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   _SpinDuration(ObjectMonitor::Knob_SpinLimit),
 278   _contentions(0),
 279   _WaitSet(nullptr),
 280   _waiters(0),
 281   _WaitSetLock(0)
 282 { }
 283 
 284 ObjectMonitor::~ObjectMonitor() {
 285   _object.release(_oop_storage);
 286 }
 287 
 288 oop ObjectMonitor::object() const {
 289   check_object_context();
 290   return _object.resolve();
 291 }
 292 
 293 oop ObjectMonitor::object_peek() const {
 294   return _object.peek();
 295 }
 296 
 297 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
 298   if (current->is_suspended()) {
 299     _om->_recursions = 0;
 300     _om->_succ = nullptr;
 301     // Don't need a full fence after clearing successor here because of the call to exit().
 302     _om->exit(current, false /* not_suspended */);
 303     _om_exited = true;
 304 
 305     current->set_current_pending_monitor(_om);
 306   }
 307 }
 308 
 309 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
 310   if (current->is_suspended()) {
 311     if (_om->_succ == current) {
 312       _om->_succ = nullptr;
 313       OrderAccess::fence(); // always do a full fence when successor is cleared
 314     }
 315   }
 316 }
 317 
 318 // -----------------------------------------------------------------------------
 319 // Enter support
 320 
 321 bool ObjectMonitor::enter_for(JavaThread* locking_thread) {
 322   // Used by ObjectSynchronizer::enter_for to enter for another thread.
 323   // The monitor is private to or already owned by locking_thread which must be suspended.
 324   // So this code may only contend with deflation.
 325   assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
 326 
 327   // Block out deflation as soon as possible.
 328   add_to_contentions(1);
 329 
 330   bool success = false;
 331   if (!is_being_async_deflated()) {
 332     void* prev_owner = try_set_owner_from(nullptr, locking_thread);
 333 
 334     if (prev_owner == nullptr) {
 335       assert(_recursions == 0, "invariant");
 336       success = true;
 337     } else if (prev_owner == owner_for(locking_thread)) {
 338       _recursions++;
 339       success = true;
 340     } else if (prev_owner == DEFLATER_MARKER) {
 341       // Racing with deflation.
 342       prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread);
 343       if (prev_owner == DEFLATER_MARKER) {
 344         // Cancelled deflation. Increment contentions as part of the deflation protocol.
 345         add_to_contentions(1);
 346         success = true;
 347       } else if (prev_owner == nullptr) {
 348         // At this point we cannot race with deflation as we have both incremented
 349         // contentions, seen contention > 0 and seen a DEFLATER_MARKER.
 350         // success will only be false if this races with something other than
 351         // deflation.
 352         prev_owner = try_set_owner_from(nullptr, locking_thread);
 353         success = prev_owner == nullptr;
 354       }





 355     }
 356     assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT
 357            ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT,
 358            p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner));
 359   } else {
 360     // Async deflation is in progress and our contentions increment
 361     // above lost the race to async deflation. Undo the work and
 362     // force the caller to retry.
 363     const oop l_object = object();
 364     if (l_object != nullptr) {
 365       // Attempt to restore the header/dmw to the object's header so that
 366       // we only retry once if the deflater thread happens to be slow.
 367       install_displaced_markword_in_object(l_object);
 368     }
 369   }
 370 
 371   add_to_contentions(-1);
 372 
 373   assert(!success || is_owner(locking_thread), "must be");
 374 
 375   return success;
 376 }
 377 
 378 bool ObjectMonitor::enter(JavaThread* current) {
 379   assert(current == JavaThread::current(), "must be");
 380   // The following code is ordered to check the most common cases first
 381   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 382 
 383   void* cur = try_set_owner_from(nullptr, current);
 384   if (cur == nullptr) {
 385     assert(_recursions == 0, "invariant");
 386     return true;
 387   }
 388 
 389   if (cur == owner_for(current)) {
 390     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 391     _recursions++;
 392     return true;
 393   }
 394 







 395   // We've encountered genuine contention.
 396 
 397   // Try one round of spinning *before* enqueueing current
 398   // and before going through the awkward and expensive state
 399   // transitions.  The following spin is strictly optional ...
 400   // Note that if we acquire the monitor from an initial spin
 401   // we forgo posting JVMTI events and firing DTRACE probes.
 402   if (TrySpin(current) > 0) {
 403     assert(owner_raw() == owner_for(current), "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
 404     assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
 405     assert(object()->mark() == markWord::encode(this),
 406            "object mark must match encoded this: mark=" INTPTR_FORMAT
 407            ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
 408            markWord::encode(this).value());
 409     return true;
 410   }
 411 
 412   assert(owner_raw() != owner_for(current), "invariant");
 413   assert(_succ != current, "invariant");
 414   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 415   assert(current->thread_state() != _thread_blocked, "invariant");
 416 
 417   // Keep track of contention for JVM/TI and M&M queries.
 418   add_to_contentions(1);
 419   if (is_being_async_deflated()) {
 420     // Async deflation is in progress and our contentions increment
 421     // above lost the race to async deflation. Undo the work and
 422     // force the caller to retry.
 423     const oop l_object = object();
 424     if (l_object != nullptr) {
 425       // Attempt to restore the header/dmw to the object's header so that
 426       // we only retry once if the deflater thread happens to be slow.
 427       install_displaced_markword_in_object(l_object);
 428     }
 429     add_to_contentions(-1);
 430     return false;
 431   }
 432 
 433   JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);)
 434   EventJavaMonitorEnter event;
 435   if (event.is_started()) {
 436     event.set_monitorClass(object()->klass());
 437     // Set an address that is 'unique enough', such that events close in
 438     // time and with the same address are likely (but not guaranteed) to
 439     // belong to the same object.
 440     event.set_address((uintptr_t)this);
 441   }
 442 
 443   { // Change java thread status to indicate blocked on monitor enter.
 444     JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
 445 
 446     assert(current->current_pending_monitor() == nullptr, "invariant");
 447     current->set_current_pending_monitor(this);
 448 
 449     DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
 450     if (JvmtiExport::should_post_monitor_contended_enter()) {
 451       JvmtiExport::post_monitor_contended_enter(current, this);
 452 
 453       // The current thread does not yet own the monitor and does not
 454       // yet appear on any queues that would get it made the successor.
 455       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 456       // handler cannot accidentally consume an unpark() meant for the
 457       // ParkEvent associated with this ObjectMonitor.
 458     }
 459 
 460 #ifdef LOOM_MONITOR_SUPPORT
 461     ContinuationEntry* ce = current->last_continuation();
 462     if (ce != nullptr && ce->is_virtual_thread() && current->is_on_monitorenter()) {
 463       int result = Continuation::try_preempt(current, ce->cont_oop(current));
 464       if (result == freeze_ok) {
 465         bool acquired = HandlePreemptedVThread(current);
 466         DEBUG_ONLY(int state = java_lang_VirtualThread::state(current->vthread()));
 467         assert((acquired && current->preemption_cancelled() && state == java_lang_VirtualThread::RUNNING) ||
 468                (!acquired && !current->preemption_cancelled() && state == java_lang_VirtualThread::BLOCKING), "invariant");
 469         return true;
 470       }
 471       if (result == freeze_pinned_native || result == freeze_pinned_cs) {
 472         EventVirtualThreadPinned e;
 473         if (e.should_commit()) {
 474           e.commit();
 475         }
 476       }
 477     }
 478 #endif
 479 
 480     OSThreadContendState osts(current->osthread());
 481 
 482     assert(current->thread_state() == _thread_in_vm, "invariant");
 483 
 484     for (;;) {
 485       ExitOnSuspend eos(this);
 486       {
 487         ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */);
 488         EnterI(current);
 489         current->set_current_pending_monitor(nullptr);
 490         // We can go to a safepoint at the end of this block. If we
 491         // do a thread dump during that safepoint, then this thread will show
 492         // as having "-locked" the monitor, but the OS and java.lang.Thread
 493         // states will still report that the thread is blocked trying to
 494         // acquire it.
 495         // If there is a suspend request, ExitOnSuspend will exit the OM
 496         // and set the OM as pending.
 497       }
 498       if (!eos.exited()) {
 499         // ExitOnSuspend did not exit the OM
 500         assert(owner_raw() == owner_for(current), "invariant");
 501         break;
 502       }
 503     }
 504 
 505     // We've just gotten past the enter-check-for-suspend dance and we now own
 506     // the monitor free and clear.
 507   }
 508 
 509   add_to_contentions(-1);
 510   assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
 511 
 512   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 513   assert(_recursions == 0, "invariant");
 514   assert(owner_raw() == owner_for(current), "invariant");
 515   assert(_succ != current, "invariant");
 516   assert(object()->mark() == markWord::encode(this), "invariant");
 517 
 518   // The thread -- now the owner -- is back in vm mode.
 519   // Report the glorious news via TI,DTrace and jvmstat.
 520   // The probe effect is non-trivial.  All the reportage occurs
 521   // while we hold the monitor, increasing the length of the critical
 522   // section.  Amdahl's parallel speedup law comes vividly into play.
 523   //
 524   // Another option might be to aggregate the events (thread local or
 525   // per-monitor aggregation) and defer reporting until a more opportune
 526   // time -- such as next time some thread encounters contention but has
 527   // yet to acquire the lock.  While spinning that thread could
 528   // spinning we could increment JVMStat counters, etc.
 529 
 530   DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
 531   if (JvmtiExport::should_post_monitor_contended_entered()) {
 532     JvmtiExport::post_monitor_contended_entered(current, this);
 533 
 534     // The current thread already owns the monitor and is not going to
 535     // call park() for the remainder of the monitor enter protocol. So
 536     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 537     // event handler consumed an unpark() issued by the thread that
 538     // just exited the monitor.
 539   }
 540   if (event.should_commit()) {
 541     event.set_previousOwner(_previous_owner_tid);
 542     event.commit();
 543   }
 544   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 545   return true;
 546 }
 547 
 548 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 549 // Callers must compensate as needed.
 550 
 551 int ObjectMonitor::TryLock(JavaThread* current) {
 552   void* own = owner_raw();
 553   if (own != nullptr) return 0;
 554   if (try_set_owner_from(nullptr, current) == nullptr) {
 555     assert(_recursions == 0, "invariant");
 556     return 1;
 557   }
 558   // The lock had been free momentarily, but we lost the race to the lock.
 559   // Interference -- the CAS failed.
 560   // We can either return -1 or retry.
