1 /*
   2  * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/vmSymbols.hpp"
  27 #include "gc/shared/oopStorage.hpp"
  28 #include "gc/shared/oopStorageSet.hpp"
  29 #include "jfr/jfrEvents.hpp"
  30 #include "jfr/support/jfrThreadId.hpp"
  31 #include "logging/log.hpp"
  32 #include "logging/logStream.hpp"
  33 #include "memory/allocation.inline.hpp"
  34 #include "memory/resourceArea.hpp"
  35 #include "oops/markWord.hpp"
  36 #include "oops/oop.inline.hpp"
  37 #include "oops/oopHandle.inline.hpp"
  38 #include "oops/weakHandle.inline.hpp"
  39 #include "prims/jvmtiDeferredUpdates.hpp"
  40 #include "prims/jvmtiExport.hpp"
  41 #include "runtime/atomic.hpp"
  42 #include "runtime/globals.hpp"
  43 #include "runtime/handles.inline.hpp"
  44 #include "runtime/interfaceSupport.inline.hpp"
  45 #include "runtime/javaThread.inline.hpp"
  46 #include "runtime/mutexLocker.hpp"
  47 #include "runtime/objectMonitor.hpp"
  48 #include "runtime/objectMonitor.inline.hpp"
  49 #include "runtime/orderAccess.hpp"
  50 #include "runtime/osThread.hpp"
  51 #include "runtime/perfData.hpp"
  52 #include "runtime/safefetch.hpp"
  53 #include "runtime/safepointMechanism.inline.hpp"
  54 #include "runtime/sharedRuntime.hpp"
  55 #include "runtime/threads.hpp"
  56 #include "services/threadService.hpp"
  57 #include "utilities/dtrace.hpp"
  58 #include "utilities/globalDefinitions.hpp"
  59 #include "utilities/macros.hpp"
  60 #include "utilities/preserveException.hpp"
  61 #if INCLUDE_JFR
  62 #include "jfr/support/jfrFlush.hpp"
  63 #endif
  64 
  65 #ifdef DTRACE_ENABLED
  66 
  67 // Only bother with this argument setup if dtrace is available
  68 // TODO-FIXME: probes should not fire when caller is _blocked.  assert() accordingly.
  69 
  70 
  71 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread)                           \
  72   char* bytes = nullptr;                                                   \
  73   int len = 0;                                                             \
  74   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  75   Symbol* klassname = obj->klass()->name();                                \
  76   if (klassname != nullptr) {                                              \
  77     bytes = (char*)klassname->bytes();                                     \
  78     len = klassname->utf8_length();                                        \
  79   }
  80 
  81 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)            \
  82   {                                                                        \
  83     if (DTraceMonitorProbes) {                                             \
  84       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  85       HOTSPOT_MONITOR_WAIT(jtid,                                           \
  86                            (monitor), bytes, len, (millis));               \
  87     }                                                                      \
  88   }
  89 
  90 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
  91 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
  92 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
  93 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
  94 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
  95 
  96 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)                  \
  97   {                                                                        \
  98     if (DTraceMonitorProbes) {                                             \
  99       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
 100       HOTSPOT_MONITOR_##probe(jtid,                                        \
 101                               (uintptr_t)(monitor), bytes, len);           \
 102     }                                                                      \
 103   }
 104 
 105 #else //  ndef DTRACE_ENABLED
 106 
 107 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon)    {;}
 108 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon)          {;}
 109 
 110 #endif // ndef DTRACE_ENABLED
 111 
 112 // Tunables ...
 113 // The knob* variables are effectively final.  Once set they should
 114 // never be modified hence.  Consider using __read_mostly with GCC.
 115 
 116 int ObjectMonitor::Knob_SpinLimit    = 5000;    // derived by an external tool -
 117 
 118 static int Knob_Bonus               = 100;     // spin success bonus
 119 static int Knob_BonusB              = 100;     // spin success bonus
 120 static int Knob_Penalty             = 200;     // spin failure penalty
 121 static int Knob_Poverty             = 1000;
 122 static int Knob_FixedSpin           = 0;
 123 static int Knob_PreSpin             = 10;      // 20-100 likely better
 124 
 125 DEBUG_ONLY(static volatile bool InitDone = false;)
 126 
 127 OopStorage* ObjectMonitor::_oop_storage = nullptr;
 128 
 129 OopHandle ObjectMonitor::_vthread_cxq_head;
 130 ParkEvent* ObjectMonitor::_vthread_unparker_ParkEvent = nullptr;
 131 
 132 // -----------------------------------------------------------------------------
 133 // Theory of operations -- Monitors lists, thread residency, etc:
 134 //
 135 // * A thread acquires ownership of a monitor by successfully
 136 //   CAS()ing the _owner field from null to non-null.
 137 //
 138 // * Invariant: A thread appears on at most one monitor list --
 139 //   cxq, EntryList or WaitSet -- at any one time.
 140 //
 141 // * Contending threads "push" themselves onto the cxq with CAS
 142 //   and then spin/park.
 143 //
 144 // * After a contending thread eventually acquires the lock it must
 145 //   dequeue itself from either the EntryList or the cxq.
 146 //
 147 // * The exiting thread identifies and unparks an "heir presumptive"
 148 //   tentative successor thread on the EntryList.  Critically, the
 149 //   exiting thread doesn't unlink the successor thread from the EntryList.
 150 //   After having been unparked, the wakee will recontend for ownership of
 151 //   the monitor.   The successor (wakee) will either acquire the lock or
 152 //   re-park itself.
 153 //
 154 //   Succession is provided for by a policy of competitive handoff.
 155 //   The exiting thread does _not_ grant or pass ownership to the
 156 //   successor thread.  (This is also referred to as "handoff" succession").
 157 //   Instead the exiting thread releases ownership and possibly wakes
 158 //   a successor, so the successor can (re)compete for ownership of the lock.
 159 //   If the EntryList is empty but the cxq is populated the exiting
 160 //   thread will drain the cxq into the EntryList.  It does so by
 161 //   by detaching the cxq (installing null with CAS) and folding
 162 //   the threads from the cxq into the EntryList.  The EntryList is
 163 //   doubly linked, while the cxq is singly linked because of the
 164 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 165 //
 166 // * Concurrency invariants:
 167 //
 168 //   -- only the monitor owner may access or mutate the EntryList.
 169 //      The mutex property of the monitor itself protects the EntryList
 170 //      from concurrent interference.
 171 //   -- Only the monitor owner may detach the cxq.
 172 //
 173 // * The monitor entry list operations avoid locks, but strictly speaking
 174 //   they're not lock-free.  Enter is lock-free, exit is not.
 175 //   For a description of 'Methods and apparatus providing non-blocking access
 176 //   to a resource,' see U.S. Pat. No. 7844973.
 177 //
 178 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 179 //   detaching thread.  This mechanism is immune from the ABA corruption.
 180 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 181 //
 182 // * Taken together, the cxq and the EntryList constitute or form a
 183 //   single logical queue of threads stalled trying to acquire the lock.
 184 //   We use two distinct lists to improve the odds of a constant-time
 185 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 186 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 187 //   A key desideratum is to minimize queue & monitor metadata manipulation
 188 //   that occurs while holding the monitor lock -- that is, we want to
 189 //   minimize monitor lock holds times.  Note that even a small amount of
 190 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 191 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 192 //   locks and monitor metadata.
 193 //
 194 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 195 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 196 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 197 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 198 //
 199 //   The EntryList is ordered by the prevailing queue discipline and
 200 //   can be organized in any convenient fashion, such as a doubly-linked list or
 201 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 202 //   to operate in constant-time.  If we need a priority queue then something akin
 203 //   to Solaris' sleepq would work nicely.  Viz.,
 204 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 205 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 206 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 207 //   EntryList accordingly.
 208 //
 209 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 210 //   somewhat similar to an elevator-scan.
 211 //
 212 // * The monitor synchronization subsystem avoids the use of native
 213 //   synchronization primitives except for the narrow platform-specific
 214 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 215 //   the semantics of park-unpark.  Put another way, this monitor implementation
 216 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 217 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 218 //   underlying OS manages the READY<->RUN transitions.
 219 //
 220 // * Waiting threads reside on the WaitSet list -- wait() puts
 221 //   the caller onto the WaitSet.
 222 //
 223 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 224 //   either the EntryList or cxq.  Subsequent exit() operations will
 225 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 226 //   it's likely the notifyee would simply impale itself on the lock held
 227 //   by the notifier.
 228 //
 229 // * An interesting alternative is to encode cxq as (List,LockByte) where
 230 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 231 //   variable, like _recursions, in the scheme.  The threads or Events that form
 232 //   the list would have to be aligned in 256-byte addresses.  A thread would
 233 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 234 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 235 //   Note that is is *not* word-tearing, but it does presume that full-word
 236 //   CAS operations are coherent with intermix with STB operations.  That's true
 237 //   on most common processors.
 238 //
 239 // * See also http://blogs.sun.com/dave
 240 
 241 
 242 // Check that object() and set_object() are called from the right context:
 243 static void check_object_context() {
 244 #ifdef ASSERT
 245   Thread* self = Thread::current();
 246   if (self->is_Java_thread()) {
 247     // Mostly called from JavaThreads so sanity check the thread state.
 248     JavaThread* jt = JavaThread::cast(self);
 249     switch (jt->thread_state()) {
 250     case _thread_in_vm:    // the usual case
 251     case _thread_in_Java:  // during deopt
 252       break;
 253     default:
 254       fatal("called from an unsafe thread state");
 255     }
 256     assert(jt->is_active_Java_thread(), "must be active JavaThread");
 257   } else {
 258     // However, ThreadService::get_current_contended_monitor()
 259     // can call here via the VMThread so sanity check it.
 260     assert(self->is_VM_thread(), "must be");
 261   }
 262 #endif // ASSERT
 263 }
 264 
 265 ObjectMonitor::ObjectMonitor(oop object) :
 266   _header(markWord::zero()),
 267   _object(_oop_storage, object),
 268   _owner(nullptr),
 269   _stack_locker(nullptr),
 270   _previous_owner_tid(0),
 271   _next_om(nullptr),
 272   _recursions(0),
 273   _EntryList(nullptr),
 274   _cxq(nullptr),
 275   _succ(nullptr),
 276   _Responsible(nullptr),
 277   _Spinner(0),
 278   _SpinDuration(ObjectMonitor::Knob_SpinLimit),
 279   _contentions(0),
 280   _WaitSet(nullptr),
 281   _waiters(0),
 282   _WaitSetLock(0)
 283 { }
 284 
 285 ObjectMonitor::~ObjectMonitor() {
 286   _object.release(_oop_storage);
 287 }
 288 
 289 oop ObjectMonitor::object() const {
 290   check_object_context();
 291   return _object.resolve();
 292 }
 293 
 294 oop ObjectMonitor::object_peek() const {
 295   return _object.peek();
 296 }
 297 
 298 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
 299   if (current->is_suspended()) {
 300     _om->_recursions = 0;
 301     _om->_succ = nullptr;
 302     // Don't need a full fence after clearing successor here because of the call to exit().
 303     _om->exit(current, false /* not_suspended */);
 304     _om_exited = true;
 305 
 306     current->set_current_pending_monitor(_om);
 307   }
 308 }
 309 
 310 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
 311   if (current->is_suspended()) {
 312     if (_om->_succ == current) {
 313       _om->_succ = nullptr;
 314       OrderAccess::fence(); // always do a full fence when successor is cleared
 315     }
 316   }
 317 }
 318 
 319 // -----------------------------------------------------------------------------
 320 // Enter support
 321 
 322 bool ObjectMonitor::enter_for(JavaThread* locking_thread) {
 323   // Used by ObjectSynchronizer::enter_for to enter for another thread.
 324   // The monitor is private to or already owned by locking_thread which must be suspended.
