1 /* 2 * Copyright (c) 1998, 2009, 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 "incls/_precompiled.incl" 26 # include "incls/_objectMonitor.cpp.incl" 27 28 #if defined(__GNUC__) && !defined(IA64) 29 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 30 #define ATTR __attribute__((noinline)) 31 #else 32 #define ATTR 33 #endif 34 35 36 #ifdef DTRACE_ENABLED 37 38 // Only bother with this argument setup if dtrace is available 39 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 40 41 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 42 jlong, uintptr_t, char*, int); 43 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 44 jlong, uintptr_t, char*, int); 45 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 46 jlong, uintptr_t, char*, int); 47 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 48 jlong, uintptr_t, char*, int); 49 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 50 jlong, uintptr_t, char*, int); 51 52 #define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \ 53 char* bytes = NULL; \ 54 int len = 0; \ 55 jlong jtid = SharedRuntime::get_java_tid(thread); \ 56 symbolOop klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \ 57 if (klassname != NULL) { \ 58 bytes = (char*)klassname->bytes(); \ 59 len = klassname->utf8_length(); \ 60 } 61 62 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ 63 { \ 64 if (DTraceMonitorProbes) { \ 65 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 66 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 67 (monitor), bytes, len, (millis)); \ 68 } \ 69 } 70 71 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 75 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 76 (uintptr_t)(monitor), bytes, len); \ 77 } \ 78 } 79 80 #else // ndef DTRACE_ENABLED 81 82 #define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;} 83 #define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;} 84 85 #endif // ndef DTRACE_ENABLED 86 87 // Tunables ... 88 // The knob* variables are effectively final. Once set they should 89 // never be modified hence. Consider using __read_mostly with GCC. 90 91 int ObjectMonitor::Knob_Verbose = 0 ; 92 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 93 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 94 static int Knob_HandOff = 0 ; 95 static int Knob_ReportSettings = 0 ; 96 97 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 98 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 99 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 100 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 101 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 102 static int Knob_SpinEarly = 1 ; 103 static int Knob_SuccEnabled = 1 ; // futile wake throttling 104 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 105 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 106 static int Knob_Bonus = 100 ; // spin success bonus 107 static int Knob_BonusB = 100 ; // spin success bonus 108 static int Knob_Penalty = 200 ; // spin failure penalty 109 static int Knob_Poverty = 1000 ; 110 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 111 static int Knob_FixedSpin = 0 ; 112 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 113 static int Knob_UsePause = 1 ; 114 static int Knob_ExitPolicy = 0 ; 115 static int Knob_PreSpin = 10 ; // 20-100 likely better 116 static int Knob_ResetEvent = 0 ; 117 static int BackOffMask = 0 ; 118 119 static int Knob_FastHSSEC = 0 ; 120 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 121 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 122 static volatile int InitDone = 0 ; 123 124 #define TrySpin TrySpin_VaryDuration 125 126 // ----------------------------------------------------------------------------- 127 // Theory of operations -- Monitors lists, thread residency, etc: 128 // 129 // * A thread acquires ownership of a monitor by successfully 130 // CAS()ing the _owner field from null to non-null. 131 // 132 // * Invariant: A thread appears on at most one monitor list -- 133 // cxq, EntryList or WaitSet -- at any one time. 134 // 135 // * Contending threads "push" themselves onto the cxq with CAS 136 // and then spin/park. 137 // 138 // * After a contending thread eventually acquires the lock it must 139 // dequeue itself from either the EntryList or the cxq. 140 // 141 // * The exiting thread identifies and unparks an "heir presumptive" 142 // tentative successor thread on the EntryList. Critically, the 143 // exiting thread doesn't unlink the successor thread from the EntryList. 144 // After having been unparked, the wakee will recontend for ownership of 145 // the monitor. The successor (wakee) will either acquire the lock or 146 // re-park itself. 147 // 148 // Succession is provided for by a policy of competitive handoff. 149 // The exiting thread does _not_ grant or pass ownership to the 150 // successor thread. (This is also referred to as "handoff" succession"). 151 // Instead the exiting thread releases ownership and possibly wakes 152 // a successor, so the successor can (re)compete for ownership of the lock. 153 // If the EntryList is empty but the cxq is populated the exiting 154 // thread will drain the cxq into the EntryList. It does so by 155 // by detaching the cxq (installing null with CAS) and folding 156 // the threads from the cxq into the EntryList. The EntryList is 157 // doubly linked, while the cxq is singly linked because of the 158 // CAS-based "push" used to enqueue recently arrived threads (RATs). 159 // 160 // * Concurrency invariants: 161 // 162 // -- only the monitor owner may access or mutate the EntryList. 163 // The mutex property of the monitor itself protects the EntryList 164 // from concurrent interference. 165 // -- Only the monitor owner may detach the cxq. 166 // 167 // * The monitor entry list operations avoid locks, but strictly speaking 168 // they're not lock-free. Enter is lock-free, exit is not. 169 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 170 // 171 // * The cxq can have multiple concurrent "pushers" but only one concurrent 172 // detaching thread. This mechanism is immune from the ABA corruption. 173 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 174 // 175 // * Taken together, the cxq and the EntryList constitute or form a 176 // single logical queue of threads stalled trying to acquire the lock. 177 // We use two distinct lists to improve the odds of a constant-time 178 // dequeue operation after acquisition (in the ::enter() epilog) and 179 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 180 // A key desideratum is to minimize queue & monitor metadata manipulation 181 // that occurs while holding the monitor lock -- that is, we want to 182 // minimize monitor lock holds times. Note that even a small amount of 183 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 184 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 185 // locks and monitor metadata. 186 // 187 // Cxq points to the the set of Recently Arrived Threads attempting entry. 188 // Because we push threads onto _cxq with CAS, the RATs must take the form of 189 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 190 // the unlocking thread notices that EntryList is null but _cxq is != null. 191 // 192 // The EntryList is ordered by the prevailing queue discipline and 193 // can be organized in any convenient fashion, such as a doubly-linked list or 194 // a circular doubly-linked list. Critically, we want insert and delete operations 195 // to operate in constant-time. If we need a priority queue then something akin 196 // to Solaris' sleepq would work nicely. Viz., 197 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 198 // Queue discipline is enforced at ::exit() time, when the unlocking thread 199 // drains the cxq into the EntryList, and orders or reorders the threads on the 200 // EntryList accordingly. 201 // 202 // Barring "lock barging", this mechanism provides fair cyclic ordering, 203 // somewhat similar to an elevator-scan. 204 // 205 // * The monitor synchronization subsystem avoids the use of native 206 // synchronization primitives except for the narrow platform-specific 207 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 208 // the semantics of park-unpark. Put another way, this monitor implementation 209 // depends only on atomic operations and park-unpark. The monitor subsystem 210 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 211 // underlying OS manages the READY<->RUN transitions. 212 // 213 // * Waiting threads reside on the WaitSet list -- wait() puts 214 // the caller onto the WaitSet. 215 // 216 // * notify() or notifyAll() simply transfers threads from the WaitSet to 217 // either the EntryList or cxq. Subsequent exit() operations will 218 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 219 // it's likely the notifyee would simply impale itself on the lock held 220 // by the notifier. 221 // 222 // * An interesting alternative is to encode cxq as (List,LockByte) where 223 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 224 // variable, like _recursions, in the scheme. The threads or Events that form 225 // the list would have to be aligned in 256-byte addresses. A thread would 226 // try to acquire the lock or enqueue itself with CAS, but exiting threads 227 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 228 // Note that is is *not* word-tearing, but it does presume that full-word 229 // CAS operations are coherent with intermix with STB operations. That's true 230 // on most common processors. 231 // 232 // * See also http://blogs.sun.com/dave 233 234 235 // ----------------------------------------------------------------------------- 236 // Enter support 237 238 bool ObjectMonitor::try_enter(Thread* THREAD) { 239 if (THREAD != _owner) { 240 if (THREAD->is_lock_owned ((address)_owner)) { 241 assert(_recursions == 0, "internal state error"); 242 _owner = THREAD ; 243 _recursions = 1 ; 244 OwnerIsThread = 1 ; 245 return true; 246 } 247 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 248 return false; 249 } 250 return true; 251 } else { 252 _recursions++; 253 return true; 254 } 255 } 256 257 void ATTR ObjectMonitor::enter(TRAPS) { 258 // The following code is ordered to check the most common cases first 259 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 260 Thread * const Self = THREAD ; 261 void * cur ; 262 263 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 264 if (cur == NULL) { 265 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 266 assert (_recursions == 0 , "invariant") ; 267 assert (_owner == Self, "invariant") ; 268 // CONSIDER: set or assert OwnerIsThread == 1 269 return ; 270 } 271 272 if (cur == Self) { 273 // TODO-FIXME: check for integer overflow! BUGID 6557169. 274 _recursions ++ ; 275 return ; 276 } 277 278 if (Self->is_lock_owned ((address)cur)) { 279 assert (_recursions == 0, "internal state error"); 280 _recursions = 1 ; 281 // Commute owner from a thread-specific on-stack BasicLockObject address to 282 // a full-fledged "Thread *". 283 _owner = Self ; 284 OwnerIsThread = 1 ; 285 return ; 286 } 287 288 // We've encountered genuine contention. 289 assert (Self->_Stalled == 0, "invariant") ; 290 Self->_Stalled = intptr_t(this) ; 291 292 // Try one round of spinning *before* enqueueing Self 293 // and before going through the awkward and expensive state 294 // transitions. The following spin is strictly optional ... 295 // Note that if we acquire the monitor from an initial spin 296 // we forgo posting JVMTI events and firing DTRACE probes. 