1 /*
   2  * Copyright (c) 1998, 2018, 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 "runtime/atomic.hpp"
  27 #include "runtime/interfaceSupport.inline.hpp"
  28 #include "runtime/mutex.hpp"
  29 #include "runtime/orderAccess.hpp"
  30 #include "runtime/osThread.hpp"
  31 #include "runtime/safepointMechanism.inline.hpp"
  32 #include "runtime/thread.inline.hpp"
  33 #include "utilities/events.hpp"
  34 #include "utilities/macros.hpp"
  35 
  36 // o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o
  37 //
  38 // Native Monitor-Mutex locking - theory of operations
  39 //
  40 // * Native Monitors are completely unrelated to Java-level monitors,
  41 //   although the "back-end" slow-path implementations share a common lineage.
  42 //   See objectMonitor:: in synchronizer.cpp.
  43 //   Native Monitors do *not* support nesting or recursion but otherwise
  44 //   they're basically Hoare-flavor monitors.
  45 //
  46 // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte
  47 //   in the _LockWord from zero to non-zero.  Note that the _Owner field
  48 //   is advisory and is used only to verify that the thread calling unlock()
  49 //   is indeed the last thread to have acquired the lock.
  50 //
  51 // * Contending threads "push" themselves onto the front of the contention
  52 //   queue -- called the cxq -- with CAS and then spin/park.
  53 //   The _LockWord contains the LockByte as well as the pointer to the head
  54 //   of the cxq.  Colocating the LockByte with the cxq precludes certain races.
  55 //
  56 // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0
  57 //   idioms.  We currently use MEMBAR in the uncontended unlock() path, as
  58 //   MEMBAR often has less latency than CAS.  If warranted, we could switch to
  59 //   a CAS:0 mode, using timers to close the resultant race, as is done
  60 //   with Java Monitors in synchronizer.cpp.
  61 //
  62 //   See the following for a discussion of the relative cost of atomics (CAS)
  63 //   MEMBAR, and ways to eliminate such instructions from the common-case paths:
  64 //   -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot
  65 //   -- http://blogs.sun.com/dave/resource/MustangSync.pdf
  66 //   -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf
  67 //   -- synchronizer.cpp
  68 //
  69 // * Overall goals - desiderata
  70 //   1. Minimize context switching
  71 //   2. Minimize lock migration
  72 //   3. Minimize CPI -- affinity and locality
  73 //   4. Minimize the execution of high-latency instructions such as CAS or MEMBAR
  74 //   5. Minimize outer lock hold times
  75 //   6. Behave gracefully on a loaded system
  76 //
  77 // * Thread flow and list residency:
  78 //
  79 //   Contention queue --> EntryList --> OnDeck --> Owner --> !Owner
  80 //   [..resident on monitor list..]
  81 //   [...........contending..................]
  82 //
  83 //   -- The contention queue (cxq) contains recently-arrived threads (RATs).
  84 //      Threads on the cxq eventually drain into the EntryList.
  85 //   -- Invariant: a thread appears on at most one list -- cxq, EntryList
  86 //      or WaitSet -- at any one time.
  87 //   -- For a given monitor there can be at most one "OnDeck" thread at any
  88 //      given time but if needbe this particular invariant could be relaxed.
  89 //
  90 // * The WaitSet and EntryList linked lists are composed of ParkEvents.
  91 //   I use ParkEvent instead of threads as ParkEvents are immortal and
  92 //   type-stable, meaning we can safely unpark() a possibly stale
  93 //   list element in the unlock()-path.  (That's benign).
  94 //
  95 // * Succession policy - providing for progress:
  96 //
  97 //   As necessary, the unlock()ing thread identifies, unlinks, and unparks
  98 //   an "heir presumptive" tentative successor thread from the EntryList.
  99 //   This becomes the so-called "OnDeck" thread, of which there can be only
 100 //   one at any given time for a given monitor.  The wakee will recontend
 101 //   for ownership of monitor.
 102 //
 103 //   Succession is provided for by a policy of competitive handoff.
 104 //   The exiting thread does _not_ grant or pass ownership to the
 105 //   successor thread.  (This is also referred to as "handoff" succession").
 106 //   Instead the exiting thread releases ownership and possibly wakes
 107 //   a successor, so the successor can (re)compete for ownership of the lock.
 108 //
 109 //   Competitive handoff provides excellent overall throughput at the expense
 110 //   of short-term fairness.  If fairness is a concern then one remedy might
 111 //   be to add an AcquireCounter field to the monitor.  After a thread acquires
 112 //   the lock it will decrement the AcquireCounter field.  When the count
 113 //   reaches 0 the thread would reset the AcquireCounter variable, abdicate
 114 //   the lock directly to some thread on the EntryList, and then move itself to the
 115 //   tail of the EntryList.
 116 //
 117 //   But in practice most threads engage or otherwise participate in resource
 118 //   bounded producer-consumer relationships, so lock domination is not usually
 119 //   a practical concern.  Recall too, that in general it's easier to construct
 120 //   a fair lock from a fast lock, but not vice-versa.
 121 //
 122 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 123 //   detaching thread.  This mechanism is immune from the ABA corruption.
 124 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 125 //   We use OnDeck as a pseudo-lock to enforce the at-most-one detaching
 126 //   thread constraint.
 127 //
 128 // * Taken together, the cxq and the EntryList constitute or form a
 129 //   single logical queue of threads stalled trying to acquire the lock.
 130 //   We use two distinct lists to reduce heat on the list ends.
 131 //   Threads in lock() enqueue onto cxq while threads in unlock() will
 132 //   dequeue from the EntryList.  (c.f. Michael Scott's "2Q" algorithm).
 133 //   A key desideratum is to minimize queue & monitor metadata manipulation
 134 //   that occurs while holding the "outer" monitor lock -- that is, we want to
 135 //   minimize monitor lock holds times.
 136 //
 137 //   The EntryList is ordered by the prevailing queue discipline and
 138 //   can be organized in any convenient fashion, such as a doubly-linked list or
 139 //   a circular doubly-linked list.  If we need a priority queue then something akin
 140 //   to Solaris' sleepq would work nicely.  Viz.,
 141 //   -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 142 //   -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c
 143 //   Queue discipline is enforced at ::unlock() time, when the unlocking thread
 144 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 145 //   EntryList accordingly.
 146 //
 147 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 148 //   somewhat similar to an elevator-scan.
 149 //
 150 // * OnDeck
 151 //   --  For a given monitor there can be at most one OnDeck thread at any given
 152 //       instant.  The OnDeck thread is contending for the lock, but has been
 153 //       unlinked from the EntryList and cxq by some previous unlock() operations.
 154 //       Once a thread has been designated the OnDeck thread it will remain so
 155 //       until it manages to acquire the lock -- being OnDeck is a stable property.
 156 //   --  Threads on the EntryList or cxq are _not allowed to attempt lock acquisition.
 157 //   --  OnDeck also serves as an "inner lock" as follows.  Threads in unlock() will, after
 158 //       having cleared the LockByte and dropped the outer lock,  attempt to "trylock"
 159 //       OnDeck by CASing the field from null to non-null.  If successful, that thread
 160 //       is then responsible for progress and succession and can use CAS to detach and
 161 //       drain the cxq into the EntryList.  By convention, only this thread, the holder of
 162 //       the OnDeck inner lock, can manipulate the EntryList or detach and drain the
 163 //       RATs on the cxq into the EntryList.  This avoids ABA corruption on the cxq as
 164 //       we allow multiple concurrent "push" operations but restrict detach concurrency
 165 //       to at most one thread.  Having selected and detached a successor, the thread then
 166 //       changes the OnDeck to refer to that successor, and then unparks the successor.