 325   // So this code may only contend with deflation.
 326   assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
 327 
 328   // Block out deflation as soon as possible.
 329   add_to_contentions(1);
 330 
 331   bool success = false;
 332   if (!is_being_async_deflated()) {
 333     void* prev_owner = try_set_owner_from(nullptr, locking_thread);
 334 
 335     if (prev_owner == nullptr) {
 336       assert(_recursions == 0, "invariant");
 337       success = true;
 338     } else if (prev_owner == owner_for(locking_thread)) {
 339       _recursions++;
 340       success = true;
 341     } else if (prev_owner == DEFLATER_MARKER) {
 342       // Racing with deflation.
 343       prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread);
 344       if (prev_owner == DEFLATER_MARKER) {
 345         // Cancelled deflation. Increment contentions as part of the deflation protocol.
 346         add_to_contentions(1);
 347         success = true;
 348       } else if (prev_owner == nullptr) {
 349         // At this point we cannot race with deflation as we have both incremented
 350         // contentions, seen contention > 0 and seen a DEFLATER_MARKER.
 351         // success will only be false if this races with something other than
 352         // deflation.
 353         prev_owner = try_set_owner_from(nullptr, locking_thread);
 354         success = prev_owner == nullptr;
 355       }
 356     }
 357     assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT
 358            ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT,
 359            p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner));
 360   } else {
 361     // Async deflation is in progress and our contentions increment
 362     // above lost the race to async deflation. Undo the work and
 363     // force the caller to retry.
 364     const oop l_object = object();
 365     if (l_object != nullptr) {
 366       // Attempt to restore the header/dmw to the object's header so that
 367       // we only retry once if the deflater thread happens to be slow.
 368       install_displaced_markword_in_object(l_object);
 369     }
 370   }
 371 
 372   add_to_contentions(-1);
 373 
 374   assert(!success || is_owner(locking_thread), "must be");
 375 
 376   return success;
 377 }
 378 
 379 bool ObjectMonitor::enter(JavaThread* current) {
 380   assert(current == JavaThread::current(), "must be");
 381   // The following code is ordered to check the most common cases first
 382   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 383 
 384   void* cur = try_set_owner_from(nullptr, current);
 385   if (cur == nullptr) {
 386     assert(_recursions == 0, "invariant");
 387     return true;
 388   }
 389 
 390   if (cur == owner_for(current)) {
 391     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 392     _recursions++;
 393     return true;
 394   }
 395 
 396   // We've encountered genuine contention.
 397   assert(current->_Stalled == 0, "invariant");
 398   current->_Stalled = intptr_t(this);
 399 
 400   // Try one round of spinning *before* enqueueing current
 401   // and before going through the awkward and expensive state
 402   // transitions.  The following spin is strictly optional ...
 403   // Note that if we acquire the monitor from an initial spin
 404   // we forgo posting JVMTI events and firing DTRACE probes.
 405   if (TrySpin(current) > 0) {
 406     assert(owner_raw() == owner_for(current), "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
 407     assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
 408     assert(object()->mark() == markWord::encode(this),
 409            "object mark must match encoded this: mark=" INTPTR_FORMAT
 410            ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
 411            markWord::encode(this).value());
 412     current->_Stalled = 0;
 413     return true;
 414   }
 415 
 416   assert(owner_raw() != owner_for(current), "invariant");
 417   assert(_succ != current, "invariant");
 418   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 419   assert(current->thread_state() != _thread_blocked, "invariant");
 420 
 421   // Keep track of contention for JVM/TI and M&M queries.
 422   add_to_contentions(1);
 423   if (is_being_async_deflated()) {
 424     // Async deflation is in progress and our contentions increment
 425     // above lost the race to async deflation. Undo the work and
 426     // force the caller to retry.
 427     const oop l_object = object();
 428     if (l_object != nullptr) {
 429       // Attempt to restore the header/dmw to the object's header so that
 430       // we only retry once if the deflater thread happens to be slow.
 431       install_displaced_markword_in_object(l_object);
 432     }
 433     current->_Stalled = 0;
 434     add_to_contentions(-1);
 435     return false;
 436   }
 437 
 438   JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);)
 439   EventJavaMonitorEnter event;
 440   if (event.is_started()) {
 441     event.set_monitorClass(object()->klass());
 442     // Set an address that is 'unique enough', such that events close in
 443     // time and with the same address are likely (but not guaranteed) to
 444     // belong to the same object.
 445     event.set_address((uintptr_t)this);
 446   }
 447 
 448   { // Change java thread status to indicate blocked on monitor enter.
 449     JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
 450 
 451     assert(current->current_pending_monitor() == nullptr, "invariant");
 452     current->set_current_pending_monitor(this);
 453 
 454     DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
 455     if (JvmtiExport::should_post_monitor_contended_enter()) {
 456       JvmtiExport::post_monitor_contended_enter(current, this);
 457 
 458       // The current thread does not yet own the monitor and does not
 459       // yet appear on any queues that would get it made the successor.
 460       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 461       // handler cannot accidentally consume an unpark() meant for the
 462       // ParkEvent associated with this ObjectMonitor.
 463     }
 464 
 465 #ifdef LOOM_MONITOR_SUPPORT
 466     ContinuationEntry* ce = current->last_continuation();
 467     if (ce != nullptr && ce->is_virtual_thread() && current->is_on_monitorenter()) {
 468       int result = Continuation::try_preempt(current, ce->cont_oop(current));
 469       if (result == freeze_ok) {
 470         bool acquired = HandlePreemptedVThread(current);
 471         DEBUG_ONLY(int state = java_lang_VirtualThread::state(current->vthread()));
 472         assert((acquired && current->preemption_cancelled() && state == java_lang_VirtualThread::RUNNING) ||
 473                (!acquired && !current->preemption_cancelled() && state == java_lang_VirtualThread::BLOCKING), "invariant");
 474         return true;
 475       }
 476       if (result == freeze_pinned_native || result == freeze_pinned_cs) {
 477         EventVirtualThreadPinned e;
 478         if (e.should_commit()) {
 479           e.commit();
 480         }
 481       }
 482     }
 483 #endif
 484 
 485     OSThreadContendState osts(current->osthread());
 486 
 487     assert(current->thread_state() == _thread_in_vm, "invariant");
 488 
 489     for (;;) {
 490       ExitOnSuspend eos(this);
 491       {
 492         ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */);
 493         EnterI(current);
 494         current->set_current_pending_monitor(nullptr);
 495         // We can go to a safepoint at the end of this block. If we
 496         // do a thread dump during that safepoint, then this thread will show
 497         // as having "-locked" the monitor, but the OS and java.lang.Thread
 498         // states will still report that the thread is blocked trying to
 499         // acquire it.
 500         // If there is a suspend request, ExitOnSuspend will exit the OM
 501         // and set the OM as pending.
 502       }
 503       if (!eos.exited()) {
 504         // ExitOnSuspend did not exit the OM
 505         assert(owner_raw() == owner_for(current), "invariant");
 506         break;
 507       }
 508     }
 509 
 510     // We've just gotten past the enter-check-for-suspend dance and we now own
 511     // the monitor free and clear.
 512   }
 513 
 514   add_to_contentions(-1);
 515   assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
 516   current->_Stalled = 0;
 517 
 518   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 519   assert(_recursions == 0, "invariant");
 520   assert(owner_raw() == owner_for(current), "invariant");
 521   assert(_succ != current, "invariant");
 522   assert(object()->mark() == markWord::encode(this), "invariant");
 523 
 524   // The thread -- now the owner -- is back in vm mode.
 525   // Report the glorious news via TI,DTrace and jvmstat.
 526   // The probe effect is non-trivial.  All the reportage occurs
 527   // while we hold the monitor, increasing the length of the critical
 528   // section.  Amdahl's parallel speedup law comes vividly into play.
 529   //
 530   // Another option might be to aggregate the events (thread local or
 531   // per-monitor aggregation) and defer reporting until a more opportune
 532   // time -- such as next time some thread encounters contention but has
 533   // yet to acquire the lock.  While spinning that thread could
 534   // spinning we could increment JVMStat counters, etc.
 535 
 536   DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
 537   if (JvmtiExport::should_post_monitor_contended_entered()) {
 538     JvmtiExport::post_monitor_contended_entered(current, this);
 539 
 540     // The current thread already owns the monitor and is not going to
 541     // call park() for the remainder of the monitor enter protocol. So
 542     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 543     // event handler consumed an unpark() issued by the thread that
 544     // just exited the monitor.
 545   }
 546   if (event.should_commit()) {
 547     event.set_previousOwner(_previous_owner_tid);
 548     event.commit();
 549   }
 550   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 551   return true;
 552 }
 553 
 554 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 555 // Callers must compensate as needed.
 556 
 557 int ObjectMonitor::TryLock(JavaThread* current) {
 558   void* own = owner_raw();
 559   if (own != nullptr) return 0;
 560   if (try_set_owner_from(nullptr, current) == nullptr) {
 561     assert(_recursions == 0, "invariant");
 562     return 1;
 563   }
 564   // The lock had been free momentarily, but we lost the race to the lock.
 565   // Interference -- the CAS failed.
 566   // We can either return -1 or retry.
 567   // Retry doesn't make as much sense because the lock was just acquired.
 568   return -1;
 569 }
 570 
 571 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
 572 // was deflated and false otherwise.
 573 //
 574 // The async deflation protocol sets owner to DEFLATER_MARKER and
 575 // makes contentions negative as signals to contending threads that
 576 // an async deflation is in progress. There are a number of checks
 577 // as part of the protocol to make sure that the calling thread has
 578 // not lost the race to a contending thread.
 579 //
 580 // The ObjectMonitor has been successfully async deflated when:
 581 //   (contentions < 0)
 582 // Contending threads that see that condition know to retry their operation.
 583 //
 584 bool ObjectMonitor::deflate_monitor() {
 585   if (is_busy()) {
 586     // Easy checks are first - the ObjectMonitor is busy so no deflation.
 587     return false;
 588   }
 589 
 590   const oop obj = object_peek();
 591 
 592   if (obj == nullptr) {
 593     // If the object died, we can recycle the monitor without racing with
 594     // Java threads. The GC already broke the association with the object.
 595     set_owner_from_raw(nullptr, DEFLATER_MARKER);
 596     assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
 597     _contentions = INT_MIN; // minimum negative int
 598   } else {
 599     // Attempt async deflation protocol.
 600 
 601     // Set a null owner to DEFLATER_MARKER to force any contending thread
 602     // through the slow path. This is just the first part of the async
 603     // deflation dance.
 604     if (try_set_owner_from_raw(nullptr, DEFLATER_MARKER) != nullptr) {
 605       // The owner field is no longer null so we lost the race since the
 606       // ObjectMonitor is now busy.
 607       return false;
 608     }
 609 
 610     if (contentions() > 0 || _waiters != 0) {
 611       // Another thread has raced to enter the ObjectMonitor after
 612       // is_busy() above or has already entered and waited on
 613       // it which makes it busy so no deflation. Restore owner to
 614       // null if it is still DEFLATER_MARKER.
 615       if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
 616         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 617         add_to_contentions(-1);
 618       }
 619       return false;
 620     }
 621 
 622     // Make a zero contentions field negative to force any contending threads
 623     // to retry. This is the second part of the async deflation dance.