297 if (Knob_SpinEarly && TrySpin (Self) > 0) { 298 assert (_owner == Self , "invariant") ; 299 assert (_recursions == 0 , "invariant") ; 300 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 301 Self->_Stalled = 0 ; 302 return ; 303 } 304 305 assert (_owner != Self , "invariant") ; 306 assert (_succ != Self , "invariant") ; 307 assert (Self->is_Java_thread() , "invariant") ; 308 JavaThread * jt = (JavaThread *) Self ; 309 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 310 assert (jt->thread_state() != _thread_blocked , "invariant") ; 311 assert (this->object() != NULL , "invariant") ; 312 assert (_count >= 0, "invariant") ; 313 314 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 315 // Ensure the object-monitor relationship remains stable while there's contention. 316 Atomic::inc_ptr(&_count); 317 318 { // Change java thread status to indicate blocked on monitor enter. 319 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 320 321 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 322 if (JvmtiExport::should_post_monitor_contended_enter()) { 323 JvmtiExport::post_monitor_contended_enter(jt, this); 324 } 325 326 OSThreadContendState osts(Self->osthread()); 327 ThreadBlockInVM tbivm(jt); 328 329 Self->set_current_pending_monitor(this); 330 331 // TODO-FIXME: change the following for(;;) loop to straight-line code. 332 for (;;) { 333 jt->set_suspend_equivalent(); 334 // cleared by handle_special_suspend_equivalent_condition() 335 // or java_suspend_self() 336 337 EnterI (THREAD) ; 338 339 if (!ExitSuspendEquivalent(jt)) break ; 340 341 // 342 // We have acquired the contended monitor, but while we were 343 // waiting another thread suspended us. We don't want to enter 344 // the monitor while suspended because that would surprise the 345 // thread that suspended us. 346 // 347 _recursions = 0 ; 348 _succ = NULL ; 349 exit (Self) ; 350 351 jt->java_suspend_self(); 352 } 353 Self->set_current_pending_monitor(NULL); 354 } 355 356 Atomic::dec_ptr(&_count); 357 assert (_count >= 0, "invariant") ; 358 Self->_Stalled = 0 ; 359 360 // Must either set _recursions = 0 or ASSERT _recursions == 0. 361 assert (_recursions == 0 , "invariant") ; 362 assert (_owner == Self , "invariant") ; 363 assert (_succ != Self , "invariant") ; 364 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 365 366 // The thread -- now the owner -- is back in vm mode. 367 // Report the glorious news via TI,DTrace and jvmstat. 368 // The probe effect is non-trivial. All the reportage occurs 369 // while we hold the monitor, increasing the length of the critical 370 // section. Amdahl's parallel speedup law comes vividly into play. 371 // 372 // Another option might be to aggregate the events (thread local or 373 // per-monitor aggregation) and defer reporting until a more opportune 374 // time -- such as next time some thread encounters contention but has 375 // yet to acquire the lock. While spinning that thread could 376 // spinning we could increment JVMStat counters, etc. 377 378 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 379 if (JvmtiExport::should_post_monitor_contended_entered()) { 380 JvmtiExport::post_monitor_contended_entered(jt, this); 381 } 382 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 383 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 384 } 385 } 386 387 388 // Caveat: TryLock() is not necessarily serializing if it returns failure. 389 // Callers must compensate as needed. 390 391 int ObjectMonitor::TryLock (Thread * Self) { 392 for (;;) { 393 void * own = _owner ; 394 if (own != NULL) return 0 ; 395 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 396 // Either guarantee _recursions == 0 or set _recursions = 0. 397 assert (_recursions == 0, "invariant") ; 398 assert (_owner == Self, "invariant") ; 399 // CONSIDER: set or assert that OwnerIsThread == 1 400 return 1 ; 401 } 402 // The lock had been free momentarily, but we lost the race to the lock. 403 // Interference -- the CAS failed. 404 // We can either return -1 or retry. 405 // Retry doesn't make as much sense because the lock was just acquired. 406 if (true) return -1 ; 407 } 408 } 409 410 void ATTR ObjectMonitor::EnterI (TRAPS) { 411 Thread * Self = THREAD ; 412 assert (Self->is_Java_thread(), "invariant") ; 413 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 414 415 // Try the lock - TATAS 416 if (TryLock (Self) > 0) { 417 assert (_succ != Self , "invariant") ; 418 assert (_owner == Self , "invariant") ; 419 assert (_Responsible != Self , "invariant") ; 420 return ; 421 } 422 423 DeferredInitialize () ; 424 425 // We try one round of spinning *before* enqueueing Self. 426 // 427 // If the _owner is ready but OFFPROC we could use a YieldTo() 428 // operation to donate the remainder of this thread's quantum 429 // to the owner. This has subtle but beneficial affinity 430 // effects. 431 432 if (TrySpin (Self) > 0) { 433 assert (_owner == Self , "invariant") ; 434 assert (_succ != Self , "invariant") ; 435 assert (_Responsible != Self , "invariant") ; 436 return ; 437 } 438 439 // The Spin failed -- Enqueue and park the thread ... 440 assert (_succ != Self , "invariant") ; 441 assert (_owner != Self , "invariant") ; 442 assert (_Responsible != Self , "invariant") ; 443 444 // Enqueue "Self" on ObjectMonitor's _cxq. 445 // 446 // Node acts as a proxy for Self. 447 // As an aside, if were to ever rewrite the synchronization code mostly 448 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 449 // Java objects. This would avoid awkward lifecycle and liveness issues, 450 // as well as eliminate a subset of ABA issues. 451 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 452 // 453 454 ObjectWaiter node(Self) ; 455 Self->_ParkEvent->reset() ; 456 node._prev = (ObjectWaiter *) 0xBAD ; 457 node.TState = ObjectWaiter::TS_CXQ ; 458 459 // Push "Self" onto the front of the _cxq. 460 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 461 // Note that spinning tends to reduce the rate at which threads 462 // enqueue and dequeue on EntryList|cxq. 463 ObjectWaiter * nxt ; 464 for (;;) { 465 node._next = nxt = _cxq ; 466 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 467 468 // Interference - the CAS failed because _cxq changed. Just retry. 469 // As an optional optimization we retry the lock. 470 if (TryLock (Self) > 0) { 471 assert (_succ != Self , "invariant") ; 472 assert (_owner == Self , "invariant") ; 473 assert (_Responsible != Self , "invariant") ; 474 return ; 475 } 476 } 477 478 // Check for cxq|EntryList edge transition to non-null. This indicates 479 // the onset of contention. While contention persists exiting threads 480 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 481 // operations revert to the faster 1-0 mode. This enter operation may interleave 482 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 483 // arrange for one of the contending thread to use a timed park() operations 484 // to detect and recover from the race. (Stranding is form of progress failure 485 // where the monitor is unlocked but all the contending threads remain parked). 486 // That is, at least one of the contended threads will periodically poll _owner. 487 // One of the contending threads will become the designated "Responsible" thread. 488 // The Responsible thread uses a timed park instead of a normal indefinite park 489 // operation -- it periodically wakes and checks for and recovers from potential 490 // strandings admitted by 1-0 exit operations. We need at most one Responsible 491 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 492 // be responsible for a monitor. 493 // 494 // Currently, one of the contended threads takes on the added role of "Responsible". 495 // A viable alternative would be to use a dedicated "stranding checker" thread 496 // that periodically iterated over all the threads (or active monitors) and unparked 497 // successors where there was risk of stranding. This would help eliminate the 498 // timer scalability issues we see on some platforms as we'd only have one thread 499 // -- the checker -- parked on a timer. 500 501 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 502 // Try to assume the role of responsible thread for the monitor. 503 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 504 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 505 } 506 507 // The lock have been released while this thread was occupied queueing 508 // itself onto _cxq. To close the race and avoid "stranding" and 509 // progress-liveness failure we must resample-retry _owner before parking. 510 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 511 // In this case the ST-MEMBAR is accomplished with CAS(). 512 // 513 // TODO: Defer all thread state transitions until park-time. 514 // Since state transitions are heavy and inefficient we'd like 515 // to defer the state transitions until absolutely necessary, 516 // and in doing so avoid some transitions ... 517 518 TEVENT (Inflated enter - Contention) ; 519 int nWakeups = 0 ; 520 int RecheckInterval = 1 ; 521 522 for (;;) { 523 524 if (TryLock (Self) > 0) break ; 525 assert (_owner != Self, "invariant") ; 526 527 if ((SyncFlags & 2) && _Responsible == NULL) { 528 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 529 } 530 531 // park self 532 if (_Responsible == Self || (SyncFlags & 1)) { 533 TEVENT (Inflated enter - park TIMED) ; 534 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 535 // Increase the RecheckInterval, but clamp the value. 536 RecheckInterval *= 8 ; 537 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 538 } else { 539 TEVENT (Inflated enter - park UNTIMED) ; 540 Self->_ParkEvent->park() ; 541 } 542 543 if (TryLock(Self) > 0) break ; 544 545 // The lock is still contested. 546 // Keep a tally of the # of futile wakeups. 547 // Note that the counter is not protected by a lock or updated by atomics. 548 // That is by design - we trade "lossy" counters which are exposed to 549 // races during updates for a lower probe effect. 550 TEVENT (Inflated enter - Futile wakeup) ; 551 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 552 ObjectMonitor::_sync_FutileWakeups->inc() ; 553 } 554 ++ nWakeups ; 555 556 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 557 // We can defer clearing _succ until after the spin completes 558 // TrySpin() must tolerate being called with _succ == Self. 559 // Try yet another round of adaptive spinning. 560 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 561 562 // We can find that we were unpark()ed and redesignated _succ while 563 // we were spinning. That's harmless. If we iterate and call park(), 564 // park() will consume the event and return immediately and we'll 565 // just spin again. This pattern can repeat, leaving _succ to simply 566 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 567 // Alternately, we can sample fired() here, and if set, forgo spinning 568 // in the next iteration. 569 570 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 571 Self->_ParkEvent->reset() ; 572 OrderAccess::fence() ; 573 } 574 if (_succ == Self) _succ = NULL ; 575 576 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 577 OrderAccess::fence() ; 578 } 579 580 // Egress : 581 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 582 // Normally we'll find Self on the EntryList . 583 // From the perspective of the lock owner (this thread), the 584 // EntryList is stable and cxq is prepend-only. 585 // The head of cxq is volatile but the interior is stable. 586 // In addition, Self.TState is stable. 587 588 assert (_owner == Self , "invariant") ; 589 assert (object() != NULL , "invariant") ; 590 // I'd like to write: 591 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 592 // but as we're at a safepoint that's not safe. 593 594 UnlinkAfterAcquire (Self, &node) ; 595 if (_succ == Self) _succ = NULL ; 596 597 assert (_succ != Self, "invariant") ; 598 if (_Responsible == Self) { 599 _Responsible = NULL ; 600 // Dekker pivot-point. 