1 /*
   2  * Copyright (c) 1998, 2012, 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 "memory/resourceArea.hpp"
  28 #include "oops/markOop.hpp"
  29 #include "oops/oop.inline.hpp"
  30 #include "runtime/handles.inline.hpp"
  31 #include "runtime/interfaceSupport.hpp"
  32 #include "runtime/mutexLocker.hpp"
  33 #include "runtime/objectMonitor.hpp"
  34 #include "runtime/objectMonitor.inline.hpp"
  35 #include "runtime/osThread.hpp"
  36 #include "runtime/stubRoutines.hpp"
  37 #include "runtime/thread.inline.hpp"
  38 #include "services/threadService.hpp"
  39 #include "utilities/dtrace.hpp"
  40 #include "utilities/preserveException.hpp"
  41 #ifdef TARGET_OS_FAMILY_linux
  42 # include "os_linux.inline.hpp"
  43 #endif
  44 #ifdef TARGET_OS_FAMILY_solaris
  45 # include "os_solaris.inline.hpp"
  46 #endif
  47 #ifdef TARGET_OS_FAMILY_windows
  48 # include "os_windows.inline.hpp"
  49 #endif
  50 #ifdef TARGET_OS_FAMILY_bsd
  51 # include "os_bsd.inline.hpp"
  52 #endif
  53 
  54 #if defined(__GNUC__) && !defined(IA64)
  55   // Need to inhibit inlining for older versions of GCC to avoid build-time failures
  56   #define ATTR __attribute__((noinline))
  57 #else
  58   #define ATTR
  59 #endif
  60 
  61 
  62 #ifdef DTRACE_ENABLED
  63 
  64 // Only bother with this argument setup if dtrace is available
  65 // TODO-FIXME: probes should not fire when caller is _blocked.  assert() accordingly.
  66 
  67 
  68 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread)                           \
  69   char* bytes = NULL;                                                      \
  70   int len = 0;                                                             \
  71   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  72   Symbol* klassname = ((oop)obj)->klass()->name();                         \
  73   if (klassname != NULL) {                                                 \
  74     bytes = (char*)klassname->bytes();                                     \
  75     len = klassname->utf8_length();                                        \
  76   }
  77 
  78 #ifndef USDT2
  79 
  80 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify,
  81   jlong, uintptr_t, char*, int);
  82 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll,
  83   jlong, uintptr_t, char*, int);
  84 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter,
  85   jlong, uintptr_t, char*, int);
  86 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered,
  87   jlong, uintptr_t, char*, int);
  88 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit,
  89   jlong, uintptr_t, char*, int);
  90 
  91 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)       \
  92   {                                                                        \
  93     if (DTraceMonitorProbes) {                                            \
  94       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                       \
  95       HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid,                       \
  96                        (monitor), bytes, len, (millis));                   \
  97     }                                                                      \
  98   }
  99 
 100 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)             \
 101   {                                                                        \
 102     if (DTraceMonitorProbes) {                                            \
 103       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                       \
 104       HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid,                    \
 105                        (uintptr_t)(monitor), bytes, len);                  \
 106     }                                                                      \
 107   }
 108 
 109 #else /* USDT2 */
 110 
 111 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)            \
 112   {                                                                        \
 113     if (DTraceMonitorProbes) {                                            \
 114       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
 115       HOTSPOT_MONITOR_WAIT(jtid,                                           \
 116                        (monitor), bytes, len, (millis));                   \
 117     }                                                                      \
 118   }
 119 
 120 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
 121 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
 122 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
 123 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
 124 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
 125 
 126 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)                  \
 127   {                                                                        \
 128     if (DTraceMonitorProbes) {                                            \
 129       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
 130       HOTSPOT_MONITOR_##probe(jtid,                                               \
 131                        (uintptr_t)(monitor), bytes, len);                  \
 132     }                                                                      \
 133   }
 134 
 135 #endif /* USDT2 */
 136 #else //  ndef DTRACE_ENABLED
 137 
 138 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon)    {;}
 139 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon)          {;}
 140 
 141 #endif // ndef DTRACE_ENABLED
 142 
 143 // Tunables ...
 144 // The knob* variables are effectively final.  Once set they should
 145 // never be modified hence.  Consider using __read_mostly with GCC.
 146 
 147 int ObjectMonitor::Knob_Verbose    = 0 ;
 148 int ObjectMonitor::Knob_SpinLimit  = 5000 ;    // derived by an external tool -
 149 static int Knob_LogSpins           = 0 ;       // enable jvmstat tally for spins
 150 static int Knob_HandOff            = 0 ;
 151 static int Knob_ReportSettings     = 0 ;
 152 
 153 static int Knob_SpinBase           = 0 ;       // Floor AKA SpinMin
 154 static int Knob_SpinBackOff        = 0 ;       // spin-loop backoff
 155 static int Knob_CASPenalty         = -1 ;      // Penalty for failed CAS
 156 static int Knob_OXPenalty          = -1 ;      // Penalty for observed _owner change
 157 static int Knob_SpinSetSucc        = 1 ;       // spinners set the _succ field
 158 static int Knob_SpinEarly          = 1 ;
 159 static int Knob_SuccEnabled        = 1 ;       // futile wake throttling
 160 static int Knob_SuccRestrict       = 0 ;       // Limit successors + spinners to at-most-one
 161 static int Knob_MaxSpinners        = -1 ;      // Should be a function of # CPUs
 162 static int Knob_Bonus              = 100 ;     // spin success bonus
 163 static int Knob_BonusB             = 100 ;     // spin success bonus
 164 static int Knob_Penalty            = 200 ;     // spin failure penalty
 165 static int Knob_Poverty            = 1000 ;
 166 static int Knob_SpinAfterFutile    = 1 ;       // Spin after returning from park()
 167 static int Knob_FixedSpin          = 0 ;
 168 static int Knob_OState             = 3 ;       // Spinner checks thread state of _owner
 169 static int Knob_UsePause           = 1 ;
 170 static int Knob_ExitPolicy         = 0 ;
 171 static int Knob_PreSpin            = 10 ;      // 20-100 likely better
 172 static int Knob_ResetEvent         = 0 ;
 173 static int BackOffMask             = 0 ;
 174 
 175 static int Knob_FastHSSEC          = 0 ;
 176 static int Knob_MoveNotifyee       = 2 ;       // notify() - disposition of notifyee
 177 static int Knob_QMode              = 0 ;       // EntryList-cxq policy - queue discipline
 178 static volatile int InitDone       = 0 ;
 179 
 180 #define TrySpin TrySpin_VaryDuration
 181 
 182 // -----------------------------------------------------------------------------
 183 // Theory of operations -- Monitors lists, thread residency, etc:
 184 //
 185 // * A thread acquires ownership of a monitor by successfully
 186 //   CAS()ing the _owner field from null to non-null.
 187 //
 188 // * Invariant: A thread appears on at most one monitor list --
 189 //   cxq, EntryList or WaitSet -- at any one time.
 190 //
 191 // * Contending threads "push" themselves onto the cxq with CAS
 192 //   and then spin/park.
 193 //
 194 // * After a contending thread eventually acquires the lock it must
 195 //   dequeue itself from either the EntryList or the cxq.
 196 //
 197 // * The exiting thread identifies and unparks an "heir presumptive"
 198 //   tentative successor thread on the EntryList.  Critically, the
 199 //   exiting thread doesn't unlink the successor thread from the EntryList.
 200 //   After having been unparked, the wakee will recontend for ownership of
 201 //   the monitor.   The successor (wakee) will either acquire the lock or
 202 //   re-park itself.
 203 //
 204 //   Succession is provided for by a policy of competitive handoff.
 205 //   The exiting thread does _not_ grant or pass ownership to the
 206 //   successor thread.  (This is also referred to as "handoff" succession").
 207 //   Instead the exiting thread releases ownership and possibly wakes
 208 //   a successor, so the successor can (re)compete for ownership of the lock.
 209 //   If the EntryList is empty but the cxq is populated the exiting
 210 //   thread will drain the cxq into the EntryList.  It does so by
 211 //   by detaching the cxq (installing null with CAS) and folding
 212 //   the threads from the cxq into the EntryList.  The EntryList is
 213 //   doubly linked, while the cxq is singly linked because of the
 214 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 215 //
 216 // * Concurrency invariants:
 217 //
 218 //   -- only the monitor owner may access or mutate the EntryList.
 219 //      The mutex property of the monitor itself protects the EntryList
 220 //      from concurrent interference.
 221 //   -- Only the monitor owner may detach the cxq.
 222 //
 223 // * The monitor entry list operations avoid locks, but strictly speaking
 224 //   they're not lock-free.  Enter is lock-free, exit is not.
 225 //   See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html
 226 //
 227 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 228 //   detaching thread.  This mechanism is immune from the ABA corruption.
 229 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 230 //
 231 // * Taken together, the cxq and the EntryList constitute or form a
 232 //   single logical queue of threads stalled trying to acquire the lock.
 233 //   We use two distinct lists to improve the odds of a constant-time
 234 //   dequeue operation after acquisition (in the ::enter() epilog) and
 235 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 236 //   A key desideratum is to minimize queue & monitor metadata manipulation
 237 //   that occurs while holding the monitor lock -- that is, we want to
 238 //   minimize monitor lock holds times.  Note that even a small amount of
 239 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 240 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 241 //   locks and monitor metadata.
 242 //
 243 //   Cxq points to the the set of Recently Arrived Threads attempting entry.
 244 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 245 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 246 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 247 //
 248 //   The EntryList is ordered by the prevailing queue discipline and
 249 //   can be organized in any convenient fashion, such as a doubly-linked list or
 250 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 251 //   to operate in constant-time.  If we need a priority queue then something akin
 252 //   to Solaris' sleepq would work nicely.  Viz.,
 253 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 254 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 255 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 256 //   EntryList accordingly.
 257 //
 258 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 259 //   somewhat similar to an elevator-scan.
 260 //
 261 // * The monitor synchronization subsystem avoids the use of native
 262 //   synchronization primitives except for the narrow platform-specific
 263 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 264 //   the semantics of park-unpark.  Put another way, this monitor implementation
 265 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 266 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 267 //   underlying OS manages the READY<->RUN transitions.
 268 //
 269 // * Waiting threads reside on the WaitSet list -- wait() puts
 270 //   the caller onto the WaitSet.
 271 //
 272 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 273 //   either the EntryList or cxq.  Subsequent exit() operations will
 274 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 275 //   it's likely the notifyee would simply impale itself on the lock held
 276 //   by the notifier.
 277 //
 278 // * An interesting alternative is to encode cxq as (List,LockByte) where
 279 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 280 //   variable, like _recursions, in the scheme.  The threads or Events that form
 281 //   the list would have to be aligned in 256-byte addresses.  A thread would
 282 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 283 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 284 //   Note that is is *not* word-tearing, but it does presume that full-word
 285 //   CAS operations are coherent with intermix with STB operations.  That's true
 286 //   on most common processors.
 287 //
 288 // * See also http://blogs.sun.com/dave
 289 
 290 
 291 // -----------------------------------------------------------------------------
 292 // Enter support
 293 
 294 bool ObjectMonitor::try_enter(Thread* THREAD) {
 295   if (THREAD != _owner) {
 296     if (THREAD->is_lock_owned ((address)_owner)) {
 297        assert(_recursions == 0, "internal state error");
 298        _owner = THREAD ;
 299        _recursions = 1 ;
 300        OwnerIsThread = 1 ;
 301        return true;
 302     }
 303     if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
 304       return false;
 305     }
 306     return true;
 307   } else {
 308     _recursions++;
 309     return true;
 310   }
 311 }
 312 
 313 void ATTR ObjectMonitor::enter(TRAPS) {
 314   // The following code is ordered to check the most common cases first
 315   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 316   Thread * const Self = THREAD ;
 317   void * cur ;
 318 
 319   cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ;
 320   if (cur == NULL) {
 321      // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
 322      assert (_recursions == 0   , "invariant") ;
 323      assert (_owner      == Self, "invariant") ;
 324      // CONSIDER: set or assert OwnerIsThread == 1
 325      return ;
 326   }
 327 
 328   if (cur == Self) {
 329      // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 330      _recursions ++ ;
 331      return ;
 332   }
 333 
 334   if (Self->is_lock_owned ((address)cur)) {
 335     assert (_recursions == 0, "internal state error");
 336     _recursions = 1 ;
 337     // Commute owner from a thread-specific on-stack BasicLockObject address to
 338     // a full-fledged "Thread *".