 624     if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) {
 625       // Contentions was no longer 0 so we lost the race since the
 626       // ObjectMonitor is now busy. Restore owner to null if it is
 627       // still DEFLATER_MARKER:
 628       if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
 629         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 630         add_to_contentions(-1);
 631       }
 632       return false;
 633     }
 634   }
 635 
 636   // Sanity checks for the races:
 637   guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
 638   guarantee(contentions() < 0, "must be negative: contentions=%d",
 639             contentions());
 640   guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
 641   guarantee(_cxq == nullptr, "must be no contending threads: cxq="
 642             INTPTR_FORMAT, p2i(_cxq));
 643   guarantee(_EntryList == nullptr,
 644             "must be no entering threads: EntryList=" INTPTR_FORMAT,
 645             p2i(_EntryList));
 646 
 647   if (obj != nullptr) {
 648     if (log_is_enabled(Trace, monitorinflation)) {
 649       ResourceMark rm;
 650       log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
 651                                   ", mark=" INTPTR_FORMAT ", type='%s'",
 652                                   p2i(obj), obj->mark().value(),
 653                                   obj->klass()->external_name());
 654     }
 655 
 656     // Install the old mark word if nobody else has already done it.
 657     install_displaced_markword_in_object(obj);
 658   }
 659 
 660   // We leave owner == DEFLATER_MARKER and contentions < 0
 661   // to force any racing threads to retry.
 662   return true;  // Success, ObjectMonitor has been deflated.
 663 }
 664 
 665 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
 666 // into the header of the object associated with the monitor. This
 667 // idempotent method is called by a thread that is deflating a
 668 // monitor and by other threads that have detected a race with the
 669 // deflation process.
 670 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
 671   // This function must only be called when (owner == DEFLATER_MARKER
 672   // && contentions <= 0), but we can't guarantee that here because
 673   // those values could change when the ObjectMonitor gets moved from
 674   // the global free list to a per-thread free list.
 675 
 676   guarantee(obj != nullptr, "must be non-null");
 677 
 678   // Separate loads in is_being_async_deflated(), which is almost always
 679   // called before this function, from the load of dmw/header below.
 680 
 681   // _contentions and dmw/header may get written by different threads.
 682   // Make sure to observe them in the same order when having several observers.
 683   OrderAccess::loadload_for_IRIW();
 684 
 685   const oop l_object = object_peek();
 686   if (l_object == nullptr) {
 687     // ObjectMonitor's object ref has already been cleared by async
 688     // deflation or GC so we're done here.
 689     return;
 690   }
 691   assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
 692          INTPTR_FORMAT, p2i(l_object), p2i(obj));
 693 
 694   markWord dmw = header();
 695   // The dmw has to be neutral (not null, not locked and not marked).
 696   assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
 697 
 698   // Install displaced mark word if the object's header still points
 699   // to this ObjectMonitor. More than one racing caller to this function
 700   // can rarely reach this point, but only one can win.
 701   markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
 702   if (res != markWord::encode(this)) {
 703     // This should be rare so log at the Info level when it happens.
 704     log_info(monitorinflation)("install_displaced_markword_in_object: "
 705                                "failed cas_set_mark: new_mark=" INTPTR_FORMAT
 706                                ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
 707                                dmw.value(), markWord::encode(this).value(),
 708                                res.value());
 709   }
 710 
 711   // Note: It does not matter which thread restored the header/dmw
 712   // into the object's header. The thread deflating the monitor just
 713   // wanted the object's header restored and it is. The threads that
 714   // detected a race with the deflation process also wanted the
 715   // object's header restored before they retry their operation and
 716   // because it is restored they will only retry once.
 717 }
 718 
 719 // Convert the fields used by is_busy() to a string that can be
 720 // used for diagnostic output.
 721 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
 722   ss->print("is_busy: waiters=%d"
 723             ", contentions=%d"
 724             ", owner=" INTPTR_FORMAT
 725             ", cxq=" PTR_FORMAT
 726             ", EntryList=" PTR_FORMAT,
 727             _waiters,
 728             (contentions() > 0 ? contentions() : 0),
 729             owner_is_DEFLATER_MARKER()
 730                 // We report null instead of DEFLATER_MARKER here because is_busy()
 731                 // ignores DEFLATER_MARKER values.
 732                 ? p2i(nullptr)
 733                 : p2i(owner_raw()),
 734             p2i(_cxq),
 735             p2i(_EntryList));
 736   return ss->base();
 737 }
 738 
 739 #define MAX_RECHECK_INTERVAL 1000
 740 
 741 void ObjectMonitor::EnterI(JavaThread* current) {
 742   assert(current->thread_state() == _thread_blocked, "invariant");
 743 
 744   // Try the lock - TATAS
 745   if (TryLock (current) > 0) {
 746     assert(_succ != current, "invariant");
 747     assert(owner_raw() == owner_for(current), "invariant");
 748     assert(_Responsible != current, "invariant");
 749     return;
 750   }
 751 
 752   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 753     // Cancelled the in-progress async deflation by changing owner from
 754     // DEFLATER_MARKER to current. As part of the contended enter protocol,
 755     // contentions was incremented to a positive value before EnterI()
 756     // was called and that prevents the deflater thread from winning the
 757     // last part of the 2-part async deflation protocol. After EnterI()
 758     // returns to enter(), contentions is decremented because the caller
 759     // now owns the monitor. We bump contentions an extra time here to
 760     // prevent the deflater thread from winning the last part of the
 761     // 2-part async deflation protocol after the regular decrement
 762     // occurs in enter(). The deflater thread will decrement contentions
 763     // after it recognizes that the async deflation was cancelled.
 764     add_to_contentions(1);
 765     assert(_succ != current, "invariant");
 766     assert(_Responsible != current, "invariant");
 767     return;
 768   }
 769 
 770   assert(InitDone, "Unexpectedly not initialized");
 771 
 772   // We try one round of spinning *before* enqueueing current.
 773   //
 774   // If the _owner is ready but OFFPROC we could use a YieldTo()
 775   // operation to donate the remainder of this thread's quantum
 776   // to the owner.  This has subtle but beneficial affinity
 777   // effects.
 778 
 779   if (TrySpin(current) > 0) {
 780     assert(owner_raw() == owner_for(current), "invariant");
 781     assert(_succ != current, "invariant");
 782     assert(_Responsible != current, "invariant");
 783     return;
 784   }
 785 
 786   // The Spin failed -- Enqueue and park the thread ...
 787   assert(_succ != current, "invariant");
 788   assert(owner_raw() != owner_for(current), "invariant");
 789   assert(_Responsible != current, "invariant");
 790 
 791   // Enqueue "current" on ObjectMonitor's _cxq.
 792   //
 793   // Node acts as a proxy for current.
 794   // As an aside, if were to ever rewrite the synchronization code mostly
 795   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 796   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 797   // as well as eliminate a subset of ABA issues.
 798   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 799 
 800   ObjectWaiter node(current);
 801   current->_ParkEvent->reset();
 802   node._prev   = (ObjectWaiter*) 0xBAD;
 803   node.TState  = ObjectWaiter::TS_CXQ;
 804 
 805   // Push "current" onto the front of the _cxq.
 806   // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
 807   // Note that spinning tends to reduce the rate at which threads
 808   // enqueue and dequeue on EntryList|cxq.
 809   ObjectWaiter* nxt;
 810   for (;;) {
 811     node._next = nxt = _cxq;
 812     if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
 813 
 814     // Interference - the CAS failed because _cxq changed.  Just retry.
 815     // As an optional optimization we retry the lock.
 816     if (TryLock (current) > 0) {
 817       assert(_succ != current, "invariant");
 818       assert(owner_raw() == owner_for(current), "invariant");
 819       assert(_Responsible != current, "invariant");
 820       return;
 821     }
 822   }
 823 
 824   // Check for cxq|EntryList edge transition to non-null.  This indicates
 825   // the onset of contention.  While contention persists exiting threads
 826   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 827   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 828   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 829   // arrange for one of the contending thread to use a timed park() operations
 830   // to detect and recover from the race.  (Stranding is form of progress failure
 831   // where the monitor is unlocked but all the contending threads remain parked).
 832   // That is, at least one of the contended threads will periodically poll _owner.
 833   // One of the contending threads will become the designated "Responsible" thread.
 834   // The Responsible thread uses a timed park instead of a normal indefinite park
 835   // operation -- it periodically wakes and checks for and recovers from potential
 836   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 837   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 838   // be responsible for a monitor.
 839   //
 840   // Currently, one of the contended threads takes on the added role of "Responsible".
 841   // A viable alternative would be to use a dedicated "stranding checker" thread
 842   // that periodically iterated over all the threads (or active monitors) and unparked
 843   // successors where there was risk of stranding.  This would help eliminate the
 844   // timer scalability issues we see on some platforms as we'd only have one thread
 845   // -- the checker -- parked on a timer.
 846 
 847   if (nxt == nullptr && _EntryList == nullptr) {
 848     // Try to assume the role of responsible thread for the monitor.
 849     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=current }
 850     Atomic::replace_if_null(&_Responsible, current);
 851   }
 852 
 853   // The lock might have been released while this thread was occupied queueing
 854   // itself onto _cxq.  To close the race and avoid "stranding" and
 855   // progress-liveness failure we must resample-retry _owner before parking.
 856   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 857   // In this case the ST-MEMBAR is accomplished with CAS().
 858   //
 859   // TODO: Defer all thread state transitions until park-time.
 860   // Since state transitions are heavy and inefficient we'd like
 861   // to defer the state transitions until absolutely necessary,
 862   // and in doing so avoid some transitions ...
 863 
 864   int nWakeups = 0;
 865   int recheckInterval = 1;
 866   bool do_timed_parked = false;
 867 
 868   ContinuationEntry* ce = current->last_continuation();
 869   if (ce != nullptr && ce->is_virtual_thread()) {
 870     do_timed_parked = true;
 871   }
 872 
 873   for (;;) {
 874 
 875     if (TryLock(current) > 0) break;
 876     assert(owner_raw() != owner_for(current), "invariant");
 877 
 878     // park self
 879     if (_Responsible == current || do_timed_parked) {
 880       current->_ParkEvent->park((jlong) recheckInterval);
 881       // Increase the recheckInterval, but clamp the value.
 882       recheckInterval *= 8;
 883       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 884         recheckInterval = MAX_RECHECK_INTERVAL;
 885       }
 886     } else {
 887       current->_ParkEvent->park();
 888     }
 889 
 890     if (TryLock(current) > 0) break;
 891 
 892     if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 893       // Cancelled the in-progress async deflation by changing owner from
 894       // DEFLATER_MARKER to current. As part of the contended enter protocol,
 895       // contentions was incremented to a positive value before EnterI()
 896       // was called and that prevents the deflater thread from winning the
 897       // last part of the 2-part async deflation protocol. After EnterI()
 898       // returns to enter(), contentions is decremented because the caller
 899       // now owns the monitor. We bump contentions an extra time here to
 900       // prevent the deflater thread from winning the last part of the
 901       // 2-part async deflation protocol after the regular decrement
 902       // occurs in enter(). The deflater thread will decrement contentions
 903       // after it recognizes that the async deflation was cancelled.
 904       add_to_contentions(1);
 905       break;
 906     }
 907 
 908     // The lock is still contested.
 909     // Keep a tally of the # of futile wakeups.
 910     // Note that the counter is not protected by a lock or updated by atomics.
 911     // That is by design - we trade "lossy" counters which are exposed to
 912     // races during updates for a lower probe effect.
 913 
 914     // This PerfData object can be used in parallel with a safepoint.
 915     // See the work around in PerfDataManager::destroy().
 916     OM_PERFDATA_OP(FutileWakeups, inc());
 917     ++nWakeups;
 918 
 919     // Assuming this is not a spurious wakeup we'll normally find _succ == current.
 920     // We can defer clearing _succ until after the spin completes
 921     // TrySpin() must tolerate being called with _succ == current.
 922     // Try yet another round of adaptive spinning.
 923     if (TrySpin(current) > 0) break;
 924 
 925     // We can find that we were unpark()ed and redesignated _succ while
 926     // we were spinning.  That's harmless.  If we iterate and call park(),
 927     // park() will consume the event and return immediately and we'll
 928     // just spin again.  This pattern can repeat, leaving _succ to simply
 929     // spin on a CPU.
