1 /* 2 * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "oops/markOop.hpp" 29 #include "oops/oop.inline.hpp" 30 #include "runtime/handles.inline.hpp" 31 #include "runtime/interfaceSupport.hpp" 32 #include "runtime/mutexLocker.hpp" 33 #include "runtime/objectMonitor.hpp" 34 #include "runtime/objectMonitor.inline.hpp" 35 #include "runtime/osThread.hpp" 36 #include "runtime/stubRoutines.hpp" 37 #include "runtime/thread.inline.hpp" 38 #include "services/threadService.hpp" 39 #include "trace/tracing.hpp" 40 #include "trace/traceMacros.hpp" 41 #include "utilities/dtrace.hpp" 42 #include "utilities/macros.hpp" 43 #include "utilities/preserveException.hpp" 44 #ifdef TARGET_OS_FAMILY_linux 45 # include "os_linux.inline.hpp" 46 #endif 47 #ifdef TARGET_OS_FAMILY_solaris 48 # include "os_solaris.inline.hpp" 49 #endif 50 #ifdef TARGET_OS_FAMILY_windows 51 # include "os_windows.inline.hpp" 52 #endif 53 #ifdef TARGET_OS_FAMILY_bsd 54 # include "os_bsd.inline.hpp" 55 #endif 56 57 #if defined(__GNUC__) && !defined(IA64) 58 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 59 #define ATTR __attribute__((noinline)) 60 #else 61 #define ATTR 62 #endif 63 64 65 #ifdef DTRACE_ENABLED 66 67 // Only bother with this argument setup if dtrace is available 68 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 69 70 71 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 72 char* bytes = NULL; \ 73 int len = 0; \ 74 jlong jtid = SharedRuntime::get_java_tid(thread); \ 75 Symbol* klassname = ((oop)obj)->klass()->name(); \ 76 if (klassname != NULL) { \ 77 bytes = (char*)klassname->bytes(); \ 78 len = klassname->utf8_length(); \ 79 } 80 81 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 82 { \ 83 if (DTraceMonitorProbes) { \ 84 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 85 HOTSPOT_MONITOR_WAIT(jtid, \ 86 (monitor), bytes, len, (millis)); \ 87 } \ 88 } 89 90 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 91 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 92 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 93 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 94 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 95 96 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 97 { \ 98 if (DTraceMonitorProbes) { \ 99 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 100 HOTSPOT_MONITOR_##probe(jtid, \ 101 (uintptr_t)(monitor), bytes, len); \ 102 } \ 103 } 104 105 #else // ndef DTRACE_ENABLED 106 107 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 108 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 109 110 #endif // ndef DTRACE_ENABLED 111 112 // Tunables ... 113 // The knob* variables are effectively final. Once set they should 114 // never be modified hence. Consider using __read_mostly with GCC. 115 116 int ObjectMonitor::Knob_Verbose = 0 ; 117 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 118 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 119 static int Knob_HandOff = 0 ; 120 static int Knob_ReportSettings = 0 ; 121 122 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 123 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 124 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 125 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 126 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 127 static int Knob_SpinEarly = 1 ; 128 static int Knob_SuccEnabled = 1 ; // futile wake throttling 129 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 130 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 131 static int Knob_Bonus = 100 ; // spin success bonus 132 static int Knob_BonusB = 100 ; // spin success bonus 133 static int Knob_Penalty = 200 ; // spin failure penalty 134 static int Knob_Poverty = 1000 ; 135 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 136 static int Knob_FixedSpin = 0 ; 137 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 138 static int Knob_UsePause = 1 ; 139 static int Knob_ExitPolicy = 0 ; 140 static int Knob_PreSpin = 10 ; // 20-100 likely better 141 static int Knob_ResetEvent = 0 ; 142 static int BackOffMask = 0 ; 143 144 static int Knob_FastHSSEC = 0 ; 145 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 146 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 147 static volatile int InitDone = 0 ; 148 149 #define TrySpin TrySpin_VaryDuration 150 151 // ----------------------------------------------------------------------------- 152 // Theory of operations -- Monitors lists, thread residency, etc: 153 // 154 // * A thread acquires ownership of a monitor by successfully 155 // CAS()ing the _owner field from null to non-null. 156 // 157 // * Invariant: A thread appears on at most one monitor list -- 158 // cxq, EntryList or WaitSet -- at any one time. 159 // 160 // * Contending threads "push" themselves onto the cxq with CAS 161 // and then spin/park. 162 // 163 // * After a contending thread eventually acquires the lock it must 164 // dequeue itself from either the EntryList or the cxq. 165 // 166 // * The exiting thread identifies and unparks an "heir presumptive" 167 // tentative successor thread on the EntryList. Critically, the 168 // exiting thread doesn't unlink the successor thread from the EntryList. 169 // After having been unparked, the wakee will recontend for ownership of 170 // the monitor. The successor (wakee) will either acquire the lock or 171 // re-park itself. 172 // 173 // Succession is provided for by a policy of competitive handoff. 174 // The exiting thread does _not_ grant or pass ownership to the 175 // successor thread. (This is also referred to as "handoff" succession"). 176 // Instead the exiting thread releases ownership and possibly wakes 177 // a successor, so the successor can (re)compete for ownership of the lock. 178 // If the EntryList is empty but the cxq is populated the exiting 179 // thread will drain the cxq into the EntryList. It does so by 180 // by detaching the cxq (installing null with CAS) and folding 181 // the threads from the cxq into the EntryList. The EntryList is 182 // doubly linked, while the cxq is singly linked because of the 183 // CAS-based "push" used to enqueue recently arrived threads (RATs). 184 // 185 // * Concurrency invariants: 186 // 187 // -- only the monitor owner may access or mutate the EntryList. 188 // The mutex property of the monitor itself protects the EntryList 189 // from concurrent interference. 190 // -- Only the monitor owner may detach the cxq. 191 // 192 // * The monitor entry list operations avoid locks, but strictly speaking 193 // they're not lock-free. Enter is lock-free, exit is not. 194 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 195 // 196 // * The cxq can have multiple concurrent "pushers" but only one concurrent 197 // detaching thread. This mechanism is immune from the ABA corruption. 198 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 199 // 200 // * Taken together, the cxq and the EntryList constitute or form a 201 // single logical queue of threads stalled trying to acquire the lock. 202 // We use two distinct lists to improve the odds of a constant-time 203 // dequeue operation after acquisition (in the ::enter() epilogue) and 204 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 205 // A key desideratum is to minimize queue & monitor metadata manipulation 206 // that occurs while holding the monitor lock -- that is, we want to 207 // minimize monitor lock holds times. Note that even a small amount of 208 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 209 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 210 // locks and monitor metadata. 211 // 212 // Cxq points to the the set of Recently Arrived Threads attempting entry. 213 // Because we push threads onto _cxq with CAS, the RATs must take the form of 214 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 215 // the unlocking thread notices that EntryList is null but _cxq is != null. 216 // 217 // The EntryList is ordered by the prevailing queue discipline and 218 // can be organized in any convenient fashion, such as a doubly-linked list or 219 // a circular doubly-linked list. Critically, we want insert and delete operations 220 // to operate in constant-time. If we need a priority queue then something akin 221 // to Solaris' sleepq would work nicely. Viz., 222 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 223 // Queue discipline is enforced at ::exit() time, when the unlocking thread 224 // drains the cxq into the EntryList, and orders or reorders the threads on the 225 // EntryList accordingly. 