 339     _owner = Self ;
 340     OwnerIsThread = 1 ;
 341     return ;
 342   }
 343 
 344   // We've encountered genuine contention.
 345   assert (Self->_Stalled == 0, "invariant") ;
 346   Self->_Stalled = intptr_t(this) ;
 347 
 348   // Try one round of spinning *before* enqueueing Self
 349   // and before going through the awkward and expensive state
 350   // transitions.  The following spin is strictly optional ...
 351   // Note that if we acquire the monitor from an initial spin
 352   // we forgo posting JVMTI events and firing DTRACE probes.
 353   if (Knob_SpinEarly && TrySpin (Self) > 0) {
 354      assert (_owner == Self      , "invariant") ;
 355      assert (_recursions == 0    , "invariant") ;
 356      assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 357      Self->_Stalled = 0 ;
 358      return ;
 359   }
 360 
 361   assert (_owner != Self          , "invariant") ;
 362   assert (_succ  != Self          , "invariant") ;
 363   assert (Self->is_Java_thread()  , "invariant") ;
 364   JavaThread * jt = (JavaThread *) Self ;
 365   assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
 366   assert (jt->thread_state() != _thread_blocked   , "invariant") ;
 367   assert (this->object() != NULL  , "invariant") ;
 368   assert (_count >= 0, "invariant") ;
 369 
 370   // Prevent deflation at STW-time.  See deflate_idle_monitors() and is_busy().
 371   // Ensure the object-monitor relationship remains stable while there's contention.
 372   Atomic::inc_ptr(&_count);
 373 
 374   { // Change java thread status to indicate blocked on monitor enter.
 375     JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
 376 
 377     DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
 378     if (JvmtiExport::should_post_monitor_contended_enter()) {
 379       JvmtiExport::post_monitor_contended_enter(jt, this);
 380     }
 381 
 382     OSThreadContendState osts(Self->osthread());
 383     ThreadBlockInVM tbivm(jt);
 384 
 385     Self->set_current_pending_monitor(this);
 386 
 387     // TODO-FIXME: change the following for(;;) loop to straight-line code.
 388     for (;;) {
 389       jt->set_suspend_equivalent();
 390       // cleared by handle_special_suspend_equivalent_condition()
 391       // or java_suspend_self()
 392 
 393       EnterI (THREAD) ;
 394 
 395       if (!ExitSuspendEquivalent(jt)) break ;
 396 
 397       //
 398       // We have acquired the contended monitor, but while we were
 399       // waiting another thread suspended us. We don't want to enter
 400       // the monitor while suspended because that would surprise the
 401       // thread that suspended us.
 402       //
 403           _recursions = 0 ;
 404       _succ = NULL ;
 405       exit (Self) ;
 406 
 407       jt->java_suspend_self();
 408     }
 409     Self->set_current_pending_monitor(NULL);
 410   }
 411 
 412   Atomic::dec_ptr(&_count);
 413   assert (_count >= 0, "invariant") ;
 414   Self->_Stalled = 0 ;
 415 
 416   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 417   assert (_recursions == 0     , "invariant") ;
 418   assert (_owner == Self       , "invariant") ;
 419   assert (_succ  != Self       , "invariant") ;
 420   assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 421 
 422   // The thread -- now the owner -- is back in vm mode.
 423   // Report the glorious news via TI,DTrace and jvmstat.
 424   // The probe effect is non-trivial.  All the reportage occurs
 425   // while we hold the monitor, increasing the length of the critical
 426   // section.  Amdahl's parallel speedup law comes vividly into play.
 427   //
 428   // Another option might be to aggregate the events (thread local or
 429   // per-monitor aggregation) and defer reporting until a more opportune
 430   // time -- such as next time some thread encounters contention but has
 431   // yet to acquire the lock.  While spinning that thread could
 432   // spinning we could increment JVMStat counters, etc.
 433 
 434   DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
 435   if (JvmtiExport::should_post_monitor_contended_entered()) {
 436     JvmtiExport::post_monitor_contended_entered(jt, this);
 437   }
 438   if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) {
 439      ObjectMonitor::_sync_ContendedLockAttempts->inc() ;
 440   }
 441 }
 442 
 443 
 444 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 445 // Callers must compensate as needed.
 446 
 447 int ObjectMonitor::TryLock (Thread * Self) {
 448    for (;;) {
 449       void * own = _owner ;
 450       if (own != NULL) return 0 ;
 451       if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) {
 452          // Either guarantee _recursions == 0 or set _recursions = 0.
 453          assert (_recursions == 0, "invariant") ;
 454          assert (_owner == Self, "invariant") ;
 455          // CONSIDER: set or assert that OwnerIsThread == 1
 456          return 1 ;
 457       }
 458       // The lock had been free momentarily, but we lost the race to the lock.
 459       // Interference -- the CAS failed.
 460       // We can either return -1 or retry.
 461       // Retry doesn't make as much sense because the lock was just acquired.
 462       if (true) return -1 ;
 463    }
 464 }
 465 
 466 void ATTR ObjectMonitor::EnterI (TRAPS) {
 467     Thread * Self = THREAD ;
 468     assert (Self->is_Java_thread(), "invariant") ;
 469     assert (((JavaThread *) Self)->thread_state() == _thread_blocked   , "invariant") ;
 470 
 471     // Try the lock - TATAS
 472     if (TryLock (Self) > 0) {
 473         assert (_succ != Self              , "invariant") ;
 474         assert (_owner == Self             , "invariant") ;
 475         assert (_Responsible != Self       , "invariant") ;
 476         return ;
 477     }
 478 
 479     DeferredInitialize () ;
 480 
 481     // We try one round of spinning *before* enqueueing Self.
 482     //
 483     // If the _owner is ready but OFFPROC we could use a YieldTo()
 484     // operation to donate the remainder of this thread's quantum
 485     // to the owner.  This has subtle but beneficial affinity
 486     // effects.
 487 
 488     if (TrySpin (Self) > 0) {
 489         assert (_owner == Self        , "invariant") ;
 490         assert (_succ != Self         , "invariant") ;
 491         assert (_Responsible != Self  , "invariant") ;
 492         return ;
 493     }
 494 
 495     // The Spin failed -- Enqueue and park the thread ...
 496     assert (_succ  != Self            , "invariant") ;
 497     assert (_owner != Self            , "invariant") ;
 498     assert (_Responsible != Self      , "invariant") ;
 499 
 500     // Enqueue "Self" on ObjectMonitor's _cxq.
 501     //
 502     // Node acts as a proxy for Self.
 503     // As an aside, if were to ever rewrite the synchronization code mostly
 504     // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 505     // Java objects.  This would avoid awkward lifecycle and liveness issues,
 506     // as well as eliminate a subset of ABA issues.
 507     // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 508     //
 509 
 510     ObjectWaiter node(Self) ;
 511     Self->_ParkEvent->reset() ;
 512     node._prev   = (ObjectWaiter *) 0xBAD ;
 513     node.TState  = ObjectWaiter::TS_CXQ ;
 514 
 515     // Push "Self" onto the front of the _cxq.
 516     // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
 517     // Note that spinning tends to reduce the rate at which threads
 518     // enqueue and dequeue on EntryList|cxq.
 519     ObjectWaiter * nxt ;
 520     for (;;) {
 521         node._next = nxt = _cxq ;
 522         if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ;
 523 
 524         // Interference - the CAS failed because _cxq changed.  Just retry.
 525         // As an optional optimization we retry the lock.
 526         if (TryLock (Self) > 0) {
 527             assert (_succ != Self         , "invariant") ;
 528             assert (_owner == Self        , "invariant") ;
 529             assert (_Responsible != Self  , "invariant") ;
 530             return ;
 531         }
 532     }
 533 
 534     // Check for cxq|EntryList edge transition to non-null.  This indicates
 535     // the onset of contention.  While contention persists exiting threads
 536     // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 537     // operations revert to the faster 1-0 mode.  This enter operation may interleave
 538     // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 539     // arrange for one of the contending thread to use a timed park() operations
 540     // to detect and recover from the race.  (Stranding is form of progress failure
 541     // where the monitor is unlocked but all the contending threads remain parked).
 542     // That is, at least one of the contended threads will periodically poll _owner.
 543     // One of the contending threads will become the designated "Responsible" thread.
 544     // The Responsible thread uses a timed park instead of a normal indefinite park
 545     // operation -- it periodically wakes and checks for and recovers from potential
 546     // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 547     // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 548     // be responsible for a monitor.
 549     //
 550     // Currently, one of the contended threads takes on the added role of "Responsible".
 551     // A viable alternative would be to use a dedicated "stranding checker" thread
 552     // that periodically iterated over all the threads (or active monitors) and unparked
 553     // successors where there was risk of stranding.  This would help eliminate the
 554     // timer scalability issues we see on some platforms as we'd only have one thread
 555     // -- the checker -- parked on a timer.
 556 
 557     if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) {
 558         // Try to assume the role of responsible thread for the monitor.
 559         // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=Self }
 560         Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ;
 561     }
 562 
 563     // The lock have been released while this thread was occupied queueing
 564     // itself onto _cxq.  To close the race and avoid "stranding" and
 565     // progress-liveness failure we must resample-retry _owner before parking.
 566     // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 567     // In this case the ST-MEMBAR is accomplished with CAS().
 568     //
 569     // TODO: Defer all thread state transitions until park-time.
 570     // Since state transitions are heavy and inefficient we'd like
 571     // to defer the state transitions until absolutely necessary,
 572     // and in doing so avoid some transitions ...
 573 
 574     TEVENT (Inflated enter - Contention) ;
 575     int nWakeups = 0 ;
 576     int RecheckInterval = 1 ;
 577 
 578     for (;;) {
 579 
 580         if (TryLock (Self) > 0) break ;
 581         assert (_owner != Self, "invariant") ;
 582 
 583         if ((SyncFlags & 2) && _Responsible == NULL) {
 584            Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ;
 585         }
 586 
 587         // park self
 588         if (_Responsible == Self || (SyncFlags & 1)) {
 589             TEVENT (Inflated enter - park TIMED) ;
 590             Self->_ParkEvent->park ((jlong) RecheckInterval) ;
 591             // Increase the RecheckInterval, but clamp the value.
 592             RecheckInterval *= 8 ;
 593             if (RecheckInterval > 1000) RecheckInterval = 1000 ;
 594         } else {
 595             TEVENT (Inflated enter - park UNTIMED) ;
 596             Self->_ParkEvent->park() ;
 597         }
 598 
 599         if (TryLock(Self) > 0) break ;
 600 
 601         // The lock is still contested.
 602         // Keep a tally of the # of futile wakeups.
 603         // Note that the counter is not protected by a lock or updated by atomics.
 604         // That is by design - we trade "lossy" counters which are exposed to
 605         // races during updates for a lower probe effect.
 606         TEVENT (Inflated enter - Futile wakeup) ;
 607         if (ObjectMonitor::_sync_FutileWakeups != NULL) {
 608            ObjectMonitor::_sync_FutileWakeups->inc() ;
 609         }
 610         ++ nWakeups ;
 611 
 612         // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
 613         // We can defer clearing _succ until after the spin completes
 614         // TrySpin() must tolerate being called with _succ == Self.
 615         // Try yet another round of adaptive spinning.
 616         if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ;
 617 
 618         // We can find that we were unpark()ed and redesignated _succ while
 619         // we were spinning.  That's harmless.  If we iterate and call park(),
 620         // park() will consume the event and return immediately and we'll
 621         // just spin again.  This pattern can repeat, leaving _succ to simply
 622         // spin on a CPU.  Enable Knob_ResetEvent to clear pending unparks().
 623         // Alternately, we can sample fired() here, and if set, forgo spinning
 624         // in the next iteration.
 625 
 626         if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
 627            Self->_ParkEvent->reset() ;
 628            OrderAccess::fence() ;
 629         }
 630         if (_succ == Self) _succ = NULL ;
 631 
 632         // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 633         OrderAccess::fence() ;
 634     }
 635 
 636     // Egress :
 637     // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
 638     // Normally we'll find Self on the EntryList .
