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 "runtime/atomic.hpp"
  27 #include "runtime/interfaceSupport.hpp"
  28 #include "runtime/mutex.hpp"
  29 #include "runtime/orderAccess.inline.hpp"
  30 #include "runtime/osThread.hpp"
  31 #include "runtime/thread.inline.hpp"
  32 #include "utilities/events.hpp"
  33 #include "utilities/macros.hpp"
  34 
  35 // 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
  36 //
  37 // Native Monitor-Mutex locking - theory of operations
  38 //
  39 // * Native Monitors are completely unrelated to Java-level monitors,
  40 //   although the "back-end" slow-path implementations share a common lineage.
  41 //   See objectMonitor:: in synchronizer.cpp.
  42 //   Native Monitors do *not* support nesting or recursion but otherwise
  43 //   they're basically Hoare-flavor monitors.
  44 //
  45 // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte
  46 //   in the _LockWord from zero to non-zero.  Note that the _Owner field
  47 //   is advisory and is used only to verify that the thread calling unlock()
  48 //   is indeed the last thread to have acquired the lock.
  49 //
  50 // * Contending threads "push" themselves onto the front of the contention
  51 //   queue -- called the cxq -- with CAS and then spin/park.
  52 //   The _LockWord contains the LockByte as well as the pointer to the head
  53 //   of the cxq.  Colocating the LockByte with the cxq precludes certain races.
  54 //
  55 // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0
  56 //   idioms.  We currently use MEMBAR in the uncontended unlock() path, as
  57 //   MEMBAR often has less latency than CAS.  If warranted, we could switch to
  58 //   a CAS:0 mode, using timers to close the resultant race, as is done
  59 //   with Java Monitors in synchronizer.cpp.
  60 //
  61 //   See the following for a discussion of the relative cost of atomics (CAS)
  62 //   MEMBAR, and ways to eliminate such instructions from the common-case paths:
  63 //   -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot
  64 //   -- http://blogs.sun.com/dave/resource/MustangSync.pdf
  65 //   -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf
  66 //   -- synchronizer.cpp
  67 //
  68 // * Overall goals - desiderata
  69 //   1. Minimize context switching
  70 //   2. Minimize lock migration
  71 //   3. Minimize CPI -- affinity and locality
  72 //   4. Minimize the execution of high-latency instructions such as CAS or MEMBAR
  73 //   5. Minimize outer lock hold times
  74 //   6. Behave gracefully on a loaded system
  75 //
  76 // * Thread flow and list residency:
  77 //
  78 //   Contention queue --> EntryList --> OnDeck --> Owner --> !Owner
  79 //   [..resident on monitor list..]
  80 //   [...........contending..................]
  81 //
  82 //   -- The contention queue (cxq) contains recently-arrived threads (RATs).
  83 //      Threads on the cxq eventually drain into the EntryList.
  84 //   -- Invariant: a thread appears on at most one list -- cxq, EntryList
  85 //      or WaitSet -- at any one time.
  86 //   -- For a given monitor there can be at most one "OnDeck" thread at any
  87 //      given time but if needbe this particular invariant could be relaxed.
  88 //
  89 // * The WaitSet and EntryList linked lists are composed of ParkEvents.
  90 //   I use ParkEvent instead of threads as ParkEvents are immortal and
  91 //   type-stable, meaning we can safely unpark() a possibly stale
  92 //   list element in the unlock()-path.  (That's benign).
  93 //
  94 // * Succession policy - providing for progress:
  95 //
  96 //   As necessary, the unlock()ing thread identifies, unlinks, and unparks
  97 //   an "heir presumptive" tentative successor thread from the EntryList.
  98 //   This becomes the so-called "OnDeck" thread, of which there can be only
  99 //   one at any given time for a given monitor.  The wakee will recontend
 100 //   for ownership of monitor.
 101 //
 102 //   Succession is provided for by a policy of competitive handoff.
 103 //   The exiting thread does _not_ grant or pass ownership to the
 104 //   successor thread.  (This is also referred to as "handoff" succession").
 105 //   Instead the exiting thread releases ownership and possibly wakes
 106 //   a successor, so the successor can (re)compete for ownership of the lock.
 107 //
 108 //   Competitive handoff provides excellent overall throughput at the expense
 109 //   of short-term fairness.  If fairness is a concern then one remedy might
 110 //   be to add an AcquireCounter field to the monitor.  After a thread acquires
 111 //   the lock it will decrement the AcquireCounter field.  When the count
 112 //   reaches 0 the thread would reset the AcquireCounter variable, abdicate
 113 //   the lock directly to some thread on the EntryList, and then move itself to the
 114 //   tail of the EntryList.
 115 //
 116 //   But in practice most threads engage or otherwise participate in resource
 117 //   bounded producer-consumer relationships, so lock domination is not usually
 118 //   a practical concern.  Recall too, that in general it's easier to construct
 119 //   a fair lock from a fast lock, but not vice-versa.
 120 //
 121 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 122 //   detaching thread.  This mechanism is immune from the ABA corruption.
 123 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 124 //   We use OnDeck as a pseudo-lock to enforce the at-most-one detaching
 125 //   thread constraint.
 126 //
 127 // * Taken together, the cxq and the EntryList constitute or form a
 128 //   single logical queue of threads stalled trying to acquire the lock.
 129 //   We use two distinct lists to reduce heat on the list ends.
 130 //   Threads in lock() enqueue onto cxq while threads in unlock() will
 131 //   dequeue from the EntryList.  (c.f. Michael Scott's "2Q" algorithm).
 132 //   A key desideratum is to minimize queue & monitor metadata manipulation
 133 //   that occurs while holding the "outer" monitor lock -- that is, we want to
 134 //   minimize monitor lock holds times.
 135 //
 136 //   The EntryList is ordered by the prevailing queue discipline and
 137 //   can be organized in any convenient fashion, such as a doubly-linked list or
 138 //   a circular doubly-linked list.  If we need a priority queue then something akin
 139 //   to Solaris' sleepq would work nicely.  Viz.,
 140 //   -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 141 //   -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c
 142 //   Queue discipline is enforced at ::unlock() time, when the unlocking thread
 143 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 144 //   EntryList accordingly.
 145 //
 146 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 147 //   somewhat similar to an elevator-scan.
 148 //
 149 // * OnDeck
 150 //   --  For a given monitor there can be at most one OnDeck thread at any given
 151 //       instant.  The OnDeck thread is contending for the lock, but has been
 152 //       unlinked from the EntryList and cxq by some previous unlock() operations.
 153 //       Once a thread has been designated the OnDeck thread it will remain so
 154 //       until it manages to acquire the lock -- being OnDeck is a stable property.
 155 //   --  Threads on the EntryList or cxq are _not allowed to attempt lock acquisition.
 156 //   --  OnDeck also serves as an "inner lock" as follows.  Threads in unlock() will, after
 157 //       having cleared the LockByte and dropped the outer lock,  attempt to "trylock"
 158 //       OnDeck by CASing the field from null to non-null.  If successful, that thread
 159 //       is then responsible for progress and succession and can use CAS to detach and
 160 //       drain the cxq into the EntryList.  By convention, only this thread, the holder of
 161 //       the OnDeck inner lock, can manipulate the EntryList or detach and drain the
 162 //       RATs on the cxq into the EntryList.  This avoids ABA corruption on the cxq as
 163 //       we allow multiple concurrent "push" operations but restrict detach concurrency
 164 //       to at most one thread.  Having selected and detached a successor, the thread then
 165 //       changes the OnDeck to refer to that successor, and then unparks the successor.