601 // Consider OrderAccess::storeload() here 602 603 // We may leave threads on cxq|EntryList without a designated 604 // "Responsible" thread. This is benign. When this thread subsequently 605 // exits the monitor it can "see" such preexisting "old" threads -- 606 // threads that arrived on the cxq|EntryList before the fence, above -- 607 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 608 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 609 // non-null and elect a new "Responsible" timer thread. 610 // 611 // This thread executes: 612 // ST Responsible=null; MEMBAR (in enter epilog - here) 613 // LD cxq|EntryList (in subsequent exit) 614 // 615 // Entering threads in the slow/contended path execute: 616 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 617 // The (ST cxq; MEMBAR) is accomplished with CAS(). 618 // 619 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 620 // exit operation from floating above the ST Responsible=null. 621 // 622 // In *practice* however, EnterI() is always followed by some atomic 623 // operation such as the decrement of _count in ::enter(). Those atomics 624 // obviate the need for the explicit MEMBAR, above. 625 } 626 627 // We've acquired ownership with CAS(). 628 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 629 // But since the CAS() this thread may have also stored into _succ, 630 // EntryList, cxq or Responsible. These meta-data updates must be 631 // visible __before this thread subsequently drops the lock. 632 // Consider what could occur if we didn't enforce this constraint -- 633 // STs to monitor meta-data and user-data could reorder with (become 634 // visible after) the ST in exit that drops ownership of the lock. 635 // Some other thread could then acquire the lock, but observe inconsistent 636 // or old monitor meta-data and heap data. That violates the JMM. 637 // To that end, the 1-0 exit() operation must have at least STST|LDST 638 // "release" barrier semantics. Specifically, there must be at least a 639 // STST|LDST barrier in exit() before the ST of null into _owner that drops 640 // the lock. The barrier ensures that changes to monitor meta-data and data 641 // protected by the lock will be visible before we release the lock, and 642 // therefore before some other thread (CPU) has a chance to acquire the lock. 643 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 644 // 645 // Critically, any prior STs to _succ or EntryList must be visible before 646 // the ST of null into _owner in the *subsequent* (following) corresponding 647 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 648 // execute a serializing instruction. 649 650 if (SyncFlags & 8) { 651 OrderAccess::fence() ; 652 } 653 return ; 654 } 655 656 // ReenterI() is a specialized inline form of the latter half of the 657 // contended slow-path from EnterI(). We use ReenterI() only for 658 // monitor reentry in wait(). 659 // 660 // In the future we should reconcile EnterI() and ReenterI(), adding 661 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 662 // loop accordingly. 663 664 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 665 assert (Self != NULL , "invariant") ; 666 assert (SelfNode != NULL , "invariant") ; 667 assert (SelfNode->_thread == Self , "invariant") ; 668 assert (_waiters > 0 , "invariant") ; 669 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 670 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 671 JavaThread * jt = (JavaThread *) Self ; 672 673 int nWakeups = 0 ; 674 for (;;) { 675 ObjectWaiter::TStates v = SelfNode->TState ; 676 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 677 assert (_owner != Self, "invariant") ; 678 679 if (TryLock (Self) > 0) break ; 680 if (TrySpin (Self) > 0) break ; 681 682 TEVENT (Wait Reentry - parking) ; 683 684 // State transition wrappers around park() ... 685 // ReenterI() wisely defers state transitions until 686 // it's clear we must park the thread. 687 { 688 OSThreadContendState osts(Self->osthread()); 689 ThreadBlockInVM tbivm(jt); 690 691 // cleared by handle_special_suspend_equivalent_condition() 692 // or java_suspend_self() 693 jt->set_suspend_equivalent(); 694 if (SyncFlags & 1) { 695 Self->_ParkEvent->park ((jlong)1000) ; 696 } else { 697 Self->_ParkEvent->park () ; 698 } 699 700 // were we externally suspended while we were waiting? 701 for (;;) { 702 if (!ExitSuspendEquivalent (jt)) break ; 703 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 704 jt->java_suspend_self(); 705 jt->set_suspend_equivalent(); 706 } 707 } 708 709 // Try again, but just so we distinguish between futile wakeups and 710 // successful wakeups. The following test isn't algorithmically 711 // necessary, but it helps us maintain sensible statistics. 712 if (TryLock(Self) > 0) break ; 713 714 // The lock is still contested. 715 // Keep a tally of the # of futile wakeups. 716 // Note that the counter is not protected by a lock or updated by atomics. 717 // That is by design - we trade "lossy" counters which are exposed to 718 // races during updates for a lower probe effect. 719 TEVENT (Wait Reentry - futile wakeup) ; 720 ++ nWakeups ; 721 722 // Assuming this is not a spurious wakeup we'll normally 723 // find that _succ == Self. 724 if (_succ == Self) _succ = NULL ; 725 726 // Invariant: after clearing _succ a contending thread 727 // *must* retry _owner before parking. 728 OrderAccess::fence() ; 729 730 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 731 ObjectMonitor::_sync_FutileWakeups->inc() ; 732 } 733 } 734 735 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 736 // Normally we'll find Self on the EntryList. 737 // Unlinking from the EntryList is constant-time and atomic-free. 738 // From the perspective of the lock owner (this thread), the 739 // EntryList is stable and cxq is prepend-only. 740 // The head of cxq is volatile but the interior is stable. 741 // In addition, Self.TState is stable. 742 743 assert (_owner == Self, "invariant") ; 744 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 745 UnlinkAfterAcquire (Self, SelfNode) ; 746 if (_succ == Self) _succ = NULL ; 747 assert (_succ != Self, "invariant") ; 748 SelfNode->TState = ObjectWaiter::TS_RUN ; 749 OrderAccess::fence() ; // see comments at the end of EnterI() 750 } 751 752 // after the thread acquires the lock in ::enter(). Equally, we could defer 753 // unlinking the thread until ::exit()-time. 754 755 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 756 { 757 assert (_owner == Self, "invariant") ; 758 assert (SelfNode->_thread == Self, "invariant") ; 759 760 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 761 // Normal case: remove Self from the DLL EntryList . 762 // This is a constant-time operation. 763 ObjectWaiter * nxt = SelfNode->_next ; 764 ObjectWaiter * prv = SelfNode->_prev ; 765 if (nxt != NULL) nxt->_prev = prv ; 766 if (prv != NULL) prv->_next = nxt ; 767 if (SelfNode == _EntryList ) _EntryList = nxt ; 768 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 769 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 770 TEVENT (Unlink from EntryList) ; 771 } else { 772 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 773 // Inopportune interleaving -- Self is still on the cxq. 774 // This usually means the enqueue of self raced an exiting thread. 775 // Normally we'll find Self near the front of the cxq, so 776 // dequeueing is typically fast. If needbe we can accelerate 777 // this with some MCS/CHL-like bidirectional list hints and advisory 778 // back-links so dequeueing from the interior will normally operate 779 // in constant-time. 780 // Dequeue Self from either the head (with CAS) or from the interior 781 // with a linear-time scan and normal non-atomic memory operations. 782 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 783 // and then unlink Self from EntryList. We have to drain eventually, 784 // so it might as well be now. 785 786 ObjectWaiter * v = _cxq ; 787 assert (v != NULL, "invariant") ; 788 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 789 // The CAS above can fail from interference IFF a "RAT" arrived. 790 // In that case Self must be in the interior and can no longer be 791 // at the head of cxq. 792 if (v == SelfNode) { 793 assert (_cxq != v, "invariant") ; 794 v = _cxq ; // CAS above failed - start scan at head of list 795 } 796 ObjectWaiter * p ; 797 ObjectWaiter * q = NULL ; 798 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 799 q = p ; 800 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 801 } 802 assert (v != SelfNode, "invariant") ; 803 assert (p == SelfNode, "Node not found on cxq") ; 804 assert (p != _cxq, "invariant") ; 805 assert (q != NULL, "invariant") ; 806 assert (q->_next == p, "invariant") ; 807 q->_next = p->_next ; 808 } 809 TEVENT (Unlink from cxq) ; 810 } 811 812 // Diagnostic hygiene ... 813 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 814 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 815 SelfNode->TState = ObjectWaiter::TS_RUN ; 816 } 817 818 // ----------------------------------------------------------------------------- 819 // Exit support 820 // 821 // exit() 822 // ~~~~~~ 823 // Note that the collector can't reclaim the objectMonitor or deflate 824 // the object out from underneath the thread calling ::exit() as the 825 // thread calling ::exit() never transitions to a stable state. 826 // This inhibits GC, which in turn inhibits asynchronous (and 827 // inopportune) reclamation of "this". 828 // 829 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 830 // There's one exception to the claim above, however. EnterI() can call 831 // exit() to drop a lock if the acquirer has been externally suspended. 832 // In that case exit() is called with _thread_state as _thread_blocked, 833 // but the monitor's _count field is > 0, which inhibits reclamation. 834 // 835 // 1-0 exit 836 // ~~~~~~~~ 837 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 838 // the fast-path operators have been optimized so the common ::exit() 839 // operation is 1-0. See i486.ad fast_unlock(), for instance. 840 // The code emitted by fast_unlock() elides the usual MEMBAR. This 841 // greatly improves latency -- MEMBAR and CAS having considerable local 842 // latency on modern processors -- but at the cost of "stranding". Absent the 843 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 844 // ::enter() path, resulting in the entering thread being stranding 845 // and a progress-liveness failure. Stranding is extremely rare. 846 // We use timers (timed park operations) & periodic polling to detect 847 // and recover from stranding. Potentially stranded threads periodically 848 // wake up and poll the lock. See the usage of the _Responsible variable. 849 // 850 // The CAS() in enter provides for safety and exclusion, while the CAS or 851 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 852 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 853 // We detect and recover from stranding with timers. 854 // 855 // If a thread transiently strands it'll park until (a) another 856 // thread acquires the lock and then drops the lock, at which time the 857 // exiting thread will notice and unpark the stranded thread, or, (b) 858 // the timer expires. If the lock is high traffic then the stranding latency 859 // will be low due to (a). If the lock is low traffic then the odds of 860 // stranding are lower, although the worst-case stranding latency 861 // is longer. Critically, we don't want to put excessive load in the 862 // platform's timer subsystem. We want to minimize both the timer injection 863 // rate (timers created/sec) as well as the number of timers active at 864 // any one time. (more precisely, we want to minimize timer-seconds, which is 865 // the integral of the # of active timers at any instant over time). 