 639     // From the perspective of the lock owner (this thread), the
 640     // EntryList is stable and cxq is prepend-only.
 641     // The head of cxq is volatile but the interior is stable.
 642     // In addition, Self.TState is stable.
 643 
 644     assert (_owner == Self      , "invariant") ;
 645     assert (object() != NULL    , "invariant") ;
 646     // I'd like to write:
 647     //   guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 648     // but as we're at a safepoint that's not safe.
 649 
 650     UnlinkAfterAcquire (Self, &node) ;
 651     if (_succ == Self) _succ = NULL ;
 652 
 653     assert (_succ != Self, "invariant") ;
 654     if (_Responsible == Self) {
 655         _Responsible = NULL ;
 656         OrderAccess::fence(); // Dekker pivot-point
 657 
 658         // We may leave threads on cxq|EntryList without a designated
 659         // "Responsible" thread.  This is benign.  When this thread subsequently
 660         // exits the monitor it can "see" such preexisting "old" threads --
 661         // threads that arrived on the cxq|EntryList before the fence, above --
 662         // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 663         // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 664         // non-null and elect a new "Responsible" timer thread.
 665         //
 666         // This thread executes:
 667         //    ST Responsible=null; MEMBAR    (in enter epilog - here)
 668         //    LD cxq|EntryList               (in subsequent exit)
 669         //
 670         // Entering threads in the slow/contended path execute:
 671         //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 672         //    The (ST cxq; MEMBAR) is accomplished with CAS().
 673         //
 674         // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 675         // exit operation from floating above the ST Responsible=null.
 676     }
 677 
 678     // We've acquired ownership with CAS().
 679     // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 680     // But since the CAS() this thread may have also stored into _succ,
 681     // EntryList, cxq or Responsible.  These meta-data updates must be
 682     // visible __before this thread subsequently drops the lock.
 683     // Consider what could occur if we didn't enforce this constraint --
 684     // STs to monitor meta-data and user-data could reorder with (become
 685     // visible after) the ST in exit that drops ownership of the lock.
 686     // Some other thread could then acquire the lock, but observe inconsistent
 687     // or old monitor meta-data and heap data.  That violates the JMM.
 688     // To that end, the 1-0 exit() operation must have at least STST|LDST
 689     // "release" barrier semantics.  Specifically, there must be at least a
 690     // STST|LDST barrier in exit() before the ST of null into _owner that drops
 691     // the lock.   The barrier ensures that changes to monitor meta-data and data
 692     // protected by the lock will be visible before we release the lock, and
 693     // therefore before some other thread (CPU) has a chance to acquire the lock.
 694     // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 695     //
 696     // Critically, any prior STs to _succ or EntryList must be visible before
 697     // the ST of null into _owner in the *subsequent* (following) corresponding
 698     // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 699     // execute a serializing instruction.
 700 
 701     if (SyncFlags & 8) {
 702        OrderAccess::fence() ;
 703     }
 704     return ;
 705 }
 706 
 707 // ReenterI() is a specialized inline form of the latter half of the
 708 // contended slow-path from EnterI().  We use ReenterI() only for
 709 // monitor reentry in wait().
 710 //
 711 // In the future we should reconcile EnterI() and ReenterI(), adding
 712 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
 713 // loop accordingly.
 714 
 715 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) {
 716     assert (Self != NULL                , "invariant") ;
 717     assert (SelfNode != NULL            , "invariant") ;
 718     assert (SelfNode->_thread == Self   , "invariant") ;
 719     assert (_waiters > 0                , "invariant") ;
 720     assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ;
 721     assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
 722     JavaThread * jt = (JavaThread *) Self ;
 723 
 724     int nWakeups = 0 ;
 725     for (;;) {
 726         ObjectWaiter::TStates v = SelfNode->TState ;
 727         guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ;
 728         assert    (_owner != Self, "invariant") ;
 729 
 730         if (TryLock (Self) > 0) break ;
 731         if (TrySpin (Self) > 0) break ;
 732 
 733         TEVENT (Wait Reentry - parking) ;
 734 
 735         // State transition wrappers around park() ...
 736         // ReenterI() wisely defers state transitions until
 737         // it's clear we must park the thread.
 738         {
 739            OSThreadContendState osts(Self->osthread());
 740            ThreadBlockInVM tbivm(jt);
 741 
 742            // cleared by handle_special_suspend_equivalent_condition()
 743            // or java_suspend_self()
 744            jt->set_suspend_equivalent();
 745            if (SyncFlags & 1) {
 746               Self->_ParkEvent->park ((jlong)1000) ;
 747            } else {
 748               Self->_ParkEvent->park () ;
 749            }
 750 
 751            // were we externally suspended while we were waiting?
 752            for (;;) {
 753               if (!ExitSuspendEquivalent (jt)) break ;
 754               if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
 755               jt->java_suspend_self();
 756               jt->set_suspend_equivalent();
 757            }
 758         }
 759 
 760         // Try again, but just so we distinguish between futile wakeups and
 761         // successful wakeups.  The following test isn't algorithmically
 762         // necessary, but it helps us maintain sensible statistics.
 763         if (TryLock(Self) > 0) break ;
 764 
 765         // The lock is still contested.
 766         // Keep a tally of the # of futile wakeups.
 767         // Note that the counter is not protected by a lock or updated by atomics.
 768         // That is by design - we trade "lossy" counters which are exposed to
 769         // races during updates for a lower probe effect.
 770         TEVENT (Wait Reentry - futile wakeup) ;
 771         ++ nWakeups ;
 772 
 773         // Assuming this is not a spurious wakeup we'll normally
 774         // find that _succ == Self.
 775         if (_succ == Self) _succ = NULL ;
 776 
 777         // Invariant: after clearing _succ a contending thread
 778         // *must* retry  _owner before parking.
 779         OrderAccess::fence() ;
 780 
 781         if (ObjectMonitor::_sync_FutileWakeups != NULL) {
 782           ObjectMonitor::_sync_FutileWakeups->inc() ;
 783         }
 784     }
 785 
 786     // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
 787     // Normally we'll find Self on the EntryList.
 788     // Unlinking from the EntryList is constant-time and atomic-free.
 789     // From the perspective of the lock owner (this thread), the
 790     // EntryList is stable and cxq is prepend-only.
 791     // The head of cxq is volatile but the interior is stable.
 792     // In addition, Self.TState is stable.
 793 
 794     assert (_owner == Self, "invariant") ;
 795     assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 796     UnlinkAfterAcquire (Self, SelfNode) ;
 797     if (_succ == Self) _succ = NULL ;
 798     assert (_succ != Self, "invariant") ;
 799     SelfNode->TState = ObjectWaiter::TS_RUN ;
 800     OrderAccess::fence() ;      // see comments at the end of EnterI()
 801 }
 802 
 803 // after the thread acquires the lock in ::enter().  Equally, we could defer
 804 // unlinking the thread until ::exit()-time.
 805 
 806 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode)
 807 {
 808     assert (_owner == Self, "invariant") ;
 809     assert (SelfNode->_thread == Self, "invariant") ;
 810 
 811     if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
 812         // Normal case: remove Self from the DLL EntryList .
 813         // This is a constant-time operation.
 814         ObjectWaiter * nxt = SelfNode->_next ;
 815         ObjectWaiter * prv = SelfNode->_prev ;
 816         if (nxt != NULL) nxt->_prev = prv ;
 817         if (prv != NULL) prv->_next = nxt ;
 818         if (SelfNode == _EntryList ) _EntryList = nxt ;
 819         assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ;
 820         assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ;
 821         TEVENT (Unlink from EntryList) ;
 822     } else {
 823         guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ;
 824         // Inopportune interleaving -- Self is still on the cxq.
 825         // This usually means the enqueue of self raced an exiting thread.
 826         // Normally we'll find Self near the front of the cxq, so
 827         // dequeueing is typically fast.  If needbe we can accelerate
 828         // this with some MCS/CHL-like bidirectional list hints and advisory
 829         // back-links so dequeueing from the interior will normally operate
 830         // in constant-time.
 831         // Dequeue Self from either the head (with CAS) or from the interior
 832         // with a linear-time scan and normal non-atomic memory operations.
 833         // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
 834         // and then unlink Self from EntryList.  We have to drain eventually,
 835         // so it might as well be now.
 836 
 837         ObjectWaiter * v = _cxq ;
 838         assert (v != NULL, "invariant") ;
 839         if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) {
 840             // The CAS above can fail from interference IFF a "RAT" arrived.
 841             // In that case Self must be in the interior and can no longer be
 842             // at the head of cxq.
 843             if (v == SelfNode) {
 844                 assert (_cxq != v, "invariant") ;
 845                 v = _cxq ;          // CAS above failed - start scan at head of list
 846             }
 847             ObjectWaiter * p ;
 848             ObjectWaiter * q = NULL ;
 849             for (p = v ; p != NULL && p != SelfNode; p = p->_next) {
 850                 q = p ;
 851                 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ;
 852             }
 853             assert (v != SelfNode,  "invariant") ;
 854             assert (p == SelfNode,  "Node not found on cxq") ;
 855             assert (p != _cxq,      "invariant") ;
 856             assert (q != NULL,      "invariant") ;
 857             assert (q->_next == p,  "invariant") ;
 858             q->_next = p->_next ;
 859         }
 860         TEVENT (Unlink from cxq) ;
 861     }
 862 
 863     // Diagnostic hygiene ...
 864     SelfNode->_prev  = (ObjectWaiter *) 0xBAD ;
 865     SelfNode->_next  = (ObjectWaiter *) 0xBAD ;
 866     SelfNode->TState = ObjectWaiter::TS_RUN ;
 867 }
 868 
 869 // -----------------------------------------------------------------------------
 870 // Exit support
 871 //
 872 // exit()
 873 // ~~~~~~
 874 // Note that the collector can't reclaim the objectMonitor or deflate
 875 // the object out from underneath the thread calling ::exit() as the
 876 // thread calling ::exit() never transitions to a stable state.
 877 // This inhibits GC, which in turn inhibits asynchronous (and
 878 // inopportune) reclamation of "this".
 879 //
 880 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
 881 // There's one exception to the claim above, however.  EnterI() can call
 882 // exit() to drop a lock if the acquirer has been externally suspended.
 883 // In that case exit() is called with _thread_state as _thread_blocked,
 884 // but the monitor's _count field is > 0, which inhibits reclamation.
 885 //
 886 // 1-0 exit
 887 // ~~~~~~~~
 888 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
 889 // the fast-path operators have been optimized so the common ::exit()
 890 // operation is 1-0.  See i486.ad fast_unlock(), for instance.
 891 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
 892 // greatly improves latency -- MEMBAR and CAS having considerable local
 893 // latency on modern processors -- but at the cost of "stranding".  Absent the
 894 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
 895 // ::enter() path, resulting in the entering thread being stranding
 896 // and a progress-liveness failure.   Stranding is extremely rare.
 897 // We use timers (timed park operations) & periodic polling to detect
 898 // and recover from stranding.  Potentially stranded threads periodically
 899 // wake up and poll the lock.  See the usage of the _Responsible variable.
 900 //
 901 // The CAS() in enter provides for safety and exclusion, while the CAS or
 902 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
 903 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding.
 904 // We detect and recover from stranding with timers.
 905 //
 906 // If a thread transiently strands it'll park until (a) another
 907 // thread acquires the lock and then drops the lock, at which time the
 908 // exiting thread will notice and unpark the stranded thread, or, (b)
 909 // the timer expires.  If the lock is high traffic then the stranding latency
 910 // will be low due to (a).  If the lock is low traffic then the odds of
 911 // stranding are lower, although the worst-case stranding latency
 912 // is longer.  Critically, we don't want to put excessive load in the
 913 // platform's timer subsystem.  We want to minimize both the timer injection
 914 // rate (timers created/sec) as well as the number of timers active at
 915 // any one time.  (more precisely, we want to minimize timer-seconds, which is
 916 // the integral of the # of active timers at any instant over time).
 917 // Both impinge on OS scalability.  Given that, at most one thread parked on
 918 // a monitor will use a timer.