 930 
 931     if (_succ == current) _succ = nullptr;
 932 
 933     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 934     OrderAccess::fence();
 935   }
 936 
 937   // Egress :
 938   // current has acquired the lock -- Unlink current from the cxq or EntryList.
 939   // Normally we'll find current on the EntryList .
 940   // From the perspective of the lock owner (this thread), the
 941   // EntryList is stable and cxq is prepend-only.
 942   // The head of cxq is volatile but the interior is stable.
 943   // In addition, current.TState is stable.
 944 
 945   assert(owner_raw() == owner_for(current), "invariant");
 946 
 947   UnlinkAfterAcquire(current, &node);
 948   if (_succ == current) _succ = nullptr;
 949 
 950   assert(_succ != current, "invariant");
 951   if (_Responsible == current) {
 952     _Responsible = nullptr;
 953     OrderAccess::fence(); // Dekker pivot-point
 954 
 955     // We may leave threads on cxq|EntryList without a designated
 956     // "Responsible" thread.  This is benign.  When this thread subsequently
 957     // exits the monitor it can "see" such preexisting "old" threads --
 958     // threads that arrived on the cxq|EntryList before the fence, above --
 959     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 960     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 961     // non-null and elect a new "Responsible" timer thread.
 962     //
 963     // This thread executes:
 964     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 965     //    LD cxq|EntryList               (in subsequent exit)
 966     //
 967     // Entering threads in the slow/contended path execute:
 968     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 969     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 970     //
 971     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 972     // exit operation from floating above the ST Responsible=null.
 973   }
 974 
 975   // We've acquired ownership with CAS().
 976   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 977   // But since the CAS() this thread may have also stored into _succ,
 978   // EntryList, cxq or Responsible.  These meta-data updates must be
 979   // visible __before this thread subsequently drops the lock.
 980   // Consider what could occur if we didn't enforce this constraint --
 981   // STs to monitor meta-data and user-data could reorder with (become
 982   // visible after) the ST in exit that drops ownership of the lock.
 983   // Some other thread could then acquire the lock, but observe inconsistent
 984   // or old monitor meta-data and heap data.  That violates the JMM.
 985   // To that end, the 1-0 exit() operation must have at least STST|LDST
 986   // "release" barrier semantics.  Specifically, there must be at least a
 987   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 988   // the lock.   The barrier ensures that changes to monitor meta-data and data
 989   // protected by the lock will be visible before we release the lock, and
 990   // therefore before some other thread (CPU) has a chance to acquire the lock.
 991   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 992   //
 993   // Critically, any prior STs to _succ or EntryList must be visible before
 994   // the ST of null into _owner in the *subsequent* (following) corresponding
 995   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 996   // execute a serializing instruction.
 997 
 998   return;
 999 }
1000 
1001 bool ObjectMonitor::HandlePreemptedVThread(JavaThread* current) {
1002   // Either because we acquire the lock below or because we will preempt the
1003   // vthread clear the _Stalled/_current_pending_monitor field from the current JavaThread.
1004   current->_Stalled = 0;
1005   current->set_current_pending_monitor(nullptr);
1006 
1007   // Try once more after freezing the continuation.
1008   if (TryLock (current) > 0) {
1009     assert(owner_raw() == owner_for(current), "invariant");
1010     assert(_succ != current, "invariant");
1011     assert(_Responsible != current, "invariant");
1012     current->set_preemption_cancelled(true);
1013     add_to_contentions(-1);
1014     return true;
1015   }
1016 
1017   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
1018     // Cancelled the in-progress async deflation by changing owner from
1019     // DEFLATER_MARKER to current. As part of the contended enter protocol,
1020     // contentions was incremented to a positive value before this call to
1021     // HandlePreemptedVThread(). We avoid decrementing contentions to
1022     // prevent the deflater thread from winning the last part of the
1023     // 2-part async deflation protocol. The deflater thread will decrement
1024     // contentions after it recognizes that the async deflation was cancelled.
1025     assert(_succ != current, "invariant");
1026     assert(_Responsible != current, "invariant");
1027     current->set_preemption_cancelled(true);
1028     return true;
1029   }
1030 
1031   oop vthread = current->vthread();
1032   assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
1033   java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING);
1034 
1035   ObjectWaiter* node = new ObjectWaiter(vthread);
1036   node->_prev   = (ObjectWaiter*) 0xBAD;
1037   node->TState  = ObjectWaiter::TS_CXQ;
1038 
1039   // Push node associated with vthread onto the front of the _cxq.
1040   ObjectWaiter* nxt;
1041   for (;;) {
1042     node->_next = nxt = _cxq;
1043     if (Atomic::cmpxchg(&_cxq, nxt, node) == nxt) break;
1044 
1045     // Interference - the CAS failed because _cxq changed.  Just retry.
1046     // As an optional optimization we retry the lock.
1047     if (TryLock (current) > 0) {
1048       assert(owner_raw() == owner_for(current), "invariant");
1049       assert(_succ != current, "invariant");
1050       assert(_Responsible != current, "invariant");
1051       current->set_preemption_cancelled(true);
1052       java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::RUNNING);
1053       add_to_contentions(-1);
1054       delete node;
1055       return true;
1056     }
1057   }
1058 
1059   // We have to try once more since owner could have exited monitor and checked
1060   // _cxq before we added the node to the queue.
1061   if (TryLock(current) > 0) {
1062     assert(owner_raw() == owner_for(current), "invariant");
1063     assert(_Responsible != current, "invariant");
1064     current->set_preemption_cancelled(true);
1065     java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::RUNNING);
1066     UnlinkAfterAcquire(current, node, vthread);
1067     delete node;
1068     if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr;
1069     add_to_contentions(-1);
1070     return true;
1071   }
1072 
1073   if (nxt == nullptr && _EntryList == nullptr) {
1074     // The C2 unlock() fast path first checks if _cxq and _EntryList are empty and
1075     // if they are it just clears the _owner field. Since we always run the risk of
1076     // having that check happening before we added the node to _cxq and the release
1077     // of the monitor happening after the last TryLock attempt we need to do something
1078     // to avoid stranding. We set the _Responsible field which results in a timed-wait.
1079     if (Atomic::replace_if_null(&_Responsible, (JavaThread*)java_lang_Thread::thread_id(vthread))) {
1080       java_lang_VirtualThread::set_recheckInterval(vthread, 1);
1081     }
1082   }
1083 
1084   return false;
1085 }
1086 
1087 // ReenterI() is a specialized inline form of the latter half of the
1088 // contended slow-path from EnterI().  We use ReenterI() only for
1089 // monitor reentry in wait().
1090 //
1091 // In the future we should reconcile EnterI() and ReenterI().
1092 
1093 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
1094   assert(current != nullptr, "invariant");
1095   assert(currentNode != nullptr, "invariant");
1096   assert(currentNode->_thread == current, "invariant");
1097   assert(_waiters > 0, "invariant");
1098   assert(object()->mark() == markWord::encode(this), "invariant");
1099 
1100   assert(current->thread_state() != _thread_blocked, "invariant");
1101 
1102   int nWakeups = 0;
1103   for (;;) {
1104     ObjectWaiter::TStates v = currentNode->TState;
1105     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1106     assert(owner_raw() != owner_for(current), "invariant");
1107 
1108     if (TryLock(current) > 0) break;
1109     if (TrySpin(current) > 0) break;
1110 
1111     {
1112       OSThreadContendState osts(current->osthread());
1113 
1114       assert(current->thread_state() == _thread_in_vm, "invariant");
1115 
1116       {
1117         ClearSuccOnSuspend csos(this);
1118         ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1119         current->_ParkEvent->park();
1120       }
1121     }
1122 
1123     // Try again, but just so we distinguish between futile wakeups and
1124     // successful wakeups.  The following test isn't algorithmically
1125     // necessary, but it helps us maintain sensible statistics.
1126     if (TryLock(current) > 0) break;
1127 
1128     // The lock is still contested.
1129     // Keep a tally of the # of futile wakeups.
1130     // Note that the counter is not protected by a lock or updated by atomics.
1131     // That is by design - we trade "lossy" counters which are exposed to
1132     // races during updates for a lower probe effect.
1133     ++nWakeups;
1134 
1135     // Assuming this is not a spurious wakeup we'll normally
1136     // find that _succ == current.
1137     if (_succ == current) _succ = nullptr;
1138 
1139     // Invariant: after clearing _succ a contending thread
1140     // *must* retry  _owner before parking.
1141     OrderAccess::fence();
1142 
1143     // This PerfData object can be used in parallel with a safepoint.
1144     // See the work around in PerfDataManager::destroy().
1145     OM_PERFDATA_OP(FutileWakeups, inc());
1146   }
1147 
1148   // current has acquired the lock -- Unlink current from the cxq or EntryList .
1149   // Normally we'll find current on the EntryList.
1150   // Unlinking from the EntryList is constant-time and atomic-free.
1151   // From the perspective of the lock owner (this thread), the
1152   // EntryList is stable and cxq is prepend-only.
1153   // The head of cxq is volatile but the interior is stable.
1154   // In addition, current.TState is stable.
1155 
1156   assert(owner_raw() == owner_for(current), "invariant");
1157   assert(object()->mark() == markWord::encode(this), "invariant");
1158   UnlinkAfterAcquire(current, currentNode);
1159   if (_succ == current) _succ = nullptr;
1160   assert(_succ != current, "invariant");
1161   currentNode->TState = ObjectWaiter::TS_RUN;
1162   OrderAccess::fence();      // see comments at the end of EnterI()
1163 }
1164 
1165 void ObjectMonitor::redo_enter(JavaThread* current) {
1166   assert(java_lang_VirtualThread::state(current->vthread()) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
1167   assert(current->is_in_VTMS_transition(), "must be");
1168 
1169   if (TryLock (current) > 0) {
1170     VThreadEpilog(current);
1171     return;
1172   }
1173 
1174   oop vthread = current->vthread();
1175   if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr;
1176 
1177   // Invariant: after clearing _succ a thread *must* retry _owner before parking.
1178   OrderAccess::fence();
1179 
1180   if (TryLock (current) > 0) {
1181     assert(owner_raw() == owner_for(current), "invariant");
1182     VThreadEpilog(current);
1183     return;
1184   }
1185 
1186   // Fast preemption. The JT will read this variable on return to the
1187   // monitorenter_redo stub and will just remove enterSpecial frame
1188   // from the stack and return to Continuation.run()
1189   current->set_preempting(true);
1190 
1191   java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING);
1192   if (_Responsible == (JavaThread*)java_lang_Thread::thread_id(vthread)) {
1193     int recheckInterval = java_lang_VirtualThread::recheckInterval(vthread);
1194     assert(recheckInterval >= 1 && recheckInterval <= 6, "invariant");
1195     if (recheckInterval < 6) {
1196       recheckInterval++;
1197       java_lang_VirtualThread::set_recheckInterval(vthread, recheckInterval);
1198     }
1199   } else if (java_lang_VirtualThread::recheckInterval(vthread) > 0) {
1200     // No need to do timed park anymore
1201     java_lang_VirtualThread::set_recheckInterval(vthread, 0);
1202   }
1203 }
1204 
1205 void ObjectMonitor::VThreadEpilog(JavaThread* current) {
1206   assert(owner_raw() == owner_for(current), "invariant");
1207   add_to_contentions(-1);
1208 
1209   oop vthread = current->vthread();
1210   if (java_lang_VirtualThread::recheckInterval(vthread) > 0) {
1211     java_lang_VirtualThread::set_recheckInterval(vthread, 0);
1212   }
1213   int64_t threadid = java_lang_Thread::thread_id(vthread);
1214   if (_succ == (JavaThread*)threadid) _succ = nullptr;
1215   if (_Responsible == (JavaThread*)threadid) {
1216     _Responsible = nullptr;
1217     OrderAccess::fence(); // Dekker pivot-point
1218   }
1219   ObjectWaiter* node = LookupWaiter(threadid);
1220   UnlinkAfterAcquire(current, node, vthread);
1221   delete node;
1222 }
1223 
1224 // By convention we unlink a contending thread from EntryList|cxq immediately
1225 // after the thread acquires the lock in ::enter().  Equally, we could defer
1226 // unlinking the thread until ::exit()-time.