 561   // Retry doesn't make as much sense because the lock was just acquired.
 562   return -1;
 563 }
 564 
 565 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
 566 // was deflated and false otherwise.
 567 //
 568 // The async deflation protocol sets owner to DEFLATER_MARKER and
 569 // makes contentions negative as signals to contending threads that
 570 // an async deflation is in progress. There are a number of checks
 571 // as part of the protocol to make sure that the calling thread has
 572 // not lost the race to a contending thread.
 573 //
 574 // The ObjectMonitor has been successfully async deflated when:
 575 //   (contentions < 0)
 576 // Contending threads that see that condition know to retry their operation.
 577 //
 578 bool ObjectMonitor::deflate_monitor() {
 579   if (is_busy()) {
 580     // Easy checks are first - the ObjectMonitor is busy so no deflation.
 581     return false;
 582   }
 583 
 584   const oop obj = object_peek();
 585 
 586   if (obj == nullptr) {
 587     // If the object died, we can recycle the monitor without racing with
 588     // Java threads. The GC already broke the association with the object.
 589     set_owner_from_raw(nullptr, DEFLATER_MARKER);
 590     assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
 591     _contentions = INT_MIN; // minimum negative int
 592   } else {
 593     // Attempt async deflation protocol.
 594 
 595     // Set a null owner to DEFLATER_MARKER to force any contending thread
 596     // through the slow path. This is just the first part of the async
 597     // deflation dance.
 598     if (try_set_owner_from_raw(nullptr, DEFLATER_MARKER) != nullptr) {
 599       // The owner field is no longer null so we lost the race since the
 600       // ObjectMonitor is now busy.
 601       return false;
 602     }
 603 
 604     if (contentions() > 0 || _waiters != 0) {
 605       // Another thread has raced to enter the ObjectMonitor after
 606       // is_busy() above or has already entered and waited on
 607       // it which makes it busy so no deflation. Restore owner to
 608       // null if it is still DEFLATER_MARKER.
 609       if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
 610         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 611         add_to_contentions(-1);
 612       }
 613       return false;
 614     }
 615 
 616     // Make a zero contentions field negative to force any contending threads
 617     // to retry. This is the second part of the async deflation dance.
 618     if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) {
 619       // Contentions was no longer 0 so we lost the race since the
 620       // ObjectMonitor is now busy. Restore owner to null if it is
 621       // still DEFLATER_MARKER:
 622       if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
 623         // Deferred decrement for the JT EnterI() that cancelled the async deflation.
 624         add_to_contentions(-1);
 625       }
 626       return false;
 627     }
 628   }
 629 
 630   // Sanity checks for the races:
 631   guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
 632   guarantee(contentions() < 0, "must be negative: contentions=%d",
 633             contentions());
 634   guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
 635   guarantee(_cxq == nullptr, "must be no contending threads: cxq="
 636             INTPTR_FORMAT, p2i(_cxq));
 637   guarantee(_EntryList == nullptr,
 638             "must be no entering threads: EntryList=" INTPTR_FORMAT,
 639             p2i(_EntryList));
 640 
 641   if (obj != nullptr) {
 642     if (log_is_enabled(Trace, monitorinflation)) {
 643       ResourceMark rm;
 644       log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
 645                                   ", mark=" INTPTR_FORMAT ", type='%s'",
 646                                   p2i(obj), obj->mark().value(),
 647                                   obj->klass()->external_name());
 648     }
 649 
 650     // Install the old mark word if nobody else has already done it.
 651     install_displaced_markword_in_object(obj);
 652   }
 653 
 654   // We leave owner == DEFLATER_MARKER and contentions < 0
 655   // to force any racing threads to retry.
 656   return true;  // Success, ObjectMonitor has been deflated.
 657 }
 658 
 659 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
 660 // into the header of the object associated with the monitor. This
 661 // idempotent method is called by a thread that is deflating a
 662 // monitor and by other threads that have detected a race with the
 663 // deflation process.
 664 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
 665   // This function must only be called when (owner == DEFLATER_MARKER
 666   // && contentions <= 0), but we can't guarantee that here because
 667   // those values could change when the ObjectMonitor gets moved from
 668   // the global free list to a per-thread free list.
 669 
 670   guarantee(obj != nullptr, "must be non-null");
 671 
 672   // Separate loads in is_being_async_deflated(), which is almost always
 673   // called before this function, from the load of dmw/header below.
 674 
 675   // _contentions and dmw/header may get written by different threads.
 676   // Make sure to observe them in the same order when having several observers.
 677   OrderAccess::loadload_for_IRIW();
 678 
 679   const oop l_object = object_peek();
 680   if (l_object == nullptr) {
 681     // ObjectMonitor's object ref has already been cleared by async
 682     // deflation or GC so we're done here.
 683     return;
 684   }
 685   assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
 686          INTPTR_FORMAT, p2i(l_object), p2i(obj));
 687 
 688   markWord dmw = header();
 689   // The dmw has to be neutral (not null, not locked and not marked).
 690   assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
 691 
 692   // Install displaced mark word if the object's header still points
 693   // to this ObjectMonitor. More than one racing caller to this function
 694   // can rarely reach this point, but only one can win.
 695   markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
 696   if (res != markWord::encode(this)) {
 697     // This should be rare so log at the Info level when it happens.
 698     log_info(monitorinflation)("install_displaced_markword_in_object: "
 699                                "failed cas_set_mark: new_mark=" INTPTR_FORMAT
 700                                ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
 701                                dmw.value(), markWord::encode(this).value(),
 702                                res.value());
 703   }
 704 
 705   // Note: It does not matter which thread restored the header/dmw
 706   // into the object's header. The thread deflating the monitor just
 707   // wanted the object's header restored and it is. The threads that
 708   // detected a race with the deflation process also wanted the
 709   // object's header restored before they retry their operation and
 710   // because it is restored they will only retry once.
 711 }
 712 
 713 // Convert the fields used by is_busy() to a string that can be
 714 // used for diagnostic output.
 715 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
 716   ss->print("is_busy: waiters=%d"
 717             ", contentions=%d"
 718             ", owner=" INTPTR_FORMAT
 719             ", cxq=" PTR_FORMAT
 720             ", EntryList=" PTR_FORMAT,
 721             _waiters,
 722             (contentions() > 0 ? contentions() : 0),
 723             owner_is_DEFLATER_MARKER()
 724                 // We report null instead of DEFLATER_MARKER here because is_busy()
 725                 // ignores DEFLATER_MARKER values.
 726                 ? p2i(nullptr)
 727                 : p2i(owner_raw()),
 728             p2i(_cxq),
 729             p2i(_EntryList));
 730   return ss->base();
 731 }
 732 
 733 #define MAX_RECHECK_INTERVAL 1000
 734 
 735 void ObjectMonitor::EnterI(JavaThread* current) {
 736   assert(current->thread_state() == _thread_blocked, "invariant");
 737 
 738   // Try the lock - TATAS
 739   if (TryLock (current) > 0) {
 740     assert(_succ != current, "invariant");
 741     assert(owner_raw() == owner_for(current), "invariant");
 742     assert(_Responsible != current, "invariant");
 743     return;
 744   }
 745 
 746   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 747     // Cancelled the in-progress async deflation by changing owner from
 748     // DEFLATER_MARKER to current. As part of the contended enter protocol,
 749     // contentions was incremented to a positive value before EnterI()
 750     // was called and that prevents the deflater thread from winning the
 751     // last part of the 2-part async deflation protocol. After EnterI()
 752     // returns to enter(), contentions is decremented because the caller
 753     // now owns the monitor. We bump contentions an extra time here to
 754     // prevent the deflater thread from winning the last part of the
 755     // 2-part async deflation protocol after the regular decrement
 756     // occurs in enter(). The deflater thread will decrement contentions
 757     // after it recognizes that the async deflation was cancelled.
 758     add_to_contentions(1);
 759     assert(_succ != current, "invariant");
 760     assert(_Responsible != current, "invariant");
 761     return;
 762   }
 763 
 764   assert(InitDone, "Unexpectedly not initialized");
 765 
 766   // We try one round of spinning *before* enqueueing current.
 767   //
 768   // If the _owner is ready but OFFPROC we could use a YieldTo()
 769   // operation to donate the remainder of this thread's quantum
 770   // to the owner.  This has subtle but beneficial affinity
 771   // effects.
 772 
 773   if (TrySpin(current) > 0) {
 774     assert(owner_raw() == owner_for(current), "invariant");
 775     assert(_succ != current, "invariant");
 776     assert(_Responsible != current, "invariant");
 777     return;
 778   }
 779 
 780   // The Spin failed -- Enqueue and park the thread ...
 781   assert(_succ != current, "invariant");
 782   assert(owner_raw() != owner_for(current), "invariant");
 783   assert(_Responsible != current, "invariant");
 784 
 785   // Enqueue "current" on ObjectMonitor's _cxq.
 786   //
 787   // Node acts as a proxy for current.
 788   // As an aside, if were to ever rewrite the synchronization code mostly
 789   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 790   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 791   // as well as eliminate a subset of ABA issues.
 792   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 793 
 794   ObjectWaiter node(current);
 795   current->_ParkEvent->reset();
 796   node._prev   = (ObjectWaiter*) 0xBAD;
 797   node.TState  = ObjectWaiter::TS_CXQ;
 798 
 799   // Push "current" onto the front of the _cxq.
 800   // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
 801   // Note that spinning tends to reduce the rate at which threads
 802   // enqueue and dequeue on EntryList|cxq.
 803   ObjectWaiter* nxt;
 804   for (;;) {
 805     node._next = nxt = _cxq;
 806     if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
 807 
 808     // Interference - the CAS failed because _cxq changed.  Just retry.
 809     // As an optional optimization we retry the lock.
 810     if (TryLock (current) > 0) {
 811       assert(_succ != current, "invariant");
 812       assert(owner_raw() == owner_for(current), "invariant");
 813       assert(_Responsible != current, "invariant");
 814       return;
 815     }
 816   }
 817 
 818   // Check for cxq|EntryList edge transition to non-null.  This indicates
 819   // the onset of contention.  While contention persists exiting threads
 820   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 821   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 822   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 823   // arrange for one of the contending thread to use a timed park() operations
 824   // to detect and recover from the race.  (Stranding is form of progress failure
 825   // where the monitor is unlocked but all the contending threads remain parked).
 826   // That is, at least one of the contended threads will periodically poll _owner.
 827   // One of the contending threads will become the designated "Responsible" thread.