 166 //       That successor will eventually acquire the lock and clear OnDeck.  Beware
 167 //       that the OnDeck usage as a lock is asymmetric.  A thread in unlock() transiently
 168 //       "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor,
 169 //       and then the successor eventually "drops" OnDeck.  Note that there's never
 170 //       any sense of contention on the inner lock, however.  Threads never contend
 171 //       or wait for the inner lock.
 172 //   --  OnDeck provides for futile wakeup throttling a described in section 3.3 of
 173 //       See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 174 //       In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter
 175 //       TState fields found in Java-level objectMonitors.  (See synchronizer.cpp).
 176 //
 177 // * Waiting threads reside on the WaitSet list -- wait() puts
 178 //   the caller onto the WaitSet.  Notify() or notifyAll() simply
 179 //   transfers threads from the WaitSet to either the EntryList or cxq.
 180 //   Subsequent unlock() operations will eventually unpark the notifyee.
 181 //   Unparking a notifee in notify() proper is inefficient - if we were to do so
 182 //   it's likely the notifyee would simply impale itself on the lock held
 183 //   by the notifier.
 184 //
 185 // * The mechanism is obstruction-free in that if the holder of the transient
 186 //   OnDeck lock in unlock() is preempted or otherwise stalls, other threads
 187 //   can still acquire and release the outer lock and continue to make progress.
 188 //   At worst, waking of already blocked contending threads may be delayed,
 189 //   but nothing worse.  (We only use "trylock" operations on the inner OnDeck
 190 //   lock).
 191 //
 192 // * Note that thread-local storage must be initialized before a thread
 193 //   uses Native monitors or mutexes.  The native monitor-mutex subsystem
 194 //   depends on Thread::current().
 195 //
 196 // * The monitor synchronization subsystem avoids the use of native
 197 //   synchronization primitives except for the narrow platform-specific
 198 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 199 //   the semantics of park-unpark.  Put another way, this monitor implementation
 200 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 201 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 202 //   underlying OS manages the READY<->RUN transitions.
 203 //
 204 // * The memory consistency model provide by lock()-unlock() is at least as
 205 //   strong or stronger than the Java Memory model defined by JSR-133.
 206 //   That is, we guarantee at least entry consistency, if not stronger.
 207 //   See http://g.oswego.edu/dl/jmm/cookbook.html.
 208 //
 209 // * Thread:: currently contains a set of purpose-specific ParkEvents:
 210 //   _MutexEvent, _ParkEvent, etc.  A better approach might be to do away with
 211 //   the purpose-specific ParkEvents and instead implement a general per-thread
 212 //   stack of available ParkEvents which we could provision on-demand.  The
 213 //   stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate()
 214 //   and ::Release().  A thread would simply pop an element from the local stack before it
 215 //   enqueued or park()ed.  When the contention was over the thread would
 216 //   push the no-longer-needed ParkEvent back onto its stack.
 217 //
 218 // * A slightly reduced form of ILock() and IUnlock() have been partially
 219 //   model-checked (Murphi) for safety and progress at T=1,2,3 and 4.
 220 //   It'd be interesting to see if TLA/TLC could be useful as well.
 221 //
 222 // * Mutex-Monitor is a low-level "leaf" subsystem.  That is, the monitor
 223 //   code should never call other code in the JVM that might itself need to
 224 //   acquire monitors or mutexes.  That's true *except* in the case of the
 225 //   ThreadBlockInVM state transition wrappers.  The ThreadBlockInVM DTOR handles
 226 //   mutator reentry (ingress) by checking for a pending safepoint in which case it will
 227 //   call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc.
 228 //   In that particular case a call to lock() for a given Monitor can end up recursively
 229 //   calling lock() on another monitor.   While distasteful, this is largely benign
 230 //   as the calls come from jacket that wraps lock(), and not from deep within lock() itself.
 231 //
 232 //   It's unfortunate that native mutexes and thread state transitions were convolved.
 233 //   They're really separate concerns and should have remained that way.  Melding
 234 //   them together was facile -- a bit too facile.   The current implementation badly
 235 //   conflates the two concerns.
 236 //
 237 // * TODO-FIXME:
 238 //
 239 //   -- Add DTRACE probes for contended acquire, contended acquired, contended unlock
 240 //      We should also add DTRACE probes in the ParkEvent subsystem for
 241 //      Park-entry, Park-exit, and Unpark.
 242 //
 243 //   -- We have an excess of mutex-like constructs in the JVM, namely:
 244 //      1. objectMonitors for Java-level synchronization (synchronizer.cpp)
 245 //      2. low-level muxAcquire and muxRelease
 246 //      3. low-level spinAcquire and spinRelease
 247 //      4. native Mutex:: and Monitor::
 248 //      5. jvm_raw_lock() and _unlock()
 249 //      6. JVMTI raw monitors -- distinct from (5) despite having a confusingly
 250 //         similar name.
 251 //
 252 // 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
 253 
 254 #define UNS(x) (uintptr_t(x))
 255 #define TRACE(m)                   \
 256   {                                \
 257     static volatile int ctr = 0;   \
 258     int x = ++ctr;                 \
 259     if ((x & (x - 1)) == 0) {      \
 260       ::printf("%d:%s\n", x, #m);  \
 261       ::fflush(stdout);            \
 262     }                              \
 263   }
 264 
 265 // Simplistic low-quality Marsaglia SHIFT-XOR RNG.
 266 // Bijective except for the trailing mask operation.
 267 // Useful for spin loops as the compiler can't optimize it away.
 268 
 269 static inline jint MarsagliaXORV(jint x) {
 270   if (x == 0) x = 1|os::random();
 271   x ^= x << 6;
 272   x ^= ((unsigned)x) >> 21;
 273   x ^= x << 7;
 274   return x & 0x7FFFFFFF;
 275 }
 276 
 277 static int Stall(int its) {
 278   static volatile jint rv = 1;
 279   volatile int OnFrame = 0;
 280   jint v = rv ^ UNS(OnFrame);
 281   while (--its >= 0) {
 282     v = MarsagliaXORV(v);
 283   }
 284   // Make this impossible for the compiler to optimize away,
 285   // but (mostly) avoid W coherency sharing on MP systems.
 286   if (v == 0x12345) rv = v;
 287   return v;
 288 }
 289 
 290 int Monitor::TryLock() {
 291   intptr_t v = _LockWord.FullWord;
 292   for (;;) {
 293     if ((v & _LBIT) != 0) return 0;
 294     const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
 295     if (v == u) return 1;
 296     v = u;
 297   }
 298 }
 299 
 300 int Monitor::TryFast() {
 301   // Optimistic fast-path form ...
 302   // Fast-path attempt for the common uncontended case.
 303   // Avoid RTS->RTO $ coherence upgrade on typical SMP systems.
 304   intptr_t v = Atomic::cmpxchg((intptr_t)_LBIT, &_LockWord.FullWord, (intptr_t)0);  // agro ...
 305   if (v == 0) return 1;
 306 
 307   for (;;) {
 308     if ((v & _LBIT) != 0) return 0;
 309     const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
 310     if (v == u) return 1;
 311     v = u;
 312   }
 313 }
 314 
 315 int Monitor::ILocked() {
 316   const intptr_t w = _LockWord.FullWord & 0xFF;
 317   assert(w == 0 || w == _LBIT, "invariant");
 318   return w == _LBIT;
 319 }
 320 
 321 // Polite TATAS spinlock with exponential backoff - bounded spin.
 322 // Ideally we'd use processor cycles, time or vtime to control
 323 // the loop, but we currently use iterations.
 324 // All the constants within were derived empirically but work over
 325 // over the spectrum of J2SE reference platforms.
 326 // On Niagara-class systems the back-off is unnecessary but
 327 // is relatively harmless.  (At worst it'll slightly retard
 328 // acquisition times).  The back-off is critical for older SMP systems
 329 // where constant fetching of the LockWord would otherwise impair
 330 // scalability.