1227 
1228 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode, oop vthread) {
1229   assert(owner_raw() == owner_for(current), "invariant");
1230   assert((currentNode->_thread == current) || (currentNode->_thread == nullptr && currentNode->vthread() == vthread), "invariant");
1231 
1232   if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1233     // Normal case: remove current from the DLL EntryList .
1234     // This is a constant-time operation.
1235     ObjectWaiter* nxt = currentNode->_next;
1236     ObjectWaiter* prv = currentNode->_prev;
1237     if (nxt != nullptr) nxt->_prev = prv;
1238     if (prv != nullptr) prv->_next = nxt;
1239     if (currentNode == _EntryList) _EntryList = nxt;
1240     assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1241     assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1242   } else {
1243     assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1244     // Inopportune interleaving -- current is still on the cxq.
1245     // This usually means the enqueue of self raced an exiting thread.
1246     // Normally we'll find current near the front of the cxq, so
1247     // dequeueing is typically fast.  If needbe we can accelerate
1248     // this with some MCS/CHL-like bidirectional list hints and advisory
1249     // back-links so dequeueing from the interior will normally operate
1250     // in constant-time.
1251     // Dequeue current from either the head (with CAS) or from the interior
1252     // with a linear-time scan and normal non-atomic memory operations.
1253     // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1254     // and then unlink current from EntryList.  We have to drain eventually,
1255     // so it might as well be now.
1256 
1257     ObjectWaiter* v = _cxq;
1258     assert(v != nullptr, "invariant");
1259     if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1260       // The CAS above can fail from interference IFF a "RAT" arrived.
1261       // In that case current must be in the interior and can no longer be
1262       // at the head of cxq.
1263       if (v == currentNode) {
1264         assert(_cxq != v, "invariant");
1265         v = _cxq;          // CAS above failed - start scan at head of list
1266       }
1267       ObjectWaiter* p;
1268       ObjectWaiter* q = nullptr;
1269       for (p = v; p != nullptr && p != currentNode; p = p->_next) {
1270         q = p;
1271         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1272       }
1273       assert(v != currentNode, "invariant");
1274       assert(p == currentNode, "Node not found on cxq");
1275       assert(p != _cxq, "invariant");
1276       assert(q != nullptr, "invariant");
1277       assert(q->_next == p, "invariant");
1278       q->_next = p->_next;
1279     }
1280   }
1281 
1282 #ifdef ASSERT
1283   // Diagnostic hygiene ...
1284   currentNode->_prev  = (ObjectWaiter*) 0xBAD;
1285   currentNode->_next  = (ObjectWaiter*) 0xBAD;
1286   currentNode->TState = ObjectWaiter::TS_RUN;
1287 #endif
1288 }
1289 
1290 // Fix this. Save ObjectWaiter* when freezing. Or use hashtable.
1291 ObjectWaiter* ObjectMonitor::LookupWaiter(int64_t threadid) {
1292   ObjectWaiter* p;
1293   for (p = _EntryList; p != nullptr && (!p->is_vthread() || java_lang_Thread::thread_id(p->vthread()) != threadid); p = p->_next) {}
1294   if (p != nullptr) return p;
1295   for (p = _cxq; p != nullptr && (!p->is_vthread() || java_lang_Thread::thread_id(p->vthread()) != threadid); p = p->_next) {}
1296   assert(p != nullptr, "should be on either _cxq or _EntryList");
1297   return p;
1298 }
1299 
1300 // -----------------------------------------------------------------------------
1301 // Exit support
1302 //
1303 // exit()
1304 // ~~~~~~
1305 // Note that the collector can't reclaim the objectMonitor or deflate
1306 // the object out from underneath the thread calling ::exit() as the
1307 // thread calling ::exit() never transitions to a stable state.
1308 // This inhibits GC, which in turn inhibits asynchronous (and
1309 // inopportune) reclamation of "this".
1310 //
1311 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1312 // There's one exception to the claim above, however.  EnterI() can call
1313 // exit() to drop a lock if the acquirer has been externally suspended.
1314 // In that case exit() is called with _thread_state == _thread_blocked,
1315 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1316 //
1317 // 1-0 exit
1318 // ~~~~~~~~
1319 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1320 // the fast-path operators have been optimized so the common ::exit()
1321 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1322 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
1323 // greatly improves latency -- MEMBAR and CAS having considerable local
1324 // latency on modern processors -- but at the cost of "stranding".  Absent the
1325 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1326 // ::enter() path, resulting in the entering thread being stranding
1327 // and a progress-liveness failure.   Stranding is extremely rare.
1328 // We use timers (timed park operations) & periodic polling to detect
1329 // and recover from stranding.  Potentially stranded threads periodically
1330 // wake up and poll the lock.  See the usage of the _Responsible variable.
1331 //
1332 // The CAS() in enter provides for safety and exclusion, while the CAS or
1333 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
1334 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1335 // We detect and recover from stranding with timers.
1336 //
1337 // If a thread transiently strands it'll park until (a) another
1338 // thread acquires the lock and then drops the lock, at which time the
1339 // exiting thread will notice and unpark the stranded thread, or, (b)
1340 // the timer expires.  If the lock is high traffic then the stranding latency
1341 // will be low due to (a).  If the lock is low traffic then the odds of
1342 // stranding are lower, although the worst-case stranding latency
1343 // is longer.  Critically, we don't want to put excessive load in the
1344 // platform's timer subsystem.  We want to minimize both the timer injection
1345 // rate (timers created/sec) as well as the number of timers active at
1346 // any one time.  (more precisely, we want to minimize timer-seconds, which is
1347 // the integral of the # of active timers at any instant over time).
1348 // Both impinge on OS scalability.  Given that, at most one thread parked on
1349 // a monitor will use a timer.
1350 //
1351 // There is also the risk of a futile wake-up. If we drop the lock
1352 // another thread can reacquire the lock immediately, and we can
1353 // then wake a thread unnecessarily. This is benign, and we've
1354 // structured the code so the windows are short and the frequency
1355 // of such futile wakups is low.
1356 
1357 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1358   void* cur = owner_raw();
1359   if (owner_for(current) != cur) {
1360     // Apparent unbalanced locking ...
1361     // Naively we'd like to throw IllegalMonitorStateException.
1362     // As a practical matter we can neither allocate nor throw an
1363     // exception as ::exit() can be called from leaf routines.
1364     // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1365     // Upon deeper reflection, however, in a properly run JVM the only
1366     // way we should encounter this situation is in the presence of
1367     // unbalanced JNI locking. TODO: CheckJNICalls.
1368     // See also: CR4414101
1369 #ifdef ASSERT
1370     LogStreamHandle(Error, monitorinflation) lsh;
1371     lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1372                   " is exiting an ObjectMonitor it does not own.", p2i(current));
1373     lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1374     print_debug_style_on(&lsh);
1375     assert(false, "Non-balanced monitor enter/exit!");
1376 #endif
1377     return;
1378   }
1379 
1380   if (_recursions != 0) {
1381     _recursions--;        // this is simple recursive enter
1382     return;
1383   }
1384 
1385   // Invariant: after setting Responsible=null an thread must execute
1386   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1387   _Responsible = nullptr;
1388 
1389 #if INCLUDE_JFR
1390   // get the owner's thread id for the MonitorEnter event
1391   // if it is enabled and the thread isn't suspended
1392   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1393     _previous_owner_tid = JFR_THREAD_ID(current);
1394   }
1395 #endif
1396 
1397   for (;;) {
1398     assert(owner_for(current) == owner_raw(), "invariant");
1399 
1400     // Drop the lock.
1401     // release semantics: prior loads and stores from within the critical section
1402     // must not float (reorder) past the following store that drops the lock.
1403     // Uses a storeload to separate release_store(owner) from the
1404     // successor check. The try_set_owner_from() below uses cmpxchg() so
1405     // we get the fence down there.
1406     release_clear_owner(current);
1407     OrderAccess::storeload();
1408 
1409     if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) {
1410       return;
1411     }
1412     // Other threads are blocked trying to acquire the lock.
1413 
1414     // Normally the exiting thread is responsible for ensuring succession,
1415     // but if other successors are ready or other entering threads are spinning
1416     // then this thread can simply store null into _owner and exit without
1417     // waking a successor.  The existence of spinners or ready successors
1418     // guarantees proper succession (liveness).  Responsibility passes to the
1419     // ready or running successors.  The exiting thread delegates the duty.
1420     // More precisely, if a successor already exists this thread is absolved
1421     // of the responsibility of waking (unparking) one.
1422     //
1423     // The _succ variable is critical to reducing futile wakeup frequency.
1424     // _succ identifies the "heir presumptive" thread that has been made
1425     // ready (unparked) but that has not yet run.  We need only one such
1426     // successor thread to guarantee progress.
1427     // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1428     // section 3.3 "Futile Wakeup Throttling" for details.
1429     //
1430     // Note that spinners in Enter() also set _succ non-null.
1431     // In the current implementation spinners opportunistically set
1432     // _succ so that exiting threads might avoid waking a successor.
1433     // Another less appealing alternative would be for the exiting thread
1434     // to drop the lock and then spin briefly to see if a spinner managed
1435     // to acquire the lock.  If so, the exiting thread could exit
1436     // immediately without waking a successor, otherwise the exiting
1437     // thread would need to dequeue and wake a successor.
1438     // (Note that we'd need to make the post-drop spin short, but no
1439     // shorter than the worst-case round-trip cache-line migration time.
1440     // The dropped lock needs to become visible to the spinner, and then
1441     // the acquisition of the lock by the spinner must become visible to
1442     // the exiting thread).
1443 
1444     // It appears that an heir-presumptive (successor) must be made ready.
1445     // Only the current lock owner can manipulate the EntryList or
1446     // drain _cxq, so we need to reacquire the lock.  If we fail
1447     // to reacquire the lock the responsibility for ensuring succession
1448     // falls to the new owner.
1449     //
1450     if (try_set_owner_from(nullptr, current) != nullptr) {
1451       return;
1452     }
1453 
1454     guarantee(owner_raw() == owner_for(current), "invariant");
1455 
1456     ObjectWaiter* w = nullptr;
1457 
1458     w = _EntryList;
1459     if (w != nullptr) {
1460       // I'd like to write: guarantee (w->_thread != current).
1461       // But in practice an exiting thread may find itself on the EntryList.
1462       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1463       // then calls exit().  Exit release the lock by setting O._owner to null.
1464       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1465       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1466       // release the lock "O".  T2 resumes immediately after the ST of null into
1467       // _owner, above.  T2 notices that the EntryList is populated, so it
1468       // reacquires the lock and then finds itself on the EntryList.
1469       // Given all that, we have to tolerate the circumstance where "w" is
1470       // associated with current.
1471       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1472       ExitEpilog(current, w);
1473       return;
1474     }
1475 
1476     // If we find that both _cxq and EntryList are null then just
1477     // re-run the exit protocol from the top.
1478     w = _cxq;
1479     if (w == nullptr) continue;
1480 
1481     // Drain _cxq into EntryList - bulk transfer.
1482     // First, detach _cxq.
1483     // The following loop is tantamount to: w = swap(&cxq, nullptr)
1484     for (;;) {
1485       assert(w != nullptr, "Invariant");
1486       ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr);
1487       if (u == w) break;
1488       w = u;
1489     }
1490 
1491     assert(w != nullptr, "invariant");
1492     assert(_EntryList == nullptr, "invariant");
1493 
1494     // Convert the LIFO SLL anchored by _cxq into a DLL.
