1 /*
   2  * Copyright (c) 1998, 2017, 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/atomic.hpp"
  31 #include "runtime/handles.inline.hpp"
  32 #include "runtime/interfaceSupport.hpp"
  33 #include "runtime/mutexLocker.hpp"
  34 #include "runtime/objectMonitor.hpp"
  35 #include "runtime/objectMonitor.inline.hpp"
  36 #include "runtime/orderAccess.inline.hpp"
  37 #include "runtime/osThread.hpp"
  38 #include "runtime/stubRoutines.hpp"
  39 #include "runtime/thread.inline.hpp"
  40 #include "services/threadService.hpp"
  41 #include "trace/tracing.hpp"
  42 #include "trace/traceMacros.hpp"
  43 #include "utilities/dtrace.hpp"
  44 #include "utilities/macros.hpp"
  45 #include "utilities/preserveException.hpp"
  46 
  47 #ifdef DTRACE_ENABLED
  48 
  49 // Only bother with this argument setup if dtrace is available
  50 // TODO-FIXME: probes should not fire when caller is _blocked.  assert() accordingly.
  51 
  52 
  53 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread)                           \
  54   char* bytes = NULL;                                                      \
  55   int len = 0;                                                             \
  56   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  57   Symbol* klassname = ((oop)obj)->klass()->name();                         \
  58   if (klassname != NULL) {                                                 \
  59     bytes = (char*)klassname->bytes();                                     \
  60     len = klassname->utf8_length();                                        \
  61   }
  62 
  63 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)            \
  64   {                                                                        \
  65     if (DTraceMonitorProbes) {                                             \
  66       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  67       HOTSPOT_MONITOR_WAIT(jtid,                                           \
  68                            (monitor), bytes, len, (millis));               \
  69     }                                                                      \
  70   }
  71 
  72 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
  73 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
  74 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
  75 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
  76 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
  77 
  78 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)                  \
  79   {                                                                        \
  80     if (DTraceMonitorProbes) {                                             \
  81       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  82       HOTSPOT_MONITOR_##probe(jtid,                                        \
  83                               (uintptr_t)(monitor), bytes, len);           \
  84     }                                                                      \
  85   }
  86 
  87 #else //  ndef DTRACE_ENABLED
  88 
  89 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon)    {;}
  90 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon)          {;}
  91 
  92 #endif // ndef DTRACE_ENABLED
  93 
  94 // Tunables ...
  95 // The knob* variables are effectively final.  Once set they should
  96 // never be modified hence.  Consider using __read_mostly with GCC.
  97 
  98 int ObjectMonitor::Knob_ExitRelease = 0;
  99 int ObjectMonitor::Knob_Verbose     = 0;
 100 int ObjectMonitor::Knob_VerifyInUse = 0;
 101 int ObjectMonitor::Knob_VerifyMatch = 0;
 102 int ObjectMonitor::Knob_SpinLimit   = 5000;    // derived by an external tool -
 103 static int Knob_LogSpins            = 0;       // enable jvmstat tally for spins
 104 static int Knob_HandOff             = 0;
 105 static int Knob_ReportSettings      = 0;
 106 
 107 static int Knob_SpinBase            = 0;       // Floor AKA SpinMin
 108 static int Knob_SpinBackOff         = 0;       // spin-loop backoff
 109 static int Knob_CASPenalty          = -1;      // Penalty for failed CAS
 110 static int Knob_OXPenalty           = -1;      // Penalty for observed _owner change
 111 static int Knob_SpinSetSucc         = 1;       // spinners set the _succ field
 112 static int Knob_SpinEarly           = 1;
 113 static int Knob_SuccEnabled         = 1;       // futile wake throttling
 114 static int Knob_SuccRestrict        = 0;       // Limit successors + spinners to at-most-one
 115 static int Knob_MaxSpinners         = -1;      // Should be a function of # CPUs
 116 static int Knob_Bonus               = 100;     // spin success bonus
 117 static int Knob_BonusB              = 100;     // spin success bonus
 118 static int Knob_Penalty             = 200;     // spin failure penalty
 119 static int Knob_Poverty             = 1000;
 120 static int Knob_SpinAfterFutile     = 1;       // Spin after returning from park()
 121 static int Knob_FixedSpin           = 0;
 122 static int Knob_OState              = 3;       // Spinner checks thread state of _owner
 123 static int Knob_UsePause            = 1;
 124 static int Knob_ExitPolicy          = 0;
 125 static int Knob_PreSpin             = 10;      // 20-100 likely better
 126 static int Knob_ResetEvent          = 0;
 127 static int BackOffMask              = 0;
 128 
 129 static int Knob_FastHSSEC           = 0;
 130 static int Knob_MoveNotifyee        = 2;       // notify() - disposition of notifyee
 131 static int Knob_QMode               = 0;       // EntryList-cxq policy - queue discipline
 132 static volatile int InitDone        = 0;
 133 
 134 // -----------------------------------------------------------------------------
 135 // Theory of operations -- Monitors lists, thread residency, etc:
 136 //
 137 // * A thread acquires ownership of a monitor by successfully
 138 //   CAS()ing the _owner field from null to non-null.
 139 //
 140 // * Invariant: A thread appears on at most one monitor list --
 141 //   cxq, EntryList or WaitSet -- at any one time.
 142 //
 143 // * Contending threads "push" themselves onto the cxq with CAS
 144 //   and then spin/park.
 145 //
 146 // * After a contending thread eventually acquires the lock it must
 147 //   dequeue itself from either the EntryList or the cxq.
 148 //
 149 // * The exiting thread identifies and unparks an "heir presumptive"
 150 //   tentative successor thread on the EntryList.  Critically, the
 151 //   exiting thread doesn't unlink the successor thread from the EntryList.
 152 //   After having been unparked, the wakee will recontend for ownership of
 153 //   the monitor.   The successor (wakee) will either acquire the lock or
 154 //   re-park itself.
 155 //
 156 //   Succession is provided for by a policy of competitive handoff.
 157 //   The exiting thread does _not_ grant or pass ownership to the
 158 //   successor thread.  (This is also referred to as "handoff" succession").
 159 //   Instead the exiting thread releases ownership and possibly wakes
 160 //   a successor, so the successor can (re)compete for ownership of the lock.
 161 //   If the EntryList is empty but the cxq is populated the exiting
 162 //   thread will drain the cxq into the EntryList.  It does so by
 163 //   by detaching the cxq (installing null with CAS) and folding
 164 //   the threads from the cxq into the EntryList.  The EntryList is
 165 //   doubly linked, while the cxq is singly linked because of the
 166 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 167 //
 168 // * Concurrency invariants:
 169 //
 170 //   -- only the monitor owner may access or mutate the EntryList.
 171 //      The mutex property of the monitor itself protects the EntryList
 172 //      from concurrent interference.
 173 //   -- Only the monitor owner may detach the cxq.
 174 //
 175 // * The monitor entry list operations avoid locks, but strictly speaking
 176 //   they're not lock-free.  Enter is lock-free, exit is not.
 177 //   For a description of 'Methods and apparatus providing non-blocking access
 178 //   to a resource,' see U.S. Pat. No. 7844973.
 179 //
 180 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 181 //   detaching thread.  This mechanism is immune from the ABA corruption.
 182 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 183 //
 184 // * Taken together, the cxq and the EntryList constitute or form a
 185 //   single logical queue of threads stalled trying to acquire the lock.
 186 //   We use two distinct lists to improve the odds of a constant-time
 187 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 188 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 189 //   A key desideratum is to minimize queue & monitor metadata manipulation
 190 //   that occurs while holding the monitor lock -- that is, we want to
 191 //   minimize monitor lock holds times.  Note that even a small amount of
 192 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 193 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 194 //   locks and monitor metadata.
 195 //
 196 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 197 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 198 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 199 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 200 //
 201 //   The EntryList is ordered by the prevailing queue discipline and
 202 //   can be organized in any convenient fashion, such as a doubly-linked list or
 203 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 204 //   to operate in constant-time.  If we need a priority queue then something akin
 205 //   to Solaris' sleepq would work nicely.  Viz.,
 206 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 207 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 208 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 209 //   EntryList accordingly.
 210 //
 211 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 212 //   somewhat similar to an elevator-scan.
 213 //
 214 // * The monitor synchronization subsystem avoids the use of native
 215 //   synchronization primitives except for the narrow platform-specific
 216 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 217 //   the semantics of park-unpark.  Put another way, this monitor implementation
 218 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 219 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 220 //   underlying OS manages the READY<->RUN transitions.
 221 //
 222 // * Waiting threads reside on the WaitSet list -- wait() puts
 223 //   the caller onto the WaitSet.
 224 //
 225 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 226 //   either the EntryList or cxq.  Subsequent exit() operations will
 227 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 228 //   it's likely the notifyee would simply impale itself on the lock held
 229 //   by the notifier.
 230 //
 231 // * An interesting alternative is to encode cxq as (List,LockByte) where
 232 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 233 //   variable, like _recursions, in the scheme.  The threads or Events that form
 234 //   the list would have to be aligned in 256-byte addresses.  A thread would
 235 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 236 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 237 //   Note that is is *not* word-tearing, but it does presume that full-word
 238 //   CAS operations are coherent with intermix with STB operations.  That's true
 239 //   on most common processors.
 240 //
 241 // * See also http://blogs.sun.com/dave
 242 
 243 
 244 // -----------------------------------------------------------------------------
 245 // Enter support
 246 
 247 void ObjectMonitor::enter(TRAPS) {
 248   // The following code is ordered to check the most common cases first
 249   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 250   Thread * const Self = THREAD;
 251 
 252   void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL);
 253   if (cur == NULL) {
 254     // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
 255     assert(_recursions == 0, "invariant");
 256     assert(_owner == Self, "invariant");
 257     return;
 258   }
 259 
 260   if (cur == Self) {
 261     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 262     _recursions++;
 263     return;
 264   }
 265 
 266   if (Self->is_lock_owned ((address)cur)) {
 267     assert(_recursions == 0, "internal state error");
 268     _recursions = 1;
 269     // Commute owner from a thread-specific on-stack BasicLockObject address to
 270     // a full-fledged "Thread *".
 271     _owner = Self;
 272     return;
 273   }
 274 
 275   // We've encountered genuine contention.
 276   assert(Self->_Stalled == 0, "invariant");
 277   Self->_Stalled = intptr_t(this);
 278 
 279   // Try one round of spinning *before* enqueueing Self
 280   // and before going through the awkward and expensive state
 281   // transitions.  The following spin is strictly optional ...
 282   // Note that if we acquire the monitor from an initial spin
 283   // we forgo posting JVMTI events and firing DTRACE probes.
 284   if (Knob_SpinEarly && TrySpin (Self) > 0) {
 285     assert(_owner == Self, "invariant");
 286     assert(_recursions == 0, "invariant");
 287     assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 288     Self->_Stalled = 0;
 289     return;
 290   }
 291 
 292   assert(_owner != Self, "invariant");
 293   assert(_succ != Self, "invariant");
 294   assert(Self->is_Java_thread(), "invariant");
 295   JavaThread * jt = (JavaThread *) Self;
 296   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 297   assert(jt->thread_state() != _thread_blocked, "invariant");
 298   assert(this->object() != NULL, "invariant");
 299   assert(_count >= 0, "invariant");
 300 
 301   // Prevent deflation at STW-time.  See deflate_idle_monitors() and is_busy().
 302   // Ensure the object-monitor relationship remains stable while there's contention.
 303   Atomic::inc(&_count);
 304 
 305   EventJavaMonitorEnter event;
 306 
 307   { // Change java thread status to indicate blocked on monitor enter.
