparent non-determinism.
1889 // We might spin when we shouldn't or vice-versa, but since the spin
1890 // count are relatively short, even in the worst case, the effect is harmless.
1891 //
1892 // Care must be taken that a low "D" value does not become an
1893 // an absorbing state.  Transient spinning failures -- when spinning
1894 // is overall profitable -- should not cause the system to converge
1895 // on low "D" values.  We want spinning to be stable and predictable
1896 // and fairly responsive to change and at the same time we don't want
1897 // it to oscillate, become metastable, be "too" non-deterministic,
1898 // or converge on or enter undesirable stable absorbing states.
1899 //
1900 // We implement a feedback-based control system -- using past behavior
1901 // to predict future behavior.  We face two issues: (a) if the
1902 // input signal is random then the spin predictor won't provide optimal
1903 // results, and (b) if the signal frequency is too high then the control
1904 // system, which has some natural response lag, will "chase" the signal.
1905 // (b) can arise from multimodal lock hold times.  Transient preemption
1906 // can also result in apparent bimodal lock hold times.
1907 // Although sub-optimal, neither condition is particularly harmful, as
1908 // in the worst-case we'll spin when we shouldn't or vice-versa.
1909 // The maximum spin duration is rather short so the failure modes aren't bad.
1910 // To be conservative, I've tuned the gain in system to bias toward
1911 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1912 // "rings" or oscillates between spinning and not spinning.  This happens
1913 // when spinning is just on the cusp of profitability, however, so the
1914 // situation is not dire.  The state is benign -- there's no need to add
1915 // hysteresis control to damp the transition rate between spinning and
1916 // not spinning.
1917 
1918 // Spinning: Fixed frequency (100%), vary duration
1919 int ObjectMonitor::TrySpin(JavaThread* current) {
1920   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1921   int ctr = Knob_FixedSpin;
1922   if (ctr != 0) {
1923     while (--ctr >= 0) {
1924       if (TryLock(current) > 0) return 1;
1925       SpinPause();
1926     }
1927     return 0;
1928   }
1929 
1930   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1931     if (TryLock(current) > 0) {
1932       // Increase _SpinDuration ...
1933       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1934       // Raising _SpurDuration to the poverty line is key.
1935       int x = _SpinDuration;
1936       if (x < Knob_SpinLimit) {
1937         if (x < Knob_Poverty) x = Knob_Poverty;
1938         _SpinDuration = x + Knob_BonusB;
1939       }
1940       return 1;
1941     }
1942     SpinPause();
1943   }
1944 
1945   // Admission control - verify preconditions for spinning
1946   //
1947   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1948   // becoming an absorbing state.  Put another way, we spin briefly to
1949   // sample, just in case the system load, parallelism, contention, or lock
1950   // modality changed.
1951   //
1952   // Consider the following alternative:
1953   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1954   // spin attempt.  "Periodically" might mean after a tally of
1955   // the # of failed spin attempts (or iterations) reaches some threshold.
1956   // This takes us into the realm of 1-out-of-N spinning, where we
1957   // hold the duration constant but vary the frequency.
1958 
1959   ctr = _SpinDuration;
1960   if (ctr <= 0) return 0;
1961 
1962   if (NotRunnable(current, (JavaThread*) owner_raw())) {
1963     return 0;
1964   }
1965 
1966   // We're good to spin ... spin ingress.
1967   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1968   // when preparing to LD...CAS _owner, etc and the CAS is likely
1969   // to succeed.
1970   if (_succ == NULL) {
1971     _succ = current;
1972   }
1973   Thread* prv = NULL;
1974 
1975   // There are three ways to exit the following loop:
1976   // 1.  A successful spin where this thread has acquired the lock.
1977   // 2.  Spin failure with prejudice
1978   // 3.  Spin failure without prejudice
1979 
1980   while (--ctr >= 0) {
1981 
1982     // Periodic polling -- Check for pending GC
1983     // Threads may spin while they're unsafe.
1984     // We don't want spinning threads to delay the JVM from reaching
1985     // a stop-the-world safepoint or to steal cycles from GC.
1986     // If we detect a pending safepoint we abort in order that
1987     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1988     // this thread, if safe, doesn't steal cycles from GC.
1989     // This is in keeping with the "no loitering in runtime" rule.
1990     // We periodically check to see if there's a safepoint pending.