 828   // The Responsible thread uses a timed park instead of a normal indefinite park
 829   // operation -- it periodically wakes and checks for and recovers from potential
 830   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 831   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 832   // be responsible for a monitor.
 833   //
 834   // Currently, one of the contended threads takes on the added role of "Responsible".
 835   // A viable alternative would be to use a dedicated "stranding checker" thread
 836   // that periodically iterated over all the threads (or active monitors) and unparked
 837   // successors where there was risk of stranding.  This would help eliminate the
 838   // timer scalability issues we see on some platforms as we'd only have one thread
 839   // -- the checker -- parked on a timer.
 840 
 841   if (nxt == nullptr && _EntryList == nullptr) {
 842     // Try to assume the role of responsible thread for the monitor.
 843     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=current }
 844     Atomic::replace_if_null(&_Responsible, current);
 845   }
 846 
 847   // The lock might have been released while this thread was occupied queueing
 848   // itself onto _cxq.  To close the race and avoid "stranding" and
 849   // progress-liveness failure we must resample-retry _owner before parking.
 850   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 851   // In this case the ST-MEMBAR is accomplished with CAS().
 852   //
 853   // TODO: Defer all thread state transitions until park-time.
 854   // Since state transitions are heavy and inefficient we'd like
 855   // to defer the state transitions until absolutely necessary,
 856   // and in doing so avoid some transitions ...
 857 
 858   int nWakeups = 0;
 859   int recheckInterval = 1;
 860   bool do_timed_parked = false;
 861 
 862   ContinuationEntry* ce = current->last_continuation();
 863   if (ce != nullptr && ce->is_virtual_thread()) {
 864     do_timed_parked = true;
 865   }
 866 
 867   for (;;) {
 868 
 869     if (TryLock(current) > 0) break;
 870     assert(owner_raw() != owner_for(current), "invariant");
 871 
 872     // park self
 873     if (_Responsible == current || do_timed_parked) {
 874       current->_ParkEvent->park((jlong) recheckInterval);
 875       // Increase the recheckInterval, but clamp the value.
 876       recheckInterval *= 8;
 877       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 878         recheckInterval = MAX_RECHECK_INTERVAL;
 879       }
 880     } else {
 881       current->_ParkEvent->park();
 882     }
 883 
 884     if (TryLock(current) > 0) break;
 885 
 886     if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
 887       // Cancelled the in-progress async deflation by changing owner from
 888       // DEFLATER_MARKER to current. As part of the contended enter protocol,
 889       // contentions was incremented to a positive value before EnterI()
 890       // was called and that prevents the deflater thread from winning the
 891       // last part of the 2-part async deflation protocol. After EnterI()
 892       // returns to enter(), contentions is decremented because the caller
 893       // now owns the monitor. We bump contentions an extra time here to
 894       // prevent the deflater thread from winning the last part of the
 895       // 2-part async deflation protocol after the regular decrement
 896       // occurs in enter(). The deflater thread will decrement contentions
 897       // after it recognizes that the async deflation was cancelled.
 898       add_to_contentions(1);
 899       break;
 900     }
 901 
 902     // The lock is still contested.
 903     // Keep a tally of the # of futile wakeups.
 904     // Note that the counter is not protected by a lock or updated by atomics.
 905     // That is by design - we trade "lossy" counters which are exposed to
 906     // races during updates for a lower probe effect.
 907 
 908     // This PerfData object can be used in parallel with a safepoint.
 909     // See the work around in PerfDataManager::destroy().
 910     OM_PERFDATA_OP(FutileWakeups, inc());
 911     ++nWakeups;
 912 
 913     // Assuming this is not a spurious wakeup we'll normally find _succ == current.
 914     // We can defer clearing _succ until after the spin completes
 915     // TrySpin() must tolerate being called with _succ == current.
 916     // Try yet another round of adaptive spinning.
 917     if (TrySpin(current) > 0) break;
 918 
 919     // We can find that we were unpark()ed and redesignated _succ while
 920     // we were spinning.  That's harmless.  If we iterate and call park(),
 921     // park() will consume the event and return immediately and we'll
 922     // just spin again.  This pattern can repeat, leaving _succ to simply
 923     // spin on a CPU.
 924 
 925     if (_succ == current) _succ = nullptr;
 926 
 927     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 928     OrderAccess::fence();
 929   }
 930 
 931   // Egress :
 932   // current has acquired the lock -- Unlink current from the cxq or EntryList.
 933   // Normally we'll find current on the EntryList .
 934   // From the perspective of the lock owner (this thread), the
 935   // EntryList is stable and cxq is prepend-only.
 936   // The head of cxq is volatile but the interior is stable.
 937   // In addition, current.TState is stable.
 938 
 939   assert(owner_raw() == owner_for(current), "invariant");
 940 
 941   UnlinkAfterAcquire(current, &node);
 942   if (_succ == current) _succ = nullptr;
 943 
 944   assert(_succ != current, "invariant");
 945   if (_Responsible == current) {
 946     _Responsible = nullptr;
 947     OrderAccess::fence(); // Dekker pivot-point
 948 
 949     // We may leave threads on cxq|EntryList without a designated
 950     // "Responsible" thread.  This is benign.  When this thread subsequently
 951     // exits the monitor it can "see" such preexisting "old" threads --
 952     // threads that arrived on the cxq|EntryList before the fence, above --
 953     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 954     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 955     // non-null and elect a new "Responsible" timer thread.
 956     //
 957     // This thread executes:
 958     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 959     //    LD cxq|EntryList               (in subsequent exit)
 960     //
 961     // Entering threads in the slow/contended path execute:
 962     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 963     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 964     //
 965     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 966     // exit operation from floating above the ST Responsible=null.
 967   }
 968 
 969   // We've acquired ownership with CAS().
 970   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 971   // But since the CAS() this thread may have also stored into _succ,
 972   // EntryList, cxq or Responsible.  These meta-data updates must be
 973   // visible __before this thread subsequently drops the lock.
 974   // Consider what could occur if we didn't enforce this constraint --
 975   // STs to monitor meta-data and user-data could reorder with (become
 976   // visible after) the ST in exit that drops ownership of the lock.
 977   // Some other thread could then acquire the lock, but observe inconsistent
 978   // or old monitor meta-data and heap data.  That violates the JMM.
 979   // To that end, the 1-0 exit() operation must have at least STST|LDST
 980   // "release" barrier semantics.  Specifically, there must be at least a
 981   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 982   // the lock.   The barrier ensures that changes to monitor meta-data and data
 983   // protected by the lock will be visible before we release the lock, and
 984   // therefore before some other thread (CPU) has a chance to acquire the lock.
 985   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 986   //
 987   // Critically, any prior STs to _succ or EntryList must be visible before
 988   // the ST of null into _owner in the *subsequent* (following) corresponding
 989   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 990   // execute a serializing instruction.
 991 
 992   return;
 993 }
 994 
 995 bool ObjectMonitor::HandlePreemptedVThread(JavaThread* current) {
 996   // Either because we acquire the lock below or because we will preempt the
 997   // vthread clear the _current_pending_monitor field from the current JavaThread.
 998   current->set_current_pending_monitor(nullptr);
 999 
1000   // Try once more after freezing the continuation.
1001   if (TryLock (current) > 0) {
1002     assert(owner_raw() == owner_for(current), "invariant");
1003     assert(_succ != current, "invariant");
1004     assert(_Responsible != current, "invariant");
1005     current->set_preemption_cancelled(true);
1006     add_to_contentions(-1);
1007     return true;
1008   }
1009 
1010   if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
1011     // Cancelled the in-progress async deflation by changing owner from
1012     // DEFLATER_MARKER to current. As part of the contended enter protocol,
1013     // contentions was incremented to a positive value before this call to
1014     // HandlePreemptedVThread(). We avoid decrementing contentions to
1015     // prevent the deflater thread from winning the last part of the
1016     // 2-part async deflation protocol. The deflater thread will decrement
1017     // contentions after it recognizes that the async deflation was cancelled.
1018     assert(_succ != current, "invariant");
1019     assert(_Responsible != current, "invariant");
1020     current->set_preemption_cancelled(true);
1021     return true;
1022   }
1023 
1024   oop vthread = current->vthread();
1025   assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
1026   java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING);
1027 
1028   ObjectWaiter* node = new ObjectWaiter(vthread);
1029   node->_prev   = (ObjectWaiter*) 0xBAD;
1030   node->TState  = ObjectWaiter::TS_CXQ;
1031 
1032   // Push node associated with vthread onto the front of the _cxq.
1033   ObjectWaiter* nxt;
1034   for (;;) {
1035     node->_next = nxt = _cxq;
1036     if (Atomic::cmpxchg(&_cxq, nxt, node) == nxt) break;
1037 
1038     // Interference - the CAS failed because _cxq changed.  Just retry.
1039     // As an optional optimization we retry the lock.
1040     if (TryLock (current) > 0) {
1041       assert(owner_raw() == owner_for(current), "invariant");
1042       assert(_succ != current, "invariant");
1043       assert(_Responsible != current, "invariant");
1044       current->set_preemption_cancelled(true);
1045       java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::RUNNING);
1046       add_to_contentions(-1);
1047       delete node;
1048       return true;
1049     }
1050   }
1051 
1052   // We have to try once more since owner could have exited monitor and checked
1053   // _cxq before we added the node to the queue.
1054   if (TryLock(current) > 0) {
1055     assert(owner_raw() == owner_for(current), "invariant");
1056     assert(_Responsible != current, "invariant");
1057     current->set_preemption_cancelled(true);
1058     java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::RUNNING);
1059     UnlinkAfterAcquire(current, node, vthread);
1060     delete node;
1061     if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr;
1062     add_to_contentions(-1);
1063     return true;
1064   }
1065 
1066   if (nxt == nullptr && _EntryList == nullptr) {
1067     // The C2 unlock() fast path first checks if _cxq and _EntryList are empty and
1068     // if they are it just clears the _owner field. Since we always run the risk of
1069     // having that check happening before we added the node to _cxq and the release
1070     // of the monitor happening after the last TryLock attempt we need to do something
1071     // to avoid stranding. We set the _Responsible field which results in a timed-wait.
1072     if (Atomic::replace_if_null(&_Responsible, (JavaThread*)java_lang_Thread::thread_id(vthread))) {
1073       java_lang_VirtualThread::set_recheckInterval(vthread, 1);
1074     }
1075   }
1076 
1077   return false;
1078 }
1079 
1080 // ReenterI() is a specialized inline form of the latter half of the
1081 // contended slow-path from EnterI().  We use ReenterI() only for
1082 // monitor reentry in wait().