 308     JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
 309 


 310     DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
 311     if (JvmtiExport::should_post_monitor_contended_enter()) {
 312       JvmtiExport::post_monitor_contended_enter(jt, this);
 313 
 314       // The current thread does not yet own the monitor and does not
 315       // yet appear on any queues that would get it made the successor.
 316       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 317       // handler cannot accidentally consume an unpark() meant for the
 318       // ParkEvent associated with this ObjectMonitor.
 319     }
 320 
 321     OSThreadContendState osts(Self->osthread());
 322     ThreadBlockInVM tbivm(jt);
 323 
 324     Self->set_current_pending_monitor(this);
 325 
 326     // TODO-FIXME: change the following for(;;) loop to straight-line code.
 327     for (;;) {
 328       jt->set_suspend_equivalent();
 329       // cleared by handle_special_suspend_equivalent_condition()
 330       // or java_suspend_self()
 331 
 332       EnterI(THREAD);
 333 
 334       if (!ExitSuspendEquivalent(jt)) break;
 335 
 336       // We have acquired the contended monitor, but while we were
 337       // waiting another thread suspended us. We don't want to enter
 338       // the monitor while suspended because that would surprise the
 339       // thread that suspended us.
 340       //
 341       _recursions = 0;
 342       _succ = NULL;
 343       exit(false, Self);
 344 
 345       jt->java_suspend_self();
 346     }
 347     Self->set_current_pending_monitor(NULL);
 348 
 349     // We cleared the pending monitor info since we've just gotten past
 350     // the enter-check-for-suspend dance and we now own the monitor free
 351     // and clear, i.e., it is no longer pending. The ThreadBlockInVM
 352     // destructor can go to a safepoint at the end of this block. If we
 353     // do a thread dump during that safepoint, then this thread will show
 354     // as having "-locked" the monitor, but the OS and java.lang.Thread
 355     // states will still report that the thread is blocked trying to
 356     // acquire it.
 357   }
 358 
 359   Atomic::dec(&_count);
 360   assert(_count >= 0, "invariant");
 361   Self->_Stalled = 0;
 362 
 363   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 364   assert(_recursions == 0, "invariant");
 365   assert(_owner == Self, "invariant");
 366   assert(_succ != Self, "invariant");
 367   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 368 
 369   // The thread -- now the owner -- is back in vm mode.
 370   // Report the glorious news via TI,DTrace and jvmstat.
 371   // The probe effect is non-trivial.  All the reportage occurs
 372   // while we hold the monitor, increasing the length of the critical
 373   // section.  Amdahl's parallel speedup law comes vividly into play.
 374   //
 375   // Another option might be to aggregate the events (thread local or
 376   // per-monitor aggregation) and defer reporting until a more opportune
 377   // time -- such as next time some thread encounters contention but has
 378   // yet to acquire the lock.  While spinning that thread could
 379   // spinning we could increment JVMStat counters, etc.
 380 
 381   DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
 382   if (JvmtiExport::should_post_monitor_contended_entered()) {
 383     JvmtiExport::post_monitor_contended_entered(jt, this);
 384 
 385     // The current thread already owns the monitor and is not going to
 386     // call park() for the remainder of the monitor enter protocol. So
 387     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 388     // event handler consumed an unpark() issued by the thread that
 389     // just exited the monitor.
 390   }
 391 
 392   if (event.should_commit()) {
 393     event.set_monitorClass(((oop)this->object())->klass());
 394     event.set_previousOwner((TYPE_THREAD)_previous_owner_tid);
 395     event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
 396     event.commit();
 397   }
 398 
 399   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 400 }
 401 
 402 
 403 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 404 // Callers must compensate as needed.
 405 
 406 int ObjectMonitor::TryLock(Thread * Self) {
 407   void * own = _owner;
 408   if (own != NULL) return 0;
 409   if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) {
 410     // Either guarantee _recursions == 0 or set _recursions = 0.
 411     assert(_recursions == 0, "invariant");
 412     assert(_owner == Self, "invariant");
 413     return 1;
 414   }
 415   // The lock had been free momentarily, but we lost the race to the lock.
 416   // Interference -- the CAS failed.
 417   // We can either return -1 or retry.
 418   // Retry doesn't make as much sense because the lock was just acquired.
 419   return -1;
 420 }
 421 
 422 #define MAX_RECHECK_INTERVAL 1000
 423 
 424 void ObjectMonitor::EnterI(TRAPS) {
 425   Thread * const Self = THREAD;
 426   assert(Self->is_Java_thread(), "invariant");
 427   assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
 428 
 429   // Try the lock - TATAS
 430   if (TryLock (Self) > 0) {
 431     assert(_succ != Self, "invariant");
 432     assert(_owner == Self, "invariant");
 433     assert(_Responsible != Self, "invariant");
 434     return;
 435   }
 436 
 437   DeferredInitialize();
 438 
 439   // We try one round of spinning *before* enqueueing Self.
 440   //
 441   // If the _owner is ready but OFFPROC we could use a YieldTo()
 442   // operation to donate the remainder of this thread's quantum
 443   // to the owner.  This has subtle but beneficial affinity
 444   // effects.
 445 
 446   if (TrySpin (Self) > 0) {
 447     assert(_owner == Self, "invariant");
 448     assert(_succ != Self, "invariant");
 449     assert(_Responsible != Self, "invariant");
 450     return;
 451   }
 452 
 453   // The Spin failed -- Enqueue and park the thread ...
 454   assert(_succ != Self, "invariant");
 455   assert(_owner != Self, "invariant");
 456   assert(_Responsible != Self, "invariant");
 457 
 458   // Enqueue "Self" on ObjectMonitor's _cxq.
 459   //
 460   // Node acts as a proxy for Self.
 461   // As an aside, if were to ever rewrite the synchronization code mostly
 462   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 463   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 464   // as well as eliminate a subset of ABA issues.
 465   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 466 
 467   ObjectWaiter node(Self);
 468   Self->_ParkEvent->reset();
 469   node._prev   = (ObjectWaiter *) 0xBAD;
 470   node.TState  = ObjectWaiter::TS_CXQ;
 471 
 472   // Push "Self" onto the front of the _cxq.
 473   // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
 474   // Note that spinning tends to reduce the rate at which threads
 475   // enqueue and dequeue on EntryList|cxq.
 476   ObjectWaiter * nxt;
 477   for (;;) {
 478     node._next = nxt = _cxq;
 479     if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break;
 480 
 481     // Interference - the CAS failed because _cxq changed.  Just retry.
 482     // As an optional optimization we retry the lock.
 483     if (TryLock (Self) > 0) {
 484       assert(_succ != Self, "invariant");
 485       assert(_owner == Self, "invariant");
 486       assert(_Responsible != Self, "invariant");
 487       return;
 488     }
 489   }
 490 
 491   // Check for cxq|EntryList edge transition to non-null.  This indicates
 492   // the onset of contention.  While contention persists exiting threads
 493   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 494   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 495   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 496   // arrange for one of the contending thread to use a timed park() operations
 497   // to detect and recover from the race.  (Stranding is form of progress failure
 498   // where the monitor is unlocked but all the contending threads remain parked).
 499   // That is, at least one of the contended threads will periodically poll _owner.
 500   // One of the contending threads will become the designated "Responsible" thread.
 501   // The Responsible thread uses a timed park instead of a normal indefinite park
 502   // operation -- it periodically wakes and checks for and recovers from potential
 503   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 504   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 505   // be responsible for a monitor.
 506   //
 507   // Currently, one of the contended threads takes on the added role of "Responsible".
 508   // A viable alternative would be to use a dedicated "stranding checker" thread
 509   // that periodically iterated over all the threads (or active monitors) and unparked
 510   // successors where there was risk of stranding.  This would help eliminate the
 511   // timer scalability issues we see on some platforms as we'd only have one thread
 512   // -- the checker -- parked on a timer.
 513 
 514   if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) {
 515     // Try to assume the role of responsible thread for the monitor.
 516     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=Self }
 517     Atomic::cmpxchg_ptr(Self, &_Responsible, NULL);
 518   }
 519 
 520   // The lock might have been released while this thread was occupied queueing
 521   // itself onto _cxq.  To close the race and avoid "stranding" and
 522   // progress-liveness failure we must resample-retry _owner before parking.
 523   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 524   // In this case the ST-MEMBAR is accomplished with CAS().
 525   //
 526   // TODO: Defer all thread state transitions until park-time.
 527   // Since state transitions are heavy and inefficient we'd like
 528   // to defer the state transitions until absolutely necessary,
 529   // and in doing so avoid some transitions ...
 530 
 531   TEVENT(Inflated enter - Contention);
 532   int nWakeups = 0;
 533   int recheckInterval = 1;
 534 
 535   for (;;) {
 536 
 537     if (TryLock(Self) > 0) break;
 538     assert(_owner != Self, "invariant");
 539 
 540     if ((SyncFlags & 2) && _Responsible == NULL) {
 541       Atomic::cmpxchg_ptr(Self, &_Responsible, NULL);
 542     }
 543 
 544     // park self
 545     if (_Responsible == Self || (SyncFlags & 1)) {
 546       TEVENT(Inflated enter - park TIMED);
 547       Self->_ParkEvent->park((jlong) recheckInterval);
 548       // Increase the recheckInterval, but clamp the value.
 549       recheckInterval *= 8;
 550       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 551         recheckInterval = MAX_RECHECK_INTERVAL;
 552       }
 553     } else {
 554       TEVENT(Inflated enter - park UNTIMED);
 555       Self->_ParkEvent->park();
 556     }
 557 
 558     if (TryLock(Self) > 0) break;
 559 
 560     // The lock is still contested.
 561     // Keep a tally of the # of futile wakeups.
 562     // Note that the counter is not protected by a lock or updated by atomics.
 563     // That is by design - we trade "lossy" counters which are exposed to
 564     // races during updates for a lower probe effect.
 565     TEVENT(Inflated enter - Futile wakeup);
 566     // This PerfData object can be used in parallel with a safepoint.
 567     // See the work around in PerfDataManager::destroy().
 568     OM_PERFDATA_OP(FutileWakeups, inc());
 569     ++nWakeups;
 570 
 571     // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
 572     // We can defer clearing _succ until after the spin completes
 573     // TrySpin() must tolerate being called with _succ == Self.
 574     // Try yet another round of adaptive spinning.
 575     if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
 576 
 577     // We can find that we were unpark()ed and redesignated _succ while
 578     // we were spinning.  That's harmless.  If we iterate and call park(),
 579     // park() will consume the event and return immediately and we'll
 580     // just spin again.  This pattern can repeat, leaving _succ to simply
 581     // spin on a CPU.  Enable Knob_ResetEvent to clear pending unparks().
 582     // Alternately, we can sample fired() here, and if set, forgo spinning
 583     // in the next iteration.
 584 
 585     if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
 586       Self->_ParkEvent->reset();
 587       OrderAccess::fence();
 588     }
 589     if (_succ == Self) _succ = NULL;
 590 
 591     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 592     OrderAccess::fence();
 593   }
 594 
 595   // Egress :
 596   // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
 597   // Normally we'll find Self on the EntryList .
 598   // From the perspective of the lock owner (this thread), the
 599   // EntryList is stable and cxq is prepend-only.
 600   // The head of cxq is volatile but the interior is stable.
 601   // In addition, Self.TState is stable.
 602 
 603   assert(_owner == Self, "invariant");
 604   assert(object() != NULL, "invariant");
 605   // I'd like to write:
 606   //   guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 607   // but as we're at a safepoint that's not safe.
 608 
 609   UnlinkAfterAcquire(Self, &node);
 610   if (_succ == Self) _succ = NULL;
 611 
 612   assert(_succ != Self, "invariant");
 613   if (_Responsible == Self) {
 614     _Responsible = NULL;
 615     OrderAccess::fence(); // Dekker pivot-point
 616 
 617     // We may leave threads on cxq|EntryList without a designated
 618     // "Responsible" thread.  This is benign.  When this thread subsequently
 619     // exits the monitor it can "see" such preexisting "old" threads --
 620     // threads that arrived on the cxq|EntryList before the fence, above --
 621     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 622     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 623     // non-null and elect a new "Responsible" timer thread.