 919 
 920 void ATTR ObjectMonitor::exit(TRAPS) {
 921    Thread * Self = THREAD ;
 922    if (THREAD != _owner) {
 923      if (THREAD->is_lock_owned((address) _owner)) {
 924        // Transmute _owner from a BasicLock pointer to a Thread address.
 925        // We don't need to hold _mutex for this transition.
 926        // Non-null to Non-null is safe as long as all readers can
 927        // tolerate either flavor.
 928        assert (_recursions == 0, "invariant") ;
 929        _owner = THREAD ;
 930        _recursions = 0 ;
 931        OwnerIsThread = 1 ;
 932      } else {
 933        // NOTE: we need to handle unbalanced monitor enter/exit
 934        // in native code by throwing an exception.
 935        // TODO: Throw an IllegalMonitorStateException ?
 936        TEVENT (Exit - Throw IMSX) ;
 937        assert(false, "Non-balanced monitor enter/exit!");
 938        if (false) {
 939           THROW(vmSymbols::java_lang_IllegalMonitorStateException());
 940        }
 941        return;
 942      }
 943    }
 944 
 945    if (_recursions != 0) {
 946      _recursions--;        // this is simple recursive enter
 947      TEVENT (Inflated exit - recursive) ;
 948      return ;
 949    }
 950 
 951    // Invariant: after setting Responsible=null an thread must execute
 952    // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
 953    if ((SyncFlags & 4) == 0) {
 954       _Responsible = NULL ;
 955    }
 956 
 957    for (;;) {
 958       assert (THREAD == _owner, "invariant") ;
 959 
 960 
 961       if (Knob_ExitPolicy == 0) {
 962          // release semantics: prior loads and stores from within the critical section
 963          // must not float (reorder) past the following store that drops the lock.
 964          // On SPARC that requires MEMBAR #loadstore|#storestore.
 965          // But of course in TSO #loadstore|#storestore is not required.
 966          // I'd like to write one of the following:
 967          // A.  OrderAccess::release() ; _owner = NULL
 968          // B.  OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
 969          // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
 970          // store into a _dummy variable.  That store is not needed, but can result
 971          // in massive wasteful coherency traffic on classic SMP systems.
 972          // Instead, I use release_store(), which is implemented as just a simple
 973          // ST on x64, x86 and SPARC.
 974          OrderAccess::release_store_ptr (&_owner, NULL) ;   // drop the lock
 975          OrderAccess::storeload() ;                         // See if we need to wake a successor
 976          if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 977             TEVENT (Inflated exit - simple egress) ;
 978             return ;
 979          }
 980          TEVENT (Inflated exit - complex egress) ;
 981 
 982          // Normally the exiting thread is responsible for ensuring succession,
 983          // but if other successors are ready or other entering threads are spinning
 984          // then this thread can simply store NULL into _owner and exit without
 985          // waking a successor.  The existence of spinners or ready successors
 986          // guarantees proper succession (liveness).  Responsibility passes to the
 987          // ready or running successors.  The exiting thread delegates the duty.
 988          // More precisely, if a successor already exists this thread is absolved
 989          // of the responsibility of waking (unparking) one.
 990          //
 991          // The _succ variable is critical to reducing futile wakeup frequency.
 992          // _succ identifies the "heir presumptive" thread that has been made
 993          // ready (unparked) but that has not yet run.  We need only one such
 994          // successor thread to guarantee progress.
 995          // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 996          // section 3.3 "Futile Wakeup Throttling" for details.
 997          //
 998          // Note that spinners in Enter() also set _succ non-null.
 999          // In the current implementation spinners opportunistically set
1000          // _succ so that exiting threads might avoid waking a successor.
1001          // Another less appealing alternative would be for the exiting thread
1002          // to drop the lock and then spin briefly to see if a spinner managed
1003          // to acquire the lock.  If so, the exiting thread could exit
1004          // immediately without waking a successor, otherwise the exiting
1005          // thread would need to dequeue and wake a successor.
1006          // (Note that we'd need to make the post-drop spin short, but no
1007          // shorter than the worst-case round-trip cache-line migration time.
1008          // The dropped lock needs to become visible to the spinner, and then
1009          // the acquisition of the lock by the spinner must become visible to
1010          // the exiting thread).
1011          //
1012 
1013          // It appears that an heir-presumptive (successor) must be made ready.
1014          // Only the current lock owner can manipulate the EntryList or
1015          // drain _cxq, so we need to reacquire the lock.  If we fail
1016          // to reacquire the lock the responsibility for ensuring succession
1017          // falls to the new owner.
1018          //
1019          if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1020             return ;
1021          }
1022          TEVENT (Exit - Reacquired) ;
1023       } else {
1024          if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1025             OrderAccess::release_store_ptr (&_owner, NULL) ;   // drop the lock
1026             OrderAccess::storeload() ;
1027             // Ratify the previously observed values.
1028             if (_cxq == NULL || _succ != NULL) {
1029                 TEVENT (Inflated exit - simple egress) ;
1030                 return ;
1031             }
1032 
1033             // inopportune interleaving -- the exiting thread (this thread)
1034             // in the fast-exit path raced an entering thread in the slow-enter
1035             // path.
1036             // We have two choices:
1037             // A.  Try to reacquire the lock.
1038             //     If the CAS() fails return immediately, otherwise
1039             //     we either restart/rerun the exit operation, or simply
1040             //     fall-through into the code below which wakes a successor.
1041             // B.  If the elements forming the EntryList|cxq are TSM
1042             //     we could simply unpark() the lead thread and return
1043             //     without having set _succ.
1044             if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1045                TEVENT (Inflated exit - reacquired succeeded) ;
1046                return ;
1047             }
1048             TEVENT (Inflated exit - reacquired failed) ;
1049          } else {
1050             TEVENT (Inflated exit - complex egress) ;
1051          }
1052       }
1053 
1054       guarantee (_owner == THREAD, "invariant") ;
1055 
1056       ObjectWaiter * w = NULL ;
1057       int QMode = Knob_QMode ;
1058 
1059       if (QMode == 2 && _cxq != NULL) {
1060           // QMode == 2 : cxq has precedence over EntryList.
1061           // Try to directly wake a successor from the cxq.
1062           // If successful, the successor will need to unlink itself from cxq.
1063           w = _cxq ;
1064           assert (w != NULL, "invariant") ;
1065           assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1066           ExitEpilog (Self, w) ;
1067           return ;
1068       }
1069 
1070       if (QMode == 3 && _cxq != NULL) {
1071           // Aggressively drain cxq into EntryList at the first opportunity.
1072           // This policy ensure that recently-run threads live at the head of EntryList.
1073           // Drain _cxq into EntryList - bulk transfer.
1074           // First, detach _cxq.
1075           // The following loop is tantamount to: w = swap (&cxq, NULL)
1076           w = _cxq ;
1077           for (;;) {
1078              assert (w != NULL, "Invariant") ;
1079              ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1080              if (u == w) break ;
1081              w = u ;
1082           }
1083           assert (w != NULL              , "invariant") ;
1084 
1085           ObjectWaiter * q = NULL ;
1086           ObjectWaiter * p ;
1087           for (p = w ; p != NULL ; p = p->_next) {
1088               guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1089               p->TState = ObjectWaiter::TS_ENTER ;
1090               p->_prev = q ;
1091               q = p ;
1092           }
1093 
1094           // Append the RATs to the EntryList
1095           // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1096           ObjectWaiter * Tail ;
1097           for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ;
1098           if (Tail == NULL) {
1099               _EntryList = w ;
1100           } else {
1101               Tail->_next = w ;
1102               w->_prev = Tail ;
1103           }
1104 
1105           // Fall thru into code that tries to wake a successor from EntryList
1106       }
1107 
1108       if (QMode == 4 && _cxq != NULL) {
1109           // Aggressively drain cxq into EntryList at the first opportunity.
1110           // This policy ensure that recently-run threads live at the head of EntryList.
1111 
1112           // Drain _cxq into EntryList - bulk transfer.
1113           // First, detach _cxq.
1114           // The following loop is tantamount to: w = swap (&cxq, NULL)
1115           w = _cxq ;
1116           for (;;) {
1117              assert (w != NULL, "Invariant") ;
1118              ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1119              if (u == w) break ;
1120              w = u ;
1121           }
1122           assert (w != NULL              , "invariant") ;
1123 
1124           ObjectWaiter * q = NULL ;
1125           ObjectWaiter * p ;
1126           for (p = w ; p != NULL ; p = p->_next) {
1127               guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1128               p->TState = ObjectWaiter::TS_ENTER ;
1129               p->_prev = q ;
1130               q = p ;
1131           }
1132 
1133           // Prepend the RATs to the EntryList
1134           if (_EntryList != NULL) {
1135               q->_next = _EntryList ;
1136               _EntryList->_prev = q ;
1137           }
1138           _EntryList = w ;
1139 
1140           // Fall thru into code that tries to wake a successor from EntryList
1141       }
1142 
1143       w = _EntryList  ;
1144       if (w != NULL) {
1145           // I'd like to write: guarantee (w->_thread != Self).
1146           // But in practice an exiting thread may find itself on the EntryList.
1147           // Lets say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1148           // then calls exit().  Exit release the lock by setting O._owner to NULL.
1149           // Lets say T1 then stalls.  T2 acquires O and calls O.notify().  The
1150           // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1151           // release the lock "O".  T2 resumes immediately after the ST of null into
1152           // _owner, above.  T2 notices that the EntryList is populated, so it
1153           // reacquires the lock and then finds itself on the EntryList.
1154           // Given all that, we have to tolerate the circumstance where "w" is
1155           // associated with Self.
1156           assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1157           ExitEpilog (Self, w) ;
1158           return ;
1159       }
1160 
1161       // If we find that both _cxq and EntryList are null then just
1162       // re-run the exit protocol from the top.
1163       w = _cxq ;
1164       if (w == NULL) continue ;
1165 
1166       // Drain _cxq into EntryList - bulk transfer.
1167       // First, detach _cxq.
1168       // The following loop is tantamount to: w = swap (&cxq, NULL)
1169       for (;;) {
1170           assert (w != NULL, "Invariant") ;
1171           ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1172           if (u == w) break ;
1173           w = u ;
1174       }
1175       TEVENT (Inflated exit - drain cxq into EntryList) ;
1176 
1177       assert (w != NULL              , "invariant") ;
1178       assert (_EntryList  == NULL    , "invariant") ;
1179 
1180       // Convert the LIFO SLL anchored by _cxq into a DLL.
1181       // The list reorganization step operates in O(LENGTH(w)) time.
1182       // It's critical that this step operate quickly as
1183       // "Self" still holds the outer-lock, restricting parallelism
1184       // and effectively lengthening the critical section.
1185       // Invariant: s chases t chases u.
1186       // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1187       // we have faster access to the tail.
1188 
1189       if (QMode == 1) {
1190          // QMode == 1 : drain cxq to EntryList, reversing order
1191          // We also reverse the order of the list.
1192          ObjectWaiter * s = NULL ;
1193          ObjectWaiter * t = w ;
1194          ObjectWaiter * u = NULL ;
1195          while (t != NULL) {
1196              guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ;
1197              t->TState = ObjectWaiter::TS_ENTER ;
1198              u = t->_next ;
1199              t->_prev = u ;
1200              t->_next = s ;
1201              s = t;
1202              t = u ;
1203          }
1204          _EntryList  = s ;
1205          assert (s != NULL, "invariant") ;
1206       } else {
1207          // QMode == 0 or QMode == 2
1208          _EntryList = w ;
1209          ObjectWaiter * q = NULL ;
1210          ObjectWaiter * p ;
1211          for (p = w ; p != NULL ; p = p->_next) {
1212              guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1213              p->TState = ObjectWaiter::TS_ENTER ;
1214              p->_prev = q ;
1215              q = p ;
1216          }
1217       }
1218 
1219       // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1220       // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1221 
1222       // See if we can abdicate to a spinner instead of waking a thread.
1223       // A primary goal of the implementation is to reduce the
1224       // context-switch rate.