1083 //
1084 // In the future we should reconcile EnterI() and ReenterI().
1085 
1086 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
1087   assert(current != nullptr, "invariant");
1088   assert(currentNode != nullptr, "invariant");
1089   assert(currentNode->_thread == current, "invariant");
1090   assert(_waiters > 0, "invariant");
1091   assert(object()->mark() == markWord::encode(this), "invariant");
1092 
1093   assert(current->thread_state() != _thread_blocked, "invariant");
1094 
1095   int nWakeups = 0;
1096   for (;;) {
1097     ObjectWaiter::TStates v = currentNode->TState;
1098     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1099     assert(owner_raw() != owner_for(current), "invariant");
1100 
1101     if (TryLock(current) > 0) break;
1102     if (TrySpin(current) > 0) break;
1103 
1104     {
1105       OSThreadContendState osts(current->osthread());
1106 
1107       assert(current->thread_state() == _thread_in_vm, "invariant");
1108 
1109       {
1110         ClearSuccOnSuspend csos(this);
1111         ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1112         current->_ParkEvent->park();
1113       }
1114     }
1115 
1116     // Try again, but just so we distinguish between futile wakeups and
1117     // successful wakeups.  The following test isn't algorithmically
1118     // necessary, but it helps us maintain sensible statistics.
1119     if (TryLock(current) > 0) break;
1120 
1121     // The lock is still contested.
1122     // Keep a tally of the # of futile wakeups.
1123     // Note that the counter is not protected by a lock or updated by atomics.
1124     // That is by design - we trade "lossy" counters which are exposed to
1125     // races during updates for a lower probe effect.
1126     ++nWakeups;
1127 
1128     // Assuming this is not a spurious wakeup we'll normally
1129     // find that _succ == current.
1130     if (_succ == current) _succ = nullptr;
1131 
1132     // Invariant: after clearing _succ a contending thread
1133     // *must* retry  _owner before parking.
1134     OrderAccess::fence();
1135 
1136     // This PerfData object can be used in parallel with a safepoint.
1137     // See the work around in PerfDataManager::destroy().
1138     OM_PERFDATA_OP(FutileWakeups, inc());
1139   }
1140 
1141   // current has acquired the lock -- Unlink current from the cxq or EntryList .
1142   // Normally we'll find current on the EntryList.
1143   // Unlinking from the EntryList is constant-time and atomic-free.
1144   // From the perspective of the lock owner (this thread), the
1145   // EntryList is stable and cxq is prepend-only.
1146   // The head of cxq is volatile but the interior is stable.
1147   // In addition, current.TState is stable.
1148 
1149   assert(owner_raw() == owner_for(current), "invariant");
1150   assert(object()->mark() == markWord::encode(this), "invariant");
1151   UnlinkAfterAcquire(current, currentNode);
1152   if (_succ == current) _succ = nullptr;
1153   assert(_succ != current, "invariant");
1154   currentNode->TState = ObjectWaiter::TS_RUN;
1155   OrderAccess::fence();      // see comments at the end of EnterI()
1156 }
1157 
1158 void ObjectMonitor::redo_enter(JavaThread* current) {
1159   assert(java_lang_VirtualThread::state(current->vthread()) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
1160   assert(current->is_in_VTMS_transition(), "must be");
1161 
1162   if (TryLock (current) > 0) {
1163     VThreadEpilog(current);
1164     return;
1165   }
1166 
1167   oop vthread = current->vthread();
1168   if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr;
1169 
1170   // Invariant: after clearing _succ a thread *must* retry _owner before parking.
1171   OrderAccess::fence();
1172 
1173   if (TryLock (current) > 0) {
1174     assert(owner_raw() == owner_for(current), "invariant");
1175     VThreadEpilog(current);
1176     return;
1177   }
1178 
1179   // Fast preemption. The JT will read this variable on return to the
1180   // monitorenter_redo stub and will just remove enterSpecial frame
1181   // from the stack and return to Continuation.run()
1182   current->set_preempting(true);
1183 
1184   java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING);
1185   if (_Responsible == (JavaThread*)java_lang_Thread::thread_id(vthread)) {
1186     int recheckInterval = java_lang_VirtualThread::recheckInterval(vthread);
1187     assert(recheckInterval >= 1 && recheckInterval <= 6, "invariant");
1188     if (recheckInterval < 6) {
1189       recheckInterval++;
1190       java_lang_VirtualThread::set_recheckInterval(vthread, recheckInterval);
1191     }
1192   } else if (java_lang_VirtualThread::recheckInterval(vthread) > 0) {
1193     // No need to do timed park anymore
1194     java_lang_VirtualThread::set_recheckInterval(vthread, 0);
1195   }
1196 }
1197 
1198 void ObjectMonitor::VThreadEpilog(JavaThread* current) {
1199   assert(owner_raw() == owner_for(current), "invariant");
1200   add_to_contentions(-1);
1201 
1202   oop vthread = current->vthread();
1203   if (java_lang_VirtualThread::recheckInterval(vthread) > 0) {
1204     java_lang_VirtualThread::set_recheckInterval(vthread, 0);
1205   }
1206   int64_t threadid = java_lang_Thread::thread_id(vthread);
1207   if (_succ == (JavaThread*)threadid) _succ = nullptr;
1208   if (_Responsible == (JavaThread*)threadid) {
1209     _Responsible = nullptr;
1210     OrderAccess::fence(); // Dekker pivot-point
1211   }
1212   ObjectWaiter* node = LookupWaiter(threadid);
1213   UnlinkAfterAcquire(current, node, vthread);
1214   delete node;
1215 }
1216 
1217 // By convention we unlink a contending thread from EntryList|cxq immediately
1218 // after the thread acquires the lock in ::enter().  Equally, we could defer
1219 // unlinking the thread until ::exit()-time.
1220 
1221 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode, oop vthread) {
1222   assert(owner_raw() == owner_for(current), "invariant");
1223   assert((currentNode->_thread == current) || (currentNode->_thread == nullptr && currentNode->vthread() == vthread), "invariant");
1224 
1225   if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1226     // Normal case: remove current from the DLL EntryList .
1227     // This is a constant-time operation.
1228     ObjectWaiter* nxt = currentNode->_next;
1229     ObjectWaiter* prv = currentNode->_prev;
1230     if (nxt != nullptr) nxt->_prev = prv;
1231     if (prv != nullptr) prv->_next = nxt;
1232     if (currentNode == _EntryList) _EntryList = nxt;
1233     assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1234     assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1235   } else {
1236     assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1237     // Inopportune interleaving -- current is still on the cxq.
1238     // This usually means the enqueue of self raced an exiting thread.
1239     // Normally we'll find current near the front of the cxq, so
1240     // dequeueing is typically fast.  If needbe we can accelerate
1241     // this with some MCS/CHL-like bidirectional list hints and advisory
1242     // back-links so dequeueing from the interior will normally operate
1243     // in constant-time.
1244     // Dequeue current from either the head (with CAS) or from the interior
1245     // with a linear-time scan and normal non-atomic memory operations.
1246     // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1247     // and then unlink current from EntryList.  We have to drain eventually,
1248     // so it might as well be now.
1249 
1250     ObjectWaiter* v = _cxq;
1251     assert(v != nullptr, "invariant");
1252     if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1253       // The CAS above can fail from interference IFF a "RAT" arrived.
1254       // In that case current must be in the interior and can no longer be
1255       // at the head of cxq.
1256       if (v == currentNode) {
1257         assert(_cxq != v, "invariant");
1258         v = _cxq;          // CAS above failed - start scan at head of list
1259       }
1260       ObjectWaiter* p;
1261       ObjectWaiter* q = nullptr;
1262       for (p = v; p != nullptr && p != currentNode; p = p->_next) {
1263         q = p;
1264         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1265       }
1266       assert(v != currentNode, "invariant");
1267       assert(p == currentNode, "Node not found on cxq");
1268       assert(p != _cxq, "invariant");
1269       assert(q != nullptr, "invariant");
1270       assert(q->_next == p, "invariant");
1271       q->_next = p->_next;
1272     }
1273   }
1274 
1275 #ifdef ASSERT
1276   // Diagnostic hygiene ...
1277   currentNode->_prev  = (ObjectWaiter*) 0xBAD;
1278   currentNode->_next  = (ObjectWaiter*) 0xBAD;
1279   currentNode->TState = ObjectWaiter::TS_RUN;
1280 #endif
1281 }
1282 
1283 // Fix this. Save ObjectWaiter* when freezing. Or use hashtable.
1284 ObjectWaiter* ObjectMonitor::LookupWaiter(int64_t threadid) {
1285   ObjectWaiter* p;
1286   for (p = _EntryList; p != nullptr && (!p->is_vthread() || java_lang_Thread::thread_id(p->vthread()) != threadid); p = p->_next) {}
1287   if (p != nullptr) return p;
1288   for (p = _cxq; p != nullptr && (!p->is_vthread() || java_lang_Thread::thread_id(p->vthread()) != threadid); p = p->_next) {}
1289   assert(p != nullptr, "should be on either _cxq or _EntryList");
1290   return p;
1291 }
1292 
1293 // -----------------------------------------------------------------------------
1294 // Exit support
1295 //
1296 // exit()
1297 // ~~~~~~
1298 // Note that the collector can't reclaim the objectMonitor or deflate
1299 // the object out from underneath the thread calling ::exit() as the
1300 // thread calling ::exit() never transitions to a stable state.
1301 // This inhibits GC, which in turn inhibits asynchronous (and
1302 // inopportune) reclamation of "this".
1303 //
1304 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1305 // There's one exception to the claim above, however.  EnterI() can call
1306 // exit() to drop a lock if the acquirer has been externally suspended.
1307 // In that case exit() is called with _thread_state == _thread_blocked,
1308 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1309 //
1310 // 1-0 exit
1311 // ~~~~~~~~
1312 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1313 // the fast-path operators have been optimized so the common ::exit()
1314 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1315 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
1316 // greatly improves latency -- MEMBAR and CAS having considerable local
1317 // latency on modern processors -- but at the cost of "stranding".  Absent the
1318 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1319 // ::enter() path, resulting in the entering thread being stranding
1320 // and a progress-liveness failure.   Stranding is extremely rare.
1321 // We use timers (timed park operations) & periodic polling to detect
1322 // and recover from stranding.  Potentially stranded threads periodically
1323 // wake up and poll the lock.  See the usage of the _Responsible variable.