226 // 227 // Barring "lock barging", this mechanism provides fair cyclic ordering, 228 // somewhat similar to an elevator-scan. 229 // 230 // * The monitor synchronization subsystem avoids the use of native 231 // synchronization primitives except for the narrow platform-specific 232 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 233 // the semantics of park-unpark. Put another way, this monitor implementation 234 // depends only on atomic operations and park-unpark. The monitor subsystem 235 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 236 // underlying OS manages the READY<->RUN transitions. 237 // 238 // * Waiting threads reside on the WaitSet list -- wait() puts 239 // the caller onto the WaitSet. 240 // 241 // * notify() or notifyAll() simply transfers threads from the WaitSet to 242 // either the EntryList or cxq. Subsequent exit() operations will 243 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 244 // it's likely the notifyee would simply impale itself on the lock held 245 // by the notifier. 246 // 247 // * An interesting alternative is to encode cxq as (List,LockByte) where 248 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 249 // variable, like _recursions, in the scheme. The threads or Events that form 250 // the list would have to be aligned in 256-byte addresses. A thread would 251 // try to acquire the lock or enqueue itself with CAS, but exiting threads 252 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 253 // Note that is is *not* word-tearing, but it does presume that full-word 254 // CAS operations are coherent with intermix with STB operations. That's true 255 // on most common processors. 256 // 257 // * See also http://blogs.sun.com/dave 258 259 260 // ----------------------------------------------------------------------------- 261 // Enter support 262 263 bool ObjectMonitor::try_enter(Thread* THREAD) { 264 if (THREAD != _owner) { 265 if (THREAD->is_lock_owned ((address)_owner)) { 266 assert(_recursions == 0, "internal state error"); 267 _owner = THREAD ; 268 _recursions = 1 ; 269 OwnerIsThread = 1 ; 270 return true; 271 } 272 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 273 return false; 274 } 275 return true; 276 } else { 277 _recursions++; 278 return true; 279 } 280 } 281 282 void ATTR ObjectMonitor::enter(TRAPS) { 283 // The following code is ordered to check the most common cases first 284 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 285 Thread * const Self = THREAD ; 286 void * cur ; 287 288 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 289 if (cur == NULL) { 290 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 291 assert (_recursions == 0 , "invariant") ; 292 assert (_owner == Self, "invariant") ; 293 // CONSIDER: set or assert OwnerIsThread == 1 294 return ; 295 } 296 297 if (cur == Self) { 298 // TODO-FIXME: check for integer overflow! BUGID 6557169. 299 _recursions ++ ; 300 return ; 301 } 302 303 if (Self->is_lock_owned ((address)cur)) { 304 assert (_recursions == 0, "internal state error"); 305 _recursions = 1 ; 306 // Commute owner from a thread-specific on-stack BasicLockObject address to 307 // a full-fledged "Thread *". 308 _owner = Self ; 309 OwnerIsThread = 1 ; 310 return ; 311 } 312 313 // We've encountered genuine contention. 314 assert (Self->_Stalled == 0, "invariant") ; 315 Self->_Stalled = intptr_t(this) ; 316 317 // Try one round of spinning *before* enqueueing Self 318 // and before going through the awkward and expensive state 319 // transitions. The following spin is strictly optional ... 320 // Note that if we acquire the monitor from an initial spin 321 // we forgo posting JVMTI events and firing DTRACE probes. 322 if (Knob_SpinEarly && TrySpin (Self) > 0) { 323 assert (_owner == Self , "invariant") ; 324 assert (_recursions == 0 , "invariant") ; 325 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 326 Self->_Stalled = 0 ; 327 return ; 328 } 329 330 assert (_owner != Self , "invariant") ; 331 assert (_succ != Self , "invariant") ; 332 assert (Self->is_Java_thread() , "invariant") ; 333 JavaThread * jt = (JavaThread *) Self ; 334 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 335 assert (jt->thread_state() != _thread_blocked , "invariant") ; 336 assert (this->object() != NULL , "invariant") ; 337 assert (_count >= 0, "invariant") ; 338 339 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 340 // Ensure the object-monitor relationship remains stable while there's contention. 