1225       if (_succ != NULL) continue;
1226 
1227       w = _EntryList  ;
1228       if (w != NULL) {
1229           guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1230           ExitEpilog (Self, w) ;
1231           return ;
1232       }
1233    }
1234 }
1235 
1236 // ExitSuspendEquivalent:
1237 // A faster alternate to handle_special_suspend_equivalent_condition()
1238 //
1239 // handle_special_suspend_equivalent_condition() unconditionally
1240 // acquires the SR_lock.  On some platforms uncontended MutexLocker()
1241 // operations have high latency.  Note that in ::enter() we call HSSEC
1242 // while holding the monitor, so we effectively lengthen the critical sections.
1243 //
1244 // There are a number of possible solutions:
1245 //
1246 // A.  To ameliorate the problem we might also defer state transitions
1247 //     to as late as possible -- just prior to parking.
1248 //     Given that, we'd call HSSEC after having returned from park(),
1249 //     but before attempting to acquire the monitor.  This is only a
1250 //     partial solution.  It avoids calling HSSEC while holding the
1251 //     monitor (good), but it still increases successor reacquisition latency --
1252 //     the interval between unparking a successor and the time the successor
1253 //     resumes and retries the lock.  See ReenterI(), which defers state transitions.
1254 //     If we use this technique we can also avoid EnterI()-exit() loop
1255 //     in ::enter() where we iteratively drop the lock and then attempt
1256 //     to reacquire it after suspending.
1257 //
1258 // B.  In the future we might fold all the suspend bits into a
1259 //     composite per-thread suspend flag and then update it with CAS().
1260 //     Alternately, a Dekker-like mechanism with multiple variables
1261 //     would suffice:
1262 //       ST Self->_suspend_equivalent = false
1263 //       MEMBAR
1264 //       LD Self_>_suspend_flags
1265 //
1266 
1267 
1268 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) {
1269    int Mode = Knob_FastHSSEC ;
1270    if (Mode && !jSelf->is_external_suspend()) {
1271       assert (jSelf->is_suspend_equivalent(), "invariant") ;
1272       jSelf->clear_suspend_equivalent() ;
1273       if (2 == Mode) OrderAccess::storeload() ;
1274       if (!jSelf->is_external_suspend()) return false ;
1275       // We raced a suspension -- fall thru into the slow path
1276       TEVENT (ExitSuspendEquivalent - raced) ;
1277       jSelf->set_suspend_equivalent() ;
1278    }
1279    return jSelf->handle_special_suspend_equivalent_condition() ;
1280 }
1281 
1282 
1283 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) {
1284    assert (_owner == Self, "invariant") ;
1285 
1286    // Exit protocol:
1287    // 1. ST _succ = wakee
1288    // 2. membar #loadstore|#storestore;
1289    // 2. ST _owner = NULL
1290    // 3. unpark(wakee)
1291 
1292    _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ;
1293    ParkEvent * Trigger = Wakee->_event ;
1294 
1295    // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1296    // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1297    // out-of-scope (non-extant).
1298    Wakee  = NULL ;
1299 
1300    // Drop the lock
1301    OrderAccess::release_store_ptr (&_owner, NULL) ;
1302    OrderAccess::fence() ;                               // ST _owner vs LD in unpark()
1303 
1304    if (SafepointSynchronize::do_call_back()) {
1305       TEVENT (unpark before SAFEPOINT) ;
1306    }
1307 
1308    DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1309    Trigger->unpark() ;
1310 
1311    // Maintain stats and report events to JVMTI
1312    if (ObjectMonitor::_sync_Parks != NULL) {
1313       ObjectMonitor::_sync_Parks->inc() ;
1314    }
1315 }
1316 
1317 
1318 // -----------------------------------------------------------------------------
1319 // Class Loader deadlock handling.
1320 //
1321 // complete_exit exits a lock returning recursion count
1322 // complete_exit/reenter operate as a wait without waiting
1323 // complete_exit requires an inflated monitor
1324 // The _owner field is not always the Thread addr even with an
1325 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1326 // thread due to contention.
1327 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1328    Thread * const Self = THREAD;
1329    assert(Self->is_Java_thread(), "Must be Java thread!");
1330    JavaThread *jt = (JavaThread *)THREAD;
1331 
1332    DeferredInitialize();
1333 
1334    if (THREAD != _owner) {
1335     if (THREAD->is_lock_owned ((address)_owner)) {
1336        assert(_recursions == 0, "internal state error");
1337        _owner = THREAD ;   /* Convert from basiclock addr to Thread addr */
1338        _recursions = 0 ;
1339        OwnerIsThread = 1 ;
1340     }
1341    }
1342 
1343    guarantee(Self == _owner, "complete_exit not owner");
1344    intptr_t save = _recursions; // record the old recursion count
1345    _recursions = 0;        // set the recursion level to be 0
1346    exit (Self) ;           // exit the monitor
1347    guarantee (_owner != Self, "invariant");
1348    return save;
1349 }
1350 
1351 // reenter() enters a lock and sets recursion count
1352 // complete_exit/reenter operate as a wait without waiting
1353 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1354    Thread * const Self = THREAD;
1355    assert(Self->is_Java_thread(), "Must be Java thread!");
1356    JavaThread *jt = (JavaThread *)THREAD;
1357 
1358    guarantee(_owner != Self, "reenter already owner");
1359    enter (THREAD);       // enter the monitor
1360    guarantee (_recursions == 0, "reenter recursion");
1361    _recursions = recursions;
1362    return;
1363 }
1364 
1365 
1366 // -----------------------------------------------------------------------------
1367 // A macro is used below because there may already be a pending
1368 // exception which should not abort the execution of the routines
1369 // which use this (which is why we don't put this into check_slow and
1370 // call it with a CHECK argument).
1371 
1372 #define CHECK_OWNER()                                                             \
1373   do {                                                                            \
1374     if (THREAD != _owner) {                                                       \
1375       if (THREAD->is_lock_owned((address) _owner)) {                              \
1376         _owner = THREAD ;  /* Convert from basiclock addr to Thread addr */       \
1377         _recursions = 0;                                                          \
1378         OwnerIsThread = 1 ;                                                       \
1379       } else {                                                                    \
1380         TEVENT (Throw IMSX) ;                                                     \
1381         THROW(vmSymbols::java_lang_IllegalMonitorStateException());               \
1382       }                                                                           \
1383     }                                                                             \
1384   } while (false)
1385 
1386 // check_slow() is a misnomer.  It's called to simply to throw an IMSX exception.
1387 // TODO-FIXME: remove check_slow() -- it's likely dead.
1388 
1389 void ObjectMonitor::check_slow(TRAPS) {
1390   TEVENT (check_slow - throw IMSX) ;
1391   assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1392   THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1393 }
1394 
1395 static int Adjust (volatile int * adr, int dx) {
1396   int v ;
1397   for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ;
1398   return v ;
1399 }
1400 // -----------------------------------------------------------------------------
1401 // Wait/Notify/NotifyAll
1402 //
1403 // Note: a subset of changes to ObjectMonitor::wait()
1404 // will need to be replicated in complete_exit above
1405 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1406    Thread * const Self = THREAD ;
1407    assert(Self->is_Java_thread(), "Must be Java thread!");
1408    JavaThread *jt = (JavaThread *)THREAD;
1409 
1410    DeferredInitialize () ;
1411 
1412    // Throw IMSX or IEX.
1413    CHECK_OWNER();
1414 
1415    // check for a pending interrupt
1416    if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1417      // post monitor waited event.  Note that this is past-tense, we are done waiting.
1418      if (JvmtiExport::should_post_monitor_waited()) {
1419         // Note: 'false' parameter is passed here because the
1420         // wait was not timed out due to thread interrupt.
1421         JvmtiExport::post_monitor_waited(jt, this, false);
1422      }
1423      TEVENT (Wait - Throw IEX) ;
1424      THROW(vmSymbols::java_lang_InterruptedException());
1425      return ;
1426    }
1427    TEVENT (Wait) ;
1428 
1429    assert (Self->_Stalled == 0, "invariant") ;
1430    Self->_Stalled = intptr_t(this) ;
1431    jt->set_current_waiting_monitor(this);
1432 
1433    // create a node to be put into the queue
1434    // Critically, after we reset() the event but prior to park(), we must check
1435    // for a pending interrupt.
1436    ObjectWaiter node(Self);
1437    node.TState = ObjectWaiter::TS_WAIT ;
1438    Self->_ParkEvent->reset() ;
1439    OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1440 
1441    // Enter the waiting queue, which is a circular doubly linked list in this case
1442    // but it could be a priority queue or any data structure.
1443    // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1444    // by the the owner of the monitor *except* in the case where park()
1445    // returns because of a timeout of interrupt.  Contention is exceptionally rare
1446    // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1447 
1448    Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ;
1449    AddWaiter (&node) ;
1450    Thread::SpinRelease (&_WaitSetLock) ;
1451 
1452    if ((SyncFlags & 4) == 0) {
1453       _Responsible = NULL ;
1454    }
1455    intptr_t save = _recursions; // record the old recursion count
1456    _waiters++;                  // increment the number of waiters
1457    _recursions = 0;             // set the recursion level to be 1
1458    exit (Self) ;                    // exit the monitor
1459    guarantee (_owner != Self, "invariant") ;
1460 
1461    // As soon as the ObjectMonitor's ownership is dropped in the exit()
1462    // call above, another thread can enter() the ObjectMonitor, do the
1463    // notify(), and exit() the ObjectMonitor. If the other thread's
1464    // exit() call chooses this thread as the successor and the unpark()
1465    // call happens to occur while this thread is posting a
1466    // MONITOR_CONTENDED_EXIT event, then we run the risk of the event
1467    // handler using RawMonitors and consuming the unpark().
1468    //
1469    // To avoid the problem, we re-post the event. This does no harm
1470    // even if the original unpark() was not consumed because we are the
1471    // chosen successor for this monitor.
1472    if (node._notified != 0 && _succ == Self) {
1473       node._event->unpark();
1474    }
1475 
1476    // The thread is on the WaitSet list - now park() it.
1477    // On MP systems it's conceivable that a brief spin before we park
1478    // could be profitable.
1479    //
1480    // TODO-FIXME: change the following logic to a loop of the form
1481    //   while (!timeout && !interrupted && _notified == 0) park()
1482 
1483    int ret = OS_OK ;
1484    int WasNotified = 0 ;
1485    { // State transition wrappers
1486      OSThread* osthread = Self->osthread();
1487      OSThreadWaitState osts(osthread, true);
1488      {
1489        ThreadBlockInVM tbivm(jt);
1490        // Thread is in thread_blocked state and oop access is unsafe.
1491        jt->set_suspend_equivalent();
1492 
1493        if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1494            // Intentionally empty
1495        } else
1496        if (node._notified == 0) {
1497          if (millis <= 0) {
1498             Self->_ParkEvent->park () ;
1499          } else {
1500             ret = Self->_ParkEvent->park (millis) ;
1501          }
1502        }
1503 
1504        // were we externally suspended while we were waiting?
1505        if (ExitSuspendEquivalent (jt)) {
1506           // TODO-FIXME: add -- if succ == Self then succ = null.
1507           jt->java_suspend_self();
1508        }
1509 
1510      } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1511 
1512 
1513      // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1514      // from the WaitSet to the EntryList.
1515      // See if we need to remove Node from the WaitSet.
1516      // We use double-checked locking to avoid grabbing _WaitSetLock
1517      // if the thread is not on the wait queue.
1518      //
1519      // Note that we don't need a fence before the fetch of TState.
1520      // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1521      // written by the is thread. (perhaps the fetch might even be satisfied
1522      // by a look-aside into the processor's own store buffer, although given
1523      // the length of the code path between the prior ST and this load that's
1524      // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1525      // then we'll acquire the lock and then re-fetch a fresh TState value.
1526      // That is, we fail toward safety.
1527 
1528      if (node.TState == ObjectWaiter::TS_WAIT) {
1529          Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ;
1530          if (node.TState == ObjectWaiter::TS_WAIT) {
1531             DequeueSpecificWaiter (&node) ;       // unlink from WaitSet
1532             assert(node._notified == 0, "invariant");
1533             node.TState = ObjectWaiter::TS_RUN ;
1534          }
1535          Thread::SpinRelease (&_WaitSetLock) ;
1536      }
1537 
1538      // The thread is now either on off-list (TS_RUN),
1539      // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1540      // The Node's TState variable is stable from the perspective of this thread.
1541      // No other threads will asynchronously modify TState.