1324 //
1325 // The CAS() in enter provides for safety and exclusion, while the CAS or
1326 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
1327 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1328 // We detect and recover from stranding with timers.
1329 //
1330 // If a thread transiently strands it'll park until (a) another
1331 // thread acquires the lock and then drops the lock, at which time the
1332 // exiting thread will notice and unpark the stranded thread, or, (b)
1333 // the timer expires.  If the lock is high traffic then the stranding latency
1334 // will be low due to (a).  If the lock is low traffic then the odds of
1335 // stranding are lower, although the worst-case stranding latency
1336 // is longer.  Critically, we don't want to put excessive load in the
1337 // platform's timer subsystem.  We want to minimize both the timer injection
1338 // rate (timers created/sec) as well as the number of timers active at
1339 // any one time.  (more precisely, we want to minimize timer-seconds, which is
1340 // the integral of the # of active timers at any instant over time).
1341 // Both impinge on OS scalability.  Given that, at most one thread parked on
1342 // a monitor will use a timer.
1343 //
1344 // There is also the risk of a futile wake-up. If we drop the lock
1345 // another thread can reacquire the lock immediately, and we can
1346 // then wake a thread unnecessarily. This is benign, and we've
1347 // structured the code so the windows are short and the frequency
1348 // of such futile wakups is low.
1349 
1350 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1351   void* cur = owner_raw();
1352   if (owner_for(current) != cur) {
1353     // Apparent unbalanced locking ...
1354     // Naively we'd like to throw IllegalMonitorStateException.
1355     // As a practical matter we can neither allocate nor throw an
1356     // exception as ::exit() can be called from leaf routines.
1357     // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1358     // Upon deeper reflection, however, in a properly run JVM the only
1359     // way we should encounter this situation is in the presence of
1360     // unbalanced JNI locking. TODO: CheckJNICalls.
1361     // See also: CR4414101





1362 #ifdef ASSERT
1363     LogStreamHandle(Error, monitorinflation) lsh;
1364     lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1365                   " is exiting an ObjectMonitor it does not own.", p2i(current));
1366     lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1367     print_debug_style_on(&lsh);
1368     assert(false, "Non-balanced monitor enter/exit!");
1369 #endif
1370     return;

1371   }
1372 
1373   if (_recursions != 0) {
1374     _recursions--;        // this is simple recursive enter
1375     return;
1376   }
1377 
1378   // Invariant: after setting Responsible=null an thread must execute
1379   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1380   _Responsible = nullptr;
1381 
1382 #if INCLUDE_JFR
1383   // get the owner's thread id for the MonitorEnter event
1384   // if it is enabled and the thread isn't suspended
1385   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1386     _previous_owner_tid = JFR_THREAD_ID(current);
1387   }
1388 #endif
1389 
1390   for (;;) {
1391     assert(owner_for(current) == owner_raw(), "invariant");
1392 
1393     // Drop the lock.
1394     // release semantics: prior loads and stores from within the critical section
1395     // must not float (reorder) past the following store that drops the lock.
1396     // Uses a storeload to separate release_store(owner) from the
1397     // successor check. The try_set_owner_from() below uses cmpxchg() so
1398     // we get the fence down there.
1399     release_clear_owner(current);
1400     OrderAccess::storeload();
1401 
1402     if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) {
1403       return;
1404     }
1405     // Other threads are blocked trying to acquire the lock.
1406 
1407     // Normally the exiting thread is responsible for ensuring succession,
1408     // but if other successors are ready or other entering threads are spinning
1409     // then this thread can simply store null into _owner and exit without
1410     // waking a successor.  The existence of spinners or ready successors
1411     // guarantees proper succession (liveness).  Responsibility passes to the
1412     // ready or running successors.  The exiting thread delegates the duty.
1413     // More precisely, if a successor already exists this thread is absolved
1414     // of the responsibility of waking (unparking) one.
1415     //
1416     // The _succ variable is critical to reducing futile wakeup frequency.
1417     // _succ identifies the "heir presumptive" thread that has been made
1418     // ready (unparked) but that has not yet run.  We need only one such
1419     // successor thread to guarantee progress.
1420     // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1421     // section 3.3 "Futile Wakeup Throttling" for details.
1422     //
1423     // Note that spinners in Enter() also set _succ non-null.
1424     // In the current implementation spinners opportunistically set
1425     // _succ so that exiting threads might avoid waking a successor.
1426     // Another less appealing alternative would be for the exiting thread
1427     // to drop the lock and then spin briefly to see if a spinner managed
1428     // to acquire the lock.  If so, the exiting thread could exit
1429     // immediately without waking a successor, otherwise the exiting
1430     // thread would need to dequeue and wake a successor.
1431     // (Note that we'd need to make the post-drop spin short, but no
1432     // shorter than the worst-case round-trip cache-line migration time.
1433     // The dropped lock needs to become visible to the spinner, and then
1434     // the acquisition of the lock by the spinner must become visible to
1435     // the exiting thread).
1436 
1437     // It appears that an heir-presumptive (successor) must be made ready.
1438     // Only the current lock owner can manipulate the EntryList or
1439     // drain _cxq, so we need to reacquire the lock.  If we fail
1440     // to reacquire the lock the responsibility for ensuring succession
1441     // falls to the new owner.
1442     //
1443     if (try_set_owner_from(nullptr, current) != nullptr) {
1444       return;
1445     }
1446 
1447     guarantee(owner_raw() == owner_for(current), "invariant");
1448 
1449     ObjectWaiter* w = nullptr;
1450 
1451     w = _EntryList;
1452     if (w != nullptr) {
1453       // I'd like to write: guarantee (w->_thread != current).
1454       // But in practice an exiting thread may find itself on the EntryList.
1455       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1456       // then calls exit().  Exit release the lock by setting O._owner to null.
1457       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1458       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1459       // release the lock "O".  T2 resumes immediately after the ST of null into
1460       // _owner, above.  T2 notices that the EntryList is populated, so it
1461       // reacquires the lock and then finds itself on the EntryList.
1462       // Given all that, we have to tolerate the circumstance where "w" is
1463       // associated with current.
1464       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1465       ExitEpilog(current, w);
1466       return;
1467     }
1468 
1469     // If we find that both _cxq and EntryList are null then just
1470     // re-run the exit protocol from the top.
1471     w = _cxq;
1472     if (w == nullptr) continue;
1473 
1474     // Drain _cxq into EntryList - bulk transfer.
1475     // First, detach _cxq.
1476     // The following loop is tantamount to: w = swap(&cxq, nullptr)
1477     for (;;) {
1478       assert(w != nullptr, "Invariant");
1479       ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr);
1480       if (u == w) break;
1481       w = u;
1482     }
1483 
1484     assert(w != nullptr, "invariant");
1485     assert(_EntryList == nullptr, "invariant");
1486 
1487     // Convert the LIFO SLL anchored by _cxq into a DLL.
1488     // The list reorganization step operates in O(LENGTH(w)) time.
1489     // It's critical that this step operate quickly as
1490     // "current" still holds the outer-lock, restricting parallelism
1491     // and effectively lengthening the critical section.
1492     // Invariant: s chases t chases u.
1493     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1494     // we have faster access to the tail.
1495 
1496     _EntryList = w;
1497     ObjectWaiter* q = nullptr;
1498     ObjectWaiter* p;
1499     for (p = w; p != nullptr; p = p->_next) {
1500       guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1501       p->TState = ObjectWaiter::TS_ENTER;
1502       p->_prev = q;
1503       q = p;
1504     }
1505 
1506     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr
1507     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1508 
1509     // See if we can abdicate to a spinner instead of waking a thread.
1510     // A primary goal of the implementation is to reduce the
1511     // context-switch rate.
1512     if (_succ != nullptr) continue;
1513 
1514     w = _EntryList;
1515     if (w != nullptr) {
1516       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1517       ExitEpilog(current, w);
1518       return;
1519     }
1520   }
1521 }
1522 
1523 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1524   assert(owner_raw() == owner_for(current), "invariant");
1525 
1526   // Exit protocol:
1527   // 1. ST _succ = wakee
1528   // 2. membar #loadstore|#storestore;
1529   // 2. ST _owner = nullptr
1530   // 3. unpark(wakee)
1531 
1532   oop vthread = nullptr;
1533   if (Wakee->_thread != nullptr) {
1534     // Platform thread case
1535     _succ = Wakee->_thread;
1536   } else {
1537     assert(Wakee->vthread() != nullptr, "invariant");
1538     vthread = Wakee->vthread();
1539     _succ = (JavaThread*)java_lang_Thread::thread_id(vthread);
1540   }
1541   ParkEvent * Trigger = Wakee->_event;
1542 
1543   // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again.
1544   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1545   // out-of-scope (non-extant).
1546   Wakee  = nullptr;
1547 
1548   // Drop the lock.
1549   // Uses a fence to separate release_store(owner) from the LD in unpark().
1550   release_clear_owner(current);
1551   OrderAccess::fence();
1552 
1553   DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1554 
1555   if (vthread == nullptr) {
1556     // Platform thread case
1557     Trigger->unpark();
1558   } else if (java_lang_VirtualThread::set_onWaitingList(vthread, _vthread_cxq_head)) {
1559     Trigger->unpark();
1560   }
1561 
1562   // Maintain stats and report events to JVMTI
1563   OM_PERFDATA_OP(Parks, inc());
1564 }
1565 
1566 // complete_exit exits a lock returning recursion count
1567 // complete_exit requires an inflated monitor
1568 // The _owner field is not always the Thread addr even with an
1569 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1570 // thread due to contention.
1571 intx ObjectMonitor::complete_exit(JavaThread* current) {
1572   assert(InitDone, "Unexpectedly not initialized");
1573 
1574   void* cur = owner_raw();
1575   if (owner_for(current) != cur) {
1576     if (LockingMode == LM_LEGACY && is_stack_locker(current)) {
1577       assert(_recursions == 0, "internal state error");
1578       set_owner_from_BasicLock(current);  // Convert from BasicLock* to Thread*.
1579       _recursions = 0;
1580     }
1581   }
1582 
1583   guarantee(owner_for(current) == owner_raw(), "complete_exit not owner");
1584   intx save = _recursions; // record the old recursion count
1585   _recursions = 0;         // set the recursion level to be 0
1586   exit(current);           // exit the monitor
1587   guarantee(owner_raw() != owner_for(current), "invariant");
1588   return save;
1589 }
1590 
1591 // Checks that the current THREAD owns this monitor and causes an
1592 // immediate return if it doesn't. We don't use the CHECK macro
1593 // because we want the IMSE to be the only exception that is thrown
1594 // from the call site when false is returned. Any other pending
1595 // exception is ignored.