1991     if ((ctr & 0xFF) == 0) {
1992       if (SafepointMechanism::should_process(current)) {
1993         goto Abort;           // abrupt spin egress
1994       }
1995       SpinPause();
1996     }
1997 
1998     // Probe _owner with TATAS
1999     // If this thread observes the monitor transition or flicker
2000     // from locked to unlocked to locked, then the odds that this
2001     // thread will acquire the lock in this spin attempt go down
2002     // considerably.  The same argument applies if the CAS fails
2003     // or if we observe _owner change from one non-null value to
2004     // another non-null value.   In such cases we might abort
2005     // the spin without prejudice or apply a "penalty" to the
2006     // spin count-down variable "ctr", reducing it by 100, say.
2007 
2008     JavaThread* ox = (JavaThread*) owner_raw();
2009     if (ox == NULL) {
2010       ox = (JavaThread*)try_set_owner_from(NULL, current);
2011       if (ox == NULL) {
2012         // The CAS succeeded -- this thread acquired ownership
2013         // Take care of some bookkeeping to exit spin state.
2014         if (_succ == current) {
2015           _succ = NULL;
2016         }
2017 
2018         // Increase _SpinDuration :
2019         // The spin was successful (profitable) so we tend toward
2020         // longer spin attempts in the future.
2021         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2022         // If we acquired the lock early in the spin cycle it
2023         // makes sense to increase _SpinDuration proportionally.
2024         // Note that we don't clamp SpinDuration precisely at SpinLimit.
2025         int x = _SpinDuration;
2026         if (x < Knob_SpinLimit) {
2027           if (x < Knob_Poverty) x = Knob_Poverty;
2028           _SpinDuration = x + Knob_Bonus;
2029         }
2030         return 1;
2031       }
2032 
2033       // The CAS failed ... we can take any of the following actions:
2034       // * penalize: ctr -= CASPenalty
2035       // * exit spin with prejudice -- goto Abort;
2036       // * exit spin without prejudice.
2037       // * Since CAS is high-latency, retry again immediately.
2038       prv = ox;
2039       goto Abort;
2040     }
2041 
2042     // Did lock ownership change hands ?
2043     if (ox != prv && prv != NULL) {
2044       goto Abort;
2045     }
2046     prv = ox;
2047 
2048     // Abort the spin if the owner is not executing.
2049     // The owner must be executing in order to drop the lock.
2050     // Spinning while the owner is OFFPROC is idiocy.
2051     // Consider: ctr -= RunnablePenalty ;
2052     if (NotRunnable(current, ox)) {
2053       goto Abort;
2054     }
2055     if (_succ == NULL) {
2056       _succ = current;
2057     }
2058   }
2059 
2060   // Spin failed with prejudice -- reduce _SpinDuration.
2061   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2062   // AIMD is globally stable.
2063   {
2064     int x = _SpinDuration;
2065     if (x > 0) {
2066       // Consider an AIMD scheme like: x -= (x >> 3) + 100
2067       // This is globally sample and tends to damp the response.
2068       x -= Knob_Penalty;
2069       if (x < 0) x = 0;
2070       _SpinDuration = x;
2071     }
2072   }
2073 
2074  Abort:
2075   if (_succ == current) {
2076     _succ = NULL;
2077     // Invariant: after setting succ=null a contending thread
2078     // must recheck-retry _owner before parking.  This usually happens
2079     // in the normal usage of TrySpin(), but it's safest
2080     // to make TrySpin() as foolproof as possible.
2081     OrderAccess::fence();
2082     if (TryLock(current) > 0) return 1;
2083   }
2084   return 0;
2085 }
2086 
2087 // NotRunnable() -- informed spinning
2088 //
2089 // Don't bother spinning if the owner is not eligible to drop the lock.
2090 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2091 // The thread must be runnable in order to drop the lock in timely fashion.
2092 // If the _owner is not runnable then spinning will not likely be
2093 // successful (profitable).
2094 //
2095 // Beware -- the thread referenced by _owner could have died
2096 // so a simply fetch from _owner->_thread_state might trap.
2097 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2098 // Because of the lifecycle issues, the _thread_state values
2099 // observed by NotRunnable() might be garbage.  NotRunnable must
2100 // tolerate this and consider the observed _thread_state value
2101 // as advisory.
2102 //
2103 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2104 // a thread pointer.
2105 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2106 // with the LSB of _owner.  Another option would be to probabilistically probe
2107 // the putative _owner->TypeTag value.