341 Atomic::inc_ptr(&_count); 342 343 EventJavaMonitorEnter event; 344 345 { // Change java thread status to indicate blocked on monitor enter. 346 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 347 348 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 349 if (JvmtiExport::should_post_monitor_contended_enter()) { 350 JvmtiExport::post_monitor_contended_enter(jt, this); 351 } 352 353 OSThreadContendState osts(Self->osthread()); 354 ThreadBlockInVM tbivm(jt); 355 356 Self->set_current_pending_monitor(this); 357 358 // TODO-FIXME: change the following for(;;) loop to straight-line code. 359 for (;;) { 360 jt->set_suspend_equivalent(); 361 // cleared by handle_special_suspend_equivalent_condition() 362 // or java_suspend_self() 363 364 EnterI (THREAD) ; 365 366 if (!ExitSuspendEquivalent(jt)) break ; 367 368 // 369 // We have acquired the contended monitor, but while we were 370 // waiting another thread suspended us. We don't want to enter 371 // the monitor while suspended because that would surprise the 372 // thread that suspended us. 373 // 374 _recursions = 0 ; 375 _succ = NULL ; 376 exit (false, Self) ; 377 378 jt->java_suspend_self(); 379 } 380 Self->set_current_pending_monitor(NULL); 381 } 382 383 Atomic::dec_ptr(&_count); 384 assert (_count >= 0, "invariant") ; 385 Self->_Stalled = 0 ; 386 387 // Must either set _recursions = 0 or ASSERT _recursions == 0. 388 assert (_recursions == 0 , "invariant") ; 389 assert (_owner == Self , "invariant") ; 390 assert (_succ != Self , "invariant") ; 391 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 392 393 // The thread -- now the owner -- is back in vm mode. 394 // Report the glorious news via TI,DTrace and jvmstat. 395 // The probe effect is non-trivial. All the reportage occurs 396 // while we hold the monitor, increasing the length of the critical 397 // section. Amdahl's parallel speedup law comes vividly into play. 398 // 399 // Another option might be to aggregate the events (thread local or 400 // per-monitor aggregation) and defer reporting until a more opportune 401 // time -- such as next time some thread encounters contention but has 402 // yet to acquire the lock. While spinning that thread could 403 // spinning we could increment JVMStat counters, etc. 404 405 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 406 if (JvmtiExport::should_post_monitor_contended_entered()) { 407 JvmtiExport::post_monitor_contended_entered(jt, this); 408 } 409 410 if (event.should_commit()) { 411 event.set_klass(((oop)this->object())->klass()); 412 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 413 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 414 event.commit(); 415 } 416 417 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 418 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 419 } 420 } 421 422 423 // Caveat: TryLock() is not necessarily serializing if it returns failure. 424 // Callers must compensate as needed. 425 426 int ObjectMonitor::TryLock (Thread * Self) { 427 for (;;) { 428 void * own = _owner ; 429 if (own != NULL) return 0 ; 430 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 431 // Either guarantee _recursions == 0 or set _recursions = 0. 432 assert (_recursions == 0, "invariant") ; 433 assert (_owner == Self, "invariant") ; 434 // CONSIDER: set or assert that OwnerIsThread == 1 435 return 1 ; 436 } 437 // The lock had been free momentarily, but we lost the race to the lock. 438 // Interference -- the CAS failed. 439 // We can either return -1 or retry. 440 // Retry doesn't make as much sense because the lock was just acquired. 441 if (true) return -1 ; 442 } 443 } 444 445 void ATTR ObjectMonitor::EnterI (TRAPS) { 446 Thread * Self = THREAD ; 447 assert (Self->is_Java_thread(), "invariant") ; 448 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 449 450 // Try the lock - TATAS 451 if (TryLock (Self) > 0) { 452 assert (_succ != Self , "invariant") ; 453 assert (_owner == Self , "invariant") ; 454 assert (_Responsible != Self , "invariant") ; 455 return ; 456 } 457 458 DeferredInitialize () ; 459 460 // We try one round of spinning *before* enqueueing Self. 461 // 462 // If the _owner is ready but OFFPROC we could use a YieldTo() 463 // operation to donate the remainder of this thread's quantum 464 // to the owner. This has subtle but beneficial affinity 465 // effects. 