866 // Both impinge on OS scalability. Given that, at most one thread parked on 867 // a monitor will use a timer. 868 869 void ATTR ObjectMonitor::exit(TRAPS) { 870 Thread * Self = THREAD ; 871 if (THREAD != _owner) { 872 if (THREAD->is_lock_owned((address) _owner)) { 873 // Transmute _owner from a BasicLock pointer to a Thread address. 874 // We don't need to hold _mutex for this transition. 875 // Non-null to Non-null is safe as long as all readers can 876 // tolerate either flavor. 877 assert (_recursions == 0, "invariant") ; 878 _owner = THREAD ; 879 _recursions = 0 ; 880 OwnerIsThread = 1 ; 881 } else { 882 // NOTE: we need to handle unbalanced monitor enter/exit 883 // in native code by throwing an exception. 884 // TODO: Throw an IllegalMonitorStateException ? 885 TEVENT (Exit - Throw IMSX) ; 886 assert(false, "Non-balanced monitor enter/exit!"); 887 if (false) { 888 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 889 } 890 return; 891 } 892 } 893 894 if (_recursions != 0) { 895 _recursions--; // this is simple recursive enter 896 TEVENT (Inflated exit - recursive) ; 897 return ; 898 } 899 900 // Invariant: after setting Responsible=null an thread must execute 901 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 902 if ((SyncFlags & 4) == 0) { 903 _Responsible = NULL ; 904 } 905 906 for (;;) { 907 assert (THREAD == _owner, "invariant") ; 908 909 910 if (Knob_ExitPolicy == 0) { 911 // release semantics: prior loads and stores from within the critical section 912 // must not float (reorder) past the following store that drops the lock. 913 // On SPARC that requires MEMBAR #loadstore|#storestore. 914 // But of course in TSO #loadstore|#storestore is not required. 915 // I'd like to write one of the following: 916 // A. OrderAccess::release() ; _owner = NULL 917 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 918 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 919 // store into a _dummy variable. That store is not needed, but can result 920 // in massive wasteful coherency traffic on classic SMP systems. 921 // Instead, I use release_store(), which is implemented as just a simple 922 // ST on x64, x86 and SPARC. 923 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 924 OrderAccess::storeload() ; // See if we need to wake a successor 925 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 926 TEVENT (Inflated exit - simple egress) ; 927 return ; 928 } 929 TEVENT (Inflated exit - complex egress) ; 930 931 // Normally the exiting thread is responsible for ensuring succession, 932 // but if other successors are ready or other entering threads are spinning 933 // then this thread can simply store NULL into _owner and exit without 934 // waking a successor. The existence of spinners or ready successors 935 // guarantees proper succession (liveness). Responsibility passes to the 936 // ready or running successors. The exiting thread delegates the duty. 937 // More precisely, if a successor already exists this thread is absolved 938 // of the responsibility of waking (unparking) one. 939 // 940 // The _succ variable is critical to reducing futile wakeup frequency. 941 // _succ identifies the "heir presumptive" thread that has been made 942 // ready (unparked) but that has not yet run. We need only one such 943 // successor thread to guarantee progress. 944 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 945 // section 3.3 "Futile Wakeup Throttling" for details. 946 // 947 // Note that spinners in Enter() also set _succ non-null. 948 // In the current implementation spinners opportunistically set 949 // _succ so that exiting threads might avoid waking a successor. 950 // Another less appealing alternative would be for the exiting thread 951 // to drop the lock and then spin briefly to see if a spinner managed 952 // to acquire the lock. If so, the exiting thread could exit 953 // immediately without waking a successor, otherwise the exiting 954 // thread would need to dequeue and wake a successor. 955 // (Note that we'd need to make the post-drop spin short, but no 956 // shorter than the worst-case round-trip cache-line migration time. 957 // The dropped lock needs to become visible to the spinner, and then 958 // the acquisition of the lock by the spinner must become visible to 959 // the exiting thread). 960 // 961 962 // It appears that an heir-presumptive (successor) must be made ready. 963 // Only the current lock owner can manipulate the EntryList or 964 // drain _cxq, so we need to reacquire the lock. If we fail 965 // to reacquire the lock the responsibility for ensuring succession 966 // falls to the new owner. 967 // 968 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 969 return ; 970 } 971 TEVENT (Exit - Reacquired) ; 972 } else { 973 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 974 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 975 OrderAccess::storeload() ; 976 // Ratify the previously observed values. 977 if (_cxq == NULL || _succ != NULL) { 978 TEVENT (Inflated exit - simple egress) ; 979 return ; 980 } 981 982 // inopportune interleaving -- the exiting thread (this thread) 983 // in the fast-exit path raced an entering thread in the slow-enter 984 // path. 985 // We have two choices: 986 // A. Try to reacquire the lock. 987 // If the CAS() fails return immediately, otherwise 988 // we either restart/rerun the exit operation, or simply 989 // fall-through into the code below which wakes a successor. 990 // B. If the elements forming the EntryList|cxq are TSM 991 // we could simply unpark() the lead thread and return 992 // without having set _succ. 993 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 994 TEVENT (Inflated exit - reacquired succeeded) ; 995 return ; 996 } 997 TEVENT (Inflated exit - reacquired failed) ; 998 } else { 999 TEVENT (Inflated exit - complex egress) ; 1000 } 1001 } 1002 1003 guarantee (_owner == THREAD, "invariant") ; 1004 1005 ObjectWaiter * w = NULL ; 1006 int QMode = Knob_QMode ; 1007 1008 if (QMode == 2 && _cxq != NULL) { 1009 // QMode == 2 : cxq has precedence over EntryList. 1010 // Try to directly wake a successor from the cxq. 1011 // If successful, the successor will need to unlink itself from cxq. 1012 w = _cxq ; 1013 assert (w != NULL, "invariant") ; 1014 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1015 ExitEpilog (Self, w) ; 1016 return ; 1017 } 1018 1019 if (QMode == 3 && _cxq != NULL) { 1020 // Aggressively drain cxq into EntryList at the first opportunity. 1021 // This policy ensure that recently-run threads live at the head of EntryList. 1022 // Drain _cxq into EntryList - bulk transfer. 1023 // First, detach _cxq. 1024 // The following loop is tantamount to: w = swap (&cxq, NULL) 1025 w = _cxq ; 1026 for (;;) { 1027 assert (w != NULL, "Invariant") ; 1028 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1029 if (u == w) break ; 1030 w = u ; 1031 } 1032 assert (w != NULL , "invariant") ; 1033 1034 ObjectWaiter * q = NULL ; 1035 ObjectWaiter * p ; 1036 for (p = w ; p != NULL ; p = p->_next) { 1037 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1038 p->TState = ObjectWaiter::TS_ENTER ; 1039 p->_prev = q ; 1040 q = p ; 1041 } 1042 1043 // Append the RATs to the EntryList 1044 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1045 ObjectWaiter * Tail ; 1046 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1047 if (Tail == NULL) { 1048 _EntryList = w ; 1049 } else { 1050 Tail->_next = w ; 1051 w->_prev = Tail ; 1052 } 1053 1054 // Fall thru into code that tries to wake a successor from EntryList 1055 } 1056 1057 if (QMode == 4 && _cxq != NULL) { 1058 // Aggressively drain cxq into EntryList at the first opportunity. 1059 // This policy ensure that recently-run threads live at the head of EntryList. 1060 1061 // Drain _cxq into EntryList - bulk transfer. 1062 // First, detach _cxq. 1063 // The following loop is tantamount to: w = swap (&cxq, NULL) 1064 w = _cxq ; 1065 for (;;) { 1066 assert (w != NULL, "Invariant") ; 1067 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1068 if (u == w) break ; 1069 w = u ; 1070 } 1071 assert (w != NULL , "invariant") ; 1072 1073 ObjectWaiter * q = NULL ; 1074 ObjectWaiter * p ; 1075 for (p = w ; p != NULL ; p = p->_next) { 1076 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1077 p->TState = ObjectWaiter::TS_ENTER ; 1078 p->_prev = q ; 1079 q = p ; 1080 } 1081 1082 // Prepend the RATs to the EntryList 1083 if (_EntryList != NULL) { 1084 q->_next = _EntryList ; 1085 _EntryList->_prev = q ; 1086 } 1087 _EntryList = w ; 1088 1089 // Fall thru into code that tries to wake a successor from EntryList 1090 } 1091 1092 w = _EntryList ; 1093 if (w != NULL) { 1094 // I'd like to write: guarantee (w->_thread != Self). 1095 // But in practice an exiting thread may find itself on the EntryList. 1096 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1097 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1098 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1099 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1100 // release the lock "O". T2 resumes immediately after the ST of null into 1101 // _owner, above. T2 notices that the EntryList is populated, so it 1102 // reacquires the lock and then finds itself on the EntryList. 1103 // Given all that, we have to tolerate the circumstance where "w" is 1104 // associated with Self. 1105 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1106 ExitEpilog (Self, w) ; 1107 return ; 1108 } 1109 1110 // If we find that both _cxq and EntryList are null then just 1111 // re-run the exit protocol from the top. 1112 w = _cxq ; 1113 if (w == NULL) continue ; 1114 1115 // Drain _cxq into EntryList - bulk transfer. 1116 // First, detach _cxq. 1117 // The following loop is tantamount to: w = swap (&cxq, NULL) 1118 for (;;) { 1119 assert (w != NULL, "Invariant") ; 1120 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1121 if (u == w) break ; 1122 w = u ; 1123 } 1124 TEVENT (Inflated exit - drain cxq into EntryList) ; 1125 1126 assert (w != NULL , "invariant") ; 1127 assert (_EntryList == NULL , "invariant") ; 1128 1129 // Convert the LIFO SLL anchored by _cxq into a DLL. 1130 // The list reorganization step operates in O(LENGTH(w)) time. 1131 // It's critical that this step operate quickly as 1132 // "Self" still holds the outer-lock, restricting parallelism 1133 // and effectively lengthening the critical section. 1134 // Invariant: s chases t chases u. 1135 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1136 // we have faster access to the tail. 1137 1138 if (QMode == 1) { 1139 // QMode == 1 : drain cxq to EntryList, reversing order 1140 // We also reverse the order of the list. 1141 ObjectWaiter * s = NULL ; 1142 ObjectWaiter * t = w ; 1143 ObjectWaiter * u = NULL ; 1144 while (t != NULL) { 1145 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1146 t->TState = ObjectWaiter::TS_ENTER ; 1147 u = t->_next ; 1148 t->_prev = u ; 1149 t->_next = s ; 1150 s = t; 1151 t = u ; 1152 } 1153 _EntryList = s ; 1154 assert (s != NULL, "invariant") ; 1155 } else { 1156 // QMode == 0 or QMode == 2 1157 _EntryList = w ; 1158 ObjectWaiter * q = NULL ; 1159 ObjectWaiter * p ; 1160 for (p = w ; p != NULL ; p = p->_next) { 1161 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1162 p->TState = ObjectWaiter::TS_ENTER ; 1163 p->_prev = q ; 1164 q = p ; 1165 } 1166 } 1167 1168 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1169 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1170 1171 // See if we can abdicate to a spinner instead of waking a thread. 