 624     //
 625     // This thread executes:
 626     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 627     //    LD cxq|EntryList               (in subsequent exit)
 628     //
 629     // Entering threads in the slow/contended path execute:
 630     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 631     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 632     //
 633     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 634     // exit operation from floating above the ST Responsible=null.
 635   }
 636 
 637   // We've acquired ownership with CAS().
 638   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 639   // But since the CAS() this thread may have also stored into _succ,
 640   // EntryList, cxq or Responsible.  These meta-data updates must be
 641   // visible __before this thread subsequently drops the lock.
 642   // Consider what could occur if we didn't enforce this constraint --
 643   // STs to monitor meta-data and user-data could reorder with (become
 644   // visible after) the ST in exit that drops ownership of the lock.
 645   // Some other thread could then acquire the lock, but observe inconsistent
 646   // or old monitor meta-data and heap data.  That violates the JMM.
 647   // To that end, the 1-0 exit() operation must have at least STST|LDST
 648   // "release" barrier semantics.  Specifically, there must be at least a
 649   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 650   // the lock.   The barrier ensures that changes to monitor meta-data and data
 651   // protected by the lock will be visible before we release the lock, and
 652   // therefore before some other thread (CPU) has a chance to acquire the lock.
 653   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 654   //
 655   // Critically, any prior STs to _succ or EntryList must be visible before
 656   // the ST of null into _owner in the *subsequent* (following) corresponding
 657   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 658   // execute a serializing instruction.
 659 
 660   if (SyncFlags & 8) {
 661     OrderAccess::fence();
 662   }
 663   return;
 664 }
 665 
 666 // ReenterI() is a specialized inline form of the latter half of the
 667 // contended slow-path from EnterI().  We use ReenterI() only for
 668 // monitor reentry in wait().
 669 //
 670 // In the future we should reconcile EnterI() and ReenterI(), adding
 671 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
 672 // loop accordingly.
 673 
 674 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
 675   assert(Self != NULL, "invariant");
 676   assert(SelfNode != NULL, "invariant");
 677   assert(SelfNode->_thread == Self, "invariant");
 678   assert(_waiters > 0, "invariant");
 679   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 680   assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
 681   JavaThread * jt = (JavaThread *) Self;
 682 
 683   int nWakeups = 0;
 684   for (;;) {
 685     ObjectWaiter::TStates v = SelfNode->TState;
 686     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
 687     assert(_owner != Self, "invariant");
 688 
 689     if (TryLock(Self) > 0) break;
 690     if (TrySpin(Self) > 0) break;
 691 
 692     TEVENT(Wait Reentry - parking);
 693 
 694     // State transition wrappers around park() ...
 695     // ReenterI() wisely defers state transitions until
 696     // it's clear we must park the thread.
 697     {
 698       OSThreadContendState osts(Self->osthread());
 699       ThreadBlockInVM tbivm(jt);
 700 
 701       // cleared by handle_special_suspend_equivalent_condition()
 702       // or java_suspend_self()
 703       jt->set_suspend_equivalent();
 704       if (SyncFlags & 1) {
 705         Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL);
 706       } else {
 707         Self->_ParkEvent->park();
 708       }
 709 
 710       // were we externally suspended while we were waiting?
 711       for (;;) {
 712         if (!ExitSuspendEquivalent(jt)) break;
 713         if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
 714         jt->java_suspend_self();
 715         jt->set_suspend_equivalent();
 716       }
 717     }
 718 
 719     // Try again, but just so we distinguish between futile wakeups and
 720     // successful wakeups.  The following test isn't algorithmically
 721     // necessary, but it helps us maintain sensible statistics.
 722     if (TryLock(Self) > 0) break;
 723 
 724     // The lock is still contested.
 725     // Keep a tally of the # of futile wakeups.
 726     // Note that the counter is not protected by a lock or updated by atomics.
 727     // That is by design - we trade "lossy" counters which are exposed to
 728     // races during updates for a lower probe effect.
 729     TEVENT(Wait Reentry - futile wakeup);
 730     ++nWakeups;
 731 
 732     // Assuming this is not a spurious wakeup we'll normally
 733     // find that _succ == Self.
 734     if (_succ == Self) _succ = NULL;
 735 
 736     // Invariant: after clearing _succ a contending thread
 737     // *must* retry  _owner before parking.
 738     OrderAccess::fence();
 739 
 740     // This PerfData object can be used in parallel with a safepoint.
 741     // See the work around in PerfDataManager::destroy().
 742     OM_PERFDATA_OP(FutileWakeups, inc());
 743   }
 744 
 745   // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
 746   // Normally we'll find Self on the EntryList.
 747   // Unlinking from the EntryList is constant-time and atomic-free.
 748   // From the perspective of the lock owner (this thread), the
 749   // EntryList is stable and cxq is prepend-only.
 750   // The head of cxq is volatile but the interior is stable.
 751   // In addition, Self.TState is stable.
 752 
 753   assert(_owner == Self, "invariant");
 754   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 755   UnlinkAfterAcquire(Self, SelfNode);
 756   if (_succ == Self) _succ = NULL;
 757   assert(_succ != Self, "invariant");
 758   SelfNode->TState = ObjectWaiter::TS_RUN;
 759   OrderAccess::fence();      // see comments at the end of EnterI()
 760 }
 761 
 762 // By convention we unlink a contending thread from EntryList|cxq immediately
 763 // after the thread acquires the lock in ::enter().  Equally, we could defer
 764 // unlinking the thread until ::exit()-time.
 765 
 766 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
 767   assert(_owner == Self, "invariant");
 768   assert(SelfNode->_thread == Self, "invariant");
 769 
 770   if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
 771     // Normal case: remove Self from the DLL EntryList .
 772     // This is a constant-time operation.
 773     ObjectWaiter * nxt = SelfNode->_next;
 774     ObjectWaiter * prv = SelfNode->_prev;
 775     if (nxt != NULL) nxt->_prev = prv;
 776     if (prv != NULL) prv->_next = nxt;
 777     if (SelfNode == _EntryList) _EntryList = nxt;
 778     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
 779     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
 780     TEVENT(Unlink from EntryList);
 781   } else {
 782     assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
 783     // Inopportune interleaving -- Self is still on the cxq.
 784     // This usually means the enqueue of self raced an exiting thread.
 785     // Normally we'll find Self near the front of the cxq, so
 786     // dequeueing is typically fast.  If needbe we can accelerate
 787     // this with some MCS/CHL-like bidirectional list hints and advisory
 788     // back-links so dequeueing from the interior will normally operate
 789     // in constant-time.
 790     // Dequeue Self from either the head (with CAS) or from the interior
 791     // with a linear-time scan and normal non-atomic memory operations.
 792     // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
 793     // and then unlink Self from EntryList.  We have to drain eventually,
 794     // so it might as well be now.
 795 
 796     ObjectWaiter * v = _cxq;
 797     assert(v != NULL, "invariant");
 798     if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) {
 799       // The CAS above can fail from interference IFF a "RAT" arrived.
 800       // In that case Self must be in the interior and can no longer be
 801       // at the head of cxq.
 802       if (v == SelfNode) {
 803         assert(_cxq != v, "invariant");
 804         v = _cxq;          // CAS above failed - start scan at head of list
 805       }
 806       ObjectWaiter * p;
 807       ObjectWaiter * q = NULL;
 808       for (p = v; p != NULL && p != SelfNode; p = p->_next) {
 809         q = p;
 810         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
 811       }
 812       assert(v != SelfNode, "invariant");
 813       assert(p == SelfNode, "Node not found on cxq");
 814       assert(p != _cxq, "invariant");
 815       assert(q != NULL, "invariant");
 816       assert(q->_next == p, "invariant");
 817       q->_next = p->_next;
 818     }
 819     TEVENT(Unlink from cxq);
 820   }
 821 
 822 #ifdef ASSERT
 823   // Diagnostic hygiene ...
 824   SelfNode->_prev  = (ObjectWaiter *) 0xBAD;
 825   SelfNode->_next  = (ObjectWaiter *) 0xBAD;
 826   SelfNode->TState = ObjectWaiter::TS_RUN;
 827 #endif
 828 }
 829 
 830 // -----------------------------------------------------------------------------
 831 // Exit support
 832 //
 833 // exit()
 834 // ~~~~~~
 835 // Note that the collector can't reclaim the objectMonitor or deflate
 836 // the object out from underneath the thread calling ::exit() as the
 837 // thread calling ::exit() never transitions to a stable state.
 838 // This inhibits GC, which in turn inhibits asynchronous (and
 839 // inopportune) reclamation of "this".
 840 //
 841 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
 842 // There's one exception to the claim above, however.  EnterI() can call
 843 // exit() to drop a lock if the acquirer has been externally suspended.
 844 // In that case exit() is called with _thread_state as _thread_blocked,
 845 // but the monitor's _count field is > 0, which inhibits reclamation.
 846 //
 847 // 1-0 exit
 848 // ~~~~~~~~
 849 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
 850 // the fast-path operators have been optimized so the common ::exit()
 851 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
 852 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
 853 // greatly improves latency -- MEMBAR and CAS having considerable local
 854 // latency on modern processors -- but at the cost of "stranding".  Absent the
 855 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
 856 // ::enter() path, resulting in the entering thread being stranding
 857 // and a progress-liveness failure.   Stranding is extremely rare.
 858 // We use timers (timed park operations) & periodic polling to detect
 859 // and recover from stranding.  Potentially stranded threads periodically
 860 // wake up and poll the lock.  See the usage of the _Responsible variable.
 861 //
 862 // The CAS() in enter provides for safety and exclusion, while the CAS or
 863 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
 864 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
 865 // We detect and recover from stranding with timers.
 866 //
 867 // If a thread transiently strands it'll park until (a) another
 868 // thread acquires the lock and then drops the lock, at which time the
 869 // exiting thread will notice and unpark the stranded thread, or, (b)
 870 // the timer expires.  If the lock is high traffic then the stranding latency
 871 // will be low due to (a).  If the lock is low traffic then the odds of
 872 // stranding are lower, although the worst-case stranding latency
 873 // is longer.  Critically, we don't want to put excessive load in the
 874 // platform's timer subsystem.  We want to minimize both the timer injection
 875 // rate (timers created/sec) as well as the number of timers active at
 876 // any one time.  (more precisely, we want to minimize timer-seconds, which is
 877 // the integral of the # of active timers at any instant over time).
 878 // Both impinge on OS scalability.  Given that, at most one thread parked on
 879 // a monitor will use a timer.
 880 //
 881 // There is also the risk of a futile wake-up. If we drop the lock
 882 // another thread can reacquire the lock immediately, and we can
 883 // then wake a thread unnecessarily. This is benign, and we've
 884 // structured the code so the windows are short and the frequency
 885 // of such futile wakups is low.
 886 
 887 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
 888   Thread * const Self = THREAD;
 889   if (THREAD != _owner) {
 890     if (THREAD->is_lock_owned((address) _owner)) {
 891       // Transmute _owner from a BasicLock pointer to a Thread address.
 892       // We don't need to hold _mutex for this transition.
 893       // Non-null to Non-null is safe as long as all readers can
 894       // tolerate either flavor.
 895       assert(_recursions == 0, "invariant");
 896       _owner = THREAD;
 897       _recursions = 0;
 898     } else {
 899       // Apparent unbalanced locking ...
 900       // Naively we'd like to throw IllegalMonitorStateException.
 901       // As a practical matter we can neither allocate nor throw an
 902       // exception as ::exit() can be called from leaf routines.
