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