466 467 if (TrySpin (Self) > 0) { 468 assert (_owner == Self , "invariant") ; 469 assert (_succ != Self , "invariant") ; 470 assert (_Responsible != Self , "invariant") ; 471 return ; 472 } 473 474 // The Spin failed -- Enqueue and park the thread ... 475 assert (_succ != Self , "invariant") ; 476 assert (_owner != Self , "invariant") ; 477 assert (_Responsible != Self , "invariant") ; 478 479 // Enqueue "Self" on ObjectMonitor's _cxq. 480 // 481 // Node acts as a proxy for Self. 482 // As an aside, if were to ever rewrite the synchronization code mostly 483 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 484 // Java objects. This would avoid awkward lifecycle and liveness issues, 485 // as well as eliminate a subset of ABA issues. 486 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 487 // 488 489 ObjectWaiter node(Self) ; 490 Self->_ParkEvent->reset() ; 491 node._prev = (ObjectWaiter *) 0xBAD ; 492 node.TState = ObjectWaiter::TS_CXQ ; 493 494 // Push "Self" onto the front of the _cxq. 495 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 496 // Note that spinning tends to reduce the rate at which threads 497 // enqueue and dequeue on EntryList|cxq. 498 ObjectWaiter * nxt ; 499 for (;;) { 500 node._next = nxt = _cxq ; 501 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 502 503 // Interference - the CAS failed because _cxq changed. Just retry. 504 // As an optional optimization we retry the lock. 505 if (TryLock (Self) > 0) { 506 assert (_succ != Self , "invariant") ; 507 assert (_owner == Self , "invariant") ; 508 assert (_Responsible != Self , "invariant") ; 509 return ; 510 } 511 } 512 513 // Check for cxq|EntryList edge transition to non-null. This indicates 514 // the onset of contention. While contention persists exiting threads 515 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 516 // operations revert to the faster 1-0 mode. This enter operation may interleave 517 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 518 // arrange for one of the contending thread to use a timed park() operations 519 // to detect and recover from the race. (Stranding is form of progress failure 520 // where the monitor is unlocked but all the contending threads remain parked). 521 // That is, at least one of the contended threads will periodically poll _owner. 522 // One of the contending threads will become the designated "Responsible" thread. 523 // The Responsible thread uses a timed park instead of a normal indefinite park 524 // operation -- it periodically wakes and checks for and recovers from potential 525 // strandings admitted by 1-0 exit operations. We need at most one Responsible 526 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 527 // be responsible for a monitor. 528 // 529 // Currently, one of the contended threads takes on the added role of "Responsible". 530 // A viable alternative would be to use a dedicated "stranding checker" thread 531 // that periodically iterated over all the threads (or active monitors) and unparked 532 // successors where there was risk of stranding. This would help eliminate the 533 // timer scalability issues we see on some platforms as we'd only have one thread 534 // -- the checker -- parked on a timer. 535 536 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 537 // Try to assume the role of responsible thread for the monitor. 538 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 539 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 540 } 541 542 // The lock have been released while this thread was occupied queueing 543 // itself onto _cxq. To close the race and avoid "stranding" and 544 // progress-liveness failure we must resample-retry _owner before parking. 545 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 546 // In this case the ST-MEMBAR is accomplished with CAS(). 547 // 548 // TODO: Defer all thread state transitions until park-time. 549 // Since state transitions are heavy and inefficient we'd like 550 // to defer the state transitions until absolutely necessary, 551 // and in doing so avoid some transitions ... 