1596 #define CHECK_OWNER()                                                  \
1597   do {                                                                 \
1598     if (!check_owner(THREAD)) {                                        \
1599        assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1600        return;                                                         \
1601      }                                                                 \
1602   } while (false)
1603 
1604 // Returns true if the specified thread owns the ObjectMonitor.
1605 // Otherwise returns false and throws IllegalMonitorStateException
1606 // (IMSE). If there is a pending exception and the specified thread
1607 // is not the owner, that exception will be replaced by the IMSE.
1608 bool ObjectMonitor::check_owner(TRAPS) {
1609   JavaThread* current = THREAD;
1610   void* cur = owner_raw();
1611   if (cur == owner_for(current)) {






1612     return true;
1613   }
1614   THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1615              "current thread is not owner", false);
1616 }
1617 
1618 static inline bool is_excluded(const Klass* monitor_klass) {
1619   assert(monitor_klass != nullptr, "invariant");
1620   NOT_JFR_RETURN_(false);
1621   JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name();)
1622 }
1623 
1624 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1625                                     ObjectMonitor* monitor,
1626                                     uint64_t notifier_tid,
1627                                     jlong timeout,
1628                                     bool timedout) {
1629   assert(event != nullptr, "invariant");
1630   assert(monitor != nullptr, "invariant");
1631   const Klass* monitor_klass = monitor->object()->klass();
1632   if (is_excluded(monitor_klass)) {
1633     return;
1634   }
1635   event->set_monitorClass(monitor_klass);
1636   event->set_timeout(timeout);
1637   // Set an address that is 'unique enough', such that events close in
1638   // time and with the same address are likely (but not guaranteed) to
1639   // belong to the same object.
1640   event->set_address((uintptr_t)monitor);
1641   event->set_notifier(notifier_tid);
1642   event->set_timedOut(timedout);
1643   event->commit();
1644 }
1645 
1646 // -----------------------------------------------------------------------------
1647 // Wait/Notify/NotifyAll
1648 //
1649 // Note: a subset of changes to ObjectMonitor::wait()
1650 // will need to be replicated in complete_exit
1651 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1652   JavaThread* current = THREAD;
1653 
1654   assert(InitDone, "Unexpectedly not initialized");
1655 
1656   CHECK_OWNER();  // Throws IMSE if not owner.
1657 
1658   ContinuationEntry* ce = current->last_continuation();
1659   if (ce != nullptr && ce->is_virtual_thread()) {
1660     EventVirtualThreadPinned e;
1661     if (e.should_commit()) {
1662       e.commit();
1663     }
1664   }
1665 
1666   EventJavaMonitorWait event;
1667 
1668   // check for a pending interrupt
1669   if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1670     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1671     if (JvmtiExport::should_post_monitor_waited()) {
1672       // Note: 'false' parameter is passed here because the
1673       // wait was not timed out due to thread interrupt.
1674       JvmtiExport::post_monitor_waited(current, this, false);
1675 
1676       // In this short circuit of the monitor wait protocol, the
1677       // current thread never drops ownership of the monitor and
1678       // never gets added to the wait queue so the current thread
1679       // cannot be made the successor. This means that the
1680       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1681       // consume an unpark() meant for the ParkEvent associated with
1682       // this ObjectMonitor.
1683     }
1684     if (event.should_commit()) {
1685       post_monitor_wait_event(&event, this, 0, millis, false);
1686     }
1687     THROW(vmSymbols::java_lang_InterruptedException());
1688     return;
1689   }
1690 
1691   current->set_current_waiting_monitor(this);
1692 
1693   // create a node to be put into the queue
1694   // Critically, after we reset() the event but prior to park(), we must check
1695   // for a pending interrupt.
1696   ObjectWaiter node(current);
1697   node.TState = ObjectWaiter::TS_WAIT;
1698   current->_ParkEvent->reset();
1699   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1700 
1701   // Enter the waiting queue, which is a circular doubly linked list in this case
1702   // but it could be a priority queue or any data structure.
1703   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1704   // by the owner of the monitor *except* in the case where park()
1705   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1706   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1707 
1708   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1709   AddWaiter(&node);
1710   Thread::SpinRelease(&_WaitSetLock);
1711 
1712   _Responsible = nullptr;
1713 
1714   intx save = _recursions;     // record the old recursion count
1715   _waiters++;                  // increment the number of waiters
1716   _recursions = 0;             // set the recursion level to be 1
1717   exit(current);               // exit the monitor
1718   guarantee(owner_raw() != owner_for(current), "invariant");
1719 
1720   // The thread is on the WaitSet list - now park() it.
1721   // On MP systems it's conceivable that a brief spin before we park
1722   // could be profitable.
1723   //
1724   // TODO-FIXME: change the following logic to a loop of the form
1725   //   while (!timeout && !interrupted && _notified == 0) park()
1726 
1727   int ret = OS_OK;
1728   int WasNotified = 0;
1729 
1730   // Need to check interrupt state whilst still _thread_in_vm
1731   bool interrupted = interruptible && current->is_interrupted(false);
1732 
1733   { // State transition wrappers
1734     OSThread* osthread = current->osthread();
1735     OSThreadWaitState osts(osthread, true);
1736 
1737     assert(current->thread_state() == _thread_in_vm, "invariant");
1738 
1739     {
1740       ClearSuccOnSuspend csos(this);
1741       ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1742       if (interrupted || HAS_PENDING_EXCEPTION) {
1743         // Intentionally empty
1744       } else if (node._notified == 0) {
1745         if (millis <= 0) {
1746           current->_ParkEvent->park();
1747         } else {
1748           ret = current->_ParkEvent->park(millis);
1749         }
1750       }
1751     }
1752 
1753     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1754     // from the WaitSet to the EntryList.
1755     // See if we need to remove Node from the WaitSet.
1756     // We use double-checked locking to avoid grabbing _WaitSetLock
1757     // if the thread is not on the wait queue.
1758     //
1759     // Note that we don't need a fence before the fetch of TState.
1760     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1761     // written by the is thread. (perhaps the fetch might even be satisfied
1762     // by a look-aside into the processor's own store buffer, although given
1763     // the length of the code path between the prior ST and this load that's
1764     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1765     // then we'll acquire the lock and then re-fetch a fresh TState value.
1766     // That is, we fail toward safety.
1767 
1768     if (node.TState == ObjectWaiter::TS_WAIT) {
1769       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1770       if (node.TState == ObjectWaiter::TS_WAIT) {
1771         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1772         assert(node._notified == 0, "invariant");
1773         node.TState = ObjectWaiter::TS_RUN;
1774       }
1775       Thread::SpinRelease(&_WaitSetLock);
1776     }
1777 
1778     // The thread is now either on off-list (TS_RUN),
1779     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1780     // The Node's TState variable is stable from the perspective of this thread.
1781     // No other threads will asynchronously modify TState.
1782     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1783     OrderAccess::loadload();
1784     if (_succ == current) _succ = nullptr;
1785     WasNotified = node._notified;
1786 
1787     // Reentry phase -- reacquire the monitor.
1788     // re-enter contended monitor after object.wait().
1789     // retain OBJECT_WAIT state until re-enter successfully completes
1790     // Thread state is thread_in_vm and oop access is again safe,
1791     // although the raw address of the object may have changed.
1792     // (Don't cache naked oops over safepoints, of course).
1793 
1794     // post monitor waited event. Note that this is past-tense, we are done waiting.
1795     if (JvmtiExport::should_post_monitor_waited()) {
1796       JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1797 
1798       if (node._notified != 0 && _succ == current) {
1799         // In this part of the monitor wait-notify-reenter protocol it
1800         // is possible (and normal) for another thread to do a fastpath
1801         // monitor enter-exit while this thread is still trying to get
1802         // to the reenter portion of the protocol.
1803         //
1804         // The ObjectMonitor was notified and the current thread is
1805         // the successor which also means that an unpark() has already
1806         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1807         // consume the unpark() that was done when the successor was
1808         // set because the same ParkEvent is shared between Java
1809         // monitors and JVM/TI RawMonitors (for now).
1810         //
1811         // We redo the unpark() to ensure forward progress, i.e., we
1812         // don't want all pending threads hanging (parked) with none
1813         // entering the unlocked monitor.
1814         node._event->unpark();
1815       }
1816     }
1817 
1818     if (event.should_commit()) {
1819       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1820     }
1821 
1822     OrderAccess::fence();
1823 
1824     assert(owner_raw() != owner_for(current), "invariant");
1825     ObjectWaiter::TStates v = node.TState;
1826     if (v == ObjectWaiter::TS_RUN) {
1827       enter(current);
1828     } else {
1829       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1830       ReenterI(current, &node);
1831       node.wait_reenter_end(this);
1832     }
1833 
1834     // current has reacquired the lock.
1835     // Lifecycle - the node representing current must not appear on any queues.
1836     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1837     // want residual elements associated with this thread left on any lists.
1838     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1839     assert(owner_raw() == owner_for(current), "invariant");
1840     assert(_succ != current, "invariant");
1841   } // OSThreadWaitState()
1842 
1843   current->set_current_waiting_monitor(nullptr);
1844 
1845   guarantee(_recursions == 0, "invariant");
1846   int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current);
1847   _recursions =   save          // restore the old recursion count
1848                 + relock_count; //  increased by the deferred relock count
1849   NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count(relock_count);) // Deopt never entered these counts.
1850   _waiters--;             // decrement the number of waiters
1851 
1852   // Verify a few postconditions
1853   assert(owner_raw() == owner_for(current), "invariant");
1854   assert(_succ != current, "invariant");
1855   assert(object()->mark() == markWord::encode(this), "invariant");
1856 
1857   // check if the notification happened
1858   if (!WasNotified) {
1859     // no, it could be timeout or Thread.interrupt() or both
1860     // check for interrupt event, otherwise it is timeout
1861     if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1862       THROW(vmSymbols::java_lang_InterruptedException());
1863     }
1864   }
1865 
1866   // NOTE: Spurious wake up will be consider as timeout.
1867   // Monitor notify has precedence over thread interrupt.