 331 //
 332 // Clamp spinning at approximately 1/2 of a context-switch round-trip.
 333 // See synchronizer.cpp for details and rationale.
 334 
 335 int Monitor::TrySpin(Thread * const Self) {
 336   if (TryLock())    return 1;
 337   if (!os::is_MP()) return 0;
 338 
 339   int Probes  = 0;
 340   int Delay   = 0;
 341   int Steps   = 0;
 342   int SpinMax = NativeMonitorSpinLimit;
 343   int flgs    = NativeMonitorFlags;
 344   for (;;) {
 345     intptr_t v = _LockWord.FullWord;
 346     if ((v & _LBIT) == 0) {
 347       if (Atomic::cmpxchg (v|_LBIT, &_LockWord.FullWord, v) == v) {
 348         return 1;
 349       }
 350       continue;
 351     }
 352 
 353     if ((flgs & 8) == 0) {
 354       SpinPause();
 355     }
 356 
 357     // Periodically increase Delay -- variable Delay form
 358     // conceptually: delay *= 1 + 1/Exponent
 359     ++Probes;
 360     if (Probes > SpinMax) return 0;
 361 
 362     if ((Probes & 0x7) == 0) {
 363       Delay = ((Delay << 1)|1) & 0x7FF;
 364       // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ;
 365     }
 366 
 367     if (flgs & 2) continue;
 368 
 369     // Consider checking _owner's schedctl state, if OFFPROC abort spin.
 370     // If the owner is OFFPROC then it's unlike that the lock will be dropped
 371     // in a timely fashion, which suggests that spinning would not be fruitful
 372     // or profitable.
 373 
 374     // Stall for "Delay" time units - iterations in the current implementation.
 375     // Avoid generating coherency traffic while stalled.
 376     // Possible ways to delay:
 377     //   PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt,
 378     //   wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ...
 379     // Note that on Niagara-class systems we want to minimize STs in the
 380     // spin loop.  N1 and brethren write-around the L1$ over the xbar into the L2$.
 381     // Furthermore, they don't have a W$ like traditional SPARC processors.
 382     // We currently use a Marsaglia Shift-Xor RNG loop.
 383     Steps += Delay;
 384     if (Self != NULL) {
 385       jint rv = Self->rng[0];
 386       for (int k = Delay; --k >= 0;) {
 387         rv = MarsagliaXORV(rv);
 388         if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0;
 389       }
 390       Self->rng[0] = rv;
 391     } else {
 392       Stall(Delay);
 393     }
 394   }
 395 }
 396 
 397 static int ParkCommon(ParkEvent * ev, jlong timo) {
 398   // Diagnostic support - periodically unwedge blocked threads
 399   intx nmt = NativeMonitorTimeout;
 400   if (nmt > 0 && (nmt < timo || timo <= 0)) {
 401     timo = nmt;
 402   }
 403   int err = OS_OK;
 404   if (0 == timo) {
 405     ev->park();
 406   } else {
 407     err = ev->park(timo);
 408   }
 409   return err;
 410 }
 411 
 412 inline int Monitor::AcquireOrPush(ParkEvent * ESelf) {
 413   intptr_t v = _LockWord.FullWord;
 414   for (;;) {
 415     if ((v & _LBIT) == 0) {
 416       const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
 417       if (u == v) return 1;        // indicate acquired
 418       v = u;
 419     } else {
 420       // Anticipate success ...
 421       ESelf->ListNext = (ParkEvent *)(v & ~_LBIT);
 422       const intptr_t u = Atomic::cmpxchg(intptr_t(ESelf)|_LBIT, &_LockWord.FullWord, v);
 423       if (u == v) return 0;        // indicate pushed onto cxq
 424       v = u;
 425     }
 426     // Interference - LockWord change - just retry
 427   }
 428 }
 429 
 430 // ILock and IWait are the lowest level primitive internal blocking
 431 // synchronization functions.  The callers of IWait and ILock must have
 432 // performed any needed state transitions beforehand.
 433 // IWait and ILock may directly call park() without any concern for thread state.
 434 // Note that ILock and IWait do *not* access _owner.
 435 // _owner is a higher-level logical concept.
 436 
 437 void Monitor::ILock(Thread * Self) {
 438   assert(_OnDeck != Self->_MutexEvent, "invariant");
 439 
 440   if (TryFast()) {
 441  Exeunt:
 442     assert(ILocked(), "invariant");
 443     return;
 444   }
 445 
 446   ParkEvent * const ESelf = Self->_MutexEvent;
 447   assert(_OnDeck != ESelf, "invariant");
 448 
 449   // As an optimization, spinners could conditionally try to set _OnDeck to _LBIT
 450   // Synchronizer.cpp uses a similar optimization.
 451   if (TrySpin(Self)) goto Exeunt;
 452 
 453   // Slow-path - the lock is contended.
 454   // Either Enqueue Self on cxq or acquire the outer lock.
 455   // LockWord encoding = (cxq,LOCKBYTE)
 456   ESelf->reset();
 457   OrderAccess::fence();
 458 
 459   // Optional optimization ... try barging on the inner lock
 460   if ((NativeMonitorFlags & 32) && Atomic::cmpxchg_if_null(ESelf, &_OnDeck)) {
 461     goto OnDeck_LOOP;
 462   }
 463 
 464   if (AcquireOrPush(ESelf)) goto Exeunt;
 465 
 466   // At any given time there is at most one ondeck thread.
 467   // ondeck implies not resident on cxq and not resident on EntryList
 468   // Only the OnDeck thread can try to acquire -- contend for -- the lock.
 469   // CONSIDER: use Self->OnDeck instead of m->OnDeck.
 470   // Deschedule Self so that others may run.
 471   while (OrderAccess::load_acquire(&_OnDeck) != ESelf) {
 472     ParkCommon(ESelf, 0);
 473   }
 474 
 475   // Self is now in the OnDeck position and will remain so until it
 476   // manages to acquire the lock.
 477  OnDeck_LOOP:
 478   for (;;) {
 479     assert(_OnDeck == ESelf, "invariant");
 480     if (TrySpin(Self)) break;
 481     // It's probably wise to spin only if we *actually* blocked
 482     // CONSIDER: check the lockbyte, if it remains set then
 483     // preemptively drain the cxq into the EntryList.
 484     // The best place and time to perform queue operations -- lock metadata --
 485     // is _before having acquired the outer lock, while waiting for the lock to drop.
 486     ParkCommon(ESelf, 0);
 487   }
 488 
 489   assert(_OnDeck == ESelf, "invariant");
 490   _OnDeck = NULL;
 491 
 492   // Note that we current drop the inner lock (clear OnDeck) in the slow-path
 493   // epilogue immediately after having acquired the outer lock.
 494   // But instead we could consider the following optimizations:
 495   // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation.
 496   //    This might avoid potential reacquisition of the inner lock in IUlock().
 497   // B. While still holding the inner lock, attempt to opportunistically select
 498   //    and unlink the next OnDeck thread from the EntryList.
 499   //    If successful, set OnDeck to refer to that thread, otherwise clear OnDeck.
 500   //    It's critical that the select-and-unlink operation run in constant-time as
 501   //    it executes when holding the outer lock and may artificially increase the
 502   //    effective length of the critical section.
 503   // Note that (A) and (B) are tantamount to succession by direct handoff for
 504   // the inner lock.