552 553 TEVENT (Inflated enter - Contention) ; 554 int nWakeups = 0 ; 555 int RecheckInterval = 1 ; 556 557 for (;;) { 558 559 if (TryLock (Self) > 0) break ; 560 assert (_owner != Self, "invariant") ; 561 562 if ((SyncFlags & 2) && _Responsible == NULL) { 563 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 564 } 565 566 // park self 567 if (_Responsible == Self || (SyncFlags & 1)) { 568 TEVENT (Inflated enter - park TIMED) ; 569 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 570 // Increase the RecheckInterval, but clamp the value. 571 RecheckInterval *= 8 ; 572 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 573 } else { 574 TEVENT (Inflated enter - park UNTIMED) ; 575 Self->_ParkEvent->park() ; 576 } 577 578 if (TryLock(Self) > 0) break ; 579 580 // The lock is still contested. 581 // Keep a tally of the # of futile wakeups. 582 // Note that the counter is not protected by a lock or updated by atomics. 583 // That is by design - we trade "lossy" counters which are exposed to 584 // races during updates for a lower probe effect. 585 TEVENT (Inflated enter - Futile wakeup) ; 586 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 587 ObjectMonitor::_sync_FutileWakeups->inc() ; 588 } 589 ++ nWakeups ; 590 591 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 592 // We can defer clearing _succ until after the spin completes 593 // TrySpin() must tolerate being called with _succ == Self. 594 // Try yet another round of adaptive spinning. 595 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 596 597 // We can find that we were unpark()ed and redesignated _succ while 598 // we were spinning. That's harmless. If we iterate and call park(), 599 // park() will consume the event and return immediately and we'll 600 // just spin again. This pattern can repeat, leaving _succ to simply 601 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 602 // Alternately, we can sample fired() here, and if set, forgo spinning 603 // in the next iteration. 604 605 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 606 Self->_ParkEvent->reset() ; 607 OrderAccess::fence() ; 608 } 609 if (_succ == Self) _succ = NULL ; 610 611 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 612 OrderAccess::fence() ; 613 } 614 615 // Egress : 616 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 617 // Normally we'll find Self on the EntryList . 618 // From the perspective of the lock owner (this thread), the 619 // EntryList is stable and cxq is prepend-only. 620 // The head of cxq is volatile but the interior is stable. 621 // In addition, Self.TState is stable. 622 623 assert (_owner == Self , "invariant") ; 624 assert (object() != NULL , "invariant") ; 625 // I'd like to write: 626 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 627 // but as we're at a safepoint that's not safe. 628 629 UnlinkAfterAcquire (Self, &node) ; 630 if (_succ == Self) _succ = NULL ; 631 632 assert (_succ != Self, "invariant") ; 633 if (_Responsible == Self) { 634 _Responsible = NULL ; 635 OrderAccess::fence(); // Dekker pivot-point 636 637 // We may leave threads on cxq|EntryList without a designated 638 // "Responsible" thread. This is benign. When this thread subsequently 639 // exits the monitor it can "see" such preexisting "old" threads -- 640 // threads that arrived on the cxq|EntryList before the fence, above -- 641 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 642 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 643 // non-null and elect a new "Responsible" timer thread. 644 // 645 // This thread executes: 646 // ST Responsible=null; MEMBAR (in enter epilogue - here) 647 // LD cxq|EntryList (in subsequent exit) 648 // 649 // Entering threads in the slow/contended path execute: 650 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 651 // The (ST cxq; MEMBAR) is accomplished with CAS(). 652 // 653 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 654 // exit operation from floating above the ST Responsible=null. 655 } 656 657 // We've acquired ownership with CAS(). 