1868 }
1869 
1870 
1871 // Consider:
1872 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1873 // then instead of transferring a thread from the WaitSet to the EntryList
1874 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1875 
1876 void ObjectMonitor::INotify(JavaThread* current) {
1877   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1878   ObjectWaiter* iterator = DequeueWaiter();
1879   if (iterator != nullptr) {
1880     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1881     guarantee(iterator->_notified == 0, "invariant");
1882     // Disposition - what might we do with iterator ?
1883     // a.  add it directly to the EntryList - either tail (policy == 1)
1884     //     or head (policy == 0).
1885     // b.  push it onto the front of the _cxq (policy == 2).
1886     // For now we use (b).
1887 
1888     iterator->TState = ObjectWaiter::TS_ENTER;
1889 
1890     iterator->_notified = 1;
1891     iterator->_notifier_tid = JFR_THREAD_ID(current);
1892 
1893     ObjectWaiter* list = _EntryList;
1894     if (list != nullptr) {
1895       assert(list->_prev == nullptr, "invariant");
1896       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1897       assert(list != iterator, "invariant");
1898     }
1899 
1900     // prepend to cxq
1901     if (list == nullptr) {
1902       iterator->_next = iterator->_prev = nullptr;
1903       _EntryList = iterator;
1904     } else {
1905       iterator->TState = ObjectWaiter::TS_CXQ;
1906       for (;;) {
1907         ObjectWaiter* front = _cxq;
1908         iterator->_next = front;
1909         if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1910           break;
1911         }
1912       }
1913     }
1914 
1915     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1916     // move the add-to-EntryList operation, above, outside the critical section
1917     // protected by _WaitSetLock.  In practice that's not useful.  With the
1918     // exception of  wait() timeouts and interrupts the monitor owner
1919     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1920     // on _WaitSetLock so it's not profitable to reduce the length of the
1921     // critical section.
1922 
1923     iterator->wait_reenter_begin(this);
1924   }
1925   Thread::SpinRelease(&_WaitSetLock);
1926 }
1927 
1928 // Consider: a not-uncommon synchronization bug is to use notify() when
1929 // notifyAll() is more appropriate, potentially resulting in stranded
1930 // threads; this is one example of a lost wakeup. A useful diagnostic
1931 // option is to force all notify() operations to behave as notifyAll().
1932 //
1933 // Note: We can also detect many such problems with a "minimum wait".
1934 // When the "minimum wait" is set to a small non-zero timeout value
1935 // and the program does not hang whereas it did absent "minimum wait",
1936 // that suggests a lost wakeup bug.
1937 
1938 void ObjectMonitor::notify(TRAPS) {
1939   JavaThread* current = THREAD;
1940   CHECK_OWNER();  // Throws IMSE if not owner.
1941   if (_WaitSet == nullptr) {
1942     return;
1943   }
1944   DTRACE_MONITOR_PROBE(notify, this, object(), current);
1945   INotify(current);
1946   OM_PERFDATA_OP(Notifications, inc(1));
1947 }
1948 
1949 
1950 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1951 // from the waitset to the EntryList. This could be done more efficiently with a
1952 // single bulk transfer but in practice it's not time-critical. Beware too,
1953 // that in prepend-mode we invert the order of the waiters. Let's say that the
1954 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1955 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1956 
1957 void ObjectMonitor::notifyAll(TRAPS) {
1958   JavaThread* current = THREAD;
1959   CHECK_OWNER();  // Throws IMSE if not owner.
1960   if (_WaitSet == nullptr) {
1961     return;
1962   }
1963 
1964   DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1965   int tally = 0;
1966   while (_WaitSet != nullptr) {
1967     tally++;
1968     INotify(current);
1969   }
1970 
1971   OM_PERFDATA_OP(Notifications, inc(tally));
1972 }
1973 
1974 // -----------------------------------------------------------------------------
1975 // Adaptive Spinning Support
1976 //
1977 // Adaptive spin-then-block - rational spinning
1978 //
1979 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1980 // algorithm.  On high order SMP systems it would be better to start with
1981 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1982 // a contending thread could enqueue itself on the cxq and then spin locally
1983 // on a thread-specific variable such as its ParkEvent._Event flag.
1984 // That's left as an exercise for the reader.  Note that global spinning is
1985 // not problematic on Niagara, as the L2 cache serves the interconnect and
1986 // has both low latency and massive bandwidth.
1987 //
1988 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1989 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1990 // (duration) or we can fix the count at approximately the duration of
1991 // a context switch and vary the frequency.   Of course we could also
1992 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1993 // For a description of 'Adaptive spin-then-block mutual exclusion in
1994 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1995 //
1996 // This implementation varies the duration "D", where D varies with
1997 // the success rate of recent spin attempts. (D is capped at approximately
1998 // length of a round-trip context switch).  The success rate for recent
1999 // spin attempts is a good predictor of the success rate of future spin
2000 // attempts.  The mechanism adapts automatically to varying critical
2001 // section length (lock modality), system load and degree of parallelism.
2002 // D is maintained per-monitor in _SpinDuration and is initialized
2003 // optimistically.  Spin frequency is fixed at 100%.
2004 //
2005 // Note that _SpinDuration is volatile, but we update it without locks
2006 // or atomics.  The code is designed so that _SpinDuration stays within
2007 // a reasonable range even in the presence of races.  The arithmetic
2008 // operations on _SpinDuration are closed over the domain of legal values,
2009 // so at worst a race will install and older but still legal value.
2010 // At the very worst this introduces some apparent non-determinism.
2011 // We might spin when we shouldn't or vice-versa, but since the spin
2012 // count are relatively short, even in the worst case, the effect is harmless.
2013 //
2014 // Care must be taken that a low "D" value does not become an
2015 // an absorbing state.  Transient spinning failures -- when spinning
2016 // is overall profitable -- should not cause the system to converge
2017 // on low "D" values.  We want spinning to be stable and predictable
2018 // and fairly responsive to change and at the same time we don't want
2019 // it to oscillate, become metastable, be "too" non-deterministic,
2020 // or converge on or enter undesirable stable absorbing states.
2021 //
2022 // We implement a feedback-based control system -- using past behavior
2023 // to predict future behavior.  We face two issues: (a) if the
2024 // input signal is random then the spin predictor won't provide optimal
2025 // results, and (b) if the signal frequency is too high then the control
2026 // system, which has some natural response lag, will "chase" the signal.
2027 // (b) can arise from multimodal lock hold times.  Transient preemption
2028 // can also result in apparent bimodal lock hold times.
2029 // Although sub-optimal, neither condition is particularly harmful, as
2030 // in the worst-case we'll spin when we shouldn't or vice-versa.
2031 // The maximum spin duration is rather short so the failure modes aren't bad.
2032 // To be conservative, I've tuned the gain in system to bias toward
2033 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
2034 // "rings" or oscillates between spinning and not spinning.  This happens
2035 // when spinning is just on the cusp of profitability, however, so the
2036 // situation is not dire.  The state is benign -- there's no need to add
2037 // hysteresis control to damp the transition rate between spinning and
2038 // not spinning.
2039 
2040 // Spinning: Fixed frequency (100%), vary duration
2041 int ObjectMonitor::TrySpin(JavaThread* current) {
2042   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
2043   int ctr = Knob_FixedSpin;
2044   if (ctr != 0) {
2045     while (--ctr >= 0) {
2046       if (TryLock(current) > 0) return 1;
2047       SpinPause();
2048     }
2049     return 0;
2050   }
2051 
2052   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
2053     if (TryLock(current) > 0) {
2054       // Increase _SpinDuration ...
2055       // Note that we don't clamp SpinDuration precisely at SpinLimit.
2056       // Raising _SpurDuration to the poverty line is key.
2057       int x = _SpinDuration;
2058       if (x < Knob_SpinLimit) {
2059         if (x < Knob_Poverty) x = Knob_Poverty;
2060         _SpinDuration = x + Knob_BonusB;
2061       }
2062       return 1;
2063     }
2064     SpinPause();
2065   }
2066 
2067   // Admission control - verify preconditions for spinning
2068   //
2069   // We always spin a little bit, just to prevent _SpinDuration == 0 from
2070   // becoming an absorbing state.  Put another way, we spin briefly to
2071   // sample, just in case the system load, parallelism, contention, or lock
2072   // modality changed.
2073   //
2074   // Consider the following alternative:
2075   // Periodically set _SpinDuration = _SpinLimit and try a long/full
2076   // spin attempt.  "Periodically" might mean after a tally of
2077   // the # of failed spin attempts (or iterations) reaches some threshold.
2078   // This takes us into the realm of 1-out-of-N spinning, where we
2079   // hold the duration constant but vary the frequency.
2080 
2081   ctr = _SpinDuration;
2082   if (ctr <= 0) return 0;
2083 
2084   // We're good to spin ... spin ingress.
2085   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
2086   // when preparing to LD...CAS _owner, etc and the CAS is likely
2087   // to succeed.
2088   if (_succ == nullptr) {
2089     _succ = current;
2090   }
2091   void* prv = nullptr;
2092 
2093   // There are three ways to exit the following loop:
2094   // 1.  A successful spin where this thread has acquired the lock.
2095   // 2.  Spin failure with prejudice
2096   // 3.  Spin failure without prejudice
2097 
2098   while (--ctr >= 0) {
2099 
2100     // Periodic polling -- Check for pending GC
2101     // Threads may spin while they're unsafe.
2102     // We don't want spinning threads to delay the JVM from reaching
2103     // a stop-the-world safepoint or to steal cycles from GC.
2104     // If we detect a pending safepoint we abort in order that
2105     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
2106     // this thread, if safe, doesn't steal cycles from GC.
2107     // This is in keeping with the "no loitering in runtime" rule.
2108     // We periodically check to see if there's a safepoint pending.
2109     if ((ctr & 0xFF) == 0) {
2110       // Can't call SafepointMechanism::should_process() since that
2111       // might update the poll values and we could be in a thread_blocked
2112       // state here which is not allowed so just check the poll.
2113       if (SafepointMechanism::local_poll_armed(current)) {
2114         goto Abort;           // abrupt spin egress
2115       }
2116       SpinPause();
2117     }
2118 
2119     // Probe _owner with TATAS
2120     // If this thread observes the monitor transition or flicker
2121     // from locked to unlocked to locked, then the odds that this
2122     // thread will acquire the lock in this spin attempt go down
2123     // considerably.  The same argument applies if the CAS fails
2124     // or if we observe _owner change from one non-null value to
2125     // another non-null value.   In such cases we might abort
2126     // the spin without prejudice or apply a "penalty" to the
2127     // spin count-down variable "ctr", reducing it by 100, say.