 505   goto Exeunt;
 506 }
 507 
 508 void Monitor::IUnlock(bool RelaxAssert) {
 509   assert(ILocked(), "invariant");
 510   // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately
 511   // before the store that releases the lock.  Crucially, all the stores and loads in the
 512   // critical section must be globally visible before the store of 0 into the lock-word
 513   // that releases the lock becomes globally visible.  That is, memory accesses in the
 514   // critical section should not be allowed to bypass or overtake the following ST that
 515   // releases the lock.  As such, to prevent accesses within the critical section
 516   // from "leaking" out, we need a release fence between the critical section and the
 517   // store that releases the lock.  In practice that release barrier is elided on
 518   // platforms with strong memory models such as TSO.
 519   //
 520   // Note that the OrderAccess::storeload() fence that appears after unlock store
 521   // provides for progress conditions and succession and is _not related to exclusion
 522   // safety or lock release consistency.
 523   OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], jbyte(0)); // drop outer lock
 524 
 525   OrderAccess::storeload();
 526   ParkEvent * const w = _OnDeck; // raw load as we will just return if non-NULL
 527   assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant");
 528   if (w != NULL) {
 529     // Either we have a valid ondeck thread or ondeck is transiently "locked"
 530     // by some exiting thread as it arranges for succession.  The LSBit of
 531     // OnDeck allows us to discriminate two cases.  If the latter, the
 532     // responsibility for progress and succession lies with that other thread.
 533     // For good performance, we also depend on the fact that redundant unpark()
 534     // operations are cheap.  That is, repeated Unpark()ing of the OnDeck thread
 535     // is inexpensive.  This approach provides implicit futile wakeup throttling.
 536     // Note that the referent "w" might be stale with respect to the lock.
 537     // In that case the following unpark() is harmless and the worst that'll happen
 538     // is a spurious return from a park() operation.  Critically, if "w" _is stale,
 539     // then progress is known to have occurred as that means the thread associated
 540     // with "w" acquired the lock.  In that case this thread need take no further
 541     // action to guarantee progress.
 542     if ((UNS(w) & _LBIT) == 0) w->unpark();
 543     return;
 544   }
 545 
 546   intptr_t cxq = _LockWord.FullWord;
 547   if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) {
 548     return;      // normal fast-path exit - cxq and EntryList both empty
 549   }
 550   if (cxq & _LBIT) {
 551     // Optional optimization ...
 552     // Some other thread acquired the lock in the window since this
 553     // thread released it.  Succession is now that thread's responsibility.
 554     return;
 555   }
 556 
 557  Succession:
 558   // Slow-path exit - this thread must ensure succession and progress.
 559   // OnDeck serves as lock to protect cxq and EntryList.
 560   // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq.
 561   // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS)
 562   // but only one concurrent consumer (detacher of RATs).
 563   // Consider protecting this critical section with schedctl on Solaris.
 564   // Unlike a normal lock, however, the exiting thread "locks" OnDeck,
 565   // picks a successor and marks that thread as OnDeck.  That successor
 566   // thread will then clear OnDeck once it eventually acquires the outer lock.
 567   if (!Atomic::cmpxchg_if_null((ParkEvent*)_LBIT, &_OnDeck)) {
 568     return;
 569   }
 570 
 571   ParkEvent * List = _EntryList;
 572   if (List != NULL) {
 573     // Transfer the head of the EntryList to the OnDeck position.
 574     // Once OnDeck, a thread stays OnDeck until it acquires the lock.
 575     // For a given lock there is at most OnDeck thread at any one instant.
 576    WakeOne:
 577     assert(List == _EntryList, "invariant");
 578     ParkEvent * const w = List;
 579     assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant");
 580     _EntryList = w->ListNext;
 581     // as a diagnostic measure consider setting w->_ListNext = BAD
 582     assert(UNS(_OnDeck) == _LBIT, "invariant");
 583 
 584     // Pass OnDeck role to w, ensuring that _EntryList has been set first.
 585     // w will clear _OnDeck once it acquires the outer lock.
 586     // Note that once we set _OnDeck that thread can acquire the mutex, proceed
 587     // with its critical section and then enter this code to unlock the mutex. So
 588     // you can have multiple threads active in IUnlock at the same time.
 589     OrderAccess::release_store(&_OnDeck, w);
 590 
 591     // Another optional optimization ...
 592     // For heavily contended locks it's not uncommon that some other
 593     // thread acquired the lock while this thread was arranging succession.
 594     // Try to defer the unpark() operation - Delegate the responsibility
 595     // for unpark()ing the OnDeck thread to the current or subsequent owners
 596     // That is, the new owner is responsible for unparking the OnDeck thread.
 597     OrderAccess::storeload();
 598     cxq = _LockWord.FullWord;
 599     if (cxq & _LBIT) return;
 600 
 601     w->unpark();
 602     return;
 603   }
 604 
 605   cxq = _LockWord.FullWord;
 606   if ((cxq & ~_LBIT) != 0) {
 607     // The EntryList is empty but the cxq is populated.
 608     // drain RATs from cxq into EntryList
 609     // Detach RATs segment with CAS and then merge into EntryList
 610     for (;;) {
 611       // optional optimization - if locked, the owner is responsible for succession
 612       if (cxq & _LBIT) goto Punt;
 613       const intptr_t vfy = Atomic::cmpxchg(cxq & _LBIT, &_LockWord.FullWord, cxq);
 614       if (vfy == cxq) break;
 615       cxq = vfy;
 616       // Interference - LockWord changed - Just retry
 617       // We can see concurrent interference from contending threads
 618       // pushing themselves onto the cxq or from lock-unlock operations.
 619       // From the perspective of this thread, EntryList is stable and
 620       // the cxq is prepend-only -- the head is volatile but the interior
 621       // of the cxq is stable.  In theory if we encounter interference from threads
 622       // pushing onto cxq we could simply break off the original cxq suffix and
 623       // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts
 624       // on the high-traffic LockWord variable.   For instance lets say the cxq is "ABCD"
 625       // when we first fetch cxq above.  Between the fetch -- where we observed "A"
 626       // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive,
 627       // yielding cxq = "PQRABCD".  In this case we could simply set A.ListNext
 628       // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList.
 629       // Note too, that it's safe for this thread to traverse the cxq
 630       // without taking any special concurrency precautions.
 631     }
 632 
 633     // We don't currently reorder the cxq segment as we move it onto
 634     // the EntryList, but it might make sense to reverse the order
 635     // or perhaps sort by thread priority.  See the comments in
 636     // synchronizer.cpp objectMonitor::exit().
 637     assert(_EntryList == NULL, "invariant");
 638     _EntryList = List = (ParkEvent *)(cxq & ~_LBIT);
 639     assert(List != NULL, "invariant");
 640     goto WakeOne;
 641   }
 642 
 643   // cxq|EntryList is empty.
 644   // w == NULL implies that cxq|EntryList == NULL in the past.
 645   // Possible race - rare inopportune interleaving.
 646   // A thread could have added itself to cxq since this thread previously checked.
 647   // Detect and recover by refetching cxq.
 648  Punt:
 649   assert(UNS(_OnDeck) == _LBIT, "invariant");
 650   _OnDeck = NULL;            // Release inner lock.
 651   OrderAccess::storeload();   // Dekker duality - pivot point
 652 
 653   // Resample LockWord/cxq to recover from possible race.
 654   // For instance, while this thread T1 held OnDeck, some other thread T2 might
 655   // acquire the outer lock.  Another thread T3 might try to acquire the outer
 656   // lock, but encounter contention and enqueue itself on cxq.  T2 then drops the
 657   // outer lock, but skips succession as this thread T1 still holds OnDeck.
 658   // T1 is and remains responsible for ensuring succession of T3.
 659   //
 660   // Note that we don't need to recheck EntryList, just cxq.
 661   // If threads moved onto EntryList since we dropped OnDeck
 662   // that implies some other thread forced succession.