 167 //       That successor will eventually acquire the lock and clear OnDeck.  Beware
 168 //       that the OnDeck usage as a lock is asymmetric.  A thread in unlock() transiently
 169 //       "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor,
 170 //       and then the successor eventually "drops" OnDeck.  Note that there's never
 171 //       any sense of contention on the inner lock, however.  Threads never contend
 172 //       or wait for the inner lock.
 173 //   --  OnDeck provides for futile wakeup throttling a described in section 3.3 of
 174 //       See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 175 //       In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter
 176 //       TState fields found in Java-level objectMonitors.  (See synchronizer.cpp).
 177 //
 178 // * Waiting threads reside on the WaitSet list -- wait() puts
 179 //   the caller onto the WaitSet.  Notify() or notifyAll() simply
 180 //   transfers threads from the WaitSet to either the EntryList or cxq.
 181 //   Subsequent unlock() operations will eventually unpark the notifyee.
 182 //   Unparking a notifee in notify() proper is inefficient - if we were to do so
 183 //   it's likely the notifyee would simply impale itself on the lock held
 184 //   by the notifier.
 185 //
 186 // * The mechanism is obstruction-free in that if the holder of the transient
 187 //   OnDeck lock in unlock() is preempted or otherwise stalls, other threads
 188 //   can still acquire and release the outer lock and continue to make progress.
 189 //   At worst, waking of already blocked contending threads may be delayed,
 190 //   but nothing worse.  (We only use "trylock" operations on the inner OnDeck
 191 //   lock).
 192 //
 193 // * Note that thread-local storage must be initialized before a thread
 194 //   uses Native monitors or mutexes.  The native monitor-mutex subsystem
 195 //   depends on Thread::current().
 196 //
 197 // * The monitor synchronization subsystem avoids the use of native
 198 //   synchronization primitives except for the narrow platform-specific
 199 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 200 //   the semantics of park-unpark.  Put another way, this monitor implementation
 201 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 202 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 203 //   underlying OS manages the READY<->RUN transitions.
 204 //
 205 // * The memory consistency model provide by lock()-unlock() is at least as
 206 //   strong or stronger than the Java Memory model defined by JSR-133.
 207 //   That is, we guarantee at least entry consistency, if not stronger.
 208 //   See http://g.oswego.edu/dl/jmm/cookbook.html.
 209 //
 210 // * Thread:: currently contains a set of purpose-specific ParkEvents:
 211 //   _MutexEvent, _ParkEvent, etc.  A better approach might be to do away with
 212 //   the purpose-specific ParkEvents and instead implement a general per-thread
 213 //   stack of available ParkEvents which we could provision on-demand.  The
 214 //   stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate()
 215 //   and ::Release().  A thread would simply pop an element from the local stack before it
 216 //   enqueued or park()ed.  When the contention was over the thread would
 217 //   push the no-longer-needed ParkEvent back onto its stack.
 218 //
 219 // * A slightly reduced form of ILock() and IUnlock() have been partially
 220 //   model-checked (Murphi) for safety and progress at T=1,2,3 and 4.
 221 //   It'd be interesting to see if TLA/TLC could be useful as well.
 222 //
 223 // * Mutex-Monitor is a low-level "leaf" subsystem.  That is, the monitor
 224 //   code should never call other code in the JVM that might itself need to
 225 //   acquire monitors or mutexes.  That's true *except* in the case of the
 226 //   ThreadBlockInVM state transition wrappers.  The ThreadBlockInVM DTOR handles
 227 //   mutator reentry (ingress) by checking for a pending safepoint in which case it will
 228 //   call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc.
 229 //   In that particular case a call to lock() for a given Monitor can end up recursively
 230 //   calling lock() on another monitor.   While distasteful, this is largely benign
 231 //   as the calls come from jacket that wraps lock(), and not from deep within lock() itself.
 232 //
 233 //   It's unfortunate that native mutexes and thread state transitions were convolved.
 234 //   They're really separate concerns and should have remained that way.  Melding
 235 //   them together was facile -- a bit too facile.   The current implementation badly
 236 //   conflates the two concerns.
 237 //
 238 // * TODO-FIXME:
 239 //
 240 //   -- Add DTRACE probes for contended acquire, contended acquired, contended unlock
 241 //      We should also add DTRACE probes in the ParkEvent subsystem for
 242 //      Park-entry, Park-exit, and Unpark.
 243 //
 244 //   -- We have an excess of mutex-like constructs in the JVM, namely:
 245 //      1. objectMonitors for Java-level synchronization (synchronizer.cpp)
 246 //      2. low-level muxAcquire and muxRelease
 247 //      3. low-level spinAcquire and spinRelease
 248 //      4. native Mutex:: and Monitor::
 249 //      5. jvm_raw_lock() and _unlock()
 250 //      6. JVMTI raw monitors -- distinct from (5) despite having a confusingly
 251 //         similar name.
 252 //
 253 // o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o
 254 
 255 #define UNS(x) (uintptr_t(x))
 256 #define TRACE(m)                   \
 257   {                                \
 258     static volatile int ctr = 0;   \
 259     int x = ++ctr;                 \
 260     if ((x & (x - 1)) == 0) {      \
 261       ::printf("%d:%s\n", x, #m);  \
 262       ::fflush(stdout);            \
 263     }                              \
 264   }
 265 
 266 const intptr_t _LBIT = 1;
 267 
 268 // Endian-ness ... index of least-significant byte in SplitWord.Bytes[]
 269 #ifdef VM_LITTLE_ENDIAN
 270  #define _LSBINDEX 0
 271 #else
 272  #define _LSBINDEX (sizeof(intptr_t)-1)
 273 #endif
 274 
 275 // Simplistic low-quality Marsaglia SHIFT-XOR RNG.
 276 // Bijective except for the trailing mask operation.
 277 // Useful for spin loops as the compiler can't optimize it away.
 278 
 279 static inline jint MarsagliaXORV(jint x) {
 280   if (x == 0) x = 1|os::random();
 281   x ^= x << 6;
 282   x ^= ((unsigned)x) >> 21;
 283   x ^= x << 7;
 284   return x & 0x7FFFFFFF;
 285 }
 286 
 287 static int Stall(int its) {
 288   static volatile jint rv = 1;
 289   volatile int OnFrame = 0;
 290   jint v = rv ^ UNS(OnFrame);
 291   while (--its >= 0) {
 292     v = MarsagliaXORV(v);
 293   }
 294   // Make this impossible for the compiler to optimize away,
 295   // but (mostly) avoid W coherency sharing on MP systems.
 296   if (v == 0x12345) rv = v;
 297   return v;
 298 }
 299 
 300 int Monitor::TryLock() {
 301   intptr_t v = _LockWord.FullWord;
 302   for (;;) {
 303     if ((v & _LBIT) != 0) return 0;
 304     const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
 305     if (v == u) return 1;
 306     v = u;
 307   }
 308 }
 309 
 310 int Monitor::TryFast() {
 311   // Optimistic fast-path form ...
 312   // Fast-path attempt for the common uncontended case.
 313   // Avoid RTS->RTO $ coherence upgrade on typical SMP systems.
 314   intptr_t v = Atomic::cmpxchg(_LBIT, &_LockWord.FullWord, (intptr_t)0);  // agro ...
 315   if (v == 0) return 1;
 316 
 317   for (;;) {
 318     if ((v & _LBIT) != 0) return 0;
 319     const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
 320     if (v == u) return 1;
 321     v = u;
 322   }
 323 }
 324 
 325 int Monitor::ILocked() {
 326   const intptr_t w = _LockWord.FullWord & 0xFF;
 327   assert(w == 0 || w == _LBIT, "invariant");
 328   return w == _LBIT;
 329 }
 330 
 331 // Polite TATAS spinlock with exponential backoff - bounded spin.