1495     // The list reorganization step operates in O(LENGTH(w)) time.
1496     // It's critical that this step operate quickly as
1497     // "current" still holds the outer-lock, restricting parallelism
1498     // and effectively lengthening the critical section.
1499     // Invariant: s chases t chases u.
1500     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1501     // we have faster access to the tail.
1502 
1503     _EntryList = w;
1504     ObjectWaiter* q = nullptr;
1505     ObjectWaiter* p;
1506     for (p = w; p != nullptr; p = p->_next) {
1507       guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1508       p->TState = ObjectWaiter::TS_ENTER;
1509       p->_prev = q;
1510       q = p;
1511     }
1512 
1513     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr
1514     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1515 
1516     // See if we can abdicate to a spinner instead of waking a thread.
1517     // A primary goal of the implementation is to reduce the
1518     // context-switch rate.
1519     if (_succ != nullptr) continue;
1520 
1521     w = _EntryList;
1522     if (w != nullptr) {
1523       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1524       ExitEpilog(current, w);
1525       return;
1526     }
1527   }
1528 }
1529 
1530 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1531   assert(owner_raw() == owner_for(current), "invariant");
1532 
1533   // Exit protocol:
1534   // 1. ST _succ = wakee
1535   // 2. membar #loadstore|#storestore;
1536   // 2. ST _owner = nullptr
1537   // 3. unpark(wakee)
1538 
1539   oop vthread = nullptr;
1540   if (Wakee->_thread != nullptr) {
1541     // Platform thread case
1542     _succ = Wakee->_thread;
1543   } else {
1544     assert(Wakee->vthread() != nullptr, "invariant");
1545     vthread = Wakee->vthread();
1546     _succ = (JavaThread*)java_lang_Thread::thread_id(vthread);
1547   }
1548   ParkEvent * Trigger = Wakee->_event;
1549 
1550   // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again.
1551   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1552   // out-of-scope (non-extant).
1553   Wakee  = nullptr;
1554 
1555   // Drop the lock.
1556   // Uses a fence to separate release_store(owner) from the LD in unpark().
1557   release_clear_owner(current);
1558   OrderAccess::fence();
1559 
1560   DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1561 
1562   if (vthread == nullptr) {
1563     // Platform thread case
1564     Trigger->unpark();
1565   } else if (java_lang_VirtualThread::set_onWaitingList(vthread, _vthread_cxq_head)) {
1566     Trigger->unpark();
1567   }
1568 
1569   // Maintain stats and report events to JVMTI
1570   OM_PERFDATA_OP(Parks, inc());
1571 }
1572 
1573 // complete_exit exits a lock returning recursion count
1574 // complete_exit requires an inflated monitor
1575 // The _owner field is not always the Thread addr even with an
1576 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1577 // thread due to contention.
1578 intx ObjectMonitor::complete_exit(JavaThread* current) {
1579   assert(InitDone, "Unexpectedly not initialized");
1580 
1581   void* cur = owner_raw();
1582   if (owner_for(current) != cur) {
1583     if (LockingMode == LM_LEGACY && is_stack_locker(current)) {
1584       assert(_recursions == 0, "internal state error");
1585       set_owner_from_BasicLock(current);  // Convert from BasicLock* to Thread*.
1586       _recursions = 0;
1587     }
1588   }
1589 
1590   guarantee(owner_for(current) == owner_raw(), "complete_exit not owner");
1591   intx save = _recursions; // record the old recursion count
1592   _recursions = 0;         // set the recursion level to be 0
1593   exit(current);           // exit the monitor
1594   guarantee(owner_raw() != owner_for(current), "invariant");
1595   return save;
1596 }
1597 
1598 // Checks that the current THREAD owns this monitor and causes an
1599 // immediate return if it doesn't. We don't use the CHECK macro
1600 // because we want the IMSE to be the only exception that is thrown
1601 // from the call site when false is returned. Any other pending
1602 // exception is ignored.
1603 #define CHECK_OWNER()                                                  \
1604   do {                                                                 \
1605     if (!check_owner(THREAD)) {                                        \
1606        assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1607        return;                                                         \
1608      }                                                                 \
1609   } while (false)
1610 
1611 // Returns true if the specified thread owns the ObjectMonitor.
1612 // Otherwise returns false and throws IllegalMonitorStateException
1613 // (IMSE). If there is a pending exception and the specified thread
1614 // is not the owner, that exception will be replaced by the IMSE.
1615 bool ObjectMonitor::check_owner(TRAPS) {
1616   JavaThread* current = THREAD;
1617   void* cur = owner_raw();
1618   if (cur == owner_for(current)) {
1619     return true;
1620   }
1621   THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1622              "current thread is not owner", false);
1623 }
1624 
1625 static inline bool is_excluded(const Klass* monitor_klass) {
1626   assert(monitor_klass != nullptr, "invariant");
1627   NOT_JFR_RETURN_(false);
1628   JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name();)
1629 }
1630 
1631 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1632                                     ObjectMonitor* monitor,
1633                                     uint64_t notifier_tid,
1634                                     jlong timeout,
1635                                     bool timedout) {
1636   assert(event != nullptr, "invariant");
1637   assert(monitor != nullptr, "invariant");
1638   const Klass* monitor_klass = monitor->object()->klass();
1639   if (is_excluded(monitor_klass)) {
1640     return;
1641   }
1642   event->set_monitorClass(monitor_klass);
1643   event->set_timeout(timeout);
1644   // Set an address that is 'unique enough', such that events close in
1645   // time and with the same address are likely (but not guaranteed) to
1646   // belong to the same object.
1647   event->set_address((uintptr_t)monitor);
1648   event->set_notifier(notifier_tid);
1649   event->set_timedOut(timedout);
1650   event->commit();
1651 }
1652 
1653 // -----------------------------------------------------------------------------
1654 // Wait/Notify/NotifyAll
1655 //
1656 // Note: a subset of changes to ObjectMonitor::wait()
1657 // will need to be replicated in complete_exit
1658 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1659   JavaThread* current = THREAD;
1660 
1661   assert(InitDone, "Unexpectedly not initialized");
1662 
1663   CHECK_OWNER();  // Throws IMSE if not owner.
1664 
1665   ContinuationEntry* ce = current->last_continuation();
1666   if (ce != nullptr && ce->is_virtual_thread()) {
1667     EventVirtualThreadPinned e;
1668     if (e.should_commit()) {
1669       e.commit();
1670     }
1671   }
1672 
1673   EventJavaMonitorWait event;
1674 
1675   // check for a pending interrupt
1676   if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1677     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1678     if (JvmtiExport::should_post_monitor_waited()) {
1679       // Note: 'false' parameter is passed here because the
1680       // wait was not timed out due to thread interrupt.
1681       JvmtiExport::post_monitor_waited(current, this, false);
1682 
1683       // In this short circuit of the monitor wait protocol, the
1684       // current thread never drops ownership of the monitor and
1685       // never gets added to the wait queue so the current thread
1686       // cannot be made the successor. This means that the
1687       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1688       // consume an unpark() meant for the ParkEvent associated with
1689       // this ObjectMonitor.
1690     }
1691     if (event.should_commit()) {
1692       post_monitor_wait_event(&event, this, 0, millis, false);
1693     }
1694     THROW(vmSymbols::java_lang_InterruptedException());
1695     return;
1696   }
1697 
1698   assert(current->_Stalled == 0, "invariant");
1699   current->_Stalled = intptr_t(this);
1700   current->set_current_waiting_monitor(this);
1701 
1702   // create a node to be put into the queue
1703   // Critically, after we reset() the event but prior to park(), we must check
1704   // for a pending interrupt.
1705   ObjectWaiter node(current);
1706   node.TState = ObjectWaiter::TS_WAIT;
1707   current->_ParkEvent->reset();
1708   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1709 
1710   // Enter the waiting queue, which is a circular doubly linked list in this case
1711   // but it could be a priority queue or any data structure.
1712   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1713   // by the owner of the monitor *except* in the case where park()
1714   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1715   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1716 
1717   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1718   AddWaiter(&node);
1719   Thread::SpinRelease(&_WaitSetLock);
1720 
1721   _Responsible = nullptr;
1722 
1723   intx save = _recursions;     // record the old recursion count
1724   _waiters++;                  // increment the number of waiters
1725   _recursions = 0;             // set the recursion level to be 1
1726   exit(current);               // exit the monitor
1727   guarantee(owner_raw() != owner_for(current), "invariant");
1728 
1729   // The thread is on the WaitSet list - now park() it.
1730   // On MP systems it's conceivable that a brief spin before we park
1731   // could be profitable.
1732   //
1733   // TODO-FIXME: change the following logic to a loop of the form
1734   //   while (!timeout && !interrupted && _notified == 0) park()
1735 
1736   int ret = OS_OK;
1737   int WasNotified = 0;
1738 
1739   // Need to check interrupt state whilst still _thread_in_vm
1740   bool interrupted = interruptible && current->is_interrupted(false);
1741 
1742   { // State transition wrappers
1743     OSThread* osthread = current->osthread();
1744     OSThreadWaitState osts(osthread, true);
1745 
1746     assert(current->thread_state() == _thread_in_vm, "invariant");
1747 
1748     {
1749       ClearSuccOnSuspend csos(this);
1750       ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1751       if (interrupted || HAS_PENDING_EXCEPTION) {
1752         // Intentionally empty
1753       } else if (node._notified == 0) {
1754         if (millis <= 0) {
1755           current->_ParkEvent->park();
1756         } else {
1757           ret = current->_ParkEvent->park(millis);
1758         }
1759       }
1760     }
1761 
1762     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1763     // from the WaitSet to the EntryList.
1764     // See if we need to remove Node from the WaitSet.
1765     // We use double-checked locking to avoid grabbing _WaitSetLock
1766     // if the thread is not on the wait queue.
1767     //
1768     // Note that we don't need a fence before the fetch of TState.
1769     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1770     // written by the is thread. (perhaps the fetch might even be satisfied
1771     // by a look-aside into the processor's own store buffer, although given
1772     // the length of the code path between the prior ST and this load that's
1773     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1774     // then we'll acquire the lock and then re-fetch a fresh TState value.
1775     // That is, we fail toward safety.
1776 
1777     if (node.TState == ObjectWaiter::TS_WAIT) {
1778       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1779       if (node.TState == ObjectWaiter::TS_WAIT) {
1780         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1781         assert(node._notified == 0, "invariant");
1782         node.TState = ObjectWaiter::TS_RUN;
1783       }
1784       Thread::SpinRelease(&_WaitSetLock);
1785     }
1786 
1787     // The thread is now either on off-list (TS_RUN),
1788     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1789     // The Node's TState variable is stable from the perspective of this thread.
1790     // No other threads will asynchronously modify TState.
1791     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1792     OrderAccess::loadload();
1793     if (_succ == current) _succ = nullptr;
1794     WasNotified = node._notified;
1795 
1796     // Reentry phase -- reacquire the monitor.
1797     // re-enter contended monitor after object.wait().
1798     // retain OBJECT_WAIT state until re-enter successfully completes
1799     // Thread state is thread_in_vm and oop access is again safe,
1800     // although the raw address of the object may have changed.
1801     // (Don't cache naked oops over safepoints, of course).
1802 
1803     // post monitor waited event. Note that this is past-tense, we are done waiting.
1804     if (JvmtiExport::should_post_monitor_waited()) {
1805       JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1806 
1807       if (node._notified != 0 && _succ == current) {
1808         // In this part of the monitor wait-notify-reenter protocol it
1809         // is possible (and normal) for another thread to do a fastpath
1810         // monitor enter-exit while this thread is still trying to get
1811         // to the reenter portion of the protocol.