 663   cxq = _LockWord.FullWord;
 664   if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) {
 665     goto Succession;         // potential race -- re-run succession
 666   }
 667   return;
 668 }
 669 
 670 bool Monitor::notify() {
 671   assert(_owner == Thread::current(), "invariant");
 672   assert(ILocked(), "invariant");
 673   if (_WaitSet == NULL) return true;
 674   NotifyCount++;
 675 
 676   // Transfer one thread from the WaitSet to the EntryList or cxq.
 677   // Currently we just unlink the head of the WaitSet and prepend to the cxq.
 678   // And of course we could just unlink it and unpark it, too, but
 679   // in that case it'd likely impale itself on the reentry.
 680   Thread::muxAcquire(_WaitLock, "notify:WaitLock");
 681   ParkEvent * nfy = _WaitSet;
 682   if (nfy != NULL) {                  // DCL idiom
 683     _WaitSet = nfy->ListNext;
 684     assert(nfy->Notified == 0, "invariant");
 685     // push nfy onto the cxq
 686     for (;;) {
 687       const intptr_t v = _LockWord.FullWord;
 688       assert((v & 0xFF) == _LBIT, "invariant");
 689       nfy->ListNext = (ParkEvent *)(v & ~_LBIT);
 690       if (Atomic::cmpxchg(intptr_t(nfy)|_LBIT, &_LockWord.FullWord, v) == v) break;
 691       // interference - _LockWord changed -- just retry
 692     }
 693     // Note that setting Notified before pushing nfy onto the cxq is
 694     // also legal and safe, but the safety properties are much more
 695     // subtle, so for the sake of code stewardship ...
 696     OrderAccess::fence();
 697     nfy->Notified = 1;
 698   }
 699   Thread::muxRelease(_WaitLock);
 700   if (nfy != NULL && (NativeMonitorFlags & 16)) {
 701     // Experimental code ... light up the wakee in the hope that this thread (the owner)
 702     // will drop the lock just about the time the wakee comes ONPROC.
 703     nfy->unpark();
 704   }
 705   assert(ILocked(), "invariant");
 706   return true;
 707 }
 708 
 709 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset
 710 // to the cxq.  This could be done more efficiently with a single bulk en-mass transfer,
 711 // but in practice notifyAll() for large #s of threads is rare and not time-critical.
 712 // Beware too, that we invert the order of the waiters.  Lets say that the
 713 // waitset is "ABCD" and the cxq is "XYZ".  After a notifyAll() the waitset
 714 // will be empty and the cxq will be "DCBAXYZ".  This is benign, of course.
 715 
 716 bool Monitor::notify_all() {
 717   assert(_owner == Thread::current(), "invariant");
 718   assert(ILocked(), "invariant");
 719   while (_WaitSet != NULL) notify();
 720   return true;
 721 }
 722 
 723 int Monitor::IWait(Thread * Self, jlong timo) {
 724   assert(ILocked(), "invariant");
 725 
 726   // Phases:
 727   // 1. Enqueue Self on WaitSet - currently prepend
 728   // 2. unlock - drop the outer lock
 729   // 3. wait for either notification or timeout
 730   // 4. lock - reentry - reacquire the outer lock
 731 
 732   ParkEvent * const ESelf = Self->_MutexEvent;
 733   ESelf->Notified = 0;
 734   ESelf->reset();
 735   OrderAccess::fence();
 736 
 737   // Add Self to WaitSet
 738   // Ideally only the holder of the outer lock would manipulate the WaitSet -
 739   // That is, the outer lock would implicitly protect the WaitSet.
 740   // But if a thread in wait() encounters a timeout it will need to dequeue itself
 741   // from the WaitSet _before it becomes the owner of the lock.  We need to dequeue
 742   // as the ParkEvent -- which serves as a proxy for the thread -- can't reside
 743   // on both the WaitSet and the EntryList|cxq at the same time..  That is, a thread
 744   // on the WaitSet can't be allowed to compete for the lock until it has managed to
 745   // unlink its ParkEvent from WaitSet.  Thus the need for WaitLock.
 746   // Contention on the WaitLock is minimal.
 747   //
 748   // Another viable approach would be add another ParkEvent, "WaitEvent" to the
 749   // thread class.  The WaitSet would be composed of WaitEvents.  Only the
 750   // owner of the outer lock would manipulate the WaitSet.  A thread in wait()
 751   // could then compete for the outer lock, and then, if necessary, unlink itself
 752   // from the WaitSet only after having acquired the outer lock.  More precisely,
 753   // there would be no WaitLock.  A thread in in wait() would enqueue its WaitEvent
 754   // on the WaitSet; release the outer lock; wait for either notification or timeout;
 755   // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet.
 756   //
 757   // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice.
 758   // One set would be for the WaitSet and one for the EntryList.
 759   // We could also deconstruct the ParkEvent into a "pure" event and add a
 760   // new immortal/TSM "ListElement" class that referred to ParkEvents.
 761   // In that case we could have one ListElement on the WaitSet and another
 762   // on the EntryList, with both referring to the same pure Event.
 763 
 764   Thread::muxAcquire(_WaitLock, "wait:WaitLock:Add");
 765   ESelf->ListNext = _WaitSet;
 766   _WaitSet = ESelf;
 767   Thread::muxRelease(_WaitLock);
 768 
 769   // Release the outer lock
 770   // We call IUnlock (RelaxAssert=true) as a thread T1 might
 771   // enqueue itself on the WaitSet, call IUnlock(), drop the lock,
 772   // and then stall before it can attempt to wake a successor.
 773   // Some other thread T2 acquires the lock, and calls notify(), moving
 774   // T1 from the WaitSet to the cxq.  T2 then drops the lock.  T1 resumes,
 775   // and then finds *itself* on the cxq.  During the course of a normal
 776   // IUnlock() call a thread should _never find itself on the EntryList
 777   // or cxq, but in the case of wait() it's possible.
 778   // See synchronizer.cpp objectMonitor::wait().
 779   IUnlock(true);
 780 
 781   // Wait for either notification or timeout
 782   // Beware that in some circumstances we might propagate
 783   // spurious wakeups back to the caller.
 784 
 785   for (;;) {
 786     if (ESelf->Notified) break;
 787     int err = ParkCommon(ESelf, timo);
 788     if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break;
 789   }
 790 
 791   // Prepare for reentry - if necessary, remove ESelf from WaitSet
 792   // ESelf can be:
 793   // 1. Still on the WaitSet.  This can happen if we exited the loop by timeout.
 794   // 2. On the cxq or EntryList
 795   // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position.
 796 
 797   OrderAccess::fence();
 798   int WasOnWaitSet = 0;
 799   if (ESelf->Notified == 0) {
 800     Thread::muxAcquire(_WaitLock, "wait:WaitLock:remove");
 801     if (ESelf->Notified == 0) {     // DCL idiom
 802       assert(_OnDeck != ESelf, "invariant");   // can't be both OnDeck and on WaitSet
 803       // ESelf is resident on the WaitSet -- unlink it.
 804       // A doubly-linked list would be better here so we can unlink in constant-time.
 805       // We have to unlink before we potentially recontend as ESelf might otherwise
 806       // end up on the cxq|EntryList -- it can't be on two lists at once.
 807       ParkEvent * p = _WaitSet;
 808       ParkEvent * q = NULL;            // classic q chases p
 809       while (p != NULL && p != ESelf) {
 810         q = p;
 811         p = p->ListNext;
 812       }
 813       assert(p == ESelf, "invariant");
 814       if (p == _WaitSet) {      // found at head
 815         assert(q == NULL, "invariant");
 816         _WaitSet = p->ListNext;
 817       } else {                  // found in interior
 818         assert(q->ListNext == p, "invariant");
 819         q->ListNext = p->ListNext;
 820       }
 821       WasOnWaitSet = 1;        // We were *not* notified but instead encountered timeout
 822     }
 823     Thread::muxRelease(_WaitLock);
 824   }
 825 
 826   // Reentry phase - reacquire the lock
 827   if (WasOnWaitSet) {
 828     // ESelf was previously on the WaitSet but we just unlinked it above
 829     // because of a timeout.  ESelf is not resident on any list and is not OnDeck
 830     assert(_OnDeck != ESelf, "invariant");
 831     ILock(Self);
 832   } else {
 833     // A prior notify() operation moved ESelf from the WaitSet to the cxq.