1542      guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ;
1543      OrderAccess::loadload() ;
1544      if (_succ == Self) _succ = NULL ;
1545      WasNotified = node._notified ;
1546 
1547      // Reentry phase -- reacquire the monitor.
1548      // re-enter contended monitor after object.wait().
1549      // retain OBJECT_WAIT state until re-enter successfully completes
1550      // Thread state is thread_in_vm and oop access is again safe,
1551      // although the raw address of the object may have changed.
1552      // (Don't cache naked oops over safepoints, of course).
1553 
1554      // post monitor waited event. Note that this is past-tense, we are done waiting.
1555      if (JvmtiExport::should_post_monitor_waited()) {
1556        JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1557      }
1558      OrderAccess::fence() ;
1559 
1560      assert (Self->_Stalled != 0, "invariant") ;
1561      Self->_Stalled = 0 ;
1562 
1563      assert (_owner != Self, "invariant") ;
1564      ObjectWaiter::TStates v = node.TState ;
1565      if (v == ObjectWaiter::TS_RUN) {
1566          enter (Self) ;
1567      } else {
1568          guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ;
1569          ReenterI (Self, &node) ;
1570          node.wait_reenter_end(this);
1571      }
1572 
1573      // Self has reacquired the lock.
1574      // Lifecycle - the node representing Self must not appear on any queues.
1575      // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1576      // want residual elements associated with this thread left on any lists.
1577      guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ;
1578      assert    (_owner == Self, "invariant") ;
1579      assert    (_succ != Self , "invariant") ;
1580    } // OSThreadWaitState()
1581 
1582    jt->set_current_waiting_monitor(NULL);
1583 
1584    guarantee (_recursions == 0, "invariant") ;
1585    _recursions = save;     // restore the old recursion count
1586    _waiters--;             // decrement the number of waiters
1587 
1588    // Verify a few postconditions
1589    assert (_owner == Self       , "invariant") ;
1590    assert (_succ  != Self       , "invariant") ;
1591    assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
1592 
1593    if (SyncFlags & 32) {
1594       OrderAccess::fence() ;
1595    }
1596 
1597    // check if the notification happened
1598    if (!WasNotified) {
1599      // no, it could be timeout or Thread.interrupt() or both
1600      // check for interrupt event, otherwise it is timeout
1601      if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1602        TEVENT (Wait - throw IEX from epilog) ;
1603        THROW(vmSymbols::java_lang_InterruptedException());
1604      }
1605    }
1606 
1607    // NOTE: Spurious wake up will be consider as timeout.
1608    // Monitor notify has precedence over thread interrupt.
1609 }
1610 
1611 
1612 // Consider:
1613 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1614 // then instead of transferring a thread from the WaitSet to the EntryList
1615 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1616 
1617 void ObjectMonitor::notify(TRAPS) {
1618   CHECK_OWNER();
1619   if (_WaitSet == NULL) {
1620      TEVENT (Empty-Notify) ;
1621      return ;
1622   }
1623   DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1624 
1625   int Policy = Knob_MoveNotifyee ;
1626 
1627   Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ;
1628   ObjectWaiter * iterator = DequeueWaiter() ;
1629   if (iterator != NULL) {
1630      TEVENT (Notify1 - Transfer) ;
1631      guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ;
1632      guarantee (iterator->_notified == 0, "invariant") ;
1633      if (Policy != 4) {
1634         iterator->TState = ObjectWaiter::TS_ENTER ;
1635      }
1636      iterator->_notified = 1 ;
1637 
1638      ObjectWaiter * List = _EntryList ;
1639      if (List != NULL) {
1640         assert (List->_prev == NULL, "invariant") ;
1641         assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1642         assert (List != iterator, "invariant") ;
1643      }
1644 
1645      if (Policy == 0) {       // prepend to EntryList
1646          if (List == NULL) {
1647              iterator->_next = iterator->_prev = NULL ;
1648              _EntryList = iterator ;
1649          } else {
1650              List->_prev = iterator ;
1651              iterator->_next = List ;
1652              iterator->_prev = NULL ;
1653              _EntryList = iterator ;
1654         }
1655      } else
1656      if (Policy == 1) {      // append to EntryList
1657          if (List == NULL) {
1658              iterator->_next = iterator->_prev = NULL ;
1659              _EntryList = iterator ;
1660          } else {
1661             // CONSIDER:  finding the tail currently requires a linear-time walk of
1662             // the EntryList.  We can make tail access constant-time by converting to
1663             // a CDLL instead of using our current DLL.
1664             ObjectWaiter * Tail ;
1665             for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ;
1666             assert (Tail != NULL && Tail->_next == NULL, "invariant") ;
1667             Tail->_next = iterator ;
1668             iterator->_prev = Tail ;
1669             iterator->_next = NULL ;
1670         }
1671      } else
1672      if (Policy == 2) {      // prepend to cxq
1673          // prepend to cxq
1674          if (List == NULL) {
1675              iterator->_next = iterator->_prev = NULL ;
1676              _EntryList = iterator ;
1677          } else {
1678             iterator->TState = ObjectWaiter::TS_CXQ ;
1679             for (;;) {
1680                 ObjectWaiter * Front = _cxq ;
1681                 iterator->_next = Front ;
1682                 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) {
1683                     break ;
1684                 }
1685             }
1686          }
1687      } else
1688      if (Policy == 3) {      // append to cxq
1689         iterator->TState = ObjectWaiter::TS_CXQ ;
1690         for (;;) {
1691             ObjectWaiter * Tail ;
1692             Tail = _cxq ;
1693             if (Tail == NULL) {
1694                 iterator->_next = NULL ;
1695                 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) {
1696                    break ;
1697                 }
1698             } else {
1699                 while (Tail->_next != NULL) Tail = Tail->_next ;
1700                 Tail->_next = iterator ;
1701                 iterator->_prev = Tail ;
1702                 iterator->_next = NULL ;
1703                 break ;
1704             }
1705         }
1706      } else {
1707         ParkEvent * ev = iterator->_event ;
1708         iterator->TState = ObjectWaiter::TS_RUN ;
1709         OrderAccess::fence() ;
1710         ev->unpark() ;
1711      }
1712 
1713      if (Policy < 4) {
1714        iterator->wait_reenter_begin(this);
1715      }
1716 
1717      // _WaitSetLock protects the wait queue, not the EntryList.  We could
1718      // move the add-to-EntryList operation, above, outside the critical section
1719      // protected by _WaitSetLock.  In practice that's not useful.  With the
1720      // exception of  wait() timeouts and interrupts the monitor owner
1721      // is the only thread that grabs _WaitSetLock.  There's almost no contention
1722      // on _WaitSetLock so it's not profitable to reduce the length of the
1723      // critical section.
1724   }
1725 
1726   Thread::SpinRelease (&_WaitSetLock) ;
1727 
1728   if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) {
1729      ObjectMonitor::_sync_Notifications->inc() ;
1730   }
1731 }
1732 
1733 
1734 void ObjectMonitor::notifyAll(TRAPS) {
1735   CHECK_OWNER();
1736   ObjectWaiter* iterator;
1737   if (_WaitSet == NULL) {
1738       TEVENT (Empty-NotifyAll) ;
1739       return ;
1740   }
1741   DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1742 
1743   int Policy = Knob_MoveNotifyee ;
1744   int Tally = 0 ;
1745   Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ;
1746 
1747   for (;;) {
1748      iterator = DequeueWaiter () ;
1749      if (iterator == NULL) break ;
1750      TEVENT (NotifyAll - Transfer1) ;
1751      ++Tally ;
1752 
1753      // Disposition - what might we do with iterator ?
1754      // a.  add it directly to the EntryList - either tail or head.
1755      // b.  push it onto the front of the _cxq.
1756      // For now we use (a).
1757 
1758      guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ;
1759      guarantee (iterator->_notified == 0, "invariant") ;
1760      iterator->_notified = 1 ;
1761      if (Policy != 4) {
1762         iterator->TState = ObjectWaiter::TS_ENTER ;
1763      }
1764 
1765      ObjectWaiter * List = _EntryList ;
1766      if (List != NULL) {
1767         assert (List->_prev == NULL, "invariant") ;
1768         assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1769         assert (List != iterator, "invariant") ;
1770      }
1771 
1772      if (Policy == 0) {       // prepend to EntryList
1773          if (List == NULL) {
1774              iterator->_next = iterator->_prev = NULL ;
1775              _EntryList = iterator ;
1776          } else {
1777              List->_prev = iterator ;
1778              iterator->_next = List ;
1779              iterator->_prev = NULL ;
1780              _EntryList = iterator ;
1781         }
1782      } else
1783      if (Policy == 1) {      // append to EntryList
1784          if (List == NULL) {
1785              iterator->_next = iterator->_prev = NULL ;
1786              _EntryList = iterator ;
1787          } else {
1788             // CONSIDER:  finding the tail currently requires a linear-time walk of
1789             // the EntryList.  We can make tail access constant-time by converting to
1790             // a CDLL instead of using our current DLL.
1791             ObjectWaiter * Tail ;
1792             for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ;
1793             assert (Tail != NULL && Tail->_next == NULL, "invariant") ;
1794             Tail->_next = iterator ;
1795             iterator->_prev = Tail ;
1796             iterator->_next = NULL ;
1797         }
1798      } else
1799      if (Policy == 2) {      // prepend to cxq
1800          // prepend to cxq
1801          iterator->TState = ObjectWaiter::TS_CXQ ;
1802          for (;;) {
1803              ObjectWaiter * Front = _cxq ;
1804              iterator->_next = Front ;
1805              if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) {
1806                  break ;
1807              }
1808          }
1809      } else
1810      if (Policy == 3) {      // append to cxq
1811         iterator->TState = ObjectWaiter::TS_CXQ ;
1812         for (;;) {
1813             ObjectWaiter * Tail ;
1814             Tail = _cxq ;
1815             if (Tail == NULL) {
1816                 iterator->_next = NULL ;
1817                 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) {
1818                    break ;
1819                 }
1820             } else {
1821                 while (Tail->_next != NULL) Tail = Tail->_next ;
1822                 Tail->_next = iterator ;
1823                 iterator->_prev = Tail ;
1824                 iterator->_next = NULL ;
1825                 break ;
1826             }
1827         }
1828      } else {
1829         ParkEvent * ev = iterator->_event ;
1830         iterator->TState = ObjectWaiter::TS_RUN ;
1831         OrderAccess::fence() ;
1832         ev->unpark() ;
1833      }
1834 
1835      if (Policy < 4) {
1836        iterator->wait_reenter_begin(this);
1837      }
1838 
1839      // _WaitSetLock protects the wait queue, not the EntryList.  We could
1840      // move the add-to-EntryList operation, above, outside the critical section
1841      // protected by _WaitSetLock.  In practice that's not useful.  With the
1842      // exception of  wait() timeouts and interrupts the monitor owner
1843      // is the only thread that grabs _WaitSetLock.  There's almost no contention
1844      // on _WaitSetLock so it's not profitable to reduce the length of the
1845      // critical section.
1846   }
1847 
1848   Thread::SpinRelease (&_WaitSetLock) ;
1849 
1850   if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) {
1851      ObjectMonitor::_sync_Notifications->inc(Tally) ;
1852   }
1853 }
1854 
1855 // -----------------------------------------------------------------------------
1856 // Adaptive Spinning Support
1857 //
1858 // Adaptive spin-then-block - rational spinning
1859 //
1860 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1861 // algorithm.  On high order SMP systems it would be better to start with
1862 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1863 // a contending thread could enqueue itself on the cxq and then spin locally
1864 // on a thread-specific variable such as its ParkEvent._Event flag.
1865 // That's left as an exercise for the reader.  Note that global spinning is
1866 // not problematic on Niagara, as the L2$ serves the interconnect and has both
1867 // low latency and massive bandwidth.
1868 //
1869 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1870 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1871 // (duration) or we can fix the count at approximately the duration of
1872 // a context switch and vary the frequency.   Of course we could also
1873 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1874 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html.
1875 //
1876 // This implementation varies the duration "D", where D varies with
1877 // the success rate of recent spin attempts. (D is capped at approximately
1878 // length of a round-trip context switch).  The success rate for recent
1879 // spin attempts is a good predictor of the success rate of future spin
1880 // attempts.  The mechanism adapts automatically to varying critical
1881 // section length (lock modality), system load and degree of parallelism.