1812         //
1813         // The ObjectMonitor was notified and the current thread is
1814         // the successor which also means that an unpark() has already
1815         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1816         // consume the unpark() that was done when the successor was
1817         // set because the same ParkEvent is shared between Java
1818         // monitors and JVM/TI RawMonitors (for now).
1819         //
1820         // We redo the unpark() to ensure forward progress, i.e., we
1821         // don't want all pending threads hanging (parked) with none
1822         // entering the unlocked monitor.
1823         node._event->unpark();
1824       }
1825     }
1826 
1827     if (event.should_commit()) {
1828       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1829     }
1830 
1831     OrderAccess::fence();
1832 
1833     assert(current->_Stalled != 0, "invariant");
1834     current->_Stalled = 0;
1835 
1836     assert(owner_raw() != owner_for(current), "invariant");
1837     ObjectWaiter::TStates v = node.TState;
1838     if (v == ObjectWaiter::TS_RUN) {
1839       enter(current);
1840     } else {
1841       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1842       ReenterI(current, &node);
1843       node.wait_reenter_end(this);
1844     }
1845 
1846     // current has reacquired the lock.
1847     // Lifecycle - the node representing current must not appear on any queues.
1848     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1849     // want residual elements associated with this thread left on any lists.
1850     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1851     assert(owner_raw() == owner_for(current), "invariant");
1852     assert(_succ != current, "invariant");
1853   } // OSThreadWaitState()
1854 
1855   current->set_current_waiting_monitor(nullptr);
1856 
1857   guarantee(_recursions == 0, "invariant");
1858   int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current);
1859   _recursions =   save          // restore the old recursion count
1860                 + relock_count; //  increased by the deferred relock count
1861   NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count(relock_count);) // Deopt never entered these counts.
1862   _waiters--;             // decrement the number of waiters
1863 
1864   // Verify a few postconditions
1865   assert(owner_raw() == owner_for(current), "invariant");
1866   assert(_succ != current, "invariant");
1867   assert(object()->mark() == markWord::encode(this), "invariant");
1868 
1869   // check if the notification happened
1870   if (!WasNotified) {
1871     // no, it could be timeout or Thread.interrupt() or both
1872     // check for interrupt event, otherwise it is timeout
1873     if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1874       THROW(vmSymbols::java_lang_InterruptedException());
1875     }
1876   }
1877 
1878   // NOTE: Spurious wake up will be consider as timeout.
1879   // Monitor notify has precedence over thread interrupt.
1880 }
1881 
1882 
1883 // Consider:
1884 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1885 // then instead of transferring a thread from the WaitSet to the EntryList
1886 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1887 
1888 void ObjectMonitor::INotify(JavaThread* current) {
1889   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1890   ObjectWaiter* iterator = DequeueWaiter();
1891   if (iterator != nullptr) {
1892     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1893     guarantee(iterator->_notified == 0, "invariant");
1894     // Disposition - what might we do with iterator ?
1895     // a.  add it directly to the EntryList - either tail (policy == 1)
1896     //     or head (policy == 0).
1897     // b.  push it onto the front of the _cxq (policy == 2).
1898     // For now we use (b).
1899 
1900     iterator->TState = ObjectWaiter::TS_ENTER;
1901 
1902     iterator->_notified = 1;
1903     iterator->_notifier_tid = JFR_THREAD_ID(current);
1904 
1905     ObjectWaiter* list = _EntryList;
1906     if (list != nullptr) {
1907       assert(list->_prev == nullptr, "invariant");
1908       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1909       assert(list != iterator, "invariant");
1910     }
1911 
1912     // prepend to cxq
1913     if (list == nullptr) {
1914       iterator->_next = iterator->_prev = nullptr;
1915       _EntryList = iterator;
1916     } else {
1917       iterator->TState = ObjectWaiter::TS_CXQ;
1918       for (;;) {
1919         ObjectWaiter* front = _cxq;
1920         iterator->_next = front;
1921         if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1922           break;
1923         }
1924       }
1925     }
1926 
1927     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1928     // move the add-to-EntryList operation, above, outside the critical section
1929     // protected by _WaitSetLock.  In practice that's not useful.  With the
1930     // exception of  wait() timeouts and interrupts the monitor owner
1931     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1932     // on _WaitSetLock so it's not profitable to reduce the length of the
1933     // critical section.
1934 
1935     iterator->wait_reenter_begin(this);
1936   }
1937   Thread::SpinRelease(&_WaitSetLock);
1938 }
1939 
1940 // Consider: a not-uncommon synchronization bug is to use notify() when
1941 // notifyAll() is more appropriate, potentially resulting in stranded
1942 // threads; this is one example of a lost wakeup. A useful diagnostic
1943 // option is to force all notify() operations to behave as notifyAll().
1944 //
1945 // Note: We can also detect many such problems with a "minimum wait".
1946 // When the "minimum wait" is set to a small non-zero timeout value
1947 // and the program does not hang whereas it did absent "minimum wait",
1948 // that suggests a lost wakeup bug.
1949 
1950 void ObjectMonitor::notify(TRAPS) {
1951   JavaThread* current = THREAD;
1952   CHECK_OWNER();  // Throws IMSE if not owner.
1953   if (_WaitSet == nullptr) {
1954     return;
1955   }
1956   DTRACE_MONITOR_PROBE(notify, this, object(), current);
1957   INotify(current);
1958   OM_PERFDATA_OP(Notifications, inc(1));
1959 }
1960 
1961 
1962 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1963 // from the waitset to the EntryList. This could be done more efficiently with a
1964 // single bulk transfer but in practice it's not time-critical. Beware too,
1965 // that in prepend-mode we invert the order of the waiters. Let's say that the
1966 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1967 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1968 
1969 void ObjectMonitor::notifyAll(TRAPS) {
1970   JavaThread* current = THREAD;
1971   CHECK_OWNER();  // Throws IMSE if not owner.
1972   if (_WaitSet == nullptr) {
1973     return;
1974   }
1975 
1976   DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1977   int tally = 0;
1978   while (_WaitSet != nullptr) {
1979     tally++;
1980     INotify(current);
1981   }
1982 
1983   OM_PERFDATA_OP(Notifications, inc(tally));
1984 }
1985 
1986 // -----------------------------------------------------------------------------
1987 // Adaptive Spinning Support
1988 //
1989 // Adaptive spin-then-block - rational spinning
1990 //
1991 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1992 // algorithm.  On high order SMP systems it would be better to start with
1993 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1994 // a contending thread could enqueue itself on the cxq and then spin locally
1995 // on a thread-specific variable such as its ParkEvent._Event flag.
1996 // That's left as an exercise for the reader.  Note that global spinning is
1997 // not problematic on Niagara, as the L2 cache serves the interconnect and
1998 // has both low latency and massive bandwidth.
1999 //
2000 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
2001 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
2002 // (duration) or we can fix the count at approximately the duration of
2003 // a context switch and vary the frequency.   Of course we could also
2004 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
2005 // For a description of 'Adaptive spin-then-block mutual exclusion in
2006 // multi-threaded processing,' see U.S. Pat. No. 8046758.
2007 //
2008 // This implementation varies the duration "D", where D varies with
2009 // the success rate of recent spin attempts. (D is capped at approximately
2010 // length of a round-trip context switch).  The success rate for recent
2011 // spin attempts is a good predictor of the success rate of future spin
2012 // attempts.  The mechanism adapts automatically to varying critical
2013 // section length (lock modality), system load and degree of parallelism.
2014 // D is maintained per-monitor in _SpinDuration and is initialized
2015 // optimistically.  Spin frequency is fixed at 100%.
2016 //
2017 // Note that _SpinDuration is volatile, but we update it without locks
2018 // or atomics.  The code is designed so that _SpinDuration stays within
2019 // a reasonable range even in the presence of races.  The arithmetic
2020 // operations on _SpinDuration are closed over the domain of legal values,
2021 // so at worst a race will install and older but still legal value.
2022 // At the very worst this introduces some apparent non-determinism.
2023 // We might spin when we shouldn't or vice-versa, but since the spin
2024 // count are relatively short, even in the worst case, the effect is harmless.
2025 //
2026 // Care must be taken that a low "D" value does not become an
2027 // an absorbing state.  Transient spinning failures -- when spinning
2028 // is overall profitable -- should not cause the system to converge
2029 // on low "D" values.  We want spinning to be stable and predictable
2030 // and fairly responsive to change and at the same time we don't want
2031 // it to oscillate, become metastable, be "too" non-deterministic,
2032 // or converge on or enter undesirable stable absorbing states.
2033 //
2034 // We implement a feedback-based control system -- using past behavior
2035 // to predict future behavior.  We face two issues: (a) if the
2036 // input signal is random then the spin predictor won't provide optimal
2037 // results, and (b) if the signal frequency is too high then the control
2038 // system, which has some natural response lag, will "chase" the signal.
2039 // (b) can arise from multimodal lock hold times.  Transient preemption
2040 // can also result in apparent bimodal lock hold times.
2041 // Although sub-optimal, neither condition is particularly harmful, as
2042 // in the worst-case we'll spin when we shouldn't or vice-versa.
2043 // The maximum spin duration is rather short so the failure modes aren't bad.
2044 // To be conservative, I've tuned the gain in system to bias toward
2045 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
2046 // "rings" or oscillates between spinning and not spinning.  This happens
2047 // when spinning is just on the cusp of profitability, however, so the
2048 // situation is not dire.  The state is benign -- there's no need to add
2049 // hysteresis control to damp the transition rate between spinning and
2050 // not spinning.
2051 
2052 // Spinning: Fixed frequency (100%), vary duration
2053 int ObjectMonitor::TrySpin(JavaThread* current) {
2054   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
2055   int ctr = Knob_FixedSpin;
2056   if (ctr != 0) {
2057     while (--ctr >= 0) {
2058       if (TryLock(current) > 0) return 1;
2059       SpinPause();
2060     }
2061     return 0;
2062   }
2063 
2064   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
2065     if (TryLock(current) > 0) {
2066       // Increase _SpinDuration ...
2067       // Note that we don't clamp SpinDuration precisely at SpinLimit.
2068       // Raising _SpurDuration to the poverty line is key.
2069       int x = _SpinDuration;
2070       if (x < Knob_SpinLimit) {
2071         if (x < Knob_Poverty) x = Knob_Poverty;
2072         _SpinDuration = x + Knob_BonusB;
2073       }
2074       return 1;
2075     }
2076     SpinPause();
2077   }
2078 
2079   // Admission control - verify preconditions for spinning
2080   //
2081   // We always spin a little bit, just to prevent _SpinDuration == 0 from
2082   // becoming an absorbing state.  Put another way, we spin briefly to
2083   // sample, just in case the system load, parallelism, contention, or lock
2084   // modality changed.
2085   //
2086   // Consider the following alternative:
2087   // Periodically set _SpinDuration = _SpinLimit and try a long/full
2088   // spin attempt.  "Periodically" might mean after a tally of
2089   // the # of failed spin attempts (or iterations) reaches some threshold.
2090   // This takes us into the realm of 1-out-of-N spinning, where we
2091   // hold the duration constant but vary the frequency.
2092 
2093   ctr = _SpinDuration;
2094   if (ctr <= 0) return 0;
2095 
2096   // We're good to spin ... spin ingress.
2097   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
2098   // when preparing to LD...CAS _owner, etc and the CAS is likely
2099   // to succeed.
2100   if (_succ == nullptr) {
2101     _succ = current;
2102   }
2103   void* prv = nullptr;
2104 
2105   // There are three ways to exit the following loop:
2106   // 1.  A successful spin where this thread has acquired the lock.
2107   // 2.  Spin failure with prejudice
2108   // 3.  Spin failure without prejudice
2109 
2110   while (--ctr >= 0) {
2111 
2112     // Periodic polling -- Check for pending GC
2113     // Threads may spin while they're unsafe.