 332 // Ideally we'd use processor cycles, time or vtime to control
 333 // the loop, but we currently use iterations.
 334 // All the constants within were derived empirically but work over
 335 // over the spectrum of J2SE reference platforms.
 336 // On Niagara-class systems the back-off is unnecessary but
 337 // is relatively harmless.  (At worst it'll slightly retard
 338 // acquisition times).  The back-off is critical for older SMP systems
 339 // where constant fetching of the LockWord would otherwise impair
 340 // scalability.
 341 //
 342 // Clamp spinning at approximately 1/2 of a context-switch round-trip.
 343 // See synchronizer.cpp for details and rationale.
 344 
 345 int Monitor::TrySpin(Thread * const Self) {
 346   if (TryLock())    return 1;
 347   if (!os::is_MP()) return 0;
 348 
 349   int Probes  = 0;
 350   int Delay   = 0;
 351   int SpinMax = 20;
 352   for (;;) {
 353     intptr_t v = _LockWord.FullWord;
 354     if ((v & _LBIT) == 0) {
 355       if (Atomic::cmpxchg (v|_LBIT, &_LockWord.FullWord, v) == v) {
 356         return 1;
 357       }
 358       continue;
 359     }
 360 
 361     SpinPause();
 362 
 363     // Periodically increase Delay -- variable Delay form
 364     // conceptually: delay *= 1 + 1/Exponent
 365     ++Probes;
 366     if (Probes > SpinMax) return 0;
 367 
 368     if ((Probes & 0x7) == 0) {
 369       Delay = ((Delay << 1)|1) & 0x7FF;
 370       // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ;
 371     }
 372 
 373     // Stall for "Delay" time units - iterations in the current implementation.
 374     // Avoid generating coherency traffic while stalled.
 375     // Possible ways to delay:
 376     //   PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt,
 377     //   wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ...
 378     // Note that on Niagara-class systems we want to minimize STs in the
 379     // spin loop.  N1 and brethren write-around the L1$ over the xbar into the L2$.
 380     // Furthermore, they don't have a W$ like traditional SPARC processors.
 381     // We currently use a Marsaglia Shift-Xor RNG loop.
 382     if (Self != NULL) {
 383       jint rv = Self->rng[0];
 384       for (int k = Delay; --k >= 0;) {
 385         rv = MarsagliaXORV(rv);
 386         if (SafepointMechanism::poll(Self)) return 0;
 387       }
 388       Self->rng[0] = rv;
 389     } else {
 390       Stall(Delay);
 391     }
 392   }
 393 }
 394 
 395 static int ParkCommon(ParkEvent * ev, jlong timo) {
 396   // Diagnostic support - periodically unwedge blocked threads
 397   int err = OS_OK;
 398   if (0 == timo) {
 399     ev->park();
 400   } else {
 401     err = ev->park(timo);
 402   }
 403   return err;
 404 }
 405 
 406 inline int Monitor::AcquireOrPush(ParkEvent * ESelf) {
 407   intptr_t v = _LockWord.FullWord;
 408   for (;;) {
 409     if ((v & _LBIT) == 0) {
 410       const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
 411       if (u == v) return 1;        // indicate acquired
 412       v = u;
 413     } else {
 414       // Anticipate success ...
 415       ESelf->ListNext = (ParkEvent *)(v & ~_LBIT);
 416       const intptr_t u = Atomic::cmpxchg(intptr_t(ESelf)|_LBIT, &_LockWord.FullWord, v);
 417       if (u == v) return 0;        // indicate pushed onto cxq
 418       v = u;
 419     }
 420     // Interference - LockWord change - just retry
 421   }
 422 }
 423 
 424 // ILock and IWait are the lowest level primitive internal blocking
 425 // synchronization functions.  The callers of IWait and ILock must have
 426 // performed any needed state transitions beforehand.
 427 // IWait and ILock may directly call park() without any concern for thread state.
 428 // Note that ILock and IWait do *not* access _owner.
 429 // _owner is a higher-level logical concept.
 430 
 431 void Monitor::ILock(Thread * Self) {
 432   assert(_OnDeck != Self->_MutexEvent, "invariant");
 433 
 434   if (TryFast()) {
 435  Exeunt:
 436     assert(ILocked(), "invariant");
 437     return;
 438   }
 439 
 440   ParkEvent * const ESelf = Self->_MutexEvent;
 441   assert(_OnDeck != ESelf, "invariant");
 442 
 443   // As an optimization, spinners could conditionally try to set _OnDeck to _LBIT
 444   // Synchronizer.cpp uses a similar optimization.
 445   if (TrySpin(Self)) goto Exeunt;
 446 
 447   // Slow-path - the lock is contended.
 448   // Either Enqueue Self on cxq or acquire the outer lock.
 449   // LockWord encoding = (cxq,LOCKBYTE)
 450   ESelf->reset();
 451   OrderAccess::fence();
 452 
 453   if (AcquireOrPush(ESelf)) goto Exeunt;
 454 
 455   // At any given time there is at most one ondeck thread.
 456   // ondeck implies not resident on cxq and not resident on EntryList
 457   // Only the OnDeck thread can try to acquire -- contend for -- the lock.
 458   // CONSIDER: use Self->OnDeck instead of m->OnDeck.
 459   // Deschedule Self so that others may run.
 460   while (OrderAccess::load_acquire(&_OnDeck) != ESelf) {
 461     ParkCommon(ESelf, 0);
 462   }
 463 
 464   // Self is now in the OnDeck position and will remain so until it
 465   // manages to acquire the lock.
 466   for (;;) {
 467     assert(_OnDeck == ESelf, "invariant");
 468     if (TrySpin(Self)) break;
 469     // It's probably wise to spin only if we *actually* blocked
 470     // CONSIDER: check the lockbyte, if it remains set then
 471     // preemptively drain the cxq into the EntryList.
 472     // The best place and time to perform queue operations -- lock metadata --
 473     // is _before having acquired the outer lock, while waiting for the lock to drop.
 474     ParkCommon(ESelf, 0);
 475   }
 476 
 477   assert(_OnDeck == ESelf, "invariant");
 478   _OnDeck = NULL;
 479 
 480   // Note that we current drop the inner lock (clear OnDeck) in the slow-path
 481   // epilogue immediately after having acquired the outer lock.
 482   // But instead we could consider the following optimizations:
 483   // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation.
 484   //    This might avoid potential reacquisition of the inner lock in IUlock().
 485   // B. While still holding the inner lock, attempt to opportunistically select
 486   //    and unlink the next OnDeck thread from the EntryList.
 487   //    If successful, set OnDeck to refer to that thread, otherwise clear OnDeck.
 488   //    It's critical that the select-and-unlink operation run in constant-time as
 489   //    it executes when holding the outer lock and may artificially increase the
 490   //    effective length of the critical section.
 491   // Note that (A) and (B) are tantamount to succession by direct handoff for
 492   // the inner lock.
 493   goto Exeunt;
 494 }
 495 
 496 void Monitor::IUnlock(bool RelaxAssert) {
 497   assert(ILocked(), "invariant");
 498   // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately
 499   // before the store that releases the lock.  Crucially, all the stores and loads in the
 500   // critical section must be globally visible before the store of 0 into the lock-word
 501   // that releases the lock becomes globally visible.  That is, memory accesses in the
 502   // critical section should not be allowed to bypass or overtake the following ST that
 503   // releases the lock.  As such, to prevent accesses within the critical section
 504   // from "leaking" out, we need a release fence between the critical section and the
 505   // store that releases the lock.  In practice that release barrier is elided on
 506   // platforms with strong memory models such as TSO.