1172 // A primary goal of the implementation is to reduce the 1173 // context-switch rate. 1174 if (_succ != NULL) continue; 1175 1176 w = _EntryList ; 1177 if (w != NULL) { 1178 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1179 ExitEpilog (Self, w) ; 1180 return ; 1181 } 1182 } 1183 } 1184 1185 // ExitSuspendEquivalent: 1186 // A faster alternate to handle_special_suspend_equivalent_condition() 1187 // 1188 // handle_special_suspend_equivalent_condition() unconditionally 1189 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1190 // operations have high latency. Note that in ::enter() we call HSSEC 1191 // while holding the monitor, so we effectively lengthen the critical sections. 1192 // 1193 // There are a number of possible solutions: 1194 // 1195 // A. To ameliorate the problem we might also defer state transitions 1196 // to as late as possible -- just prior to parking. 1197 // Given that, we'd call HSSEC after having returned from park(), 1198 // but before attempting to acquire the monitor. This is only a 1199 // partial solution. It avoids calling HSSEC while holding the 1200 // monitor (good), but it still increases successor reacquisition latency -- 1201 // the interval between unparking a successor and the time the successor 1202 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1203 // If we use this technique we can also avoid EnterI()-exit() loop 1204 // in ::enter() where we iteratively drop the lock and then attempt 1205 // to reacquire it after suspending. 1206 // 1207 // B. In the future we might fold all the suspend bits into a 1208 // composite per-thread suspend flag and then update it with CAS(). 1209 // Alternately, a Dekker-like mechanism with multiple variables 1210 // would suffice: 1211 // ST Self->_suspend_equivalent = false 1212 // MEMBAR 1213 // LD Self_>_suspend_flags 1214 // 1215 1216 1217 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1218 int Mode = Knob_FastHSSEC ; 1219 if (Mode && !jSelf->is_external_suspend()) { 1220 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1221 jSelf->clear_suspend_equivalent() ; 1222 if (2 == Mode) OrderAccess::storeload() ; 1223 if (!jSelf->is_external_suspend()) return false ; 1224 // We raced a suspension -- fall thru into the slow path 1225 TEVENT (ExitSuspendEquivalent - raced) ; 1226 jSelf->set_suspend_equivalent() ; 1227 } 1228 return jSelf->handle_special_suspend_equivalent_condition() ; 1229 } 1230 1231 1232 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1233 assert (_owner == Self, "invariant") ; 1234 1235 // Exit protocol: 1236 // 1. ST _succ = wakee 1237 // 2. membar #loadstore|#storestore; 1238 // 2. ST _owner = NULL 1239 // 3. unpark(wakee) 1240 1241 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1242 ParkEvent * Trigger = Wakee->_event ; 1243 1244 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1245 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1246 // out-of-scope (non-extant). 1247 Wakee = NULL ; 1248 1249 // Drop the lock 1250 OrderAccess::release_store_ptr (&_owner, NULL) ; 1251 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1252 1253 if (SafepointSynchronize::do_call_back()) { 1254 TEVENT (unpark before SAFEPOINT) ; 1255 } 1256 1257 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1258 Trigger->unpark() ; 1259 1260 // Maintain stats and report events to JVMTI 1261 if (ObjectMonitor::_sync_Parks != NULL) { 1262 ObjectMonitor::_sync_Parks->inc() ; 1263 } 1264 } 1265 1266 1267 // ----------------------------------------------------------------------------- 1268 // Class Loader deadlock handling. 1269 // 1270 // complete_exit exits a lock returning recursion count 1271 // complete_exit/reenter operate as a wait without waiting 1272 // complete_exit requires an inflated monitor 1273 // The _owner field is not always the Thread addr even with an 1274 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1275 // thread due to contention. 1276 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1277 Thread * const Self = THREAD; 1278 assert(Self->is_Java_thread(), "Must be Java thread!"); 1279 JavaThread *jt = (JavaThread *)THREAD; 1280 1281 DeferredInitialize(); 1282 1283 if (THREAD != _owner) { 1284 if (THREAD->is_lock_owned ((address)_owner)) { 1285 assert(_recursions == 0, "internal state error"); 1286 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1287 _recursions = 0 ; 1288 OwnerIsThread = 1 ; 1289 } 1290 } 1291 1292 guarantee(Self == _owner, "complete_exit not owner"); 1293 intptr_t save = _recursions; // record the old recursion count 1294 _recursions = 0; // set the recursion level to be 0 1295 exit (Self) ; // exit the monitor 1296 guarantee (_owner != Self, "invariant"); 1297 return save; 1298 } 1299 1300 // reenter() enters a lock and sets recursion count 1301 // complete_exit/reenter operate as a wait without waiting 1302 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1303 Thread * const Self = THREAD; 1304 assert(Self->is_Java_thread(), "Must be Java thread!"); 1305 JavaThread *jt = (JavaThread *)THREAD; 1306 1307 guarantee(_owner != Self, "reenter already owner"); 1308 enter (THREAD); // enter the monitor 1309 guarantee (_recursions == 0, "reenter recursion"); 1310 _recursions = recursions; 1311 return; 1312 } 1313 1314 1315 // ----------------------------------------------------------------------------- 1316 // A macro is used below because there may already be a pending 1317 // exception which should not abort the execution of the routines 1318 // which use this (which is why we don't put this into check_slow and 1319 // call it with a CHECK argument). 1320 1321 #define CHECK_OWNER() \ 1322 do { \ 1323 if (THREAD != _owner) { \ 1324 if (THREAD->is_lock_owned((address) _owner)) { \ 1325 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1326 _recursions = 0; \ 1327 OwnerIsThread = 1 ; \ 1328 } else { \ 1329 TEVENT (Throw IMSX) ; \ 1330 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1331 } \ 1332 } \ 1333 } while (false) 1334 1335 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1336 // TODO-FIXME: remove check_slow() -- it's likely dead. 1337 1338 void ObjectMonitor::check_slow(TRAPS) { 1339 TEVENT (check_slow - throw IMSX) ; 1340 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1341 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1342 } 1343 1344 static int Adjust (volatile int * adr, int dx) { 1345 int v ; 1346 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1347 return v ; 1348 } 1349 // ----------------------------------------------------------------------------- 1350 // Wait/Notify/NotifyAll 1351 // 1352 // Note: a subset of changes to ObjectMonitor::wait() 1353 // will need to be replicated in complete_exit above 1354 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1355 Thread * const Self = THREAD ; 1356 assert(Self->is_Java_thread(), "Must be Java thread!"); 1357 JavaThread *jt = (JavaThread *)THREAD; 1358 1359 DeferredInitialize () ; 1360 1361 // Throw IMSX or IEX. 1362 CHECK_OWNER(); 1363 1364 // check for a pending interrupt 1365 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1366 // post monitor waited event. Note that this is past-tense, we are done waiting. 1367 if (JvmtiExport::should_post_monitor_waited()) { 1368 // Note: 'false' parameter is passed here because the 1369 // wait was not timed out due to thread interrupt. 1370 JvmtiExport::post_monitor_waited(jt, this, false); 1371 } 1372 TEVENT (Wait - Throw IEX) ; 1373 THROW(vmSymbols::java_lang_InterruptedException()); 1374 return ; 1375 } 1376 TEVENT (Wait) ; 1377 1378 assert (Self->_Stalled == 0, "invariant") ; 1379 Self->_Stalled = intptr_t(this) ; 1380 jt->set_current_waiting_monitor(this); 1381 1382 // create a node to be put into the queue 1383 // Critically, after we reset() the event but prior to park(), we must check 1384 // for a pending interrupt. 1385 ObjectWaiter node(Self); 1386 node.TState = ObjectWaiter::TS_WAIT ; 1387 Self->_ParkEvent->reset() ; 1388 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1389 1390 // Enter the waiting queue, which is a circular doubly linked list in this case 1391 // but it could be a priority queue or any data structure. 1392 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1393 // by the the owner of the monitor *except* in the case where park() 1394 // returns because of a timeout of interrupt. Contention is exceptionally rare 1395 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1396 1397 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1398 AddWaiter (&node) ; 1399 Thread::SpinRelease (&_WaitSetLock) ; 1400 1401 if ((SyncFlags & 4) == 0) { 1402 _Responsible = NULL ; 1403 } 1404 intptr_t save = _recursions; // record the old recursion count 1405 _waiters++; // increment the number of waiters 1406 _recursions = 0; // set the recursion level to be 1 1407 exit (Self) ; // exit the monitor 1408 guarantee (_owner != Self, "invariant") ; 1409 1410 // As soon as the ObjectMonitor's ownership is dropped in the exit() 1411 // call above, another thread can enter() the ObjectMonitor, do the 1412 // notify(), and exit() the ObjectMonitor. If the other thread's 1413 // exit() call chooses this thread as the successor and the unpark() 1414 // call happens to occur while this thread is posting a 1415 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event 1416 // handler using RawMonitors and consuming the unpark(). 1417 // 1418 // To avoid the problem, we re-post the event. This does no harm 1419 // even if the original unpark() was not consumed because we are the 1420 // chosen successor for this monitor. 1421 if (node._notified != 0 && _succ == Self) { 1422 node._event->unpark(); 1423 } 1424 1425 // The thread is on the WaitSet list - now park() it. 1426 // On MP systems it's conceivable that a brief spin before we park 1427 // could be profitable. 1428 // 1429 // TODO-FIXME: change the following logic to a loop of the form 1430 // while (!timeout && !interrupted && _notified == 0) park() 1431 1432 int ret = OS_OK ; 1433 int WasNotified = 0 ; 1434 { // State transition wrappers 1435 OSThread* osthread = Self->osthread(); 1436 OSThreadWaitState osts(osthread, true); 1437 { 1438 ThreadBlockInVM tbivm(jt); 1439 // Thread is in thread_blocked state and oop access is unsafe. 1440 jt->set_suspend_equivalent(); 1441 1442 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1443 // Intentionally empty 1444 } else 1445 if (node._notified == 0) { 1446 if (millis <= 0) { 1447 Self->_ParkEvent->park () ; 1448 } else { 1449 ret = Self->_ParkEvent->park (millis) ; 1450 } 1451 } 1452 1453 // were we externally suspended while we were waiting? 1454 if (ExitSuspendEquivalent (jt)) { 1455 // TODO-FIXME: add -- if succ == Self then succ = null. 1456 jt->java_suspend_self(); 1457 } 1458 1459 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1460 1461 1462 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1463 // from the WaitSet to the EntryList. 1464 // See if we need to remove Node from the WaitSet. 1465 // We use double-checked locking to avoid grabbing _WaitSetLock 1466 // if the thread is not on the wait queue. 1467 // 1468 // Note that we don't need a fence before the fetch of TState. 