 903       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
 904       // Upon deeper reflection, however, in a properly run JVM the only
 905       // way we should encounter this situation is in the presence of
 906       // unbalanced JNI locking. TODO: CheckJNICalls.
 907       // See also: CR4414101
 908       TEVENT(Exit - Throw IMSX);
 909       assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
 910       return;
 911     }
 912   }
 913 
 914   if (_recursions != 0) {
 915     _recursions--;        // this is simple recursive enter
 916     TEVENT(Inflated exit - recursive);
 917     return;
 918   }
 919 
 920   // Invariant: after setting Responsible=null an thread must execute
 921   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
 922   if ((SyncFlags & 4) == 0) {
 923     _Responsible = NULL;
 924   }
 925 
 926 #if INCLUDE_TRACE
 927   // get the owner's thread id for the MonitorEnter event
 928   // if it is enabled and the thread isn't suspended
 929   if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) {
 930     _previous_owner_tid = THREAD_TRACE_ID(Self);
 931   }
 932 #endif
 933 
 934   for (;;) {
 935     assert(THREAD == _owner, "invariant");
 936 
 937     if (Knob_ExitPolicy == 0) {
 938       // release semantics: prior loads and stores from within the critical section
 939       // must not float (reorder) past the following store that drops the lock.
 940       // On SPARC that requires MEMBAR #loadstore|#storestore.
 941       // But of course in TSO #loadstore|#storestore is not required.
 942       // I'd like to write one of the following:
 943       // A.  OrderAccess::release() ; _owner = NULL
 944       // B.  OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
 945       // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
 946       // store into a _dummy variable.  That store is not needed, but can result
 947       // in massive wasteful coherency traffic on classic SMP systems.
 948       // Instead, I use release_store(), which is implemented as just a simple
 949       // ST on x64, x86 and SPARC.
 950       OrderAccess::release_store_ptr(&_owner, NULL);   // drop the lock
 951       OrderAccess::storeload();                        // See if we need to wake a successor
 952       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 953         TEVENT(Inflated exit - simple egress);
 954         return;
 955       }
 956       TEVENT(Inflated exit - complex egress);
 957       // Other threads are blocked trying to acquire the lock.
 958 
 959       // Normally the exiting thread is responsible for ensuring succession,
 960       // but if other successors are ready or other entering threads are spinning
 961       // then this thread can simply store NULL into _owner and exit without
 962       // waking a successor.  The existence of spinners or ready successors
 963       // guarantees proper succession (liveness).  Responsibility passes to the
 964       // ready or running successors.  The exiting thread delegates the duty.
 965       // More precisely, if a successor already exists this thread is absolved
 966       // of the responsibility of waking (unparking) one.
 967       //
 968       // The _succ variable is critical to reducing futile wakeup frequency.
 969       // _succ identifies the "heir presumptive" thread that has been made
 970       // ready (unparked) but that has not yet run.  We need only one such
 971       // successor thread to guarantee progress.
 972       // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 973       // section 3.3 "Futile Wakeup Throttling" for details.
 974       //
 975       // Note that spinners in Enter() also set _succ non-null.
 976       // In the current implementation spinners opportunistically set
 977       // _succ so that exiting threads might avoid waking a successor.
 978       // Another less appealing alternative would be for the exiting thread
 979       // to drop the lock and then spin briefly to see if a spinner managed
 980       // to acquire the lock.  If so, the exiting thread could exit
 981       // immediately without waking a successor, otherwise the exiting
 982       // thread would need to dequeue and wake a successor.
 983       // (Note that we'd need to make the post-drop spin short, but no
 984       // shorter than the worst-case round-trip cache-line migration time.
 985       // The dropped lock needs to become visible to the spinner, and then
 986       // the acquisition of the lock by the spinner must become visible to
 987       // the exiting thread).
 988 
 989       // It appears that an heir-presumptive (successor) must be made ready.
 990       // Only the current lock owner can manipulate the EntryList or
 991       // drain _cxq, so we need to reacquire the lock.  If we fail
 992       // to reacquire the lock the responsibility for ensuring succession
 993       // falls to the new owner.
 994       //
 995       if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
 996         return;
 997       }
 998       TEVENT(Exit - Reacquired);
 999     } else {
1000       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1001         OrderAccess::release_store_ptr(&_owner, NULL);   // drop the lock
1002         OrderAccess::storeload();
1003         // Ratify the previously observed values.
1004         if (_cxq == NULL || _succ != NULL) {
1005           TEVENT(Inflated exit - simple egress);
1006           return;
1007         }
1008 
1009         // inopportune interleaving -- the exiting thread (this thread)
1010         // in the fast-exit path raced an entering thread in the slow-enter
1011         // path.
1012         // We have two choices:
1013         // A.  Try to reacquire the lock.
1014         //     If the CAS() fails return immediately, otherwise
1015         //     we either restart/rerun the exit operation, or simply
1016         //     fall-through into the code below which wakes a successor.
1017         // B.  If the elements forming the EntryList|cxq are TSM
1018         //     we could simply unpark() the lead thread and return
1019         //     without having set _succ.
1020         if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1021           TEVENT(Inflated exit - reacquired succeeded);
1022           return;
1023         }
1024         TEVENT(Inflated exit - reacquired failed);
1025       } else {
1026         TEVENT(Inflated exit - complex egress);
1027       }
1028     }
1029 
1030     guarantee(_owner == THREAD, "invariant");
1031 
1032     ObjectWaiter * w = NULL;
1033     int QMode = Knob_QMode;
1034 
1035     if (QMode == 2 && _cxq != NULL) {
1036       // QMode == 2 : cxq has precedence over EntryList.
1037       // Try to directly wake a successor from the cxq.
1038       // If successful, the successor will need to unlink itself from cxq.
1039       w = _cxq;
1040       assert(w != NULL, "invariant");
1041       assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1042       ExitEpilog(Self, w);
1043       return;
1044     }
1045 
1046     if (QMode == 3 && _cxq != NULL) {
1047       // Aggressively drain cxq into EntryList at the first opportunity.
1048       // This policy ensure that recently-run threads live at the head of EntryList.
1049       // Drain _cxq into EntryList - bulk transfer.
1050       // First, detach _cxq.
1051       // The following loop is tantamount to: w = swap(&cxq, NULL)
1052       w = _cxq;
1053       for (;;) {
1054         assert(w != NULL, "Invariant");
1055         ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w);
1056         if (u == w) break;
1057         w = u;
1058       }
1059       assert(w != NULL, "invariant");
1060 
1061       ObjectWaiter * q = NULL;
1062       ObjectWaiter * p;
1063       for (p = w; p != NULL; p = p->_next) {
1064         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1065         p->TState = ObjectWaiter::TS_ENTER;
1066         p->_prev = q;
1067         q = p;
1068       }
1069 
1070       // Append the RATs to the EntryList
1071       // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1072       ObjectWaiter * Tail;
1073       for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1074            Tail = Tail->_next)
1075         /* empty */;
1076       if (Tail == NULL) {
1077         _EntryList = w;
1078       } else {
1079         Tail->_next = w;
1080         w->_prev = Tail;
1081       }
1082 
1083       // Fall thru into code that tries to wake a successor from EntryList
1084     }
1085 
1086     if (QMode == 4 && _cxq != NULL) {
1087       // Aggressively drain cxq into EntryList at the first opportunity.
1088       // This policy ensure that recently-run threads live at the head of EntryList.
1089 
1090       // Drain _cxq into EntryList - bulk transfer.
1091       // First, detach _cxq.
1092       // The following loop is tantamount to: w = swap(&cxq, NULL)
1093       w = _cxq;
1094       for (;;) {
1095         assert(w != NULL, "Invariant");
1096         ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w);
1097         if (u == w) break;
1098         w = u;
1099       }
1100       assert(w != NULL, "invariant");
1101 
1102       ObjectWaiter * q = NULL;
1103       ObjectWaiter * p;
1104       for (p = w; p != NULL; p = p->_next) {
1105         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1106         p->TState = ObjectWaiter::TS_ENTER;
1107         p->_prev = q;
1108         q = p;
1109       }
1110 
1111       // Prepend the RATs to the EntryList
1112       if (_EntryList != NULL) {
1113         q->_next = _EntryList;
1114         _EntryList->_prev = q;
1115       }
1116       _EntryList = w;
1117 
1118       // Fall thru into code that tries to wake a successor from EntryList
1119     }
1120 
1121     w = _EntryList;
1122     if (w != NULL) {
1123       // I'd like to write: guarantee (w->_thread != Self).
1124       // But in practice an exiting thread may find itself on the EntryList.
1125       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1126       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1127       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1128       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1129       // release the lock "O".  T2 resumes immediately after the ST of null into
1130       // _owner, above.  T2 notices that the EntryList is populated, so it
1131       // reacquires the lock and then finds itself on the EntryList.
1132       // Given all that, we have to tolerate the circumstance where "w" is
1133       // associated with Self.
1134       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1135       ExitEpilog(Self, w);
1136       return;
1137     }
1138 
1139     // If we find that both _cxq and EntryList are null then just
1140     // re-run the exit protocol from the top.
1141     w = _cxq;
1142     if (w == NULL) continue;
1143 
1144     // Drain _cxq into EntryList - bulk transfer.
1145     // First, detach _cxq.
1146     // The following loop is tantamount to: w = swap(&cxq, NULL)
1147     for (;;) {
1148       assert(w != NULL, "Invariant");
1149       ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w);
1150       if (u == w) break;
1151       w = u;
1152     }
1153     TEVENT(Inflated exit - drain cxq into EntryList);
1154 
1155     assert(w != NULL, "invariant");
1156     assert(_EntryList == NULL, "invariant");
1157 
1158     // Convert the LIFO SLL anchored by _cxq into a DLL.
1159     // The list reorganization step operates in O(LENGTH(w)) time.
1160     // It's critical that this step operate quickly as
1161     // "Self" still holds the outer-lock, restricting parallelism
1162     // and effectively lengthening the critical section.
1163     // Invariant: s chases t chases u.
1164     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1165     // we have faster access to the tail.
1166 
1167     if (QMode == 1) {
1168       // QMode == 1 : drain cxq to EntryList, reversing order
1169       // We also reverse the order of the list.
1170       ObjectWaiter * s = NULL;
1171       ObjectWaiter * t = w;
1172       ObjectWaiter * u = NULL;
1173       while (t != NULL) {
1174         guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1175         t->TState = ObjectWaiter::TS_ENTER;
1176         u = t->_next;
1177         t->_prev = u;
1178         t->_next = s;
1179         s = t;
1180         t = u;
1181       }
1182       _EntryList  = s;
1183       assert(s != NULL, "invariant");
1184     } else {
1185       // QMode == 0 or QMode == 2
1186       _EntryList = w;
1187       ObjectWaiter * q = NULL;
1188       ObjectWaiter * p;
1189       for (p = w; p != NULL; p = p->_next) {
1190         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1191         p->TState = ObjectWaiter::TS_ENTER;
1192         p->_prev = q;
1193         q = p;
1194       }
1195     }
1196 
1197     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1198     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1199 
1200     // See if we can abdicate to a spinner instead of waking a thread.
1201     // A primary goal of the implementation is to reduce the
1202     // context-switch rate.
1203     if (_succ != NULL) continue;
1204 
1205     w = _EntryList;
1206     if (w != NULL) {
1207       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1208       ExitEpilog(Self, w);
1209       return;
1210     }
1211   }
1212 }
1213 
1214 // ExitSuspendEquivalent:
1215 // A faster alternate to handle_special_suspend_equivalent_condition()
1216 //
1217 // handle_special_suspend_equivalent_condition() unconditionally
1218 // acquires the SR_lock.  On some platforms uncontended MutexLocker()
1219 // operations have high latency.  Note that in ::enter() we call HSSEC
1220 // while holding the monitor, so we effectively lengthen the critical sections.