 507   //
 508   // Note that the OrderAccess::storeload() fence that appears after unlock store
 509   // provides for progress conditions and succession and is _not related to exclusion
 510   // safety or lock release consistency.
 511   OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], jbyte(0)); // drop outer lock
 512 
 513   OrderAccess::storeload();
 514   ParkEvent * const w = _OnDeck; // raw load as we will just return if non-NULL
 515   assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant");
 516   if (w != NULL) {
 517     // Either we have a valid ondeck thread or ondeck is transiently "locked"
 518     // by some exiting thread as it arranges for succession.  The LSBit of
 519     // OnDeck allows us to discriminate two cases.  If the latter, the
 520     // responsibility for progress and succession lies with that other thread.
 521     // For good performance, we also depend on the fact that redundant unpark()
 522     // operations are cheap.  That is, repeated Unpark()ing of the OnDeck thread
 523     // is inexpensive.  This approach provides implicit futile wakeup throttling.
 524     // Note that the referent "w" might be stale with respect to the lock.
 525     // In that case the following unpark() is harmless and the worst that'll happen
 526     // is a spurious return from a park() operation.  Critically, if "w" _is stale,
 527     // then progress is known to have occurred as that means the thread associated
 528     // with "w" acquired the lock.  In that case this thread need take no further
 529     // action to guarantee progress.
 530     if ((UNS(w) & _LBIT) == 0) w->unpark();
 531     return;
 532   }
 533 
 534   intptr_t cxq = _LockWord.FullWord;
 535   if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) {
 536     return;      // normal fast-path exit - cxq and EntryList both empty
 537   }
 538   if (cxq & _LBIT) {
 539     // Optional optimization ...
 540     // Some other thread acquired the lock in the window since this
 541     // thread released it.  Succession is now that thread's responsibility.
 542     return;
 543   }
 544 
 545  Succession:
 546   // Slow-path exit - this thread must ensure succession and progress.
 547   // OnDeck serves as lock to protect cxq and EntryList.
 548   // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq.
 549   // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS)
 550   // but only one concurrent consumer (detacher of RATs).
 551   // Unlike a normal lock, however, the exiting thread "locks" OnDeck,
 552   // picks a successor and marks that thread as OnDeck.  That successor
 553   // thread will then clear OnDeck once it eventually acquires the outer lock.
 554   if (!Atomic::replace_if_null((ParkEvent*)_LBIT, &_OnDeck)) {
 555     return;
 556   }
 557 
 558   ParkEvent * List = _EntryList;
 559   if (List != NULL) {
 560     // Transfer the head of the EntryList to the OnDeck position.
 561     // Once OnDeck, a thread stays OnDeck until it acquires the lock.
 562     // For a given lock there is at most OnDeck thread at any one instant.
 563    WakeOne:
 564     assert(List == _EntryList, "invariant");
 565     ParkEvent * const w = List;
 566     assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant");
 567     _EntryList = w->ListNext;
 568     // as a diagnostic measure consider setting w->_ListNext = BAD
 569     assert(intptr_t(_OnDeck) == _LBIT, "invariant");
 570 
 571     // Pass OnDeck role to w, ensuring that _EntryList has been set first.
 572     // w will clear _OnDeck once it acquires the outer lock.
 573     // Note that once we set _OnDeck that thread can acquire the mutex, proceed
 574     // with its critical section and then enter this code to unlock the mutex. So
 575     // you can have multiple threads active in IUnlock at the same time.
 576     OrderAccess::release_store(&_OnDeck, w);
 577 
 578     // Another optional optimization ...
 579     // For heavily contended locks it's not uncommon that some other
 580     // thread acquired the lock while this thread was arranging succession.
 581     // Try to defer the unpark() operation - Delegate the responsibility
 582     // for unpark()ing the OnDeck thread to the current or subsequent owners
 583     // That is, the new owner is responsible for unparking the OnDeck thread.
 584     OrderAccess::storeload();
 585     cxq = _LockWord.FullWord;
 586     if (cxq & _LBIT) return;
 587 
 588     w->unpark();
 589     return;
 590   }
 591 
 592   cxq = _LockWord.FullWord;
 593   if ((cxq & ~_LBIT) != 0) {
 594     // The EntryList is empty but the cxq is populated.
 595     // drain RATs from cxq into EntryList
 596     // Detach RATs segment with CAS and then merge into EntryList
 597     for (;;) {
 598       // optional optimization - if locked, the owner is responsible for succession
 599       if (cxq & _LBIT) goto Punt;
 600       const intptr_t vfy = Atomic::cmpxchg(cxq & _LBIT, &_LockWord.FullWord, cxq);
 601       if (vfy == cxq) break;
 602       cxq = vfy;
 603       // Interference - LockWord changed - Just retry
 604       // We can see concurrent interference from contending threads
 605       // pushing themselves onto the cxq or from lock-unlock operations.
 606       // From the perspective of this thread, EntryList is stable and
 607       // the cxq is prepend-only -- the head is volatile but the interior
 608       // of the cxq is stable.  In theory if we encounter interference from threads
 609       // pushing onto cxq we could simply break off the original cxq suffix and
 610       // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts
 611       // on the high-traffic LockWord variable.   For instance lets say the cxq is "ABCD"
 612       // when we first fetch cxq above.  Between the fetch -- where we observed "A"
 613       // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive,
 614       // yielding cxq = "PQRABCD".  In this case we could simply set A.ListNext
 615       // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList.
 616       // Note too, that it's safe for this thread to traverse the cxq
 617       // without taking any special concurrency precautions.
 618     }
 619 
 620     // We don't currently reorder the cxq segment as we move it onto
 621     // the EntryList, but it might make sense to reverse the order
 622     // or perhaps sort by thread priority.  See the comments in
 623     // synchronizer.cpp objectMonitor::exit().
 624     assert(_EntryList == NULL, "invariant");
 625     _EntryList = List = (ParkEvent *)(cxq & ~_LBIT);
 626     assert(List != NULL, "invariant");
 627     goto WakeOne;
 628   }
 629 
 630   // cxq|EntryList is empty.
 631   // w == NULL implies that cxq|EntryList == NULL in the past.
 632   // Possible race - rare inopportune interleaving.
 633   // A thread could have added itself to cxq since this thread previously checked.
 634   // Detect and recover by refetching cxq.
 635  Punt:
 636   assert(intptr_t(_OnDeck) == _LBIT, "invariant");
 637   _OnDeck = NULL;            // Release inner lock.
 638   OrderAccess::storeload();   // Dekker duality - pivot point
 639 
 640   // Resample LockWord/cxq to recover from possible race.
 641   // For instance, while this thread T1 held OnDeck, some other thread T2 might
 642   // acquire the outer lock.  Another thread T3 might try to acquire the outer
 643   // lock, but encounter contention and enqueue itself on cxq.  T2 then drops the
 644   // outer lock, but skips succession as this thread T1 still holds OnDeck.
 645   // T1 is and remains responsible for ensuring succession of T3.
 646   //
 647   // Note that we don't need to recheck EntryList, just cxq.
 648   // If threads moved onto EntryList since we dropped OnDeck
 649   // that implies some other thread forced succession.
 650   cxq = _LockWord.FullWord;
 651   if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) {
 652     goto Succession;         // potential race -- re-run succession
 653   }
 654   return;
 655 }
 656 
 657 bool Monitor::notify() {
 658   assert(_owner == Thread::current(), "invariant");
 659   assert(ILocked(), "invariant");
 660   if (_WaitSet == NULL) return true;
 661 
 662   // Transfer one thread from the WaitSet to the EntryList or cxq.