1469 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1470 // written by the is thread. (perhaps the fetch might even be satisfied 1471 // by a look-aside into the processor's own store buffer, although given 1472 // the length of the code path between the prior ST and this load that's 1473 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1474 // then we'll acquire the lock and then re-fetch a fresh TState value. 1475 // That is, we fail toward safety. 1476 1477 if (node.TState == ObjectWaiter::TS_WAIT) { 1478 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1479 if (node.TState == ObjectWaiter::TS_WAIT) { 1480 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1481 assert(node._notified == 0, "invariant"); 1482 node.TState = ObjectWaiter::TS_RUN ; 1483 } 1484 Thread::SpinRelease (&_WaitSetLock) ; 1485 } 1486 1487 // The thread is now either on off-list (TS_RUN), 1488 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1489 // The Node's TState variable is stable from the perspective of this thread. 1490 // No other threads will asynchronously modify TState. 1491 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1492 OrderAccess::loadload() ; 1493 if (_succ == Self) _succ = NULL ; 1494 WasNotified = node._notified ; 1495 1496 // Reentry phase -- reacquire the monitor. 1497 // re-enter contended monitor after object.wait(). 1498 // retain OBJECT_WAIT state until re-enter successfully completes 1499 // Thread state is thread_in_vm and oop access is again safe, 1500 // although the raw address of the object may have changed. 1501 // (Don't cache naked oops over safepoints, of course). 1502 1503 // post monitor waited event. Note that this is past-tense, we are done waiting. 1504 if (JvmtiExport::should_post_monitor_waited()) { 1505 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1506 } 1507 OrderAccess::fence() ; 1508 1509 assert (Self->_Stalled != 0, "invariant") ; 1510 Self->_Stalled = 0 ; 1511 1512 assert (_owner != Self, "invariant") ; 1513 ObjectWaiter::TStates v = node.TState ; 1514 if (v == ObjectWaiter::TS_RUN) { 1515 enter (Self) ; 1516 } else { 1517 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1518 ReenterI (Self, &node) ; 1519 node.wait_reenter_end(this); 1520 } 1521 1522 // Self has reacquired the lock. 1523 // Lifecycle - the node representing Self must not appear on any queues. 1524 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1525 // want residual elements associated with this thread left on any lists. 1526 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1527 assert (_owner == Self, "invariant") ; 1528 assert (_succ != Self , "invariant") ; 1529 } // OSThreadWaitState() 1530 1531 jt->set_current_waiting_monitor(NULL); 1532 1533 guarantee (_recursions == 0, "invariant") ; 1534 _recursions = save; // restore the old recursion count 1535 _waiters--; // decrement the number of waiters 1536 1537 // Verify a few postconditions 1538 assert (_owner == Self , "invariant") ; 1539 assert (_succ != Self , "invariant") ; 1540 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1541 1542 if (SyncFlags & 32) { 1543 OrderAccess::fence() ; 1544 } 1545 1546 // check if the notification happened 1547 if (!WasNotified) { 1548 // no, it could be timeout or Thread.interrupt() or both 1549 // check for interrupt event, otherwise it is timeout 1550 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1551 TEVENT (Wait - throw IEX from epilog) ; 1552 THROW(vmSymbols::java_lang_InterruptedException()); 1553 } 1554 } 1555 1556 // NOTE: Spurious wake up will be consider as timeout. 1557 // Monitor notify has precedence over thread interrupt. 1558 } 1559 1560 1561 // Consider: 1562 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1563 // then instead of transferring a thread from the WaitSet to the EntryList 1564 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1565 1566 void ObjectMonitor::notify(TRAPS) { 1567 CHECK_OWNER(); 1568 if (_WaitSet == NULL) { 1569 TEVENT (Empty-Notify) ; 1570 return ; 1571 } 1572 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1573 1574 int Policy = Knob_MoveNotifyee ; 1575 1576 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1577 ObjectWaiter * iterator = DequeueWaiter() ; 1578 if (iterator != NULL) { 1579 TEVENT (Notify1 - Transfer) ; 1580 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1581 guarantee (iterator->_notified == 0, "invariant") ; 1582 if (Policy != 4) { 1583 iterator->TState = ObjectWaiter::TS_ENTER ; 1584 } 1585 iterator->_notified = 1 ; 1586 1587 ObjectWaiter * List = _EntryList ; 1588 if (List != NULL) { 1589 assert (List->_prev == NULL, "invariant") ; 1590 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1591 assert (List != iterator, "invariant") ; 1592 } 1593 1594 if (Policy == 0) { // prepend to EntryList 1595 if (List == NULL) { 1596 iterator->_next = iterator->_prev = NULL ; 1597 _EntryList = iterator ; 1598 } else { 1599 List->_prev = iterator ; 1600 iterator->_next = List ; 1601 iterator->_prev = NULL ; 1602 _EntryList = iterator ; 1603 } 1604 } else 1605 if (Policy == 1) { // append to EntryList 1606 if (List == NULL) { 1607 iterator->_next = iterator->_prev = NULL ; 1608 _EntryList = iterator ; 1609 } else { 1610 // CONSIDER: finding the tail currently requires a linear-time walk of 1611 // the EntryList. We can make tail access constant-time by converting to 1612 // a CDLL instead of using our current DLL. 1613 ObjectWaiter * Tail ; 1614 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1615 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1616 Tail->_next = iterator ; 1617 iterator->_prev = Tail ; 1618 iterator->_next = NULL ; 1619 } 1620 } else 1621 if (Policy == 2) { // prepend to cxq 1622 // prepend to cxq 1623 if (List == NULL) { 1624 iterator->_next = iterator->_prev = NULL ; 1625 _EntryList = iterator ; 1626 } else { 1627 iterator->TState = ObjectWaiter::TS_CXQ ; 1628 for (;;) { 1629 ObjectWaiter * Front = _cxq ; 1630 iterator->_next = Front ; 1631 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1632 break ; 1633 } 1634 } 1635 } 1636 } else 1637 if (Policy == 3) { // append to cxq 1638 iterator->TState = ObjectWaiter::TS_CXQ ; 1639 for (;;) { 1640 ObjectWaiter * Tail ; 1641 Tail = _cxq ; 1642 if (Tail == NULL) { 1643 iterator->_next = NULL ; 1644 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1645 break ; 1646 } 1647 } else { 1648 while (Tail->_next != NULL) Tail = Tail->_next ; 1649 Tail->_next = iterator ; 1650 iterator->_prev = Tail ; 1651 iterator->_next = NULL ; 1652 break ; 1653 } 1654 } 1655 } else { 1656 ParkEvent * ev = iterator->_event ; 1657 iterator->TState = ObjectWaiter::TS_RUN ; 1658 OrderAccess::fence() ; 1659 ev->unpark() ; 1660 } 1661 1662 if (Policy < 4) { 1663 iterator->wait_reenter_begin(this); 1664 } 1665 1666 // _WaitSetLock protects the wait queue, not the EntryList. We could 1667 // move the add-to-EntryList operation, above, outside the critical section 1668 // protected by _WaitSetLock. In practice that's not useful. With the 1669 // exception of wait() timeouts and interrupts the monitor owner 1670 // is the only thread that grabs _WaitSetLock. There's almost no contention 1671 // on _WaitSetLock so it's not profitable to reduce the length of the 1672 // critical section. 1673 } 1674 1675 Thread::SpinRelease (&_WaitSetLock) ; 1676 1677 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1678 ObjectMonitor::_sync_Notifications->inc() ; 1679 } 1680 } 1681 1682 1683 void ObjectMonitor::notifyAll(TRAPS) { 1684 CHECK_OWNER(); 1685 ObjectWaiter* iterator; 1686 if (_WaitSet == NULL) { 1687 TEVENT (Empty-NotifyAll) ; 1688 return ; 1689 } 1690 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1691 1692 int Policy = Knob_MoveNotifyee ; 1693 int Tally = 0 ; 1694 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1695 1696 for (;;) { 1697 iterator = DequeueWaiter () ; 1698 if (iterator == NULL) break ; 1699 TEVENT (NotifyAll - Transfer1) ; 1700 ++Tally ; 1701 1702 // Disposition - what might we do with iterator ? 1703 // a. add it directly to the EntryList - either tail or head. 1704 // b. push it onto the front of the _cxq. 1705 // For now we use (a). 1706 1707 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1708 guarantee (iterator->_notified == 0, "invariant") ; 1709 iterator->_notified = 1 ; 1710 if (Policy != 4) { 1711 iterator->TState = ObjectWaiter::TS_ENTER ; 1712 } 1713 1714 ObjectWaiter * List = _EntryList ; 1715 if (List != NULL) { 1716 assert (List->_prev == NULL, "invariant") ; 1717 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1718 assert (List != iterator, "invariant") ; 1719 } 1720 1721 if (Policy == 0) { // prepend to EntryList 1722 if (List == NULL) { 1723 iterator->_next = iterator->_prev = NULL ; 1724 _EntryList = iterator ; 1725 } else { 1726 List->_prev = iterator ; 1727 iterator->_next = List ; 1728 iterator->_prev = NULL ; 1729 _EntryList = iterator ; 1730 } 1731 } else 1732 if (Policy == 1) { // append to EntryList 1733 if (List == NULL) { 1734 iterator->_next = iterator->_prev = NULL ; 1735 _EntryList = iterator ; 1736 } else { 1737 // CONSIDER: finding the tail currently requires a linear-time walk of 1738 // the EntryList. We can make tail access constant-time by converting to 1739 // a CDLL instead of using our current DLL. 1740 ObjectWaiter * Tail ; 1741 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1742 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1743 Tail->_next = iterator ; 1744 iterator->_prev = Tail ; 1745 iterator->_next = NULL ; 1746 } 1747 } else 1748 if (Policy == 2) { // prepend to cxq 1749 // prepend to cxq 1750 iterator->TState = ObjectWaiter::TS_CXQ ; 1751 for (;;) { 1752 ObjectWaiter * Front = _cxq ; 1753 iterator->_next = Front ; 1754 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1755 break ; 1756 } 1757 } 1758 } else 1759 if (Policy == 3) { // append to cxq 1760 iterator->TState = ObjectWaiter::TS_CXQ ; 1761 for (;;) { 1762 ObjectWaiter * Tail ; 1763 Tail = _cxq ; 1764 if (Tail == NULL) { 1765 iterator->_next = NULL ; 1766 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1767 break ; 1768 } 1769 } else { 1770 while (Tail->_next != NULL) Tail = Tail->_next ; 1771 Tail->_next = iterator ; 1772 iterator->_prev = Tail ; 1773 iterator->_next = NULL ; 1774 break ; 1775 } 1776 } 1777 } else { 1778 ParkEvent * ev = iterator->_event ; 1779 iterator->TState = ObjectWaiter::TS_RUN ; 1780 OrderAccess::fence() ; 1781 ev->unpark() ; 1782 } 1783 1784 if (Policy < 4) { 1785 iterator->wait_reenter_begin(this); 1786 } 1787 1788 // _WaitSetLock protects the wait queue, not the EntryList. We could 1789 // move the add-to-EntryList operation, above, outside the critical section 1790 // protected by _WaitSetLock. In practice that's not useful. With the 1791 // exception of wait() timeouts and interrupts the monitor owner 1792 // is the only thread that grabs _WaitSetLock. There's almost no contention 1793 // on _WaitSetLock so it's not profitable to reduce the length of the 1794 // critical section. 1795 } 1796 1797 Thread::SpinRelease (&_WaitSetLock) ; 1798 1799 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1800 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1801 } 1802 } 1803 1804 // ----------------------------------------------------------------------------- 1805 // Adaptive Spinning Support 1806 // 1807 // Adaptive spin-then-block - rational spinning 1808 // 1809 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1810 // algorithm. On high order SMP systems it would be better to start with 1811 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1812 // a contending thread could enqueue itself on the cxq and then spin locally 1813 // on a thread-specific variable such as its ParkEvent._Event flag. 1814 // That's left as an exercise for the reader. Note that global spinning is 1815 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1816 // low latency and massive bandwidth. 