1221 //
1222 // There are a number of possible solutions:
1223 //
1224 // A.  To ameliorate the problem we might also defer state transitions
1225 //     to as late as possible -- just prior to parking.
1226 //     Given that, we'd call HSSEC after having returned from park(),
1227 //     but before attempting to acquire the monitor.  This is only a
1228 //     partial solution.  It avoids calling HSSEC while holding the
1229 //     monitor (good), but it still increases successor reacquisition latency --
1230 //     the interval between unparking a successor and the time the successor
1231 //     resumes and retries the lock.  See ReenterI(), which defers state transitions.
1232 //     If we use this technique we can also avoid EnterI()-exit() loop
1233 //     in ::enter() where we iteratively drop the lock and then attempt
1234 //     to reacquire it after suspending.
1235 //
1236 // B.  In the future we might fold all the suspend bits into a
1237 //     composite per-thread suspend flag and then update it with CAS().
1238 //     Alternately, a Dekker-like mechanism with multiple variables
1239 //     would suffice:
1240 //       ST Self->_suspend_equivalent = false
1241 //       MEMBAR
1242 //       LD Self_>_suspend_flags
1243 //
1244 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1245 // with a more efficient implementation, the need to use "FastHSSEC" has
1246 // decreased. - Dave
1247 
1248 
1249 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1250   const int Mode = Knob_FastHSSEC;
1251   if (Mode && !jSelf->is_external_suspend()) {
1252     assert(jSelf->is_suspend_equivalent(), "invariant");
1253     jSelf->clear_suspend_equivalent();
1254     if (2 == Mode) OrderAccess::storeload();
1255     if (!jSelf->is_external_suspend()) return false;
1256     // We raced a suspension -- fall thru into the slow path
1257     TEVENT(ExitSuspendEquivalent - raced);
1258     jSelf->set_suspend_equivalent();
1259   }
1260   return jSelf->handle_special_suspend_equivalent_condition();
1261 }
1262 
1263 
1264 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1265   assert(_owner == Self, "invariant");
1266 
1267   // Exit protocol:
1268   // 1. ST _succ = wakee
1269   // 2. membar #loadstore|#storestore;
1270   // 2. ST _owner = NULL
1271   // 3. unpark(wakee)
1272 
1273   _succ = Knob_SuccEnabled ? Wakee->_thread : NULL;
1274   ParkEvent * Trigger = Wakee->_event;
1275 
1276   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1277   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1278   // out-of-scope (non-extant).
1279   Wakee  = NULL;
1280 
1281   // Drop the lock
1282   OrderAccess::release_store_ptr(&_owner, NULL);
1283   OrderAccess::fence();                               // ST _owner vs LD in unpark()
1284 
1285   if (SafepointSynchronize::do_call_back()) {
1286     TEVENT(unpark before SAFEPOINT);
1287   }
1288 
1289   DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1290   Trigger->unpark();
1291 
1292   // Maintain stats and report events to JVMTI
1293   OM_PERFDATA_OP(Parks, inc());
1294 }
1295 
1296 
1297 // -----------------------------------------------------------------------------
1298 // Class Loader deadlock handling.
1299 //
1300 // complete_exit exits a lock returning recursion count
1301 // complete_exit/reenter operate as a wait without waiting
1302 // complete_exit requires an inflated monitor
1303 // The _owner field is not always the Thread addr even with an
1304 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1305 // thread due to contention.
1306 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1307   Thread * const Self = THREAD;
1308   assert(Self->is_Java_thread(), "Must be Java thread!");
1309   JavaThread *jt = (JavaThread *)THREAD;
1310 
1311   DeferredInitialize();
1312 
1313   if (THREAD != _owner) {
1314     if (THREAD->is_lock_owned ((address)_owner)) {
1315       assert(_recursions == 0, "internal state error");
1316       _owner = THREAD;   // Convert from basiclock addr to Thread addr
1317       _recursions = 0;
1318     }
1319   }
1320 
1321   guarantee(Self == _owner, "complete_exit not owner");
1322   intptr_t save = _recursions; // record the old recursion count
1323   _recursions = 0;        // set the recursion level to be 0
1324   exit(true, Self);           // exit the monitor
1325   guarantee(_owner != Self, "invariant");
1326   return save;
1327 }
1328 
1329 // reenter() enters a lock and sets recursion count
1330 // complete_exit/reenter operate as a wait without waiting
1331 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1332   Thread * const Self = THREAD;
1333   assert(Self->is_Java_thread(), "Must be Java thread!");
1334   JavaThread *jt = (JavaThread *)THREAD;
1335 
1336   guarantee(_owner != Self, "reenter already owner");
1337   enter(THREAD);       // enter the monitor
1338   guarantee(_recursions == 0, "reenter recursion");
1339   _recursions = recursions;
1340   return;
1341 }
1342 
1343 
1344 // -----------------------------------------------------------------------------
1345 // A macro is used below because there may already be a pending
1346 // exception which should not abort the execution of the routines
1347 // which use this (which is why we don't put this into check_slow and
1348 // call it with a CHECK argument).
1349 
1350 #define CHECK_OWNER()                                                       \
1351   do {                                                                      \
1352     if (THREAD != _owner) {                                                 \
1353       if (THREAD->is_lock_owned((address) _owner)) {                        \
1354         _owner = THREAD;  /* Convert from basiclock addr to Thread addr */  \
1355         _recursions = 0;                                                    \
1356       } else {                                                              \
1357         TEVENT(Throw IMSX);                                                 \
1358         THROW(vmSymbols::java_lang_IllegalMonitorStateException());         \
1359       }                                                                     \
1360     }                                                                       \
1361   } while (false)
1362 
1363 // check_slow() is a misnomer.  It's called to simply to throw an IMSX exception.
1364 // TODO-FIXME: remove check_slow() -- it's likely dead.
1365 
1366 void ObjectMonitor::check_slow(TRAPS) {
1367   TEVENT(check_slow - throw IMSX);
1368   assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1369   THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1370 }
1371 
1372 static int Adjust(volatile int * adr, int dx) {
1373   int v;
1374   for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1375   return v;
1376 }
1377 
1378 // helper method for posting a monitor wait event
1379 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event,
1380                                             jlong notifier_tid,
1381                                             jlong timeout,
1382                                             bool timedout) {
1383   assert(event != NULL, "invariant");
1384   event->set_monitorClass(((oop)this->object())->klass());
1385   event->set_timeout(timeout);
1386   event->set_address((TYPE_ADDRESS)this->object_addr());
1387   event->set_notifier(notifier_tid);
1388   event->set_timedOut(timedout);
1389   event->commit();
1390 }
1391 
1392 // -----------------------------------------------------------------------------
1393 // Wait/Notify/NotifyAll
1394 //
1395 // Note: a subset of changes to ObjectMonitor::wait()
1396 // will need to be replicated in complete_exit
1397 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1398   Thread * const Self = THREAD;
1399   assert(Self->is_Java_thread(), "Must be Java thread!");
1400   JavaThread *jt = (JavaThread *)THREAD;
1401 
1402   DeferredInitialize();
1403 
1404   // Throw IMSX or IEX.
1405   CHECK_OWNER();
1406 
1407   EventJavaMonitorWait event;
1408 
1409   // check for a pending interrupt
1410   if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1411     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1412     if (JvmtiExport::should_post_monitor_waited()) {
1413       // Note: 'false' parameter is passed here because the
1414       // wait was not timed out due to thread interrupt.
1415       JvmtiExport::post_monitor_waited(jt, this, false);
1416 
1417       // In this short circuit of the monitor wait protocol, the
1418       // current thread never drops ownership of the monitor and
1419       // never gets added to the wait queue so the current thread
1420       // cannot be made the successor. This means that the
1421       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1422       // consume an unpark() meant for the ParkEvent associated with
1423       // this ObjectMonitor.
1424     }
1425     if (event.should_commit()) {
1426       post_monitor_wait_event(&event, 0, millis, false);
1427     }
1428     TEVENT(Wait - Throw IEX);
1429     THROW(vmSymbols::java_lang_InterruptedException());
1430     return;
1431   }
1432 
1433   TEVENT(Wait);
1434 
1435   assert(Self->_Stalled == 0, "invariant");
1436   Self->_Stalled = intptr_t(this);
1437   jt->set_current_waiting_monitor(this);
1438 
1439   // create a node to be put into the queue
1440   // Critically, after we reset() the event but prior to park(), we must check
1441   // for a pending interrupt.
1442   ObjectWaiter node(Self);
1443   node.TState = ObjectWaiter::TS_WAIT;
1444   Self->_ParkEvent->reset();
1445   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1446 
1447   // Enter the waiting queue, which is a circular doubly linked list in this case
1448   // but it could be a priority queue or any data structure.
1449   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1450   // by the the owner of the monitor *except* in the case where park()
1451   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1452   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1453 
1454   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1455   AddWaiter(&node);
1456   Thread::SpinRelease(&_WaitSetLock);
1457 
1458   if ((SyncFlags & 4) == 0) {
1459     _Responsible = NULL;
1460   }
1461   intptr_t save = _recursions; // record the old recursion count
1462   _waiters++;                  // increment the number of waiters
1463   _recursions = 0;             // set the recursion level to be 1
1464   exit(true, Self);                    // exit the monitor
1465   guarantee(_owner != Self, "invariant");
1466 
1467   // The thread is on the WaitSet list - now park() it.
1468   // On MP systems it's conceivable that a brief spin before we park
1469   // could be profitable.
1470   //
1471   // TODO-FIXME: change the following logic to a loop of the form
1472   //   while (!timeout && !interrupted && _notified == 0) park()
1473 
1474   int ret = OS_OK;
1475   int WasNotified = 0;
1476   { // State transition wrappers
1477     OSThread* osthread = Self->osthread();
1478     OSThreadWaitState osts(osthread, true);
1479     {
1480       ThreadBlockInVM tbivm(jt);
1481       // Thread is in thread_blocked state and oop access is unsafe.
1482       jt->set_suspend_equivalent();
1483 
1484       if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1485         // Intentionally empty
1486       } else if (node._notified == 0) {
1487         if (millis <= 0) {
1488           Self->_ParkEvent->park();
1489         } else {
1490           ret = Self->_ParkEvent->park(millis);
1491         }
1492       }
1493 
1494       // were we externally suspended while we were waiting?
1495       if (ExitSuspendEquivalent (jt)) {
1496         // TODO-FIXME: add -- if succ == Self then succ = null.
1497         jt->java_suspend_self();
1498       }
1499 
1500     } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1501 
1502     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1503     // from the WaitSet to the EntryList.
1504     // See if we need to remove Node from the WaitSet.
1505     // We use double-checked locking to avoid grabbing _WaitSetLock
1506     // if the thread is not on the wait queue.
1507     //
1508     // Note that we don't need a fence before the fetch of TState.
1509     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1510     // written by the is thread. (perhaps the fetch might even be satisfied
1511     // by a look-aside into the processor's own store buffer, although given
1512     // the length of the code path between the prior ST and this load that's
1513     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1514     // then we'll acquire the lock and then re-fetch a fresh TState value.
1515     // That is, we fail toward safety.
1516 
1517     if (node.TState == ObjectWaiter::TS_WAIT) {
1518       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1519       if (node.TState == ObjectWaiter::TS_WAIT) {
1520         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1521         assert(node._notified == 0, "invariant");
1522         node.TState = ObjectWaiter::TS_RUN;
1523       }
1524       Thread::SpinRelease(&_WaitSetLock);
1525     }
1526 
1527     // The thread is now either on off-list (TS_RUN),
1528     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1529     // The Node's TState variable is stable from the perspective of this thread.
1530     // No other threads will asynchronously modify TState.