 834     // ESelf is now on the cxq, EntryList or at the OnDeck position.
 835     // The following fragment is extracted from Monitor::ILock()
 836     for (;;) {
 837       if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(Self)) break;
 838       ParkCommon(ESelf, 0);
 839     }
 840     assert(_OnDeck == ESelf, "invariant");
 841     _OnDeck = NULL;
 842   }
 843 
 844   assert(ILocked(), "invariant");
 845   return WasOnWaitSet != 0;        // return true IFF timeout
 846 }
 847 
 848 
 849 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS:
 850 // In particular, there are certain types of global lock that may be held
 851 // by a Java thread while it is blocked at a safepoint but before it has
 852 // written the _owner field. These locks may be sneakily acquired by the
 853 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should
 854 // identify all such locks, and ensure that Java threads never block at
 855 // safepoints while holding them (_no_safepoint_check_flag). While it
 856 // seems as though this could increase the time to reach a safepoint
 857 // (or at least increase the mean, if not the variance), the latter
 858 // approach might make for a cleaner, more maintainable JVM design.
 859 //
 860 // Sneaking is vile and reprehensible and should be excised at the 1st
 861 // opportunity.  It's possible that the need for sneaking could be obviated
 862 // as follows.  Currently, a thread might (a) while TBIVM, call pthread_mutex_lock
 863 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex.
 864 // (b) stall at the TBIVM exit point as a safepoint is in effect.  Critically,
 865 // it'll stall at the TBIVM reentry state transition after having acquired the
 866 // underlying lock, but before having set _owner and having entered the actual
 867 // critical section.  The lock-sneaking facility leverages that fact and allowed the
 868 // VM thread to logically acquire locks that had already be physically locked by mutators
 869 // but where mutators were known blocked by the reentry thread state transition.
 870 //
 871 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly
 872 // wrapped calls to park(), then we could likely do away with sneaking.  We'd
 873 // decouple lock acquisition and parking.  The critical invariant  to eliminating
 874 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM.
 875 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket.
 876 // One difficulty with this approach is that the TBIVM wrapper could recurse and
 877 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued.
 878 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem.
 879 //
 880 // But of course the proper ultimate approach is to avoid schemes that require explicit
 881 // sneaking or dependence on any any clever invariants or subtle implementation properties
 882 // of Mutex-Monitor and instead directly address the underlying design flaw.
 883 
 884 void Monitor::lock(Thread * Self) {
 885   // Ensure that the Monitor requires/allows safepoint checks.
 886   assert(_safepoint_check_required != Monitor::_safepoint_check_never,
 887          "This lock should never have a safepoint check: %s", name());
 888 
 889 #ifdef CHECK_UNHANDLED_OOPS
 890   // Clear unhandled oops so we get a crash right away.  Only clear for non-vm
 891   // or GC threads.
 892   if (Self->is_Java_thread()) {
 893     Self->clear_unhandled_oops();
 894   }
 895 #endif // CHECK_UNHANDLED_OOPS
 896 
 897   debug_only(check_prelock_state(Self));
 898   assert(_owner != Self, "invariant");
 899   assert(_OnDeck != Self->_MutexEvent, "invariant");
 900 
 901   if (TryFast()) {
 902  Exeunt:
 903     assert(ILocked(), "invariant");
 904     assert(owner() == NULL, "invariant");
 905     set_owner(Self);
 906     return;
 907   }
 908 
 909   // The lock is contended ...
 910 
 911   bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
 912   if (can_sneak && _owner == NULL) {
 913     // a java thread has locked the lock but has not entered the
 914     // critical region -- let's just pretend we've locked the lock
 915     // and go on.  we note this with _snuck so we can also
 916     // pretend to unlock when the time comes.
 917     _snuck = true;
 918     goto Exeunt;
 919   }
 920 
 921   // Try a brief spin to avoid passing thru thread state transition ...
 922   if (TrySpin(Self)) goto Exeunt;
 923 
 924   check_block_state(Self);
 925   if (Self->is_Java_thread()) {
 926     // Horrible dictu - we suffer through a state transition
 927     assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex");
 928     ThreadBlockInVM tbivm((JavaThread *) Self);
 929     ILock(Self);
 930   } else {
 931     // Mirabile dictu
 932     ILock(Self);
 933   }
 934   goto Exeunt;
 935 }
 936 
 937 void Monitor::lock() {
 938   this->lock(Thread::current());
 939 }
 940 
 941 // Lock without safepoint check - a degenerate variant of lock().
 942 // Should ONLY be used by safepoint code and other code
 943 // that is guaranteed not to block while running inside the VM. If this is called with
 944 // thread state set to be in VM, the safepoint synchronization code will deadlock!
 945 
 946 void Monitor::lock_without_safepoint_check(Thread * Self) {
 947   // Ensure that the Monitor does not require or allow safepoint checks.
 948   assert(_safepoint_check_required != Monitor::_safepoint_check_always,
 949          "This lock should always have a safepoint check: %s", name());
 950   assert(_owner != Self, "invariant");
 951   ILock(Self);
 952   assert(_owner == NULL, "invariant");
 953   set_owner(Self);
 954 }
 955 
 956 void Monitor::lock_without_safepoint_check() {
 957   lock_without_safepoint_check(Thread::current());
 958 }
 959 
 960 
 961 // Returns true if thread succeeds in grabbing the lock, otherwise false.
 962 
 963 bool Monitor::try_lock() {
 964   Thread * const Self = Thread::current();
 965   debug_only(check_prelock_state(Self));
 966   // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler");
 967 
 968   // Special case, where all Java threads are stopped.
 969   // The lock may have been acquired but _owner is not yet set.
 970   // In that case the VM thread can safely grab the lock.
 971   // It strikes me this should appear _after the TryLock() fails, below.
 972   bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
 973   if (can_sneak && _owner == NULL) {
 974     set_owner(Self); // Do not need to be atomic, since we are at a safepoint
 975     _snuck = true;
 976     return true;
 977   }
 978 
 979   if (TryLock()) {
 980     // We got the lock
 981     assert(_owner == NULL, "invariant");
 982     set_owner(Self);
 983     return true;
 984   }
 985   return false;
 986 }
 987 
 988 void Monitor::unlock() {
 989   assert(_owner == Thread::current(), "invariant");
 990   assert(_OnDeck != Thread::current()->_MutexEvent, "invariant");
 991   set_owner(NULL);
 992   if (_snuck) {
 993     assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
 994     _snuck = false;
 995     return;
 996   }
 997   IUnlock(false);
 998 }
 999 
1000 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check()
1001 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter.
1002 //
1003 // There's no expectation that JVM_RawMonitors will interoperate properly with the native
1004 // Mutex-Monitor constructs.  We happen to implement JVM_RawMonitors in terms of
1005 // native Mutex-Monitors simply as a matter of convenience.  A simple abstraction layer
1006 // over a pthread_mutex_t would work equally as well, but require more platform-specific
1007 // code -- a "PlatformMutex".  Alternatively, a simply layer over muxAcquire-muxRelease
1008 // would work too.