 663   // Currently we just unlink the head of the WaitSet and prepend to the cxq.
 664   // And of course we could just unlink it and unpark it, too, but
 665   // in that case it'd likely impale itself on the reentry.
 666   Thread::muxAcquire(_WaitLock, "notify:WaitLock");
 667   ParkEvent * nfy = _WaitSet;
 668   if (nfy != NULL) {                  // DCL idiom
 669     _WaitSet = nfy->ListNext;
 670     assert(nfy->Notified == 0, "invariant");
 671     // push nfy onto the cxq
 672     for (;;) {
 673       const intptr_t v = _LockWord.FullWord;
 674       assert((v & 0xFF) == _LBIT, "invariant");
 675       nfy->ListNext = (ParkEvent *)(v & ~_LBIT);
 676       if (Atomic::cmpxchg(intptr_t(nfy)|_LBIT, &_LockWord.FullWord, v) == v) break;
 677       // interference - _LockWord changed -- just retry
 678     }
 679     // Note that setting Notified before pushing nfy onto the cxq is
 680     // also legal and safe, but the safety properties are much more
 681     // subtle, so for the sake of code stewardship ...
 682     OrderAccess::fence();
 683     nfy->Notified = 1;
 684   }
 685   Thread::muxRelease(_WaitLock);
 686   assert(ILocked(), "invariant");
 687   return true;
 688 }
 689 
 690 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset
 691 // to the cxq.  This could be done more efficiently with a single bulk en-mass transfer,
 692 // but in practice notifyAll() for large #s of threads is rare and not time-critical.
 693 // Beware too, that we invert the order of the waiters.  Lets say that the
 694 // waitset is "ABCD" and the cxq is "XYZ".  After a notifyAll() the waitset
 695 // will be empty and the cxq will be "DCBAXYZ".  This is benign, of course.
 696 
 697 bool Monitor::notify_all() {
 698   assert(_owner == Thread::current(), "invariant");
 699   assert(ILocked(), "invariant");
 700   while (_WaitSet != NULL) notify();
 701   return true;
 702 }
 703 
 704 int Monitor::IWait(Thread * Self, jlong timo) {
 705   assert(ILocked(), "invariant");
 706 
 707   // Phases:
 708   // 1. Enqueue Self on WaitSet - currently prepend
 709   // 2. unlock - drop the outer lock
 710   // 3. wait for either notification or timeout
 711   // 4. lock - reentry - reacquire the outer lock
 712 
 713   ParkEvent * const ESelf = Self->_MutexEvent;
 714   ESelf->Notified = 0;
 715   ESelf->reset();
 716   OrderAccess::fence();
 717 
 718   // Add Self to WaitSet
 719   // Ideally only the holder of the outer lock would manipulate the WaitSet -
 720   // That is, the outer lock would implicitly protect the WaitSet.
 721   // But if a thread in wait() encounters a timeout it will need to dequeue itself
 722   // from the WaitSet _before it becomes the owner of the lock.  We need to dequeue
 723   // as the ParkEvent -- which serves as a proxy for the thread -- can't reside
 724   // on both the WaitSet and the EntryList|cxq at the same time..  That is, a thread
 725   // on the WaitSet can't be allowed to compete for the lock until it has managed to
 726   // unlink its ParkEvent from WaitSet.  Thus the need for WaitLock.
 727   // Contention on the WaitLock is minimal.
 728   //
 729   // Another viable approach would be add another ParkEvent, "WaitEvent" to the
 730   // thread class.  The WaitSet would be composed of WaitEvents.  Only the
 731   // owner of the outer lock would manipulate the WaitSet.  A thread in wait()
 732   // could then compete for the outer lock, and then, if necessary, unlink itself
 733   // from the WaitSet only after having acquired the outer lock.  More precisely,
 734   // there would be no WaitLock.  A thread in in wait() would enqueue its WaitEvent
 735   // on the WaitSet; release the outer lock; wait for either notification or timeout;
 736   // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet.
 737   //
 738   // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice.
 739   // One set would be for the WaitSet and one for the EntryList.
 740   // We could also deconstruct the ParkEvent into a "pure" event and add a
 741   // new immortal/TSM "ListElement" class that referred to ParkEvents.
 742   // In that case we could have one ListElement on the WaitSet and another
 743   // on the EntryList, with both referring to the same pure Event.
 744 
 745   Thread::muxAcquire(_WaitLock, "wait:WaitLock:Add");
 746   ESelf->ListNext = _WaitSet;
 747   _WaitSet = ESelf;
 748   Thread::muxRelease(_WaitLock);
 749 
 750   // Release the outer lock
 751   // We call IUnlock (RelaxAssert=true) as a thread T1 might
 752   // enqueue itself on the WaitSet, call IUnlock(), drop the lock,
 753   // and then stall before it can attempt to wake a successor.
 754   // Some other thread T2 acquires the lock, and calls notify(), moving
 755   // T1 from the WaitSet to the cxq.  T2 then drops the lock.  T1 resumes,
 756   // and then finds *itself* on the cxq.  During the course of a normal
 757   // IUnlock() call a thread should _never find itself on the EntryList
 758   // or cxq, but in the case of wait() it's possible.
 759   // See synchronizer.cpp objectMonitor::wait().
 760   IUnlock(true);
 761 
 762   // Wait for either notification or timeout
 763   // Beware that in some circumstances we might propagate
 764   // spurious wakeups back to the caller.
 765 
 766   for (;;) {
 767     if (ESelf->Notified) break;
 768     int err = ParkCommon(ESelf, timo);
 769     if (err == OS_TIMEOUT) break;
 770   }
 771 
 772   // Prepare for reentry - if necessary, remove ESelf from WaitSet
 773   // ESelf can be:
 774   // 1. Still on the WaitSet.  This can happen if we exited the loop by timeout.
 775   // 2. On the cxq or EntryList
 776   // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position.
 777 
 778   OrderAccess::fence();
 779   int WasOnWaitSet = 0;
 780   if (ESelf->Notified == 0) {
 781     Thread::muxAcquire(_WaitLock, "wait:WaitLock:remove");
 782     if (ESelf->Notified == 0) {     // DCL idiom
 783       assert(_OnDeck != ESelf, "invariant");   // can't be both OnDeck and on WaitSet
 784       // ESelf is resident on the WaitSet -- unlink it.
 785       // A doubly-linked list would be better here so we can unlink in constant-time.
 786       // We have to unlink before we potentially recontend as ESelf might otherwise
 787       // end up on the cxq|EntryList -- it can't be on two lists at once.
 788       ParkEvent * p = _WaitSet;
 789       ParkEvent * q = NULL;            // classic q chases p
 790       while (p != NULL && p != ESelf) {
 791         q = p;
 792         p = p->ListNext;
 793       }
 794       assert(p == ESelf, "invariant");
 795       if (p == _WaitSet) {      // found at head
 796         assert(q == NULL, "invariant");
 797         _WaitSet = p->ListNext;
 798       } else {                  // found in interior
 799         assert(q->ListNext == p, "invariant");
 800         q->ListNext = p->ListNext;
 801       }
 802       WasOnWaitSet = 1;        // We were *not* notified but instead encountered timeout
 803     }
 804     Thread::muxRelease(_WaitLock);
 805   }
 806 
 807   // Reentry phase - reacquire the lock
 808   if (WasOnWaitSet) {
 809     // ESelf was previously on the WaitSet but we just unlinked it above
 810     // because of a timeout.  ESelf is not resident on any list and is not OnDeck
 811     assert(_OnDeck != ESelf, "invariant");
 812     ILock(Self);
 813   } else {
 814     // A prior notify() operation moved ESelf from the WaitSet to the cxq.