1817 // 1818 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1819 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1820 // (duration) or we can fix the count at approximately the duration of 1821 // a context switch and vary the frequency. Of course we could also 1822 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1823 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1824 // 1825 // This implementation varies the duration "D", where D varies with 1826 // the success rate of recent spin attempts. (D is capped at approximately 1827 // length of a round-trip context switch). The success rate for recent 1828 // spin attempts is a good predictor of the success rate of future spin 1829 // attempts. The mechanism adapts automatically to varying critical 1830 // section length (lock modality), system load and degree of parallelism. 1831 // D is maintained per-monitor in _SpinDuration and is initialized 1832 // optimistically. Spin frequency is fixed at 100%. 1833 // 1834 // Note that _SpinDuration is volatile, but we update it without locks 1835 // or atomics. The code is designed so that _SpinDuration stays within 1836 // a reasonable range even in the presence of races. The arithmetic 1837 // operations on _SpinDuration are closed over the domain of legal values, 1838 // so at worst a race will install and older but still legal value. 1839 // At the very worst this introduces some apparent non-determinism. 1840 // We might spin when we shouldn't or vice-versa, but since the spin 1841 // count are relatively short, even in the worst case, the effect is harmless. 1842 // 1843 // Care must be taken that a low "D" value does not become an 1844 // an absorbing state. Transient spinning failures -- when spinning 1845 // is overall profitable -- should not cause the system to converge 1846 // on low "D" values. We want spinning to be stable and predictable 1847 // and fairly responsive to change and at the same time we don't want 1848 // it to oscillate, become metastable, be "too" non-deterministic, 1849 // or converge on or enter undesirable stable absorbing states. 1850 // 1851 // We implement a feedback-based control system -- using past behavior 1852 // to predict future behavior. We face two issues: (a) if the 1853 // input signal is random then the spin predictor won't provide optimal 1854 // results, and (b) if the signal frequency is too high then the control 1855 // system, which has some natural response lag, will "chase" the signal. 1856 // (b) can arise from multimodal lock hold times. Transient preemption 1857 // can also result in apparent bimodal lock hold times. 1858 // Although sub-optimal, neither condition is particularly harmful, as 1859 // in the worst-case we'll spin when we shouldn't or vice-versa. 1860 // The maximum spin duration is rather short so the failure modes aren't bad. 1861 // To be conservative, I've tuned the gain in system to bias toward 1862 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1863 // "rings" or oscillates between spinning and not spinning. This happens 1864 // when spinning is just on the cusp of profitability, however, so the 1865 // situation is not dire. The state is benign -- there's no need to add 1866 // hysteresis control to damp the transition rate between spinning and 1867 // not spinning. 1868 // 1869 1870 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1871 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1872 1873 // Spinning: Fixed frequency (100%), vary duration 1874 1875 1876 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1877 1878 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1879 int ctr = Knob_FixedSpin ; 1880 if (ctr != 0) { 1881 while (--ctr >= 0) { 1882 if (TryLock (Self) > 0) return 1 ; 1883 SpinPause () ; 1884 } 1885 return 0 ; 1886 } 1887 1888 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 1889 if (TryLock(Self) > 0) { 1890 // Increase _SpinDuration ... 1891 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1892 // Raising _SpurDuration to the poverty line is key. 1893 int x = _SpinDuration ; 1894 if (x < Knob_SpinLimit) { 1895 if (x < Knob_Poverty) x = Knob_Poverty ; 1896 _SpinDuration = x + Knob_BonusB ; 1897 } 1898 return 1 ; 1899 } 1900 SpinPause () ; 1901 } 1902 1903 // Admission control - verify preconditions for spinning 1904 // 1905 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1906 // becoming an absorbing state. Put another way, we spin briefly to 1907 // sample, just in case the system load, parallelism, contention, or lock 1908 // modality changed. 1909 // 1910 // Consider the following alternative: 1911 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1912 // spin attempt. "Periodically" might mean after a tally of 1913 // the # of failed spin attempts (or iterations) reaches some threshold. 1914 // This takes us into the realm of 1-out-of-N spinning, where we 1915 // hold the duration constant but vary the frequency. 1916 1917 ctr = _SpinDuration ; 1918 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 1919 if (ctr <= 0) return 0 ; 1920 1921 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 1922 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1923 TEVENT (Spin abort - notrunnable [TOP]); 1924 return 0 ; 1925 } 1926 1927 int MaxSpin = Knob_MaxSpinners ; 1928 if (MaxSpin >= 0) { 1929 if (_Spinner > MaxSpin) { 1930 TEVENT (Spin abort -- too many spinners) ; 1931 return 0 ; 1932 } 1933 // Slighty racy, but benign ... 1934 Adjust (&_Spinner, 1) ; 1935 } 1936 1937 // We're good to spin ... spin ingress. 1938 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1939 // when preparing to LD...CAS _owner, etc and the CAS is likely 1940 // to succeed. 1941 int hits = 0 ; 1942 int msk = 0 ; 1943 int caspty = Knob_CASPenalty ; 1944 int oxpty = Knob_OXPenalty ; 1945 int sss = Knob_SpinSetSucc ; 1946 if (sss && _succ == NULL ) _succ = Self ; 1947 Thread * prv = NULL ; 1948 1949 // There are three ways to exit the following loop: 1950 // 1. A successful spin where this thread has acquired the lock. 1951 // 2. Spin failure with prejudice 1952 // 3. Spin failure without prejudice 1953 1954 while (--ctr >= 0) { 1955 1956 // Periodic polling -- Check for pending GC 1957 // Threads may spin while they're unsafe. 1958 // We don't want spinning threads to delay the JVM from reaching 1959 // a stop-the-world safepoint or to steal cycles from GC. 1960 // If we detect a pending safepoint we abort in order that 1961 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1962 // this thread, if safe, doesn't steal cycles from GC. 1963 // This is in keeping with the "no loitering in runtime" rule. 1964 // We periodically check to see if there's a safepoint pending. 1965 if ((ctr & 0xFF) == 0) { 1966 if (SafepointSynchronize::do_call_back()) { 1967 TEVENT (Spin: safepoint) ; 1968 goto Abort ; // abrupt spin egress 1969 } 1970 if (Knob_UsePause & 1) SpinPause () ; 1971 1972 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 1973 if (hits > 50 && scb != NULL) { 1974 int abend = (*scb)(SpinCallbackArgument, 0) ; 1975 } 1976 } 1977 1978 if (Knob_UsePause & 2) SpinPause() ; 1979 1980 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1981 // This is useful on classic SMP systems, but is of less utility on 1982 // N1-style CMT platforms. 1983 // 1984 // Trade-off: lock acquisition latency vs coherency bandwidth. 1985 // Lock hold times are typically short. A histogram 1986 // of successful spin attempts shows that we usually acquire 1987 // the lock early in the spin. That suggests we want to 1988 // sample _owner frequently in the early phase of the spin, 1989 // but then back-off and sample less frequently as the spin 1990 // progresses. The back-off makes a good citizen on SMP big 1991 // SMP systems. Oversampling _owner can consume excessive 1992 // coherency bandwidth. Relatedly, if we _oversample _owner we 1993 // can inadvertently interfere with the the ST m->owner=null. 1994 // executed by the lock owner. 1995 if (ctr & msk) continue ; 1996 ++hits ; 1997 if ((hits & 0xF) == 0) { 1998 // The 0xF, above, corresponds to the exponent. 1999 // Consider: (msk+1)|msk 2000 msk = ((msk << 2)|3) & BackOffMask ; 2001 } 2002 2003 // Probe _owner with TATAS 2004 // If this thread observes the monitor transition or flicker 2005 // from locked to unlocked to locked, then the odds that this 2006 // thread will acquire the lock in this spin attempt go down 2007 // considerably. The same argument applies if the CAS fails 2008 // or if we observe _owner change from one non-null value to 2009 // another non-null value. In such cases we might abort 2010 // the spin without prejudice or apply a "penalty" to the 2011 // spin count-down variable "ctr", reducing it by 100, say. 2012 2013 Thread * ox = (Thread *) _owner ; 2014 if (ox == NULL) { 2015 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2016 if (ox == NULL) { 2017 // The CAS succeeded -- this thread acquired ownership 2018 // Take care of some bookkeeping to exit spin state. 2019 if (sss && _succ == Self) { 2020 _succ = NULL ; 2021 } 2022 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2023 2024 // Increase _SpinDuration : 2025 // The spin was successful (profitable) so we tend toward 2026 // longer spin attempts in the future. 2027 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2028 // If we acquired the lock early in the spin cycle it 2029 // makes sense to increase _SpinDuration proportionally. 2030 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2031 int x = _SpinDuration ; 2032 if (x < Knob_SpinLimit) { 2033 if (x < Knob_Poverty) x = Knob_Poverty ; 2034 _SpinDuration = x + Knob_Bonus ; 2035 } 2036 return 1 ; 2037 } 2038 2039 // The CAS failed ... we can take any of the following actions: 2040 // * penalize: ctr -= Knob_CASPenalty 2041 // * exit spin with prejudice -- goto Abort; 2042 // * exit spin without prejudice. 2043 // * Since CAS is high-latency, retry again immediately. 2044 prv = ox ; 2045 TEVENT (Spin: cas failed) ; 2046 if (caspty == -2) break ; 2047 if (caspty == -1) goto Abort ; 2048 ctr -= caspty ; 2049 continue ; 2050 } 2051 2052 // Did lock ownership change hands ? 2053 if (ox != prv && prv != NULL ) { 2054 TEVENT (spin: Owner changed) 2055 if (oxpty == -2) break ; 2056 if (oxpty == -1) goto Abort ; 2057 ctr -= oxpty ; 2058 } 2059 prv = ox ; 2060 2061 // Abort the spin if the owner is not executing. 2062 // The owner must be executing in order to drop the lock. 2063 // Spinning while the owner is OFFPROC is idiocy. 2064 // Consider: ctr -= RunnablePenalty ; 2065 if (Knob_OState && NotRunnable (Self, ox)) { 2066 TEVENT (Spin abort - notrunnable); 2067 goto Abort ; 2068 } 2069 if (sss && _succ == NULL ) _succ = Self ; 2070 } 2071 2072 // Spin failed with prejudice -- reduce _SpinDuration. 2073 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2074 // AIMD is globally stable. 2075 TEVENT (Spin failure) ; 2076 { 2077 int x = _SpinDuration ; 2078 if (x > 0) { 2079 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2080 // This is globally sample and tends to damp the response. 2081 x -= Knob_Penalty ; 2082 if (x < 0) x = 0 ; 2083 _SpinDuration = x ; 2084 } 2085 } 2086 2087 Abort: 2088 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2089 if (sss && _succ == Self) { 2090 _succ = NULL ; 2091 // Invariant: after setting succ=null a contending thread 2092 // must recheck-retry _owner before parking. This usually happens 2093 // in the normal usage of TrySpin(), but it's safest 2094 // to make TrySpin() as foolproof as possible. 