2108 //
2109 // Checking _thread_state isn't perfect.  Even if the thread is
2110 // in_java it might be blocked on a page-fault or have been preempted
2111 // and sitting on a ready/dispatch queue.
2112 //
2113 // The return value from NotRunnable() is *advisory* -- the
2114 // result is based on sampling and is not necessarily coherent.
2115 // The caller must tolerate false-negative and false-positive errors.
2116 // Spinning, in general, is probabilistic anyway.
2117 
2118 
2119 int ObjectMonitor::NotRunnable(JavaThread* current, JavaThread* ox) {
2120   // Check ox->TypeTag == 2BAD.
2121   if (ox == NULL) return 0;
2122 
2123   // Avoid transitive spinning ...
2124   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2125   // Immediately after T1 acquires L it's possible that T2, also
2126   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2127   // This occurs transiently after T1 acquired L but before
2128   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2129   // its spin in this circumstance.
2130   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2131 
2132   if (BlockedOn == 1) return 1;
2133   if (BlockedOn != 0) {
2134     return BlockedOn != intptr_t(this) && owner_raw() == ox;
2135   }
2136 
2137   assert(sizeof(ox->_thread_state == sizeof(int)), "invariant");
2138   int jst = SafeFetch32((int *) &ox->_thread_state, -1);;
2139   // consider also: jst != _thread_in_Java -- but that's overspecific.
2140   return jst == _thread_blocked || jst == _thread_in_native;
2141 }
2142 
2143 
2144 // -----------------------------------------------------------------------------
2145 // WaitSet management ...
2146 
2147 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2148   _next     = NULL;
2149   _prev     = NULL;
2150   _notified = 0;
2151   _notifier_tid = 0;
2152   TState    = TS_RUN;
2153   _thread   = current;
2154   _event    = _thread->_ParkEvent;
2155   _active   = false;
2156   assert(_event != NULL, "invariant");
2157 }
2158 
2159 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2160   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2161 }
2162 
2163 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2164   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2165 }
2166 
2167 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2168   assert(node != NULL, "should not add NULL node");
2169   assert(node->_prev == NULL, "node already in list");
2170   assert(node->_next == NULL, "node already in list");
2171   // put node at end of queue (circular doubly linked list)
2172   if (_WaitSet == NULL) {
2173     _WaitSet = node;
2174     node->_prev = node;
2175     node->_next = node;
2176   } else {
2177     ObjectWaiter* head = _WaitSet;
2178     ObjectWaiter* tail = head->_prev;
2179     assert(tail->_next == head, "invariant check");
2180     tail->_next = node;
2181     head->_prev = node;
2182     node->_next = head;
2183     node->_prev = tail;
2184   }
2185 }
2186 
2187 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2188   // dequeue the very first waiter
2189   ObjectWaiter* waiter = _WaitSet;
2190   if (waiter) {
2191     DequeueSpecificWaiter(waiter);
2192   }
2193   return waiter;
2194 }
2195 
2196 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2197   assert(node != NULL, "should not dequeue NULL node");
2198   assert(node->_prev != NULL, "node already removed from list");
2199   assert(node->_next != NULL, "node already removed from list");
2200   // when the waiter has woken up because of interrupt,
2201   // timeout or other spurious wake-up, dequeue the
2202   // waiter from waiting list
2203   ObjectWaiter* next = node->_next;
2204   if (next == node) {
2205     assert(node->_prev == node, "invariant check");
2206     _WaitSet = NULL;
2207   } else {
2208     ObjectWaiter* prev = node->_prev;
2209     assert(prev->_next == node, "invariant check");
2210     assert(next->_prev == node, "invariant check");
2211     next->_prev = prev;
2212     prev->_next = next;
2213     if (_WaitSet == node) {
2214       _WaitSet = next;
2215     }
2216   }
2217   node->_next = NULL;
2218   node->_prev = NULL;
2219 }
2220 
2221 // -----------------------------------------------------------------------------
2222 // PerfData support
2223 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2224 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2225 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2226 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2227 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2228 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2229 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2230 
2231 // One-shot global initialization for the sync subsystem.
2232 // We could also defer initialization and initialize on-demand
2233 // the first time we call ObjectSynchronizer::inflate().
2234 // Initialization would be protected - like so many things - by
2235 // the MonitorCache_lock.