 815     // ESelf is now on the cxq, EntryList or at the OnDeck position.
 816     // The following fragment is extracted from Monitor::ILock()
 817     for (;;) {
 818       if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(Self)) break;
 819       ParkCommon(ESelf, 0);
 820     }
 821     assert(_OnDeck == ESelf, "invariant");
 822     _OnDeck = NULL;
 823   }
 824 
 825   assert(ILocked(), "invariant");
 826   return WasOnWaitSet != 0;        // return true IFF timeout
 827 }
 828 
 829 
 830 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS:
 831 // In particular, there are certain types of global lock that may be held
 832 // by a Java thread while it is blocked at a safepoint but before it has
 833 // written the _owner field. These locks may be sneakily acquired by the
 834 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should
 835 // identify all such locks, and ensure that Java threads never block at
 836 // safepoints while holding them (_no_safepoint_check_flag). While it
 837 // seems as though this could increase the time to reach a safepoint
 838 // (or at least increase the mean, if not the variance), the latter
 839 // approach might make for a cleaner, more maintainable JVM design.
 840 //
 841 // Sneaking is vile and reprehensible and should be excised at the 1st
 842 // opportunity.  It's possible that the need for sneaking could be obviated
 843 // as follows.  Currently, a thread might (a) while TBIVM, call pthread_mutex_lock
 844 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex.
 845 // (b) stall at the TBIVM exit point as a safepoint is in effect.  Critically,
 846 // it'll stall at the TBIVM reentry state transition after having acquired the
 847 // underlying lock, but before having set _owner and having entered the actual
 848 // critical section.  The lock-sneaking facility leverages that fact and allowed the
 849 // VM thread to logically acquire locks that had already be physically locked by mutators
 850 // but where mutators were known blocked by the reentry thread state transition.
 851 //
 852 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly
 853 // wrapped calls to park(), then we could likely do away with sneaking.  We'd
 854 // decouple lock acquisition and parking.  The critical invariant  to eliminating
 855 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM.
 856 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket.
 857 // One difficulty with this approach is that the TBIVM wrapper could recurse and
 858 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued.
 859 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem.
 860 //
 861 // But of course the proper ultimate approach is to avoid schemes that require explicit
 862 // sneaking or dependence on any any clever invariants or subtle implementation properties
 863 // of Mutex-Monitor and instead directly address the underlying design flaw.
 864 
 865 void Monitor::lock(Thread * Self) {
 866   // Ensure that the Monitor requires/allows safepoint checks.
 867   assert(_safepoint_check_required != Monitor::_safepoint_check_never,
 868          "This lock should never have a safepoint check: %s", name());
 869 
 870 #ifdef CHECK_UNHANDLED_OOPS
 871   // Clear unhandled oops so we get a crash right away.  Only clear for non-vm
 872   // or GC threads.
 873   if (Self->is_Java_thread()) {
 874     Self->clear_unhandled_oops();
 875   }
 876 #endif // CHECK_UNHANDLED_OOPS
 877 
 878   debug_only(check_prelock_state(Self, StrictSafepointChecks));
 879   assert(_owner != Self, "invariant");
 880   assert(_OnDeck != Self->_MutexEvent, "invariant");
 881 
 882   if (TryFast()) {
 883  Exeunt:
 884     assert(ILocked(), "invariant");
 885     assert(owner() == NULL, "invariant");
 886     set_owner(Self);
 887     return;
 888   }
 889 
 890   // The lock is contended ...
 891 
 892   bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
 893   if (can_sneak && _owner == NULL) {
 894     // a java thread has locked the lock but has not entered the
 895     // critical region -- let's just pretend we've locked the lock
 896     // and go on.  we note this with _snuck so we can also
 897     // pretend to unlock when the time comes.
 898     _snuck = true;
 899     goto Exeunt;
 900   }
 901 
 902   // Try a brief spin to avoid passing thru thread state transition ...
 903   if (TrySpin(Self)) goto Exeunt;
 904 
 905   check_block_state(Self);
 906   if (Self->is_Java_thread()) {
 907     // Horrible dictu - we suffer through a state transition
 908     assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex");
 909     ThreadBlockInVM tbivm((JavaThread *) Self);
 910     ILock(Self);
 911   } else {
 912     // Mirabile dictu
 913     ILock(Self);
 914   }
 915   goto Exeunt;
 916 }
 917 
 918 void Monitor::lock() {
 919   this->lock(Thread::current());
 920 }
 921 
 922 // Lock without safepoint check - a degenerate variant of lock().
 923 // Should ONLY be used by safepoint code and other code
 924 // that is guaranteed not to block while running inside the VM. If this is called with
 925 // thread state set to be in VM, the safepoint synchronization code will deadlock!
 926 
 927 void Monitor::lock_without_safepoint_check(Thread * Self) {
 928   // Ensure that the Monitor does not require or allow safepoint checks.
 929   assert(_safepoint_check_required != Monitor::_safepoint_check_always,
 930          "This lock should always have a safepoint check: %s", name());
 931   assert(_owner != Self, "invariant");
 932   ILock(Self);
 933   assert(_owner == NULL, "invariant");
 934   set_owner(Self);
 935 }
 936 
 937 void Monitor::lock_without_safepoint_check() {
 938   lock_without_safepoint_check(Thread::current());
 939 }
 940 
 941 
 942 // Returns true if thread succeeds in grabbing the lock, otherwise false.
 943 
 944 bool Monitor::try_lock() {
 945   Thread * const Self = Thread::current();
 946   debug_only(check_prelock_state(Self, false));
 947   // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler");
 948 
 949   // Special case, where all Java threads are stopped.
 950   // The lock may have been acquired but _owner is not yet set.
 951   // In that case the VM thread can safely grab the lock.
 952   // It strikes me this should appear _after the TryLock() fails, below.
 953   bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
 954   if (can_sneak && _owner == NULL) {
 955     set_owner(Self); // Do not need to be atomic, since we are at a safepoint
 956     _snuck = true;
 957     return true;
 958   }
 959 
 960   if (TryLock()) {
 961     // We got the lock
 962     assert(_owner == NULL, "invariant");
 963     set_owner(Self);
 964     return true;
 965   }
 966   return false;
 967 }
 968 
 969 void Monitor::unlock() {
 970   assert(_owner == Thread::current(), "invariant");
 971   assert(_OnDeck != Thread::current()->_MutexEvent, "invariant");
 972   set_owner(NULL);
 973   if (_snuck) {
 974     assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
 975     _snuck = false;
 976     return;
 977   }
 978   IUnlock(false);
 979 }
 980 
 981 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check()
 982 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter.
 983 //
 984 // There's no expectation that JVM_RawMonitors will interoperate properly with the native
 985 // Mutex-Monitor constructs.  We happen to implement JVM_RawMonitors in terms of
 986 // native Mutex-Monitors simply as a matter of convenience.  A simple abstraction layer
 987 // over a pthread_mutex_t would work equally as well, but require more platform-specific
 988 // code -- a "PlatformMutex".  Alternatively, a simply layer over muxAcquire-muxRelease
 989 // would work too.
 990 //
 991 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent
 992 // instance available.  Instead, we transiently allocate a ParkEvent on-demand if
 993 // we encounter contention.  That ParkEvent remains associated with the thread
 994 // until it manages to acquire the lock, at which time we return the ParkEvent
 995 // to the global ParkEvent free list.  This is correct and suffices for our purposes.