1009 //
1010 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent
1011 // instance available.  Instead, we transiently allocate a ParkEvent on-demand if
1012 // we encounter contention.  That ParkEvent remains associated with the thread
1013 // until it manages to acquire the lock, at which time we return the ParkEvent
1014 // to the global ParkEvent free list.  This is correct and suffices for our purposes.
1015 //
1016 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that
1017 // jvm_raw_lock() didn't have the corresponding test.  I suspect that's an
1018 // oversight, but I've replicated the original suspect logic in the new code ...
1019 
1020 void Monitor::jvm_raw_lock() {
1021   assert(rank() == native, "invariant");
1022 
1023   if (TryLock()) {
1024  Exeunt:
1025     assert(ILocked(), "invariant");
1026     assert(_owner == NULL, "invariant");
1027     // This can potentially be called by non-java Threads. Thus, the Thread::current_or_null()
1028     // might return NULL. Don't call set_owner since it will break on an NULL owner
1029     // Consider installing a non-null "ANON" distinguished value instead of just NULL.
1030     _owner = Thread::current_or_null();
1031     return;
1032   }
1033 
1034   if (TrySpin(NULL)) goto Exeunt;
1035 
1036   // slow-path - apparent contention
1037   // Allocate a ParkEvent for transient use.
1038   // The ParkEvent remains associated with this thread until
1039   // the time the thread manages to acquire the lock.
1040   ParkEvent * const ESelf = ParkEvent::Allocate(NULL);
1041   ESelf->reset();
1042   OrderAccess::storeload();
1043 
1044   // Either Enqueue Self on cxq or acquire the outer lock.
1045   if (AcquireOrPush (ESelf)) {
1046     ParkEvent::Release(ESelf);      // surrender the ParkEvent
1047     goto Exeunt;
1048   }
1049 
1050   // At any given time there is at most one ondeck thread.
1051   // ondeck implies not resident on cxq and not resident on EntryList
1052   // Only the OnDeck thread can try to acquire -- contend for -- the lock.
1053   // CONSIDER: use Self->OnDeck instead of m->OnDeck.
1054   for (;;) {
1055     if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(NULL)) break;
1056     ParkCommon(ESelf, 0);
1057   }
1058 
1059   assert(_OnDeck == ESelf, "invariant");
1060   _OnDeck = NULL;
1061   ParkEvent::Release(ESelf);      // surrender the ParkEvent
1062   goto Exeunt;
1063 }
1064 
1065 void Monitor::jvm_raw_unlock() {
1066   // Nearly the same as Monitor::unlock() ...
1067   // directly set _owner instead of using set_owner(null)
1068   _owner = NULL;
1069   if (_snuck) {         // ???
1070     assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1071     _snuck = false;
1072     return;
1073   }
1074   IUnlock(false);
1075 }
1076 
1077 bool Monitor::wait(bool no_safepoint_check, long timeout,
1078                    bool as_suspend_equivalent) {
1079   // Make sure safepoint checking is used properly.
1080   assert(!(_safepoint_check_required == Monitor::_safepoint_check_never && no_safepoint_check == false),
1081          "This lock should never have a safepoint check: %s", name());
1082   assert(!(_safepoint_check_required == Monitor::_safepoint_check_always && no_safepoint_check == true),
1083          "This lock should always have a safepoint check: %s", name());
1084 
1085   Thread * const Self = Thread::current();
1086   assert(_owner == Self, "invariant");
1087   assert(ILocked(), "invariant");
1088 
1089   // as_suspend_equivalent logically implies !no_safepoint_check
1090   guarantee(!as_suspend_equivalent || !no_safepoint_check, "invariant");
1091   // !no_safepoint_check logically implies java_thread
1092   guarantee(no_safepoint_check || Self->is_Java_thread(), "invariant");
1093 
1094   #ifdef ASSERT
1095   Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks());
1096   assert(least != this, "Specification of get_least_... call above");
1097   if (least != NULL && least->rank() <= special) {
1098     tty->print("Attempting to wait on monitor %s/%d while holding"
1099                " lock %s/%d -- possible deadlock",
1100                name(), rank(), least->name(), least->rank());
1101     assert(false, "Shouldn't block(wait) while holding a lock of rank special");
1102   }
1103   #endif // ASSERT
1104 
1105   int wait_status;
1106   // conceptually set the owner to NULL in anticipation of
1107   // abdicating the lock in wait
1108   set_owner(NULL);
1109   if (no_safepoint_check) {
1110     wait_status = IWait(Self, timeout);
1111   } else {
1112     assert(Self->is_Java_thread(), "invariant");
1113     JavaThread *jt = (JavaThread *)Self;
1114 
1115     // Enter safepoint region - ornate and Rococo ...
1116     ThreadBlockInVM tbivm(jt);
1117     OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */);
1118 
1119     if (as_suspend_equivalent) {
1120       jt->set_suspend_equivalent();
1121       // cleared by handle_special_suspend_equivalent_condition() or
1122       // java_suspend_self()
1123     }
1124 
1125     wait_status = IWait(Self, timeout);
1126 
1127     // were we externally suspended while we were waiting?
1128     if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) {
1129       // Our event wait has finished and we own the lock, but
1130       // while we were waiting another thread suspended us. We don't
1131       // want to hold the lock while suspended because that
1132       // would surprise the thread that suspended us.
1133       assert(ILocked(), "invariant");
1134       IUnlock(true);
1135       jt->java_suspend_self();
1136       ILock(Self);
1137       assert(ILocked(), "invariant");
1138     }
1139   }
1140 
1141   // Conceptually reestablish ownership of the lock.
1142   // The "real" lock -- the LockByte -- was reacquired by IWait().
1143   assert(ILocked(), "invariant");
1144   assert(_owner == NULL, "invariant");
1145   set_owner(Self);
1146   return wait_status != 0;          // return true IFF timeout
1147 }
1148 
1149 Monitor::~Monitor() {
1150 #ifdef ASSERT
1151   uintptr_t owner = UNS(_owner);
1152   uintptr_t lockword = UNS(_LockWord.FullWord);
1153   uintptr_t entrylist = UNS(_EntryList);
1154   uintptr_t waitset = UNS(_WaitSet);
1155   uintptr_t ondeck = UNS(_OnDeck);
1156   // Print _name with precision limit, in case failure is due to memory
1157   // corruption that also trashed _name.