1531     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1532     OrderAccess::loadload();
1533     if (_succ == Self) _succ = NULL;
1534     WasNotified = node._notified;
1535 
1536     // Reentry phase -- reacquire the monitor.
1537     // re-enter contended monitor after object.wait().
1538     // retain OBJECT_WAIT state until re-enter successfully completes
1539     // Thread state is thread_in_vm and oop access is again safe,
1540     // although the raw address of the object may have changed.
1541     // (Don't cache naked oops over safepoints, of course).
1542 
1543     // post monitor waited event. Note that this is past-tense, we are done waiting.
1544     if (JvmtiExport::should_post_monitor_waited()) {
1545       JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1546 
1547       if (node._notified != 0 && _succ == Self) {
1548         // In this part of the monitor wait-notify-reenter protocol it
1549         // is possible (and normal) for another thread to do a fastpath
1550         // monitor enter-exit while this thread is still trying to get
1551         // to the reenter portion of the protocol.
1552         //
1553         // The ObjectMonitor was notified and the current thread is
1554         // the successor which also means that an unpark() has already
1555         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1556         // consume the unpark() that was done when the successor was
1557         // set because the same ParkEvent is shared between Java
1558         // monitors and JVM/TI RawMonitors (for now).
1559         //
1560         // We redo the unpark() to ensure forward progress, i.e., we
1561         // don't want all pending threads hanging (parked) with none
1562         // entering the unlocked monitor.
1563         node._event->unpark();
1564       }
1565     }
1566 
1567     if (event.should_commit()) {
1568       post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT);
1569     }
1570 
1571     OrderAccess::fence();
1572 
1573     assert(Self->_Stalled != 0, "invariant");
1574     Self->_Stalled = 0;
1575 
1576     assert(_owner != Self, "invariant");
1577     ObjectWaiter::TStates v = node.TState;
1578     if (v == ObjectWaiter::TS_RUN) {
1579       enter(Self);
1580     } else {
1581       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1582       ReenterI(Self, &node);
1583       node.wait_reenter_end(this);
1584     }
1585 
1586     // Self has reacquired the lock.
1587     // Lifecycle - the node representing Self must not appear on any queues.
1588     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1589     // want residual elements associated with this thread left on any lists.
1590     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1591     assert(_owner == Self, "invariant");
1592     assert(_succ != Self, "invariant");
1593   } // OSThreadWaitState()
1594 
1595   jt->set_current_waiting_monitor(NULL);
1596 
1597   guarantee(_recursions == 0, "invariant");
1598   _recursions = save;     // restore the old recursion count
1599   _waiters--;             // decrement the number of waiters
1600 
1601   // Verify a few postconditions
1602   assert(_owner == Self, "invariant");
1603   assert(_succ != Self, "invariant");
1604   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1605 
1606   if (SyncFlags & 32) {
1607     OrderAccess::fence();
1608   }
1609 
1610   // check if the notification happened
1611   if (!WasNotified) {
1612     // no, it could be timeout or Thread.interrupt() or both
1613     // check for interrupt event, otherwise it is timeout
1614     if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1615       TEVENT(Wait - throw IEX from epilog);
1616       THROW(vmSymbols::java_lang_InterruptedException());
1617     }
1618   }
1619 
1620   // NOTE: Spurious wake up will be consider as timeout.
1621   // Monitor notify has precedence over thread interrupt.
1622 }
1623 
1624 
1625 // Consider:
1626 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1627 // then instead of transferring a thread from the WaitSet to the EntryList
1628 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1629 
1630 void ObjectMonitor::INotify(Thread * Self) {
1631   const int policy = Knob_MoveNotifyee;
1632 
1633   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1634   ObjectWaiter * iterator = DequeueWaiter();
1635   if (iterator != NULL) {
1636     TEVENT(Notify1 - Transfer);
1637     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1638     guarantee(iterator->_notified == 0, "invariant");
1639     // Disposition - what might we do with iterator ?
1640     // a.  add it directly to the EntryList - either tail (policy == 1)
1641     //     or head (policy == 0).
1642     // b.  push it onto the front of the _cxq (policy == 2).
1643     // For now we use (b).
1644     if (policy != 4) {
1645       iterator->TState = ObjectWaiter::TS_ENTER;
1646     }
1647     iterator->_notified = 1;
1648     iterator->_notifier_tid = THREAD_TRACE_ID(Self);
1649 
1650     ObjectWaiter * list = _EntryList;
1651     if (list != NULL) {
1652       assert(list->_prev == NULL, "invariant");
1653       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1654       assert(list != iterator, "invariant");
1655     }
1656 
1657     if (policy == 0) {       // prepend to EntryList
1658       if (list == NULL) {
1659         iterator->_next = iterator->_prev = NULL;
1660         _EntryList = iterator;
1661       } else {
1662         list->_prev = iterator;
1663         iterator->_next = list;
1664         iterator->_prev = NULL;
1665         _EntryList = iterator;
1666       }
1667     } else if (policy == 1) {      // append to EntryList
1668       if (list == NULL) {
1669         iterator->_next = iterator->_prev = NULL;
1670         _EntryList = iterator;
1671       } else {
1672         // CONSIDER:  finding the tail currently requires a linear-time walk of
1673         // the EntryList.  We can make tail access constant-time by converting to
1674         // a CDLL instead of using our current DLL.
1675         ObjectWaiter * tail;
1676         for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1677         assert(tail != NULL && tail->_next == NULL, "invariant");
1678         tail->_next = iterator;
1679         iterator->_prev = tail;
1680         iterator->_next = NULL;
1681       }
1682     } else if (policy == 2) {      // prepend to cxq
1683       if (list == NULL) {
1684         iterator->_next = iterator->_prev = NULL;
1685         _EntryList = iterator;
1686       } else {
1687         iterator->TState = ObjectWaiter::TS_CXQ;
1688         for (;;) {
1689           ObjectWaiter * front = _cxq;
1690           iterator->_next = front;
1691           if (Atomic::cmpxchg_ptr(iterator, &_cxq, front) == front) {
1692             break;
1693           }
1694         }
1695       }
1696     } else if (policy == 3) {      // append to cxq
1697       iterator->TState = ObjectWaiter::TS_CXQ;
1698       for (;;) {
1699         ObjectWaiter * tail = _cxq;
1700         if (tail == NULL) {
1701           iterator->_next = NULL;
1702           if (Atomic::cmpxchg_ptr(iterator, &_cxq, NULL) == NULL) {
1703             break;
1704           }
1705         } else {
1706           while (tail->_next != NULL) tail = tail->_next;
1707           tail->_next = iterator;
1708           iterator->_prev = tail;
1709           iterator->_next = NULL;
1710           break;
1711         }
1712       }
1713     } else {
1714       ParkEvent * ev = iterator->_event;
1715       iterator->TState = ObjectWaiter::TS_RUN;
1716       OrderAccess::fence();
1717       ev->unpark();
1718     }
1719 
1720     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1721     // move the add-to-EntryList operation, above, outside the critical section
1722     // protected by _WaitSetLock.  In practice that's not useful.  With the
1723     // exception of  wait() timeouts and interrupts the monitor owner
1724     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1725     // on _WaitSetLock so it's not profitable to reduce the length of the
1726     // critical section.
1727 
1728     if (policy < 4) {
1729       iterator->wait_reenter_begin(this);
1730     }
1731   }
1732   Thread::SpinRelease(&_WaitSetLock);
1733 }
1734 
1735 // Consider: a not-uncommon synchronization bug is to use notify() when
1736 // notifyAll() is more appropriate, potentially resulting in stranded
1737 // threads; this is one example of a lost wakeup. A useful diagnostic
1738 // option is to force all notify() operations to behave as notifyAll().
1739 //
1740 // Note: We can also detect many such problems with a "minimum wait".
1741 // When the "minimum wait" is set to a small non-zero timeout value
1742 // and the program does not hang whereas it did absent "minimum wait",
1743 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses
1744 // a "minimum wait" for all park() operations; see the recheckInterval
1745 // variable and MAX_RECHECK_INTERVAL.
1746 
1747 void ObjectMonitor::notify(TRAPS) {
1748   CHECK_OWNER();
1749   if (_WaitSet == NULL) {
1750     TEVENT(Empty-Notify);
1751     return;
1752   }
1753   DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1754   INotify(THREAD);
1755   OM_PERFDATA_OP(Notifications, inc(1));
1756 }
1757 
1758 
1759 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1760 // from the waitset to the EntryList. This could be done more efficiently with a
1761 // single bulk transfer but in practice it's not time-critical. Beware too,
1762 // that in prepend-mode we invert the order of the waiters. Let's say that the
1763 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1764 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1765 
1766 void ObjectMonitor::notifyAll(TRAPS) {
1767   CHECK_OWNER();
1768   if (_WaitSet == NULL) {
1769     TEVENT(Empty-NotifyAll);
1770     return;
1771   }
1772 
1773   DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1774   int tally = 0;
1775   while (_WaitSet != NULL) {
1776     tally++;
1777     INotify(THREAD);
1778   }
1779 
1780   OM_PERFDATA_OP(Notifications, inc(tally));
1781 }
1782 
1783 // -----------------------------------------------------------------------------
1784 // Adaptive Spinning Support
1785 //
1786 // Adaptive spin-then-block - rational spinning
1787 //
1788 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1789 // algorithm.  On high order SMP systems it would be better to start with
1790 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1791 // a contending thread could enqueue itself on the cxq and then spin locally
1792 // on a thread-specific variable such as its ParkEvent._Event flag.
1793 // That's left as an exercise for the reader.  Note that global spinning is
1794 // not problematic on Niagara, as the L2 cache serves the interconnect and
1795 // has both low latency and massive bandwidth.
1796 //
1797 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1798 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1799 // (duration) or we can fix the count at approximately the duration of
1800 // a context switch and vary the frequency.   Of course we could also
1801 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1802 // For a description of 'Adaptive spin-then-block mutual exclusion in
1803 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1804 //
1805 // This implementation varies the duration "D", where D varies with
1806 // the success rate of recent spin attempts. (D is capped at approximately
1807 // length of a round-trip context switch).  The success rate for recent
1808 // spin attempts is a good predictor of the success rate of future spin
1809 // attempts.  The mechanism adapts automatically to varying critical
1810 // section length (lock modality), system load and degree of parallelism.
1811 // D is maintained per-monitor in _SpinDuration and is initialized
1812 // optimistically.  Spin frequency is fixed at 100%.
1813 //
1814 // Note that _SpinDuration is volatile, but we update it without locks
1815 // or atomics.  The code is designed so that _SpinDuration stays within
1816 // a reasonable range even in the presence of races.  The arithmetic
1817 // operations on _SpinDuration are closed over the domain of legal values,
1818 // so at worst a race will install and older but still legal value.
1819 // At the very worst this introduces some apparent non-determinism.
1820 // We might spin when we shouldn't or vice-versa, but since the spin
1821 // count are relatively short, even in the worst case, the effect is harmless.
1822 //
1823 // Care must be taken that a low "D" value does not become an
1824 // an absorbing state.  Transient spinning failures -- when spinning
1825 // is overall profitable -- should not cause the system to converge
1826 // on low "D" values.  We want spinning to be stable and predictable
1827 // and fairly responsive to change and at the same time we don't want
1828 // it to oscillate, become metastable, be "too" non-deterministic,
1829 // or converge on or enter undesirable stable absorbing states.
1830 //
1831 // We implement a feedback-based control system -- using past behavior
1832 // to predict future behavior.  We face two issues: (a) if the
1833 // input signal is random then the spin predictor won't provide optimal
1834 // results, and (b) if the signal frequency is too high then the control
1835 // system, which has some natural response lag, will "chase" the signal.
1836 // (b) can arise from multimodal lock hold times.  Transient preemption
1837 // can also result in apparent bimodal lock hold times.