2095 OrderAccess::fence() ; 2096 if (TryLock(Self) > 0) return 1 ; 2097 } 2098 return 0 ; 2099 } 2100 2101 // NotRunnable() -- informed spinning 2102 // 2103 // Don't bother spinning if the owner is not eligible to drop the lock. 2104 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2105 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2106 // The thread must be runnable in order to drop the lock in timely fashion. 2107 // If the _owner is not runnable then spinning will not likely be 2108 // successful (profitable). 2109 // 2110 // Beware -- the thread referenced by _owner could have died 2111 // so a simply fetch from _owner->_thread_state might trap. 2112 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2113 // Because of the lifecycle issues the schedctl and _thread_state values 2114 // observed by NotRunnable() might be garbage. NotRunnable must 2115 // tolerate this and consider the observed _thread_state value 2116 // as advisory. 2117 // 2118 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2119 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2120 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2121 // with the LSB of _owner. Another option would be to probablistically probe 2122 // the putative _owner->TypeTag value. 2123 // 2124 // Checking _thread_state isn't perfect. Even if the thread is 2125 // in_java it might be blocked on a page-fault or have been preempted 2126 // and sitting on a ready/dispatch queue. _thread state in conjunction 2127 // with schedctl.sc_state gives us a good picture of what the 2128 // thread is doing, however. 2129 // 2130 // TODO: check schedctl.sc_state. 2131 // We'll need to use SafeFetch32() to read from the schedctl block. 2132 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2133 // 2134 // The return value from NotRunnable() is *advisory* -- the 2135 // result is based on sampling and is not necessarily coherent. 2136 // The caller must tolerate false-negative and false-positive errors. 2137 // Spinning, in general, is probabilistic anyway. 2138 2139 2140 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2141 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2142 if (!OwnerIsThread) return 0 ; 2143 2144 if (ox == NULL) return 0 ; 2145 2146 // Avoid transitive spinning ... 2147 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2148 // Immediately after T1 acquires L it's possible that T2, also 2149 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2150 // This occurs transiently after T1 acquired L but before 2151 // T1 managed to clear T1.Stalled. T2 does not need to abort 2152 // its spin in this circumstance. 2153 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2154 2155 if (BlockedOn == 1) return 1 ; 2156 if (BlockedOn != 0) { 2157 return BlockedOn != intptr_t(this) && _owner == ox ; 2158 } 2159 2160 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2161 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2162 // consider also: jst != _thread_in_Java -- but that's overspecific. 2163 return jst == _thread_blocked || jst == _thread_in_native ; 2164 } 2165 2166 2167 // ----------------------------------------------------------------------------- 2168 // WaitSet management ... 2169 2170 ObjectWaiter::ObjectWaiter(Thread* thread) { 2171 _next = NULL; 2172 _prev = NULL; 2173 _notified = 0; 2174 TState = TS_RUN ; 2175 _thread = thread; 2176 _event = thread->_ParkEvent ; 2177 _active = false; 2178 assert (_event != NULL, "invariant") ; 2179 } 2180 2181 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2182 JavaThread *jt = (JavaThread *)this->_thread; 2183 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2184 } 2185 2186 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2187 JavaThread *jt = (JavaThread *)this->_thread; 2188 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2189 } 2190 2191 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2192 assert(node != NULL, "should not dequeue NULL node"); 2193 assert(node->_prev == NULL, "node already in list"); 2194 assert(node->_next == NULL, "node already in list"); 2195 // put node at end of queue (circular doubly linked list) 2196 if (_WaitSet == NULL) { 2197 _WaitSet = node; 2198 node->_prev = node; 2199 node->_next = node; 2200 } else { 2201 ObjectWaiter* head = _WaitSet ; 2202 ObjectWaiter* tail = head->_prev; 2203 assert(tail->_next == head, "invariant check"); 2204 tail->_next = node; 2205 head->_prev = node; 2206 node->_next = head; 2207 node->_prev = tail; 2208 } 2209 } 2210 2211 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2212 // dequeue the very first waiter 2213 ObjectWaiter* waiter = _WaitSet; 2214 if (waiter) { 2215 DequeueSpecificWaiter(waiter); 2216 } 2217 return waiter; 2218 } 2219 2220 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2221 assert(node != NULL, "should not dequeue NULL node"); 2222 assert(node->_prev != NULL, "node already removed from list"); 2223 assert(node->_next != NULL, "node already removed from list"); 2224 // when the waiter has woken up because of interrupt, 2225 // timeout or other spurious wake-up, dequeue the 2226 // waiter from waiting list 2227 ObjectWaiter* next = node->_next; 2228 if (next == node) { 2229 assert(node->_prev == node, "invariant check"); 2230 _WaitSet = NULL; 2231 } else { 2232 ObjectWaiter* prev = node->_prev; 2233 assert(prev->_next == node, "invariant check"); 2234 assert(next->_prev == node, "invariant check"); 2235 next->_prev = prev; 2236 prev->_next = next; 2237 if (_WaitSet == node) { 2238 _WaitSet = next; 2239 } 2240 } 2241 node->_next = NULL; 2242 node->_prev = NULL; 2243 } 2244 2245 // ----------------------------------------------------------------------------- 2246 // PerfData support 2247 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2248 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2249 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2250 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2251 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2252 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2253 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2254 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2255 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2256 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2257 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2258 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2259 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2260 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2261 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2262 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2263 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2264 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2265 2266 // One-shot global initialization for the sync subsystem. 2267 // We could also defer initialization and initialize on-demand 2268 // the first time we call inflate(). Initialization would 2269 // be protected - like so many things - by the MonitorCache_lock. 2270 2271 void ObjectMonitor::Initialize () { 2272 static int InitializationCompleted = 0 ; 2273 assert (InitializationCompleted == 0, "invariant") ; 2274 InitializationCompleted = 1 ; 2275 if (UsePerfData) { 2276 EXCEPTION_MARK ; 2277 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2278 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2279 NEWPERFCOUNTER(_sync_Inflations) ; 2280 NEWPERFCOUNTER(_sync_Deflations) ; 2281 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2282 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2283 NEWPERFCOUNTER(_sync_Parks) ; 2284 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2285 NEWPERFCOUNTER(_sync_Notifications) ; 2286 NEWPERFCOUNTER(_sync_SlowEnter) ; 2287 NEWPERFCOUNTER(_sync_SlowExit) ; 2288 NEWPERFCOUNTER(_sync_SlowNotify) ; 2289 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2290 NEWPERFCOUNTER(_sync_FailedSpins) ; 2291 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2292 NEWPERFCOUNTER(_sync_PrivateA) ; 2293 NEWPERFCOUNTER(_sync_PrivateB) ; 2294 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2295 NEWPERFCOUNTER(_sync_MonScavenged) ; 2296 NEWPERFVARIABLE(_sync_MonExtant) ; 2297 #undef NEWPERFCOUNTER 2298 } 2299 } 2300 2301 2302 // Compile-time asserts 2303 // When possible, it's better to catch errors deterministically at 2304 // compile-time than at runtime. The down-side to using compile-time 2305 // asserts is that error message -- often something about negative array 2306 // indices -- is opaque. 2307 2308 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2309 2310 void ObjectMonitor::ctAsserts() { 2311 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2312 } 2313 2314 2315 static char * kvGet (char * kvList, const char * Key) { 2316 if (kvList == NULL) return NULL ; 2317 size_t n = strlen (Key) ; 2318 char * Search ; 2319 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2320 if (strncmp (Search, Key, n) == 0) { 2321 if (Search[n] == '=') return Search + n + 1 ; 2322 if (Search[n] == 0) return (char *) "1" ; 2323 } 2324 } 2325 return NULL ; 2326 } 2327 2328 static int kvGetInt (char * kvList, const char * Key, int Default) { 2329 char * v = kvGet (kvList, Key) ; 2330 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2331 if (Knob_ReportSettings && v != NULL) { 2332 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2333 ::fflush (stdout) ; 2334 } 2335 return rslt ; 2336 } 2337 2338 void ObjectMonitor::DeferredInitialize () { 2339 if (InitDone > 0) return ; 2340 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2341 while (InitDone != 1) ; 2342 return ; 2343 } 2344 2345 // One-shot global initialization ... 2346 // The initialization is idempotent, so we don't need locks. 2347 // In the future consider doing this via os::init_2(). 2348 // SyncKnobs consist of <Key>=<Value> pairs in the style 2349 // of environment variables. Start by converting ':' to NUL. 2350 2351 if (SyncKnobs == NULL) SyncKnobs = "" ; 2352 2353 size_t sz = strlen (SyncKnobs) ; 2354 char * knobs = (char *) malloc (sz + 2) ; 2355 if (knobs == NULL) { 2356 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; 2357 guarantee (0, "invariant") ; 2358 } 2359 strcpy (knobs, SyncKnobs) ; 2360 knobs[sz+1] = 0 ; 2361 for (char * p = knobs ; *p ; p++) { 2362 if (*p == ':') *p = 0 ; 2363 } 2364 2365 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2366 SETKNOB(ReportSettings) ; 2367 SETKNOB(Verbose) ; 2368 SETKNOB(FixedSpin) ; 2369 SETKNOB(SpinLimit) ; 2370 SETKNOB(SpinBase) ; 2371 SETKNOB(SpinBackOff); 2372 SETKNOB(CASPenalty) ; 2373 SETKNOB(OXPenalty) ; 2374 SETKNOB(LogSpins) ; 2375 SETKNOB(SpinSetSucc) ; 2376 SETKNOB(SuccEnabled) ; 2377 SETKNOB(SuccRestrict) ; 2378 SETKNOB(Penalty) ; 2379 SETKNOB(Bonus) ; 2380 SETKNOB(BonusB) ; 2381 SETKNOB(Poverty) ; 2382 SETKNOB(SpinAfterFutile) ; 2383 SETKNOB(UsePause) ; 2384 SETKNOB(SpinEarly) ; 2385 SETKNOB(OState) ; 2386 SETKNOB(MaxSpinners) ; 2387 SETKNOB(PreSpin) ; 2388 SETKNOB(ExitPolicy) ; 2389 SETKNOB(QMode); 2390 SETKNOB(ResetEvent) ; 2391 SETKNOB(MoveNotifyee) ; 2392 SETKNOB(FastHSSEC) ; 2393 #undef SETKNOB 2394 2395 if (os::is_MP()) { 2396 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2397 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2398 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2399 } else { 2400 Knob_SpinLimit = 0 ; 2401 Knob_SpinBase = 0 ; 2402 Knob_PreSpin = 0 ; 2403 Knob_FixedSpin = -1 ; 2404 } 2405 2406 if (Knob_LogSpins == 0) { 2407 ObjectMonitor::_sync_FailedSpins = NULL ; 2408 } 2409 2410 free (knobs) ; 2411 OrderAccess::fence() ; 2412 InitDone = 1 ; 2413 } 2414 2415 #ifndef PRODUCT 2416 void ObjectMonitor::verify() { 2417 } 2418 2419 void ObjectMonitor::print() { 2420 } 2421 #endif