2236 
2237 void ObjectMonitor::Initialize() {
2238   assert(!InitDone, "invariant");
2239 
2240   if (!os::is_MP()) {
2241     Knob_SpinLimit = 0;
2242     Knob_PreSpin   = 0;
2243     Knob_FixedSpin = -1;
2244   }
2245 
2246   if (UsePerfData) {
2247     EXCEPTION_MARK;
2248 #define NEWPERFCOUNTER(n)                                                \
2249   {                                                                      \
2250     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2251                                         CHECK);                          \
2252   }
2253 #define NEWPERFVARIABLE(n)                                                \
2254   {                                                                       \
2255     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2256                                          CHECK);                          \
2257   }
2258     NEWPERFCOUNTER(_sync_Inflations);
2259     NEWPERFCOUNTER(_sync_Deflations);
2260     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2261     NEWPERFCOUNTER(_sync_FutileWakeups);
2262     NEWPERFCOUNTER(_sync_Parks);
2263     NEWPERFCOUNTER(_sync_Notifications);
2264     NEWPERFVARIABLE(_sync_MonExtant);
2265 #undef NEWPERFCOUNTER
2266 #undef NEWPERFVARIABLE
2267   }
2268 
2269   _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2270 
2271   DEBUG_ONLY(InitDone = true;)
2272 }
2273 
2274 void ObjectMonitor::print_on(outputStream* st) const {
2275   // The minimal things to print for markWord printing, more can be added for debugging and logging.
2276   st->print("{contentions=0x%08x,waiters=0x%08x"
2277             ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2278             contentions(), waiters(), recursions(),
2279             p2i(owner()));
2280 }
2281 void ObjectMonitor::print() const { print_on(tty); }
2282 
2283 #ifdef ASSERT
2284 // Print the ObjectMonitor like a debugger would:
2285 //
2286 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2287 //   _header = 0x0000000000000001
2288 //   _object = 0x000000070ff45fd0
2289 //   _pad_buf0 = {
2290 //     [0] = '\0'
2291 //     ...
2292 //     [43] = '\0'
2293 //   }
2294 //   _owner = 0x0000000000000000
2295 //   _previous_owner_tid = 0
2296 //   _pad_buf1 = {
2297 //     [0] = '\0'
2298 //     ...
2299 //     [47] = '\0'
2300 //   }
2301 //   _next_om = 0x0000000000000000
2302 //   _recursions = 0
2303 //   _EntryList = 0x0000000000000000
2304 //   _cxq = 0x0000000000000000
2305 //   _succ = 0x0000000000000000
2306 //   _Responsible = 0x0000000000000000
2307 //   _Spinner = 0
2308 //   _SpinDuration = 5000
2309 //   _contentions = 0
2310 //   _WaitSet = 0x0000700009756248
2311 //   _waiters = 1
2312 //   _WaitSetLock = 0
2313 // }
2314 //
2315 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2316   st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2317   st->print_cr("  _header = " INTPTR_FORMAT, header().value());
2318   st->print_cr("  _object = " INTPTR_FORMAT, p2i(object_peek()));
2319   st->print_cr("  _pad_buf0 = {");
2320   st->print_cr("    [0] = '\\0'");
2321   st->print_cr("    ...");
2322   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2323   st->print_cr("  }");
2324   st->print_cr("  _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2325   st->print_cr("  _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid);
2326   st->print_cr("  _pad_buf1 = {");
2327   st->print_cr("    [0] = '\\0'");
2328   st->print_cr("    ...");
2329   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2330   st->print_cr("  }");
2331   st->print_cr("  _next_om = " INTPTR_FORMAT, p2i(next_om()));
2332   st->print_cr("  _recursions = " INTX_FORMAT, _recursions);
2333   st->print_cr("  _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2334   st->print_cr("  _cxq = " INTPTR_FORMAT, p2i(_cxq));
2335   st->print_cr("  _succ = " INTPTR_FORMAT, p2i(_succ));
2336   st->print_cr("  _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2337   st->print_cr("  _Spinner = %d", _Spinner);
2338   st->print_cr("  _SpinDuration = %d", _SpinDuration);
2339   st->print_cr("  _contentions = %d", contentions());
2340   st->print_cr("  _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2341   st->print_cr("  _waiters = %d", _waiters);
2342   st->print_cr("  _WaitSetLock = %d", _WaitSetLock);
2343   st->print_cr("}");
2344 }
2345 #endif