1838 // Although sub-optimal, neither condition is particularly harmful, as
1839 // in the worst-case we'll spin when we shouldn't or vice-versa.
1840 // The maximum spin duration is rather short so the failure modes aren't bad.
1841 // To be conservative, I've tuned the gain in system to bias toward
1842 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1843 // "rings" or oscillates between spinning and not spinning.  This happens
1844 // when spinning is just on the cusp of profitability, however, so the
1845 // situation is not dire.  The state is benign -- there's no need to add
1846 // hysteresis control to damp the transition rate between spinning and
1847 // not spinning.
1848 
1849 // Spinning: Fixed frequency (100%), vary duration
1850 int ObjectMonitor::TrySpin(Thread * Self) {
1851   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1852   int ctr = Knob_FixedSpin;
1853   if (ctr != 0) {
1854     while (--ctr >= 0) {
1855       if (TryLock(Self) > 0) return 1;
1856       SpinPause();
1857     }
1858     return 0;
1859   }
1860 
1861   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1862     if (TryLock(Self) > 0) {
1863       // Increase _SpinDuration ...
1864       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1865       // Raising _SpurDuration to the poverty line is key.
1866       int x = _SpinDuration;
1867       if (x < Knob_SpinLimit) {
1868         if (x < Knob_Poverty) x = Knob_Poverty;
1869         _SpinDuration = x + Knob_BonusB;
1870       }
1871       return 1;
1872     }
1873     SpinPause();
1874   }
1875 
1876   // Admission control - verify preconditions for spinning
1877   //
1878   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1879   // becoming an absorbing state.  Put another way, we spin briefly to
1880   // sample, just in case the system load, parallelism, contention, or lock
1881   // modality changed.
1882   //
1883   // Consider the following alternative:
1884   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1885   // spin attempt.  "Periodically" might mean after a tally of
1886   // the # of failed spin attempts (or iterations) reaches some threshold.
1887   // This takes us into the realm of 1-out-of-N spinning, where we
1888   // hold the duration constant but vary the frequency.
1889 
1890   ctr = _SpinDuration;
1891   if (ctr < Knob_SpinBase) ctr = Knob_SpinBase;
1892   if (ctr <= 0) return 0;
1893 
1894   if (Knob_SuccRestrict && _succ != NULL) return 0;
1895   if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1896     TEVENT(Spin abort - notrunnable [TOP]);
1897     return 0;
1898   }
1899 
1900   int MaxSpin = Knob_MaxSpinners;
1901   if (MaxSpin >= 0) {
1902     if (_Spinner > MaxSpin) {
1903       TEVENT(Spin abort -- too many spinners);
1904       return 0;
1905     }
1906     // Slightly racy, but benign ...
1907     Adjust(&_Spinner, 1);
1908   }
1909 
1910   // We're good to spin ... spin ingress.
1911   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1912   // when preparing to LD...CAS _owner, etc and the CAS is likely
1913   // to succeed.
1914   int hits    = 0;
1915   int msk     = 0;
1916   int caspty  = Knob_CASPenalty;
1917   int oxpty   = Knob_OXPenalty;
1918   int sss     = Knob_SpinSetSucc;
1919   if (sss && _succ == NULL) _succ = Self;
1920   Thread * prv = NULL;
1921 
1922   // There are three ways to exit the following loop:
1923   // 1.  A successful spin where this thread has acquired the lock.
1924   // 2.  Spin failure with prejudice
1925   // 3.  Spin failure without prejudice
1926 
1927   while (--ctr >= 0) {
1928 
1929     // Periodic polling -- Check for pending GC
1930     // Threads may spin while they're unsafe.
1931     // We don't want spinning threads to delay the JVM from reaching
1932     // a stop-the-world safepoint or to steal cycles from GC.
1933     // If we detect a pending safepoint we abort in order that
1934     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1935     // this thread, if safe, doesn't steal cycles from GC.
1936     // This is in keeping with the "no loitering in runtime" rule.
1937     // We periodically check to see if there's a safepoint pending.
1938     if ((ctr & 0xFF) == 0) {
1939       if (SafepointSynchronize::do_call_back()) {
1940         TEVENT(Spin: safepoint);
1941         goto Abort;           // abrupt spin egress
1942       }
1943       if (Knob_UsePause & 1) SpinPause();
1944     }
1945 
1946     if (Knob_UsePause & 2) SpinPause();
1947 
1948     // Exponential back-off ...  Stay off the bus to reduce coherency traffic.
1949     // This is useful on classic SMP systems, but is of less utility on
1950     // N1-style CMT platforms.
1951     //
1952     // Trade-off: lock acquisition latency vs coherency bandwidth.
1953     // Lock hold times are typically short.  A histogram
1954     // of successful spin attempts shows that we usually acquire
1955     // the lock early in the spin.  That suggests we want to
1956     // sample _owner frequently in the early phase of the spin,
1957     // but then back-off and sample less frequently as the spin
1958     // progresses.  The back-off makes a good citizen on SMP big
1959     // SMP systems.  Oversampling _owner can consume excessive
1960     // coherency bandwidth.  Relatedly, if we _oversample _owner we
1961     // can inadvertently interfere with the the ST m->owner=null.
1962     // executed by the lock owner.
1963     if (ctr & msk) continue;
1964     ++hits;
1965     if ((hits & 0xF) == 0) {
1966       // The 0xF, above, corresponds to the exponent.
1967       // Consider: (msk+1)|msk
1968       msk = ((msk << 2)|3) & BackOffMask;
1969     }
1970 
1971     // Probe _owner with TATAS
1972     // If this thread observes the monitor transition or flicker
1973     // from locked to unlocked to locked, then the odds that this
1974     // thread will acquire the lock in this spin attempt go down
1975     // considerably.  The same argument applies if the CAS fails
1976     // or if we observe _owner change from one non-null value to
1977     // another non-null value.   In such cases we might abort
1978     // the spin without prejudice or apply a "penalty" to the
1979     // spin count-down variable "ctr", reducing it by 100, say.
1980 
1981     Thread * ox = (Thread *) _owner;
1982     if (ox == NULL) {
1983       ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL);
1984       if (ox == NULL) {
1985         // The CAS succeeded -- this thread acquired ownership
1986         // Take care of some bookkeeping to exit spin state.
1987         if (sss && _succ == Self) {
1988           _succ = NULL;
1989         }
1990         if (MaxSpin > 0) Adjust(&_Spinner, -1);
1991 
1992         // Increase _SpinDuration :
1993         // The spin was successful (profitable) so we tend toward
1994         // longer spin attempts in the future.
1995         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1996         // If we acquired the lock early in the spin cycle it
1997         // makes sense to increase _SpinDuration proportionally.
1998         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1999         int x = _SpinDuration;
2000         if (x < Knob_SpinLimit) {
2001           if (x < Knob_Poverty) x = Knob_Poverty;
2002           _SpinDuration = x + Knob_Bonus;
2003         }
2004         return 1;
2005       }
2006 
2007       // The CAS failed ... we can take any of the following actions:
2008       // * penalize: ctr -= Knob_CASPenalty
2009       // * exit spin with prejudice -- goto Abort;
2010       // * exit spin without prejudice.
2011       // * Since CAS is high-latency, retry again immediately.
2012       prv = ox;
2013       TEVENT(Spin: cas failed);
2014       if (caspty == -2) break;
2015       if (caspty == -1) goto Abort;
2016       ctr -= caspty;
2017       continue;
2018     }
2019 
2020     // Did lock ownership change hands ?
2021     if (ox != prv && prv != NULL) {
2022       TEVENT(spin: Owner changed)
2023       if (oxpty == -2) break;
2024       if (oxpty == -1) goto Abort;
2025       ctr -= oxpty;
2026     }
2027     prv = ox;
2028 
2029     // Abort the spin if the owner is not executing.
2030     // The owner must be executing in order to drop the lock.
2031     // Spinning while the owner is OFFPROC is idiocy.
2032     // Consider: ctr -= RunnablePenalty ;
2033     if (Knob_OState && NotRunnable (Self, ox)) {
2034       TEVENT(Spin abort - notrunnable);
2035       goto Abort;
2036     }
2037     if (sss && _succ == NULL) _succ = Self;
2038   }
2039 
2040   // Spin failed with prejudice -- reduce _SpinDuration.
2041   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2042   // AIMD is globally stable.
2043   TEVENT(Spin failure);
2044   {
2045     int x = _SpinDuration;
2046     if (x > 0) {
2047       // Consider an AIMD scheme like: x -= (x >> 3) + 100
2048       // This is globally sample and tends to damp the response.
2049       x -= Knob_Penalty;
2050       if (x < 0) x = 0;
2051       _SpinDuration = x;
2052     }
2053   }
2054 
2055  Abort:
2056   if (MaxSpin >= 0) Adjust(&_Spinner, -1);
2057   if (sss && _succ == Self) {
2058     _succ = NULL;
2059     // Invariant: after setting succ=null a contending thread
2060     // must recheck-retry _owner before parking.  This usually happens
2061     // in the normal usage of TrySpin(), but it's safest
2062     // to make TrySpin() as foolproof as possible.
2063     OrderAccess::fence();
2064     if (TryLock(Self) > 0) return 1;
2065   }
2066   return 0;
2067 }
2068 
2069 // NotRunnable() -- informed spinning
2070 //
2071 // Don't bother spinning if the owner is not eligible to drop the lock.
2072 // Peek at the owner's schedctl.sc_state and Thread._thread_values and
2073 // spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2074 // The thread must be runnable in order to drop the lock in timely fashion.
2075 // If the _owner is not runnable then spinning will not likely be
2076 // successful (profitable).
2077 //
2078 // Beware -- the thread referenced by _owner could have died
2079 // so a simply fetch from _owner->_thread_state might trap.
2080 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2081 // Because of the lifecycle issues the schedctl and _thread_state values
2082 // observed by NotRunnable() might be garbage.  NotRunnable must
2083 // tolerate this and consider the observed _thread_state value
2084 // as advisory.
2085 //
2086 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2087 // a thread pointer.
2088 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2089 // with the LSB of _owner.  Another option would be to probablistically probe
2090 // the putative _owner->TypeTag value.
2091 //
2092 // Checking _thread_state isn't perfect.  Even if the thread is
2093 // in_java it might be blocked on a page-fault or have been preempted
2094 // and sitting on a ready/dispatch queue.  _thread state in conjunction
2095 // with schedctl.sc_state gives us a good picture of what the
2096 // thread is doing, however.
2097 //
2098 // TODO: check schedctl.sc_state.
2099 // We'll need to use SafeFetch32() to read from the schedctl block.
2100 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/
2101 //
2102 // The return value from NotRunnable() is *advisory* -- the
2103 // result is based on sampling and is not necessarily coherent.
2104 // The caller must tolerate false-negative and false-positive errors.
2105 // Spinning, in general, is probabilistic anyway.
2106 
2107 
2108 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2109   // Check ox->TypeTag == 2BAD.
2110   if (ox == NULL) return 0;
2111 
2112   // Avoid transitive spinning ...
2113   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2114   // Immediately after T1 acquires L it's possible that T2, also
2115   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2116   // This occurs transiently after T1 acquired L but before
2117   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2118   // its spin in this circumstance.
2119   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2120 
2121   if (BlockedOn == 1) return 1;
2122   if (BlockedOn != 0) {
2123     return BlockedOn != intptr_t(this) && _owner == ox;
2124   }
2125 
2126   assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2127   int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2128   // consider also: jst != _thread_in_Java -- but that's overspecific.
2129   return jst == _thread_blocked || jst == _thread_in_native;
2130 }
2131 
2132 
2133 // -----------------------------------------------------------------------------
2134 // WaitSet management ...