2128 
2129     void* ox = owner_raw();
2130     if (ox == nullptr) {
2131       ox = try_set_owner_from(nullptr, current);
2132       if (ox == nullptr) {
2133         // The CAS succeeded -- this thread acquired ownership
2134         // Take care of some bookkeeping to exit spin state.
2135         if (_succ == current) {
2136           _succ = nullptr;
2137         }
2138 
2139         // Increase _SpinDuration :
2140         // The spin was successful (profitable) so we tend toward
2141         // longer spin attempts in the future.
2142         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2143         // If we acquired the lock early in the spin cycle it
2144         // makes sense to increase _SpinDuration proportionally.
2145         // Note that we don't clamp SpinDuration precisely at SpinLimit.
2146         int x = _SpinDuration;
2147         if (x < Knob_SpinLimit) {
2148           if (x < Knob_Poverty) x = Knob_Poverty;
2149           _SpinDuration = x + Knob_Bonus;
2150         }
2151         return 1;
2152       }
2153 
2154       // The CAS failed ... we can take any of the following actions:
2155       // * penalize: ctr -= CASPenalty
2156       // * exit spin with prejudice -- goto Abort;
2157       // * exit spin without prejudice.
2158       // * Since CAS is high-latency, retry again immediately.
2159       prv = ox;
2160       goto Abort;
2161     }
2162 
2163     // Did lock ownership change hands ?
2164     if (ox != prv && prv != nullptr) {
2165       goto Abort;
2166     }
2167     prv = ox;
2168 
2169     if (_succ == nullptr) {
2170       _succ = current;
2171     }
2172   }
2173 
2174   // Spin failed with prejudice -- reduce _SpinDuration.
2175   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2176   // AIMD is globally stable.
2177   {
2178     int x = _SpinDuration;
2179     if (x > 0) {
2180       // Consider an AIMD scheme like: x -= (x >> 3) + 100
2181       // This is globally sample and tends to damp the response.
2182       x -= Knob_Penalty;
2183       if (x < 0) x = 0;
2184       _SpinDuration = x;
2185     }
2186   }
2187 
2188  Abort:
2189   if (_succ == current) {
2190     _succ = nullptr;
2191     // Invariant: after setting succ=null a contending thread
2192     // must recheck-retry _owner before parking.  This usually happens
2193     // in the normal usage of TrySpin(), but it's safest
2194     // to make TrySpin() as foolproof as possible.
2195     OrderAccess::fence();
2196     if (TryLock(current) > 0) return 1;
2197   }
2198   return 0;
2199 }
2200 
2201 
2202 // -----------------------------------------------------------------------------
2203 // WaitSet management ...
2204 
2205 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2206   _next     = nullptr;
2207   _prev     = nullptr;
2208   _notified = 0;
2209   _notifier_tid = 0;
2210   TState    = TS_RUN;
2211   _thread   = current;
2212   _event    = _thread != nullptr ? _thread->_ParkEvent : ObjectMonitor::vthread_unparker_ParkEvent();
2213   _active   = false;
2214   assert(_event != nullptr, "invariant");
2215 }
2216 
2217 ObjectWaiter::ObjectWaiter(oop vthread) : ObjectWaiter((JavaThread*)nullptr) {
2218   _vthread = OopHandle(JavaThread::thread_oop_storage(), vthread);
2219 }
2220 
2221 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2222   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2223 }
2224 
2225 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2226   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2227 }
2228 
2229 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2230   assert(node != nullptr, "should not add null node");
2231   assert(node->_prev == nullptr, "node already in list");
2232   assert(node->_next == nullptr, "node already in list");
2233   // put node at end of queue (circular doubly linked list)
2234   if (_WaitSet == nullptr) {
2235     _WaitSet = node;
2236     node->_prev = node;
2237     node->_next = node;
2238   } else {
2239     ObjectWaiter* head = _WaitSet;
2240     ObjectWaiter* tail = head->_prev;
2241     assert(tail->_next == head, "invariant check");
2242     tail->_next = node;
2243     head->_prev = node;
2244     node->_next = head;
2245     node->_prev = tail;
2246   }
2247 }
2248 
2249 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2250   // dequeue the very first waiter
2251   ObjectWaiter* waiter = _WaitSet;
2252   if (waiter) {
2253     DequeueSpecificWaiter(waiter);
2254   }
2255   return waiter;
2256 }
2257 
2258 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2259   assert(node != nullptr, "should not dequeue nullptr node");
2260   assert(node->_prev != nullptr, "node already removed from list");
2261   assert(node->_next != nullptr, "node already removed from list");
2262   // when the waiter has woken up because of interrupt,
2263   // timeout or other spurious wake-up, dequeue the
2264   // waiter from waiting list
2265   ObjectWaiter* next = node->_next;
2266   if (next == node) {
2267     assert(node->_prev == node, "invariant check");
2268     _WaitSet = nullptr;
2269   } else {
2270     ObjectWaiter* prev = node->_prev;
2271     assert(prev->_next == node, "invariant check");
2272     assert(next->_prev == node, "invariant check");
2273     next->_prev = prev;
2274     prev->_next = next;
2275     if (_WaitSet == node) {
2276       _WaitSet = next;
2277     }
2278   }
2279   node->_next = nullptr;
2280   node->_prev = nullptr;
2281 }
2282 
2283 // -----------------------------------------------------------------------------
2284 // PerfData support
2285 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = nullptr;
2286 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = nullptr;
2287 PerfCounter * ObjectMonitor::_sync_Parks                       = nullptr;
2288 PerfCounter * ObjectMonitor::_sync_Notifications               = nullptr;
2289 PerfCounter * ObjectMonitor::_sync_Inflations                  = nullptr;
2290 PerfCounter * ObjectMonitor::_sync_Deflations                  = nullptr;
2291 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = nullptr;
2292 
2293 // One-shot global initialization for the sync subsystem.
2294 // We could also defer initialization and initialize on-demand
2295 // the first time we call ObjectSynchronizer::inflate().
2296 // Initialization would be protected - like so many things - by
2297 // the MonitorCache_lock.
2298 
2299 void ObjectMonitor::Initialize() {
2300   assert(!InitDone, "invariant");
2301 
2302   if (!os::is_MP()) {
2303     Knob_SpinLimit = 0;
2304     Knob_PreSpin   = 0;
2305     Knob_FixedSpin = -1;
2306   }
2307 
2308   if (UsePerfData) {
2309     EXCEPTION_MARK;
2310 #define NEWPERFCOUNTER(n)                                                \
2311   {                                                                      \
2312     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2313                                         CHECK);                          \
2314   }
2315 #define NEWPERFVARIABLE(n)                                                \
2316   {                                                                       \
2317     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2318                                          CHECK);                          \
2319   }
2320     NEWPERFCOUNTER(_sync_Inflations);
2321     NEWPERFCOUNTER(_sync_Deflations);
2322     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2323     NEWPERFCOUNTER(_sync_FutileWakeups);
2324     NEWPERFCOUNTER(_sync_Parks);
2325     NEWPERFCOUNTER(_sync_Notifications);
2326     NEWPERFVARIABLE(_sync_MonExtant);
2327 #undef NEWPERFCOUNTER
2328 #undef NEWPERFVARIABLE
2329   }
2330 
2331   _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2332 
2333   DEBUG_ONLY(InitDone = true;)
2334 }
2335 
2336 void ObjectMonitor::Initialize2() {
2337   _vthread_cxq_head = OopHandle(JavaThread::thread_oop_storage(), nullptr);
2338   _vthread_unparker_ParkEvent = ParkEvent::Allocate(nullptr);
2339 }
2340 
2341 void ObjectMonitor::print_on(outputStream* st) const {
2342   // The minimal things to print for markWord printing, more can be added for debugging and logging.
2343   st->print("{contentions=0x%08x,waiters=0x%08x"
2344             ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2345             contentions(), waiters(), recursions(),
2346             p2i(owner()));
2347 }
2348 void ObjectMonitor::print() const { print_on(tty); }
2349 
2350 #ifdef ASSERT
2351 // Print the ObjectMonitor like a debugger would:
2352 //
2353 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2354 //   _header = 0x0000000000000001
2355 //   _object = 0x000000070ff45fd0
2356 //   _pad_buf0 = {
2357 //     [0] = '\0'
2358 //     ...
2359 //     [43] = '\0'
2360 //   }
2361 //   _owner = 0x0000000000000000
2362 //   _previous_owner_tid = 0
2363 //   _pad_buf1 = {
2364 //     [0] = '\0'
2365 //     ...
2366 //     [47] = '\0'
2367 //   }
2368 //   _next_om = 0x0000000000000000
2369 //   _recursions = 0
2370 //   _EntryList = 0x0000000000000000
2371 //   _cxq = 0x0000000000000000
2372 //   _succ = 0x0000000000000000
2373 //   _Responsible = 0x0000000000000000
2374 //   _SpinDuration = 5000
2375 //   _contentions = 0
2376 //   _WaitSet = 0x0000700009756248
2377 //   _waiters = 1
2378 //   _WaitSetLock = 0
2379 // }
2380 //
2381 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2382   st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2383   st->print_cr("  _header = " INTPTR_FORMAT, header().value());
2384   st->print_cr("  _object = " INTPTR_FORMAT, p2i(object_peek()));
2385   st->print_cr("  _pad_buf0 = {");
2386   st->print_cr("    [0] = '\\0'");
2387   st->print_cr("    ...");
2388   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2389   st->print_cr("  }");
2390   st->print_cr("  _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2391   st->print_cr("  _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid);
2392   st->print_cr("  _pad_buf1 = {");
2393   st->print_cr("    [0] = '\\0'");
2394   st->print_cr("    ...");
2395   st->print_cr("    [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2396   st->print_cr("  }");
2397   st->print_cr("  _next_om = " INTPTR_FORMAT, p2i(next_om()));
2398   st->print_cr("  _recursions = " INTX_FORMAT, _recursions);
2399   st->print_cr("  _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2400   st->print_cr("  _cxq = " INTPTR_FORMAT, p2i(_cxq));
2401   st->print_cr("  _succ = " INTPTR_FORMAT, p2i(_succ));
2402   st->print_cr("  _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2403   st->print_cr("  _SpinDuration = %d", _SpinDuration);
2404   st->print_cr("  _contentions = %d", contentions());
2405   st->print_cr("  _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2406   st->print_cr("  _waiters = %d", _waiters);
2407   st->print_cr("  _WaitSetLock = %d", _WaitSetLock);
2408   st->print_cr("}");
2409 }
2410 #endif
--- EOF ---