 996 //
 997 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that
 998 // jvm_raw_lock() didn't have the corresponding test.  I suspect that's an
 999 // oversight, but I've replicated the original suspect logic in the new code ...
1000 
1001 void Monitor::jvm_raw_lock() {
1002   assert(rank() == native, "invariant");
1003 
1004   if (TryLock()) {
1005  Exeunt:
1006     assert(ILocked(), "invariant");
1007     assert(_owner == NULL, "invariant");
1008     // This can potentially be called by non-java Threads. Thus, the Thread::current_or_null()
1009     // might return NULL. Don't call set_owner since it will break on an NULL owner
1010     // Consider installing a non-null "ANON" distinguished value instead of just NULL.
1011     _owner = Thread::current_or_null();
1012     return;
1013   }
1014 
1015   if (TrySpin(NULL)) goto Exeunt;
1016 
1017   // slow-path - apparent contention
1018   // Allocate a ParkEvent for transient use.
1019   // The ParkEvent remains associated with this thread until
1020   // the time the thread manages to acquire the lock.
1021   ParkEvent * const ESelf = ParkEvent::Allocate(NULL);
1022   ESelf->reset();
1023   OrderAccess::storeload();
1024 
1025   // Either Enqueue Self on cxq or acquire the outer lock.
1026   if (AcquireOrPush (ESelf)) {
1027     ParkEvent::Release(ESelf);      // surrender the ParkEvent
1028     goto Exeunt;
1029   }
1030 
1031   // At any given time there is at most one ondeck thread.
1032   // ondeck implies not resident on cxq and not resident on EntryList
1033   // Only the OnDeck thread can try to acquire -- contend for -- the lock.
1034   // CONSIDER: use Self->OnDeck instead of m->OnDeck.
1035   for (;;) {
1036     if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(NULL)) break;
1037     ParkCommon(ESelf, 0);
1038   }
1039 
1040   assert(_OnDeck == ESelf, "invariant");
1041   _OnDeck = NULL;
1042   ParkEvent::Release(ESelf);      // surrender the ParkEvent
1043   goto Exeunt;
1044 }
1045 
1046 void Monitor::jvm_raw_unlock() {
1047   // Nearly the same as Monitor::unlock() ...
1048   // directly set _owner instead of using set_owner(null)
1049   _owner = NULL;
1050   if (_snuck) {         // ???
1051     assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1052     _snuck = false;
1053     return;
1054   }
1055   IUnlock(false);
1056 }
1057 
1058 bool Monitor::wait(bool no_safepoint_check, long timeout,
1059                    bool as_suspend_equivalent) {
1060   // Make sure safepoint checking is used properly.
1061   assert(!(_safepoint_check_required == Monitor::_safepoint_check_never && no_safepoint_check == false),
1062          "This lock should never have a safepoint check: %s", name());
1063   assert(!(_safepoint_check_required == Monitor::_safepoint_check_always && no_safepoint_check == true),
1064          "This lock should always have a safepoint check: %s", name());
1065 
1066   Thread * const Self = Thread::current();
1067   assert(_owner == Self, "invariant");
1068   assert(ILocked(), "invariant");
1069 
1070   // as_suspend_equivalent logically implies !no_safepoint_check
1071   guarantee(!as_suspend_equivalent || !no_safepoint_check, "invariant");
1072   // !no_safepoint_check logically implies java_thread
1073   guarantee(no_safepoint_check || Self->is_Java_thread(), "invariant");
1074 
1075   #ifdef ASSERT
1076   Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks());
1077   assert(least != this, "Specification of get_least_... call above");
1078   if (least != NULL && least->rank() <= special) {
1079     tty->print("Attempting to wait on monitor %s/%d while holding"
1080                " lock %s/%d -- possible deadlock",
1081                name(), rank(), least->name(), least->rank());
1082     assert(false, "Shouldn't block(wait) while holding a lock of rank special");
1083   }
1084   #endif // ASSERT
1085 
1086   int wait_status;
1087   // conceptually set the owner to NULL in anticipation of
1088   // abdicating the lock in wait
1089   set_owner(NULL);
1090   if (no_safepoint_check) {
1091     wait_status = IWait(Self, timeout);
1092   } else {
1093     assert(Self->is_Java_thread(), "invariant");
1094     JavaThread *jt = (JavaThread *)Self;
1095 
1096     // Enter safepoint region - ornate and Rococo ...
1097     ThreadBlockInVM tbivm(jt);
1098     OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */);
1099 
1100     if (as_suspend_equivalent) {
1101       jt->set_suspend_equivalent();
1102       // cleared by handle_special_suspend_equivalent_condition() or
1103       // java_suspend_self()
1104     }
1105 
1106     wait_status = IWait(Self, timeout);
1107 
1108     // were we externally suspended while we were waiting?
1109     if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) {
1110       // Our event wait has finished and we own the lock, but
1111       // while we were waiting another thread suspended us. We don't
1112       // want to hold the lock while suspended because that
1113       // would surprise the thread that suspended us.
1114       assert(ILocked(), "invariant");
1115       IUnlock(true);
1116       jt->java_suspend_self();
1117       ILock(Self);
1118       assert(ILocked(), "invariant");
1119     }
1120   }
1121 
1122   // Conceptually reestablish ownership of the lock.
1123   // The "real" lock -- the LockByte -- was reacquired by IWait().
1124   assert(ILocked(), "invariant");
1125   assert(_owner == NULL, "invariant");
1126   set_owner(Self);
1127   return wait_status != 0;          // return true IFF timeout
1128 }
1129 
1130 Monitor::~Monitor() {
1131 #ifdef ASSERT
1132   uintptr_t owner = UNS(_owner);
1133   uintptr_t lockword = UNS(_LockWord.FullWord);
1134   uintptr_t entrylist = UNS(_EntryList);
1135   uintptr_t waitset = UNS(_WaitSet);
1136   uintptr_t ondeck = UNS(_OnDeck);
1137   // Print _name with precision limit, in case failure is due to memory
1138   // corruption that also trashed _name.