1882 // D is maintained per-monitor in _SpinDuration and is initialized
1883 // optimistically.  Spin frequency is fixed at 100%.
1884 //
1885 // Note that _SpinDuration is volatile, but we update it without locks
1886 // or atomics.  The code is designed so that _SpinDuration stays within
1887 // a reasonable range even in the presence of races.  The arithmetic
1888 // operations on _SpinDuration are closed over the domain of legal values,
1889 // so at worst a race will install and older but still legal value.
1890 // At the very worst this introduces some apparent non-determinism.
1891 // We might spin when we shouldn't or vice-versa, but since the spin
1892 // count are relatively short, even in the worst case, the effect is harmless.
1893 //
1894 // Care must be taken that a low "D" value does not become an
1895 // an absorbing state.  Transient spinning failures -- when spinning
1896 // is overall profitable -- should not cause the system to converge
1897 // on low "D" values.  We want spinning to be stable and predictable
1898 // and fairly responsive to change and at the same time we don't want
1899 // it to oscillate, become metastable, be "too" non-deterministic,
1900 // or converge on or enter undesirable stable absorbing states.
1901 //
1902 // We implement a feedback-based control system -- using past behavior
1903 // to predict future behavior.  We face two issues: (a) if the
1904 // input signal is random then the spin predictor won't provide optimal
1905 // results, and (b) if the signal frequency is too high then the control
1906 // system, which has some natural response lag, will "chase" the signal.
1907 // (b) can arise from multimodal lock hold times.  Transient preemption
1908 // can also result in apparent bimodal lock hold times.
1909 // Although sub-optimal, neither condition is particularly harmful, as
1910 // in the worst-case we'll spin when we shouldn't or vice-versa.
1911 // The maximum spin duration is rather short so the failure modes aren't bad.
1912 // To be conservative, I've tuned the gain in system to bias toward
1913 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1914 // "rings" or oscillates between spinning and not spinning.  This happens
1915 // when spinning is just on the cusp of profitability, however, so the
1916 // situation is not dire.  The state is benign -- there's no need to add
1917 // hysteresis control to damp the transition rate between spinning and
1918 // not spinning.
1919 //
1920 
1921 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ;
1922 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ;
1923 
1924 // Spinning: Fixed frequency (100%), vary duration
1925 
1926 
1927 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) {
1928 
1929     // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1930     int ctr = Knob_FixedSpin ;
1931     if (ctr != 0) {
1932         while (--ctr >= 0) {
1933             if (TryLock (Self) > 0) return 1 ;
1934             SpinPause () ;
1935         }
1936         return 0 ;
1937     }
1938 
1939     for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) {
1940       if (TryLock(Self) > 0) {
1941         // Increase _SpinDuration ...
1942         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1943         // Raising _SpurDuration to the poverty line is key.
1944         int x = _SpinDuration ;
1945         if (x < Knob_SpinLimit) {
1946            if (x < Knob_Poverty) x = Knob_Poverty ;
1947            _SpinDuration = x + Knob_BonusB ;
1948         }
1949         return 1 ;
1950       }
1951       SpinPause () ;
1952     }
1953 
1954     // Admission control - verify preconditions for spinning
1955     //
1956     // We always spin a little bit, just to prevent _SpinDuration == 0 from
1957     // becoming an absorbing state.  Put another way, we spin briefly to
1958     // sample, just in case the system load, parallelism, contention, or lock
1959     // modality changed.
1960     //
1961     // Consider the following alternative:
1962     // Periodically set _SpinDuration = _SpinLimit and try a long/full
1963     // spin attempt.  "Periodically" might mean after a tally of
1964     // the # of failed spin attempts (or iterations) reaches some threshold.
1965     // This takes us into the realm of 1-out-of-N spinning, where we
1966     // hold the duration constant but vary the frequency.
1967 
1968     ctr = _SpinDuration  ;
1969     if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ;
1970     if (ctr <= 0) return 0 ;
1971 
1972     if (Knob_SuccRestrict && _succ != NULL) return 0 ;
1973     if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1974        TEVENT (Spin abort - notrunnable [TOP]);
1975        return 0 ;
1976     }
1977 
1978     int MaxSpin = Knob_MaxSpinners ;
1979     if (MaxSpin >= 0) {
1980        if (_Spinner > MaxSpin) {
1981           TEVENT (Spin abort -- too many spinners) ;
1982           return 0 ;
1983        }
1984        // Slighty racy, but benign ...
1985        Adjust (&_Spinner, 1) ;
1986     }
1987 
1988     // We're good to spin ... spin ingress.
1989     // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1990     // when preparing to LD...CAS _owner, etc and the CAS is likely
1991     // to succeed.
1992     int hits    = 0 ;
1993     int msk     = 0 ;
1994     int caspty  = Knob_CASPenalty ;
1995     int oxpty   = Knob_OXPenalty ;
1996     int sss     = Knob_SpinSetSucc ;
1997     if (sss && _succ == NULL ) _succ = Self ;
1998     Thread * prv = NULL ;
1999 
2000     // There are three ways to exit the following loop:
2001     // 1.  A successful spin where this thread has acquired the lock.
2002     // 2.  Spin failure with prejudice
2003     // 3.  Spin failure without prejudice
2004 
2005     while (--ctr >= 0) {
2006 
2007       // Periodic polling -- Check for pending GC
2008       // Threads may spin while they're unsafe.
2009       // We don't want spinning threads to delay the JVM from reaching
2010       // a stop-the-world safepoint or to steal cycles from GC.
2011       // If we detect a pending safepoint we abort in order that
2012       // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
2013       // this thread, if safe, doesn't steal cycles from GC.
2014       // This is in keeping with the "no loitering in runtime" rule.
2015       // We periodically check to see if there's a safepoint pending.
2016       if ((ctr & 0xFF) == 0) {
2017          if (SafepointSynchronize::do_call_back()) {
2018             TEVENT (Spin: safepoint) ;
2019             goto Abort ;           // abrupt spin egress
2020          }
2021          if (Knob_UsePause & 1) SpinPause () ;
2022 
2023          int (*scb)(intptr_t,int) = SpinCallbackFunction ;
2024          if (hits > 50 && scb != NULL) {
2025             int abend = (*scb)(SpinCallbackArgument, 0) ;
2026          }
2027       }
2028 
2029       if (Knob_UsePause & 2) SpinPause() ;
2030 
2031       // Exponential back-off ...  Stay off the bus to reduce coherency traffic.
2032       // This is useful on classic SMP systems, but is of less utility on
2033       // N1-style CMT platforms.
2034       //
2035       // Trade-off: lock acquisition latency vs coherency bandwidth.
2036       // Lock hold times are typically short.  A histogram
2037       // of successful spin attempts shows that we usually acquire
2038       // the lock early in the spin.  That suggests we want to
2039       // sample _owner frequently in the early phase of the spin,
2040       // but then back-off and sample less frequently as the spin
2041       // progresses.  The back-off makes a good citizen on SMP big
2042       // SMP systems.  Oversampling _owner can consume excessive
2043       // coherency bandwidth.  Relatedly, if we _oversample _owner we
2044       // can inadvertently interfere with the the ST m->owner=null.
2045       // executed by the lock owner.
2046       if (ctr & msk) continue ;
2047       ++hits ;
2048       if ((hits & 0xF) == 0) {
2049         // The 0xF, above, corresponds to the exponent.
2050         // Consider: (msk+1)|msk
2051         msk = ((msk << 2)|3) & BackOffMask ;
2052       }
2053 
2054       // Probe _owner with TATAS
2055       // If this thread observes the monitor transition or flicker
2056       // from locked to unlocked to locked, then the odds that this
2057       // thread will acquire the lock in this spin attempt go down
2058       // considerably.  The same argument applies if the CAS fails
2059       // or if we observe _owner change from one non-null value to
2060       // another non-null value.   In such cases we might abort
2061       // the spin without prejudice or apply a "penalty" to the
2062       // spin count-down variable "ctr", reducing it by 100, say.
2063 
2064       Thread * ox = (Thread *) _owner ;
2065       if (ox == NULL) {
2066          ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ;
2067          if (ox == NULL) {
2068             // The CAS succeeded -- this thread acquired ownership
2069             // Take care of some bookkeeping to exit spin state.
2070             if (sss && _succ == Self) {
2071                _succ = NULL ;
2072             }
2073             if (MaxSpin > 0) Adjust (&_Spinner, -1) ;
2074 
2075             // Increase _SpinDuration :
2076             // The spin was successful (profitable) so we tend toward
2077             // longer spin attempts in the future.
2078             // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2079             // If we acquired the lock early in the spin cycle it
2080             // makes sense to increase _SpinDuration proportionally.
2081             // Note that we don't clamp SpinDuration precisely at SpinLimit.
2082             int x = _SpinDuration ;
2083             if (x < Knob_SpinLimit) {
2084                 if (x < Knob_Poverty) x = Knob_Poverty ;
2085                 _SpinDuration = x + Knob_Bonus ;
2086             }
2087             return 1 ;
2088          }
2089 
2090          // The CAS failed ... we can take any of the following actions:
2091          // * penalize: ctr -= Knob_CASPenalty
2092          // * exit spin with prejudice -- goto Abort;
2093          // * exit spin without prejudice.
2094          // * Since CAS is high-latency, retry again immediately.
2095          prv = ox ;
2096          TEVENT (Spin: cas failed) ;
2097          if (caspty == -2) break ;
2098          if (caspty == -1) goto Abort ;
2099          ctr -= caspty ;
2100          continue ;
2101       }
2102 
2103       // Did lock ownership change hands ?
2104       if (ox != prv && prv != NULL ) {
2105           TEVENT (spin: Owner changed)
2106           if (oxpty == -2) break ;
2107           if (oxpty == -1) goto Abort ;
2108           ctr -= oxpty ;
2109       }
2110       prv = ox ;
2111 
2112       // Abort the spin if the owner is not executing.
2113       // The owner must be executing in order to drop the lock.
2114       // Spinning while the owner is OFFPROC is idiocy.
2115       // Consider: ctr -= RunnablePenalty ;
2116       if (Knob_OState && NotRunnable (Self, ox)) {
2117          TEVENT (Spin abort - notrunnable);
2118          goto Abort ;
2119       }
2120       if (sss && _succ == NULL ) _succ = Self ;
2121    }
2122 
2123    // Spin failed with prejudice -- reduce _SpinDuration.
2124    // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2125    // AIMD is globally stable.
2126    TEVENT (Spin failure) ;
2127    {
2128      int x = _SpinDuration ;
2129      if (x > 0) {
2130         // Consider an AIMD scheme like: x -= (x >> 3) + 100
2131         // This is globally sample and tends to damp the response.
2132         x -= Knob_Penalty ;
2133         if (x < 0) x = 0 ;
2134         _SpinDuration = x ;
2135      }
2136    }
2137 
2138  Abort:
2139    if (MaxSpin >= 0) Adjust (&_Spinner, -1) ;
2140    if (sss && _succ == Self) {
2141       _succ = NULL ;
2142       // Invariant: after setting succ=null a contending thread
2143       // must recheck-retry _owner before parking.  This usually happens
2144       // in the normal usage of TrySpin(), but it's safest
2145       // to make TrySpin() as foolproof as possible.
2146       OrderAccess::fence() ;
2147       if (TryLock(Self) > 0) return 1 ;
2148    }
2149    return 0 ;
2150 }
2151 
2152 // NotRunnable() -- informed spinning
2153 //
2154 // Don't bother spinning if the owner is not eligible to drop the lock.
2155 // Peek at the owner's schedctl.sc_state and Thread._thread_values and
2156 // spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2157 // The thread must be runnable in order to drop the lock in timely fashion.
2158 // If the _owner is not runnable then spinning will not likely be
2159 // successful (profitable).
2160 //
2161 // Beware -- the thread referenced by _owner could have died
2162 // so a simply fetch from _owner->_thread_state might trap.
2163 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2164 // Because of the lifecycle issues the schedctl and _thread_state values
2165 // observed by NotRunnable() might be garbage.  NotRunnable must
2166 // tolerate this and consider the observed _thread_state value
2167 // as advisory.