2114     // We don't want spinning threads to delay the JVM from reaching
2115     // a stop-the-world safepoint or to steal cycles from GC.
2116     // If we detect a pending safepoint we abort in order that
2117     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
2118     // this thread, if safe, doesn't steal cycles from GC.
2119     // This is in keeping with the "no loitering in runtime" rule.
2120     // We periodically check to see if there's a safepoint pending.
2121     if ((ctr & 0xFF) == 0) {
2122       // Can't call SafepointMechanism::should_process() since that
2123       // might update the poll values and we could be in a thread_blocked
2124       // state here which is not allowed so just check the poll.
2125       if (SafepointMechanism::local_poll_armed(current)) {
2126         goto Abort;           // abrupt spin egress
2127       }
2128       SpinPause();
2129     }
2130 
2131     // Probe _owner with TATAS
2132     // If this thread observes the monitor transition or flicker
2133     // from locked to unlocked to locked, then the odds that this
2134     // thread will acquire the lock in this spin attempt go down
2135     // considerably.  The same argument applies if the CAS fails
2136     // or if we observe _owner change from one non-null value to
2137     // another non-null value.   In such cases we might abort
2138     // the spin without prejudice or apply a "penalty" to the
2139     // spin count-down variable "ctr", reducing it by 100, say.
2140 
2141     void* ox = owner_raw();
2142     if (ox == nullptr) {
2143       ox = try_set_owner_from(nullptr, current);
2144       if (ox == nullptr) {
2145         // The CAS succeeded -- this thread acquired ownership
2146         // Take care of some bookkeeping to exit spin state.
2147         if (_succ == current) {
2148           _succ = nullptr;
2149         }
2150 
2151         // Increase _SpinDuration :
2152         // The spin was successful (profitable) so we tend toward
2153         // longer spin attempts in the future.
2154         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2155         // If we acquired the lock early in the spin cycle it
2156         // makes sense to increase _SpinDuration proportionally.
2157         // Note that we don't clamp SpinDuration precisely at SpinLimit.
2158         int x = _SpinDuration;
2159         if (x < Knob_SpinLimit) {
2160           if (x < Knob_Poverty) x = Knob_Poverty;
2161           _SpinDuration = x + Knob_Bonus;
2162         }
2163         return 1;
2164       }
2165 
2166       // The CAS failed ... we can take any of the following actions:
2167       // * penalize: ctr -= CASPenalty
2168       // * exit spin with prejudice -- goto Abort;
2169       // * exit spin without prejudice.
2170       // * Since CAS is high-latency, retry again immediately.
2171       prv = ox;
2172       goto Abort;
2173     }
2174 
2175     // Did lock ownership change hands ?
2176     if (ox != prv && prv != nullptr) {
2177       goto Abort;
2178     }
2179     prv = ox;
2180 
2181     if (_succ == nullptr) {
2182       _succ = current;
2183     }
2184   }
2185 
2186   // Spin failed with prejudice -- reduce _SpinDuration.
2187   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2188   // AIMD is globally stable.
2189   {
2190     int x = _SpinDuration;
2191     if (x > 0) {
2192       // Consider an AIMD scheme like: x -= (x >> 3) + 100
2193       // This is globally sample and tends to damp the response.
2194       x -= Knob_Penalty;
2195       if (x < 0) x = 0;
2196       _SpinDuration = x;
2197     }
2198   }
2199 
2200  Abort:
2201   if (_succ == current) {
2202     _succ = nullptr;
2203     // Invariant: after setting succ=null a contending thread
2204     // must recheck-retry _owner before parking.  This usually happens
2205     // in the normal usage of TrySpin(), but it's safest
2206     // to make TrySpin() as foolproof as possible.
2207     OrderAccess::fence();
2208     if (TryLock(current) > 0) return 1;
2209   }
2210   return 0;
2211 }
2212 
2213 
2214 // -----------------------------------------------------------------------------
2215 // WaitSet management ...
2216 
2217 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2218   _next     = nullptr;
2219   _prev     = nullptr;
2220   _notified = 0;
2221   _notifier_tid = 0;
2222   TState    = TS_RUN;
2223   _thread   = current;
2224   _event    = _thread != nullptr ? _thread->_ParkEvent : ObjectMonitor::vthread_unparker_ParkEvent();
2225   _active   = false;
2226   assert(_event != nullptr, "invariant");
2227 }
2228 
2229 ObjectWaiter::ObjectWaiter(oop vthread) : ObjectWaiter((JavaThread*)nullptr) {
2230   _vthread = OopHandle(JavaThread::thread_oop_storage(), vthread);
2231 }
2232 
2233 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2234   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2235 }
2236 
2237 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2238   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2239 }
2240 
2241 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2242   assert(node != nullptr, "should not add null node");
2243   assert(node->_prev == nullptr, "node already in list");
2244   assert(node->_next == nullptr, "node already in list");
2245   // put node at end of queue (circular doubly linked list)
2246   if (_WaitSet == nullptr) {
2247     _WaitSet = node;
2248     node->_prev = node;
2249     node->_next = node;
2250   } else {
2251     ObjectWaiter* head = _WaitSet;
2252     ObjectWaiter* tail = head->_prev;
2253     assert(tail->_next == head, "invariant check");
2254     tail->_next = node;
2255     head->_prev = node;
2256     node->_next = head;
2257     node->_prev = tail;
2258   }
2259 }
2260 
2261 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2262   // dequeue the very first waiter
2263   ObjectWaiter* waiter = _WaitSet;
2264   if (waiter) {
2265     DequeueSpecificWaiter(waiter);
2266   }
2267   return waiter;
2268 }
2269 
2270 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2271   assert(node != nullptr, "should not dequeue nullptr node");
2272   assert(node->_prev != nullptr, "node already removed from list");
2273   assert(node->_next != nullptr, "node already removed from list");
2274   // when the waiter has woken up because of interrupt,
2275   // timeout or other spurious wake-up, dequeue the
2276   // waiter from waiting list
2277   ObjectWaiter* next = node->_next;
2278   if (next == node) {
2279     assert(node->_prev == node, "invariant check");
2280     _WaitSet = nullptr;
2281   } else {
2282     ObjectWaiter* prev = node->_prev;
2283     assert(prev->_next == node, "invariant check");
2284     assert(next->_prev == node, "invariant check");
2285     next->_prev = prev;
2286     prev->_next = next;
2287     if (_WaitSet == node) {
2288       _WaitSet = next;
2289     }
2290   }
2291   node->_next = nullptr;
2292   node->_prev = nullptr;
2293 }
2294 
2295 // -----------------------------------------------------------------------------
2296 // PerfData support
2297 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = nullptr;
2298 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = nullptr;
2299 PerfCounter * ObjectMonitor::_sync_Parks                       = nullptr;
2300 PerfCounter * ObjectMonitor::_sync_Notifications               = nullptr;
2301 PerfCounter * ObjectMonitor::_sync_Inflations                  = nullptr;
2302 PerfCounter * ObjectMonitor::_sync_Deflations                  = nullptr;
2303 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = nullptr;
2304 
2305 // One-shot global initialization for the sync subsystem.
2306 // We could also defer initialization and initialize on-demand
2307 // the first time we call ObjectSynchronizer::inflate().
2308 // Initialization would be protected - like so many things - by
2309 // the MonitorCache_lock.
2310 
2311 void ObjectMonitor::Initialize() {
2312   assert(!InitDone, "invariant");
2313 
2314   if (!os::is_MP()) {
2315     Knob_SpinLimit = 0;
2316     Knob_PreSpin   = 0;
2317     Knob_FixedSpin = -1;
2318   }
2319 
2320   if (UsePerfData) {
2321     EXCEPTION_MARK;
2322 #define NEWPERFCOUNTER(n)                                                \
2323   {                                                                      \
2324     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2325                                         CHECK);                          \
2326   }
2327 #define NEWPERFVARIABLE(n)                                                \
2328   {                                                                       \
2329     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2330                                          CHECK);                          \
2331   }
2332     NEWPERFCOUNTER(_sync_Inflations);
2333     NEWPERFCOUNTER(_sync_Deflations);
2334     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2335     NEWPERFCOUNTER(_sync_FutileWakeups);
2336     NEWPERFCOUNTER(_sync_Parks);
2337     NEWPERFCOUNTER(_sync_Notifications);
2338     NEWPERFVARIABLE(_sync_MonExtant);
2339 #undef NEWPERFCOUNTER
2340 #undef NEWPERFVARIABLE
2341   }
2342 
2343   _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2344 
2345   DEBUG_ONLY(InitDone = true;)
2346 }
2347 
2348 void ObjectMonitor::Initialize2() {
2349   _vthread_cxq_head = OopHandle(JavaThread::thread_oop_storage(), nullptr);
2350   _vthread_unparker_ParkEvent = ParkEvent::Allocate(nullptr);
2351 }
2352 
2353 void ObjectMonitor::print_on(outputStream* st) const {
2354   // The minimal things to print for markWord printing, more can be added for debugging and logging.
2355   st->print("{contentions=0x%08x,waiters=0x%08x"
2356             ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2357             contentions(), waiters(), recursions(),
2358             p2i(owner()));
2359 }
2360 void ObjectMonitor::print() const { print_on(tty); }
2361 
2362 #ifdef ASSERT
2363 // Print the ObjectMonitor like a debugger would:
2364 //
2365 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2366 //   _header = 0x0000000000000001
2367 //   _object = 0x000000070ff45fd0
2368 //   _pad_buf0 = {
2369 //     [0] = '\0'
2370 //     ...
2371 //     [43] = '\0'
2372 //   }
2373 //   _owner = 0x0000000000000000
2374 //   _previous_owner_tid = 0
2375 //   _pad_buf1 = {
2376 //     [0] = '\0'
2377 //     ...
2378 //     [47] = '\0'
2379 //   }
2380 //   _next_om = 0x0000000000000000
2381 //   _recursions = 0
2382 //   _EntryList = 0x0000000000000000
2383 //   _cxq = 0x0000000000000000
2384 //   _succ = 0x0000000000000000
2385 //   _Responsible = 0x0000000000000000
2386 //   _Spinner = 0
2387 //   _SpinDuration = 5000
2388 //   _contentions = 0
2389 //   _WaitSet = 0x0000700009756248
2390 //   _waiters = 1
2391 //   _WaitSetLock = 0
2392 // }
2393 //
2394 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2395   st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2396   st->print_cr("  _header = " INTPTR_FORMAT, header().value());
2397   st->print_cr("  _object = " INTPTR_FORMAT, p2i(object_peek()));
2398   st->print_cr("  _pad_buf0 = {");
2399   st->print_cr("    [0] = '\\0'");
2400   st->print_cr("    ...");
2401   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2402   st->print_cr("  }");
2403   st->print_cr("  _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2404   st->print_cr("  _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid);
2405   st->print_cr("  _pad_buf1 = {");
2406   st->print_cr("    [0] = '\\0'");
2407   st->print_cr("    ...");
2408   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2409   st->print_cr("  }");
2410   st->print_cr("  _next_om = " INTPTR_FORMAT, p2i(next_om()));
2411   st->print_cr("  _recursions = " INTX_FORMAT, _recursions);
2412   st->print_cr("  _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2413   st->print_cr("  _cxq = " INTPTR_FORMAT, p2i(_cxq));
2414   st->print_cr("  _succ = " INTPTR_FORMAT, p2i(_succ));
2415   st->print_cr("  _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2416   st->print_cr("  _Spinner = %d", _Spinner);
2417   st->print_cr("  _SpinDuration = %d", _SpinDuration);
2418   st->print_cr("  _contentions = %d", contentions());
2419   st->print_cr("  _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2420   st->print_cr("  _waiters = %d", _waiters);
2421   st->print_cr("  _WaitSetLock = %d", _WaitSetLock);
2422   st->print_cr("}");
2423 }
2424 #endif