2135 
2136 ObjectWaiter::ObjectWaiter(Thread* thread) {
2137   _next     = NULL;
2138   _prev     = NULL;
2139   _notified = 0;
2140   TState    = TS_RUN;
2141   _thread   = thread;
2142   _event    = thread->_ParkEvent;
2143   _active   = false;
2144   assert(_event != NULL, "invariant");
2145 }
2146 
2147 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2148   JavaThread *jt = (JavaThread *)this->_thread;
2149   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2150 }
2151 
2152 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2153   JavaThread *jt = (JavaThread *)this->_thread;
2154   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2155 }
2156 
2157 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2158   assert(node != NULL, "should not add NULL node");
2159   assert(node->_prev == NULL, "node already in list");
2160   assert(node->_next == NULL, "node already in list");
2161   // put node at end of queue (circular doubly linked list)
2162   if (_WaitSet == NULL) {
2163     _WaitSet = node;
2164     node->_prev = node;
2165     node->_next = node;
2166   } else {
2167     ObjectWaiter* head = _WaitSet;
2168     ObjectWaiter* tail = head->_prev;
2169     assert(tail->_next == head, "invariant check");
2170     tail->_next = node;
2171     head->_prev = node;
2172     node->_next = head;
2173     node->_prev = tail;
2174   }
2175 }
2176 
2177 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2178   // dequeue the very first waiter
2179   ObjectWaiter* waiter = _WaitSet;
2180   if (waiter) {
2181     DequeueSpecificWaiter(waiter);
2182   }
2183   return waiter;
2184 }
2185 
2186 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2187   assert(node != NULL, "should not dequeue NULL node");
2188   assert(node->_prev != NULL, "node already removed from list");
2189   assert(node->_next != NULL, "node already removed from list");
2190   // when the waiter has woken up because of interrupt,
2191   // timeout or other spurious wake-up, dequeue the
2192   // waiter from waiting list
2193   ObjectWaiter* next = node->_next;
2194   if (next == node) {
2195     assert(node->_prev == node, "invariant check");
2196     _WaitSet = NULL;
2197   } else {
2198     ObjectWaiter* prev = node->_prev;
2199     assert(prev->_next == node, "invariant check");
2200     assert(next->_prev == node, "invariant check");
2201     next->_prev = prev;
2202     prev->_next = next;
2203     if (_WaitSet == node) {
2204       _WaitSet = next;
2205     }
2206   }
2207   node->_next = NULL;
2208   node->_prev = NULL;
2209 }
2210 
2211 // -----------------------------------------------------------------------------
2212 // PerfData support
2213 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2214 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2215 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2216 PerfCounter * ObjectMonitor::_sync_EmptyNotifications          = NULL;
2217 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2218 PerfCounter * ObjectMonitor::_sync_PrivateA                    = NULL;
2219 PerfCounter * ObjectMonitor::_sync_PrivateB                    = NULL;
2220 PerfCounter * ObjectMonitor::_sync_SlowExit                    = NULL;
2221 PerfCounter * ObjectMonitor::_sync_SlowEnter                   = NULL;
2222 PerfCounter * ObjectMonitor::_sync_SlowNotify                  = NULL;
2223 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll               = NULL;
2224 PerfCounter * ObjectMonitor::_sync_FailedSpins                 = NULL;
2225 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins             = NULL;
2226 PerfCounter * ObjectMonitor::_sync_MonInCirculation            = NULL;
2227 PerfCounter * ObjectMonitor::_sync_MonScavenged                = NULL;
2228 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2229 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2230 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2231 
2232 // One-shot global initialization for the sync subsystem.
2233 // We could also defer initialization and initialize on-demand
2234 // the first time we call inflate().  Initialization would
2235 // be protected - like so many things - by the MonitorCache_lock.
2236 
2237 void ObjectMonitor::Initialize() {
2238   static int InitializationCompleted = 0;
2239   assert(InitializationCompleted == 0, "invariant");
2240   InitializationCompleted = 1;
2241   if (UsePerfData) {
2242     EXCEPTION_MARK;
2243 #define NEWPERFCOUNTER(n)                                                \
2244   {                                                                      \
2245     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2246                                         CHECK);                          \
2247   }
2248 #define NEWPERFVARIABLE(n)                                                \
2249   {                                                                       \
2250     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2251                                          CHECK);                          \
2252   }
2253     NEWPERFCOUNTER(_sync_Inflations);
2254     NEWPERFCOUNTER(_sync_Deflations);
2255     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2256     NEWPERFCOUNTER(_sync_FutileWakeups);
2257     NEWPERFCOUNTER(_sync_Parks);
2258     NEWPERFCOUNTER(_sync_EmptyNotifications);
2259     NEWPERFCOUNTER(_sync_Notifications);
2260     NEWPERFCOUNTER(_sync_SlowEnter);
2261     NEWPERFCOUNTER(_sync_SlowExit);
2262     NEWPERFCOUNTER(_sync_SlowNotify);
2263     NEWPERFCOUNTER(_sync_SlowNotifyAll);
2264     NEWPERFCOUNTER(_sync_FailedSpins);
2265     NEWPERFCOUNTER(_sync_SuccessfulSpins);
2266     NEWPERFCOUNTER(_sync_PrivateA);
2267     NEWPERFCOUNTER(_sync_PrivateB);
2268     NEWPERFCOUNTER(_sync_MonInCirculation);
2269     NEWPERFCOUNTER(_sync_MonScavenged);
2270     NEWPERFVARIABLE(_sync_MonExtant);
2271 #undef NEWPERFCOUNTER
2272 #undef NEWPERFVARIABLE
2273   }
2274 }
2275 
2276 static char * kvGet(char * kvList, const char * Key) {
2277   if (kvList == NULL) return NULL;
2278   size_t n = strlen(Key);
2279   char * Search;
2280   for (Search = kvList; *Search; Search += strlen(Search) + 1) {
2281     if (strncmp (Search, Key, n) == 0) {
2282       if (Search[n] == '=') return Search + n + 1;
2283       if (Search[n] == 0)   return(char *) "1";
2284     }
2285   }
2286   return NULL;
2287 }
2288 
2289 static int kvGetInt(char * kvList, const char * Key, int Default) {
2290   char * v = kvGet(kvList, Key);
2291   int rslt = v ? ::strtol(v, NULL, 0) : Default;
2292   if (Knob_ReportSettings && v != NULL) {
2293     tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ;
2294     tty->flush();
2295   }
2296   return rslt;
2297 }
2298 
2299 void ObjectMonitor::DeferredInitialize() {
2300   if (InitDone > 0) return;
2301   if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2302     while (InitDone != 1) /* empty */;
2303     return;
2304   }
2305 
2306   // One-shot global initialization ...
2307   // The initialization is idempotent, so we don't need locks.
2308   // In the future consider doing this via os::init_2().
2309   // SyncKnobs consist of <Key>=<Value> pairs in the style
2310   // of environment variables.  Start by converting ':' to NUL.
2311 
2312   if (SyncKnobs == NULL) SyncKnobs = "";
2313 
2314   size_t sz = strlen(SyncKnobs);
2315   char * knobs = (char *) malloc(sz + 2);
2316   if (knobs == NULL) {
2317     vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs");
2318     guarantee(0, "invariant");
2319   }
2320   strcpy(knobs, SyncKnobs);
2321   knobs[sz+1] = 0;
2322   for (char * p = knobs; *p; p++) {
2323     if (*p == ':') *p = 0;
2324   }
2325 
2326   #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); }
2327   SETKNOB(ReportSettings);
2328   SETKNOB(ExitRelease);
2329   SETKNOB(Verbose);
2330   SETKNOB(VerifyInUse);
2331   SETKNOB(VerifyMatch);
2332   SETKNOB(FixedSpin);
2333   SETKNOB(SpinLimit);
2334   SETKNOB(SpinBase);
2335   SETKNOB(SpinBackOff);
2336   SETKNOB(CASPenalty);
2337   SETKNOB(OXPenalty);
2338   SETKNOB(LogSpins);
2339   SETKNOB(SpinSetSucc);
2340   SETKNOB(SuccEnabled);
2341   SETKNOB(SuccRestrict);
2342   SETKNOB(Penalty);
2343   SETKNOB(Bonus);
2344   SETKNOB(BonusB);
2345   SETKNOB(Poverty);
2346   SETKNOB(SpinAfterFutile);
2347   SETKNOB(UsePause);
2348   SETKNOB(SpinEarly);
2349   SETKNOB(OState);
2350   SETKNOB(MaxSpinners);
2351   SETKNOB(PreSpin);
2352   SETKNOB(ExitPolicy);
2353   SETKNOB(QMode);
2354   SETKNOB(ResetEvent);
2355   SETKNOB(MoveNotifyee);
2356   SETKNOB(FastHSSEC);
2357   #undef SETKNOB
2358 
2359   if (Knob_Verbose) {
2360     sanity_checks();
2361   }
2362 
2363   if (os::is_MP()) {
2364     BackOffMask = (1 << Knob_SpinBackOff) - 1;
2365     if (Knob_ReportSettings) {
2366       tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask);
2367     }
2368     // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2369   } else {
2370     Knob_SpinLimit = 0;
2371     Knob_SpinBase  = 0;
2372     Knob_PreSpin   = 0;
2373     Knob_FixedSpin = -1;
2374   }
2375 
2376   if (Knob_LogSpins == 0) {
2377     ObjectMonitor::_sync_FailedSpins = NULL;
2378   }
2379 
2380   free(knobs);
2381   OrderAccess::fence();
2382   InitDone = 1;
2383 }
2384 
2385 void ObjectMonitor::sanity_checks() {
2386   int error_cnt = 0;
2387   int warning_cnt = 0;
2388   bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests);
2389 
2390   if (verbose) {
2391     tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT,
2392                   sizeof(ObjectMonitor));
2393     tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT,
2394                   sizeof(PaddedEnd<ObjectMonitor>));
2395   }
2396 
2397   uint cache_line_size = VM_Version::L1_data_cache_line_size();
2398   if (verbose) {
2399     tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size);
2400   }
2401 
2402   ObjectMonitor dummy;
2403   u_char *addr_begin  = (u_char*)&dummy;
2404   u_char *addr_header = (u_char*)&dummy._header;
2405   u_char *addr_owner  = (u_char*)&dummy._owner;
2406 
2407   uint offset_header = (uint)(addr_header - addr_begin);
2408   if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header);
2409 
2410   uint offset_owner = (uint)(addr_owner - addr_begin);
2411   if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner);
2412 
2413   if ((uint)(addr_header - addr_begin) != 0) {
2414     tty->print_cr("ERROR: offset(_header) must be zero (0).");
2415     error_cnt++;
2416   }
2417 
2418   if (cache_line_size != 0) {
2419     // We were able to determine the L1 data cache line size so
2420     // do some cache line specific sanity checks
2421 
2422     if ((offset_owner - offset_header) < cache_line_size) {
2423       tty->print_cr("WARNING: the _header and _owner fields are closer "
2424                     "than a cache line which permits false sharing.");
2425       warning_cnt++;
2426     }
2427 
2428     if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) {
2429       tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a "
2430                     "multiple of a cache line which permits false sharing.");
2431       warning_cnt++;
2432     }
2433   }
2434 
2435   ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt,
2436                                     &warning_cnt);
2437 
2438   if (verbose || error_cnt != 0 || warning_cnt != 0) {
2439     tty->print_cr("INFO: error_cnt=%d", error_cnt);
2440     tty->print_cr("INFO: warning_cnt=%d", warning_cnt);
2441   }
2442 
2443   guarantee(error_cnt == 0,
2444             "Fatal error(s) found in ObjectMonitor::sanity_checks()");
2445 }
2446 
2447 #ifndef PRODUCT
2448 void ObjectMonitor_test() {
2449   ObjectMonitor::sanity_checks();
2450 }
2451 #endif
--- EOF ---