2168 //
2169 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2170 // a thread pointer.  We differentiate the two cases with OwnerIsThread.
2171 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2172 // with the LSB of _owner.  Another option would be to probablistically probe
2173 // the putative _owner->TypeTag value.
2174 //
2175 // Checking _thread_state isn't perfect.  Even if the thread is
2176 // in_java it might be blocked on a page-fault or have been preempted
2177 // and sitting on a ready/dispatch queue.  _thread state in conjunction
2178 // with schedctl.sc_state gives us a good picture of what the
2179 // thread is doing, however.
2180 //
2181 // TODO: check schedctl.sc_state.
2182 // We'll need to use SafeFetch32() to read from the schedctl block.
2183 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/
2184 //
2185 // The return value from NotRunnable() is *advisory* -- the
2186 // result is based on sampling and is not necessarily coherent.
2187 // The caller must tolerate false-negative and false-positive errors.
2188 // Spinning, in general, is probabilistic anyway.
2189 
2190 
2191 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) {
2192     // Check either OwnerIsThread or ox->TypeTag == 2BAD.
2193     if (!OwnerIsThread) return 0 ;
2194 
2195     if (ox == NULL) return 0 ;
2196 
2197     // Avoid transitive spinning ...
2198     // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2199     // Immediately after T1 acquires L it's possible that T2, also
2200     // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2201     // This occurs transiently after T1 acquired L but before
2202     // T1 managed to clear T1.Stalled.  T2 does not need to abort
2203     // its spin in this circumstance.
2204     intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ;
2205 
2206     if (BlockedOn == 1) return 1 ;
2207     if (BlockedOn != 0) {
2208       return BlockedOn != intptr_t(this) && _owner == ox ;
2209     }
2210 
2211     assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ;
2212     int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ;
2213     // consider also: jst != _thread_in_Java -- but that's overspecific.
2214     return jst == _thread_blocked || jst == _thread_in_native ;
2215 }
2216 
2217 
2218 // -----------------------------------------------------------------------------
2219 // WaitSet management ...
2220 
2221 ObjectWaiter::ObjectWaiter(Thread* thread) {
2222   _next     = NULL;
2223   _prev     = NULL;
2224   _notified = 0;
2225   TState    = TS_RUN ;
2226   _thread   = thread;
2227   _event    = thread->_ParkEvent ;
2228   _active   = false;
2229   assert (_event != NULL, "invariant") ;
2230 }
2231 
2232 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) {
2233   JavaThread *jt = (JavaThread *)this->_thread;
2234   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2235 }
2236 
2237 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) {
2238   JavaThread *jt = (JavaThread *)this->_thread;
2239   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2240 }
2241 
2242 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2243   assert(node != NULL, "should not dequeue NULL node");
2244   assert(node->_prev == NULL, "node already in list");
2245   assert(node->_next == NULL, "node already in list");
2246   // put node at end of queue (circular doubly linked list)
2247   if (_WaitSet == NULL) {
2248     _WaitSet = node;
2249     node->_prev = node;
2250     node->_next = node;
2251   } else {
2252     ObjectWaiter* head = _WaitSet ;
2253     ObjectWaiter* tail = head->_prev;
2254     assert(tail->_next == head, "invariant check");
2255     tail->_next = node;
2256     head->_prev = node;
2257     node->_next = head;
2258     node->_prev = tail;
2259   }
2260 }
2261 
2262 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2263   // dequeue the very first waiter
2264   ObjectWaiter* waiter = _WaitSet;
2265   if (waiter) {
2266     DequeueSpecificWaiter(waiter);
2267   }
2268   return waiter;
2269 }
2270 
2271 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2272   assert(node != NULL, "should not dequeue NULL node");
2273   assert(node->_prev != NULL, "node already removed from list");
2274   assert(node->_next != NULL, "node already removed from list");
2275   // when the waiter has woken up because of interrupt,
2276   // timeout or other spurious wake-up, dequeue the
2277   // waiter from waiting list
2278   ObjectWaiter* next = node->_next;
2279   if (next == node) {
2280     assert(node->_prev == node, "invariant check");
2281     _WaitSet = NULL;
2282   } else {
2283     ObjectWaiter* prev = node->_prev;
2284     assert(prev->_next == node, "invariant check");
2285     assert(next->_prev == node, "invariant check");
2286     next->_prev = prev;
2287     prev->_next = next;
2288     if (_WaitSet == node) {
2289       _WaitSet = next;
2290     }
2291   }
2292   node->_next = NULL;
2293   node->_prev = NULL;
2294 }
2295 
2296 // -----------------------------------------------------------------------------
2297 // PerfData support
2298 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL ;
2299 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL ;
2300 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL ;
2301 PerfCounter * ObjectMonitor::_sync_EmptyNotifications          = NULL ;
2302 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL ;
2303 PerfCounter * ObjectMonitor::_sync_PrivateA                    = NULL ;
2304 PerfCounter * ObjectMonitor::_sync_PrivateB                    = NULL ;
2305 PerfCounter * ObjectMonitor::_sync_SlowExit                    = NULL ;
2306 PerfCounter * ObjectMonitor::_sync_SlowEnter                   = NULL ;
2307 PerfCounter * ObjectMonitor::_sync_SlowNotify                  = NULL ;
2308 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll               = NULL ;
2309 PerfCounter * ObjectMonitor::_sync_FailedSpins                 = NULL ;
2310 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins             = NULL ;
2311 PerfCounter * ObjectMonitor::_sync_MonInCirculation            = NULL ;
2312 PerfCounter * ObjectMonitor::_sync_MonScavenged                = NULL ;
2313 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL ;
2314 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL ;
2315 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL ;
2316 
2317 // One-shot global initialization for the sync subsystem.
2318 // We could also defer initialization and initialize on-demand
2319 // the first time we call inflate().  Initialization would
2320 // be protected - like so many things - by the MonitorCache_lock.
2321 
2322 void ObjectMonitor::Initialize () {
2323   static int InitializationCompleted = 0 ;
2324   assert (InitializationCompleted == 0, "invariant") ;
2325   InitializationCompleted = 1 ;
2326   if (UsePerfData) {
2327       EXCEPTION_MARK ;
2328       #define NEWPERFCOUNTER(n)   {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); }
2329       #define NEWPERFVARIABLE(n)  {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); }
2330       NEWPERFCOUNTER(_sync_Inflations) ;
2331       NEWPERFCOUNTER(_sync_Deflations) ;
2332       NEWPERFCOUNTER(_sync_ContendedLockAttempts) ;
2333       NEWPERFCOUNTER(_sync_FutileWakeups) ;
2334       NEWPERFCOUNTER(_sync_Parks) ;
2335       NEWPERFCOUNTER(_sync_EmptyNotifications) ;
2336       NEWPERFCOUNTER(_sync_Notifications) ;
2337       NEWPERFCOUNTER(_sync_SlowEnter) ;
2338       NEWPERFCOUNTER(_sync_SlowExit) ;
2339       NEWPERFCOUNTER(_sync_SlowNotify) ;
2340       NEWPERFCOUNTER(_sync_SlowNotifyAll) ;
2341       NEWPERFCOUNTER(_sync_FailedSpins) ;
2342       NEWPERFCOUNTER(_sync_SuccessfulSpins) ;
2343       NEWPERFCOUNTER(_sync_PrivateA) ;
2344       NEWPERFCOUNTER(_sync_PrivateB) ;
2345       NEWPERFCOUNTER(_sync_MonInCirculation) ;
2346       NEWPERFCOUNTER(_sync_MonScavenged) ;
2347       NEWPERFVARIABLE(_sync_MonExtant) ;
2348       #undef NEWPERFCOUNTER
2349   }
2350 }
2351 
2352 
2353 // Compile-time asserts
2354 // When possible, it's better to catch errors deterministically at
2355 // compile-time than at runtime.  The down-side to using compile-time
2356 // asserts is that error message -- often something about negative array
2357 // indices -- is opaque.
2358 
2359 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); }
2360 
2361 void ObjectMonitor::ctAsserts() {
2362   CTASSERT(offset_of (ObjectMonitor, _header) == 0);
2363 }
2364 
2365 
2366 static char * kvGet (char * kvList, const char * Key) {
2367     if (kvList == NULL) return NULL ;
2368     size_t n = strlen (Key) ;
2369     char * Search ;
2370     for (Search = kvList ; *Search ; Search += strlen(Search) + 1) {
2371         if (strncmp (Search, Key, n) == 0) {
2372             if (Search[n] == '=') return Search + n + 1 ;
2373             if (Search[n] == 0)   return (char *) "1" ;
2374         }
2375     }
2376     return NULL ;
2377 }
2378 
2379 static int kvGetInt (char * kvList, const char * Key, int Default) {
2380     char * v = kvGet (kvList, Key) ;
2381     int rslt = v ? ::strtol (v, NULL, 0) : Default ;
2382     if (Knob_ReportSettings && v != NULL) {
2383         ::printf ("  SyncKnob: %s %d(%d)\n", Key, rslt, Default) ;
2384         ::fflush (stdout) ;
2385     }
2386     return rslt ;
2387 }
2388 
2389 void ObjectMonitor::DeferredInitialize () {
2390   if (InitDone > 0) return ;
2391   if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2392       while (InitDone != 1) ;
2393       return ;
2394   }
2395 
2396   // One-shot global initialization ...
2397   // The initialization is idempotent, so we don't need locks.
2398   // In the future consider doing this via os::init_2().
2399   // SyncKnobs consist of <Key>=<Value> pairs in the style
2400   // of environment variables.  Start by converting ':' to NUL.
2401 
2402   if (SyncKnobs == NULL) SyncKnobs = "" ;
2403 
2404   size_t sz = strlen (SyncKnobs) ;
2405   char * knobs = (char *) malloc (sz + 2) ;
2406   if (knobs == NULL) {
2407      vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ;
2408      guarantee (0, "invariant") ;
2409   }
2410   strcpy (knobs, SyncKnobs) ;
2411   knobs[sz+1] = 0 ;
2412   for (char * p = knobs ; *p ; p++) {
2413      if (*p == ':') *p = 0 ;
2414   }
2415 
2416   #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); }
2417   SETKNOB(ReportSettings) ;
2418   SETKNOB(Verbose) ;
2419   SETKNOB(FixedSpin) ;
2420   SETKNOB(SpinLimit) ;
2421   SETKNOB(SpinBase) ;
2422   SETKNOB(SpinBackOff);
2423   SETKNOB(CASPenalty) ;
2424   SETKNOB(OXPenalty) ;
2425   SETKNOB(LogSpins) ;
2426   SETKNOB(SpinSetSucc) ;
2427   SETKNOB(SuccEnabled) ;
2428   SETKNOB(SuccRestrict) ;
2429   SETKNOB(Penalty) ;
2430   SETKNOB(Bonus) ;
2431   SETKNOB(BonusB) ;
2432   SETKNOB(Poverty) ;
2433   SETKNOB(SpinAfterFutile) ;
2434   SETKNOB(UsePause) ;
2435   SETKNOB(SpinEarly) ;
2436   SETKNOB(OState) ;
2437   SETKNOB(MaxSpinners) ;
2438   SETKNOB(PreSpin) ;
2439   SETKNOB(ExitPolicy) ;
2440   SETKNOB(QMode);
2441   SETKNOB(ResetEvent) ;
2442   SETKNOB(MoveNotifyee) ;
2443   SETKNOB(FastHSSEC) ;
2444   #undef SETKNOB
2445 
2446   if (os::is_MP()) {
2447      BackOffMask = (1 << Knob_SpinBackOff) - 1 ;
2448      if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ;
2449      // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2450   } else {
2451      Knob_SpinLimit = 0 ;
2452      Knob_SpinBase  = 0 ;
2453      Knob_PreSpin   = 0 ;
2454      Knob_FixedSpin = -1 ;
2455   }
2456 
2457   if (Knob_LogSpins == 0) {
2458      ObjectMonitor::_sync_FailedSpins = NULL ;
2459   }
2460 
2461   free (knobs) ;
2462   OrderAccess::fence() ;
2463   InitDone = 1 ;
2464 }
2465 
2466 #ifndef PRODUCT
2467 void ObjectMonitor::verify() {
2468 }
2469 
2470 void ObjectMonitor::print() {
2471 }
2472 #endif