1139   assert((owner|lockword|entrylist|waitset|ondeck) == 0,
1140          "%.*s: _owner(" INTPTR_FORMAT ")|_LockWord(" INTPTR_FORMAT ")|_EntryList(" INTPTR_FORMAT ")|_WaitSet("
1141          INTPTR_FORMAT ")|_OnDeck(" INTPTR_FORMAT ") != 0",
1142          MONITOR_NAME_LEN, _name, owner, lockword, entrylist, waitset, ondeck);
1143 #endif
1144 }
1145 
1146 void Monitor::ClearMonitor(Monitor * m, const char *name) {
1147   m->_owner             = NULL;
1148   m->_snuck             = false;
1149   if (name == NULL) {
1150     strcpy(m->_name, "UNKNOWN");
1151   } else {
1152     strncpy(m->_name, name, MONITOR_NAME_LEN - 1);
1153     m->_name[MONITOR_NAME_LEN - 1] = '\0';
1154   }
1155   m->_LockWord.FullWord = 0;
1156   m->_EntryList         = NULL;
1157   m->_OnDeck            = NULL;
1158   m->_WaitSet           = NULL;
1159   m->_WaitLock[0]       = 0;
1160 }
1161 
1162 Monitor::Monitor() { ClearMonitor(this); }
1163 
1164 Monitor::Monitor(int Rank, const char * name, bool allow_vm_block,
1165                  SafepointCheckRequired safepoint_check_required) {
1166   ClearMonitor(this, name);
1167 #ifdef ASSERT
1168   _allow_vm_block  = allow_vm_block;
1169   _rank            = Rank;
1170   NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;)
1171 #endif
1172 }
1173 
1174 Mutex::Mutex(int Rank, const char * name, bool allow_vm_block,
1175              SafepointCheckRequired safepoint_check_required) {
1176   ClearMonitor((Monitor *) this, name);
1177 #ifdef ASSERT
1178   _allow_vm_block   = allow_vm_block;
1179   _rank             = Rank;
1180   NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;)
1181 #endif
1182 }
1183 
1184 bool Monitor::owned_by_self() const {
1185   bool ret = _owner == Thread::current();
1186   assert(!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant");
1187   return ret;
1188 }
1189 
1190 void Monitor::print_on_error(outputStream* st) const {
1191   st->print("[" PTR_FORMAT, p2i(this));
1192   st->print("] %s", _name);
1193   st->print(" - owner thread: " PTR_FORMAT, p2i(_owner));
1194 }
1195 
1196 
1197 
1198 
1199 // ----------------------------------------------------------------------------------
1200 // Non-product code
1201 
1202 #ifndef PRODUCT
1203 void Monitor::print_on(outputStream* st) const {
1204   st->print_cr("Mutex: [" PTR_FORMAT "/" PTR_FORMAT "] %s - owner: " PTR_FORMAT,
1205                p2i(this), _LockWord.FullWord, _name, p2i(_owner));
1206 }
1207 #endif
1208 
1209 #ifndef PRODUCT
1210 #ifdef ASSERT
1211 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) {
1212   Monitor *res, *tmp;
1213   for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) {
1214     if (tmp->rank() < res->rank()) {
1215       res = tmp;
1216     }
1217   }
1218   if (!SafepointSynchronize::is_at_safepoint()) {
1219     // In this case, we expect the held locks to be
1220     // in increasing rank order (modulo any native ranks)
1221     for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1222       if (tmp->next() != NULL) {
1223         assert(tmp->rank() == Mutex::native ||
1224                tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1225       }
1226     }
1227   }
1228   return res;
1229 }
1230 
1231 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) {
1232   Monitor *res, *tmp;
1233   for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) {
1234     if (tmp != this && (res == NULL || tmp->rank() < res->rank())) {
1235       res = tmp;
1236     }
1237   }
1238   if (!SafepointSynchronize::is_at_safepoint()) {
1239     // In this case, we expect the held locks to be
1240     // in increasing rank order (modulo any native ranks)
1241     for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1242       if (tmp->next() != NULL) {
1243         assert(tmp->rank() == Mutex::native ||
1244                tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1245       }
1246     }
1247   }
1248   return res;
1249 }
1250 
1251 
1252 bool Monitor::contains(Monitor* locks, Monitor * lock) {
1253   for (; locks != NULL; locks = locks->next()) {
1254     if (locks == lock) {
1255       return true;
1256     }
1257   }
1258   return false;
1259 }
1260 #endif
1261 
1262 // Called immediately after lock acquisition or release as a diagnostic
1263 // to track the lock-set of the thread and test for rank violations that
1264 // might indicate exposure to deadlock.
1265 // Rather like an EventListener for _owner (:>).
1266 
1267 void Monitor::set_owner_implementation(Thread *new_owner) {
1268   // This function is solely responsible for maintaining
1269   // and checking the invariant that threads and locks
1270   // are in a 1/N relation, with some some locks unowned.
1271   // It uses the Mutex::_owner, Mutex::_next, and
1272   // Thread::_owned_locks fields, and no other function
1273   // changes those fields.
1274   // It is illegal to set the mutex from one non-NULL
1275   // owner to another--it must be owned by NULL as an
1276   // intermediate state.
1277 
1278   if (new_owner != NULL) {
1279     // the thread is acquiring this lock
1280 
1281     assert(new_owner == Thread::current(), "Should I be doing this?");
1282     assert(_owner == NULL, "setting the owner thread of an already owned mutex");
1283     _owner = new_owner; // set the owner
1284 
1285     // link "this" into the owned locks list
1286 
1287 #ifdef ASSERT  // Thread::_owned_locks is under the same ifdef
1288     Monitor* locks = get_least_ranked_lock(new_owner->owned_locks());
1289     // Mutex::set_owner_implementation is a friend of Thread
1290 
1291     assert(this->rank() >= 0, "bad lock rank");
1292 
1293     // Deadlock avoidance rules require us to acquire Mutexes only in
1294     // a global total order. For example m1 is the lowest ranked mutex
1295     // that the thread holds and m2 is the mutex the thread is trying
1296     // to acquire, then  deadlock avoidance rules require that the rank
1297     // of m2 be less  than the rank of m1.
1298     // The rank Mutex::native  is an exception in that it is not subject
1299     // to the verification rules.
1300     // Here are some further notes relating to mutex acquisition anomalies:
1301     // . it is also ok to acquire Safepoint_lock at the very end while we
1302     //   already hold Terminator_lock - may happen because of periodic safepoints
1303     if (this->rank() != Mutex::native &&
1304         this->rank() != Mutex::suspend_resume &&
1305         locks != NULL && locks->rank() <= this->rank() &&
1306         !SafepointSynchronize::is_at_safepoint() &&
1307         !(this == Safepoint_lock && contains(locks, Terminator_lock) &&
1308         SafepointSynchronize::is_synchronizing())) {
1309       new_owner->print_owned_locks();
1310       fatal("acquiring lock %s/%d out of order with lock %s/%d -- "
1311             "possible deadlock", this->name(), this->rank(),
1312             locks->name(), locks->rank());
1313     }
1314 
1315     this->_next = new_owner->_owned_locks;
1316     new_owner->_owned_locks = this;
1317 #endif
1318 
1319   } else {
1320     // the thread is releasing this lock
1321 
1322     Thread* old_owner = _owner;
1323     debug_only(_last_owner = old_owner);
1324 
1325     assert(old_owner != NULL, "removing the owner thread of an unowned mutex");
1326     assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex");
1327 
1328     _owner = NULL; // set the owner
1329 
1330 #ifdef ASSERT
1331     Monitor *locks = old_owner->owned_locks();
1332 
1333     // remove "this" from the owned locks list
1334 
1335     Monitor *prev = NULL;
1336     bool found = false;
1337     for (; locks != NULL; prev = locks, locks = locks->next()) {
1338       if (locks == this) {
1339         found = true;
1340         break;
1341       }
1342     }
1343     assert(found, "Removing a lock not owned");
1344     if (prev == NULL) {
1345       old_owner->_owned_locks = _next;
1346     } else {
1347       prev->_next = _next;
1348     }
1349     _next = NULL;
1350 #endif
1351   }
1352 }
1353 
1354 
1355 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock()
1356 void Monitor::check_prelock_state(Thread *thread, bool safepoint_check) {
1357   if (safepoint_check) {
1358     assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm)
1359            || rank() == Mutex::special, "wrong thread state for using locks");
1360     if (thread->is_VM_thread() && !allow_vm_block()) {
1361       fatal("VM thread using lock %s (not allowed to block on)", name());
1362     }
1363     debug_only(if (rank() != Mutex::special) \
1364                thread->check_for_valid_safepoint_state(false);)
1365   }
1366   assert(!os::ThreadCrashProtection::is_crash_protected(thread),
1367          "locking not allowed when crash protection is set");
1368 }
1369 
1370 void Monitor::check_block_state(Thread *thread) {
1371   if (!_allow_vm_block && thread->is_VM_thread()) {
1372     warning("VM thread blocked on lock");
1373     print();
1374     BREAKPOINT;
1375   }
1376   assert(_owner != thread, "deadlock: blocking on monitor owned by current thread");
1377 }
1378 
1379 #endif // PRODUCT