1158   assert((owner|lockword|entrylist|waitset|ondeck) == 0,
1159          "%.*s: _owner(" INTPTR_FORMAT ")|_LockWord(" INTPTR_FORMAT ")|_EntryList(" INTPTR_FORMAT ")|_WaitSet("
1160          INTPTR_FORMAT ")|_OnDeck(" INTPTR_FORMAT ") != 0",
1161          MONITOR_NAME_LEN, _name, owner, lockword, entrylist, waitset, ondeck);
1162 #endif
1163 }
1164 
1165 void Monitor::ClearMonitor(Monitor * m, const char *name) {
1166   m->_owner             = NULL;
1167   m->_snuck             = false;
1168   if (name == NULL) {
1169     strcpy(m->_name, "UNKNOWN");
1170   } else {
1171     strncpy(m->_name, name, MONITOR_NAME_LEN - 1);
1172     m->_name[MONITOR_NAME_LEN - 1] = '\0';
1173   }
1174   m->_LockWord.FullWord = 0;
1175   m->_EntryList         = NULL;
1176   m->_OnDeck            = NULL;
1177   m->_WaitSet           = NULL;
1178   m->_WaitLock[0]       = 0;
1179 }
1180 
1181 Monitor::Monitor() { ClearMonitor(this); }
1182 
1183 Monitor::Monitor(int Rank, const char * name, bool allow_vm_block,
1184                  SafepointCheckRequired safepoint_check_required) {
1185   ClearMonitor(this, name);
1186 #ifdef ASSERT
1187   _allow_vm_block  = allow_vm_block;
1188   _rank            = Rank;
1189   NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;)
1190 #endif
1191 }
1192 
1193 Mutex::Mutex(int Rank, const char * name, bool allow_vm_block,
1194              SafepointCheckRequired safepoint_check_required) {
1195   ClearMonitor((Monitor *) this, name);
1196 #ifdef ASSERT
1197   _allow_vm_block   = allow_vm_block;
1198   _rank             = Rank;
1199   NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;)
1200 #endif
1201 }
1202 
1203 bool Monitor::owned_by_self() const {
1204   bool ret = _owner == Thread::current();
1205   assert(!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant");
1206   return ret;
1207 }
1208 
1209 void Monitor::print_on_error(outputStream* st) const {
1210   st->print("[" PTR_FORMAT, p2i(this));
1211   st->print("] %s", _name);
1212   st->print(" - owner thread: " PTR_FORMAT, p2i(_owner));
1213 }
1214 
1215 
1216 
1217 
1218 // ----------------------------------------------------------------------------------
1219 // Non-product code
1220 
1221 #ifndef PRODUCT
1222 void Monitor::print_on(outputStream* st) const {
1223   st->print_cr("Mutex: [" PTR_FORMAT "/" PTR_FORMAT "] %s - owner: " PTR_FORMAT,
1224                p2i(this), _LockWord.FullWord, _name, p2i(_owner));
1225 }
1226 #endif
1227 
1228 #ifndef PRODUCT
1229 #ifdef ASSERT
1230 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) {
1231   Monitor *res, *tmp;
1232   for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) {
1233     if (tmp->rank() < res->rank()) {
1234       res = tmp;
1235     }
1236   }
1237   if (!SafepointSynchronize::is_at_safepoint()) {
1238     // In this case, we expect the held locks to be
1239     // in increasing rank order (modulo any native ranks)
1240     for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1241       if (tmp->next() != NULL) {
1242         assert(tmp->rank() == Mutex::native ||
1243                tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1244       }
1245     }
1246   }
1247   return res;
1248 }
1249 
1250 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) {
1251   Monitor *res, *tmp;
1252   for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) {
1253     if (tmp != this && (res == NULL || tmp->rank() < res->rank())) {
1254       res = tmp;
1255     }
1256   }
1257   if (!SafepointSynchronize::is_at_safepoint()) {
1258     // In this case, we expect the held locks to be
1259     // in increasing rank order (modulo any native ranks)
1260     for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1261       if (tmp->next() != NULL) {
1262         assert(tmp->rank() == Mutex::native ||
1263                tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1264       }
1265     }
1266   }
1267   return res;
1268 }
1269 
1270 
1271 bool Monitor::contains(Monitor* locks, Monitor * lock) {
1272   for (; locks != NULL; locks = locks->next()) {
1273     if (locks == lock) {
1274       return true;
1275     }
1276   }
1277   return false;
1278 }
1279 #endif
1280 
1281 // Called immediately after lock acquisition or release as a diagnostic
1282 // to track the lock-set of the thread and test for rank violations that
1283 // might indicate exposure to deadlock.
1284 // Rather like an EventListener for _owner (:>).
1285 
1286 void Monitor::set_owner_implementation(Thread *new_owner) {
1287   // This function is solely responsible for maintaining
1288   // and checking the invariant that threads and locks
1289   // are in a 1/N relation, with some some locks unowned.
1290   // It uses the Mutex::_owner, Mutex::_next, and
1291   // Thread::_owned_locks fields, and no other function
1292   // changes those fields.
1293   // It is illegal to set the mutex from one non-NULL
1294   // owner to another--it must be owned by NULL as an
1295   // intermediate state.
1296 
1297   if (new_owner != NULL) {
1298     // the thread is acquiring this lock
1299 
1300     assert(new_owner == Thread::current(), "Should I be doing this?");
1301     assert(_owner == NULL, "setting the owner thread of an already owned mutex");
1302     _owner = new_owner; // set the owner
1303 
1304     // link "this" into the owned locks list
1305 
1306 #ifdef ASSERT  // Thread::_owned_locks is under the same ifdef
1307     Monitor* locks = get_least_ranked_lock(new_owner->owned_locks());
1308     // Mutex::set_owner_implementation is a friend of Thread
1309 
1310     assert(this->rank() >= 0, "bad lock rank");
1311 
1312     // Deadlock avoidance rules require us to acquire Mutexes only in
1313     // a global total order. For example m1 is the lowest ranked mutex
1314     // that the thread holds and m2 is the mutex the thread is trying
1315     // to acquire, then  deadlock avoidance rules require that the rank
1316     // of m2 be less  than the rank of m1.
1317     // The rank Mutex::native  is an exception in that it is not subject
1318     // to the verification rules.
1319     // Here are some further notes relating to mutex acquisition anomalies:
1320     // . it is also ok to acquire Safepoint_lock at the very end while we
1321     //   already hold Terminator_lock - may happen because of periodic safepoints
1322     if (this->rank() != Mutex::native &&
1323         this->rank() != Mutex::suspend_resume &&
1324         locks != NULL && locks->rank() <= this->rank() &&
1325         !SafepointSynchronize::is_at_safepoint() &&
1326         !(this == Safepoint_lock && contains(locks, Terminator_lock) &&
1327         SafepointSynchronize::is_synchronizing())) {
1328       new_owner->print_owned_locks();
1329       fatal("acquiring lock %s/%d out of order with lock %s/%d -- "
1330             "possible deadlock", this->name(), this->rank(),
1331             locks->name(), locks->rank());
1332     }
1333 
1334     this->_next = new_owner->_owned_locks;
1335     new_owner->_owned_locks = this;
1336 #endif
1337 
1338   } else {
1339     // the thread is releasing this lock
1340 
1341     Thread* old_owner = _owner;
1342     debug_only(_last_owner = old_owner);
1343 
1344     assert(old_owner != NULL, "removing the owner thread of an unowned mutex");
1345     assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex");
1346 
1347     _owner = NULL; // set the owner
1348 
1349 #ifdef ASSERT
1350     Monitor *locks = old_owner->owned_locks();
1351 
1352     // remove "this" from the owned locks list
1353 
1354     Monitor *prev = NULL;
1355     bool found = false;
1356     for (; locks != NULL; prev = locks, locks = locks->next()) {
1357       if (locks == this) {
1358         found = true;
1359         break;
1360       }
1361     }
1362     assert(found, "Removing a lock not owned");
1363     if (prev == NULL) {
1364       old_owner->_owned_locks = _next;
1365     } else {
1366       prev->_next = _next;
1367     }
1368     _next = NULL;
1369 #endif
1370   }
1371 }
1372 
1373 
1374 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock()
1375 void Monitor::check_prelock_state(Thread *thread) {
1376   assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm)
1377          || rank() == Mutex::special, "wrong thread state for using locks");
1378   if (StrictSafepointChecks) {
1379     if (thread->is_VM_thread() && !allow_vm_block()) {
1380       fatal("VM thread using lock %s (not allowed to block on)", name());
1381     }
1382     debug_only(if (rank() != Mutex::special) \
1383                thread->check_for_valid_safepoint_state(false);)
1384   }
1385   assert(!os::ThreadCrashProtection::is_crash_protected(thread),
1386          "locking not allowed when crash protection is set");
1387 }
1388 
1389 void Monitor::check_block_state(Thread *thread) {
1390   if (!_allow_vm_block && thread->is_VM_thread()) {
1391     warning("VM thread blocked on lock");
1392     print();
1393     BREAKPOINT;
1394   }
1395   assert(_owner != thread, "deadlock: blocking on monitor owned by current thread");
1396 }
1397 
1398 #endif // PRODUCT