658 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 659 // But since the CAS() this thread may have also stored into _succ, 660 // EntryList, cxq or Responsible. These meta-data updates must be 661 // visible __before this thread subsequently drops the lock. 662 // Consider what could occur if we didn't enforce this constraint -- 663 // STs to monitor meta-data and user-data could reorder with (become 664 // visible after) the ST in exit that drops ownership of the lock. 665 // Some other thread could then acquire the lock, but observe inconsistent 666 // or old monitor meta-data and heap data. That violates the JMM. 667 // To that end, the 1-0 exit() operation must have at least STST|LDST 668 // "release" barrier semantics. Specifically, there must be at least a 669 // STST|LDST barrier in exit() before the ST of null into _owner that drops 670 // the lock. The barrier ensures that changes to monitor meta-data and data 671 // protected by the lock will be visible before we release the lock, and 672 // therefore before some other thread (CPU) has a chance to acquire the lock. 673 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 674 // 675 // Critically, any prior STs to _succ or EntryList must be visible before 676 // the ST of null into _owner in the *subsequent* (following) corresponding 677 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 678 // execute a serializing instruction. 679 680 if (SyncFlags & 8) { 681 OrderAccess::fence() ; 682 } 683 return ; 684 } 685 686 // ReenterI() is a specialized inline form of the latter half of the 687 // contended slow-path from EnterI(). We use ReenterI() only for 688 // monitor reentry in wait(). 689 // 690 // In the future we should reconcile EnterI() and ReenterI(), adding 691 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 692 // loop accordingly. 693 694 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 695 assert (Self != NULL , "invariant") ; 696 assert (SelfNode != NULL , "invariant") ; 697 assert (SelfNode->_thread == Self , "invariant") ; 698 assert (_waiters > 0 , "invariant") ; 699 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 700 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 701 JavaThread * jt = (JavaThread *) Self ; 702 703 int nWakeups = 0 ; 704 for (;;) { 705 ObjectWaiter::TStates v = SelfNode->TState ; 706 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 707 assert (_owner != Self, "invariant") ; 708 709 if (TryLock (Self) > 0) break ; 710 if (TrySpin (Self) > 0) break ; 711 712 TEVENT (Wait Reentry - parking) ; 713 714 // State transition wrappers around park() ... 715 // ReenterI() wisely defers state transitions until 716 // it's clear we must park the thread. 717 { 718 OSThreadContendState osts(Self->osthread()); 719 ThreadBlockInVM tbivm(jt); 720 721 // cleared by handle_special_suspend_equivalent_condition() 722 // or java_suspend_self() 723 jt->set_suspend_equivalent(); 724 if (SyncFlags & 1) { 725 Self->_ParkEvent->park ((jlong)1000) ; 726 } else { 727 Self->_ParkEvent->park () ; 728 } 729 730 // were we externally suspended while we were waiting? 731 for (;;) { 732 if (!ExitSuspendEquivalent (jt)) break ; 733 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 734 jt->java_suspend_self(); 735 jt->set_suspend_equivalent(); 736 } 737 } 738 739 // Try again, but just so we distinguish between futile wakeups and 740 // successful wakeups. The following test isn't algorithmically 741 // necessary, but it helps us maintain sensible statistics. 742 if (TryLock(Self) > 0) break ; 743 744 // The lock is still contested. 745 // Keep a tally of the # of futile wakeups. 746 // Note that the counter is not protected by a lock or updated by atomics. 747 // That is by design - we trade "lossy" counters which are exposed to 748 // races during updates for a lower probe effect. 749 TEVENT (Wait Reentry - futile wakeup) ; 750 ++ nWakeups ; 751 752 // Assuming this is not a spurious wakeup we'll normally 753 // find that _succ == Self. 754 if (_succ == Self) _succ = NULL ; 755 756 // Invariant: after clearing _succ a contending thread 757 // *must* retry _owner before parking. 758 OrderAccess::fence() ; 759 760 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 761 ObjectMonitor::_sync_FutileWakeups->inc() ; 762 } 763 } 764 765 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 766 // Normally we'll find Self on the EntryList. 767 // Unlinking from the EntryList is constant-time and atomic-free. 768 // From the perspective of the lock owner (this thread), the 769 // EntryList is stable and cxq is prepend-only. 770 // The head of cxq is volatile but the interior is stable. 771 // In addition, Self.TState is stable. 772 773 assert (_owner == Self, "invariant") ; 774 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 775 UnlinkAfterAcquire (Self, SelfNode) ; 776 if (_succ == Self) _succ = NULL ; 777 assert (_succ != Self, "invariant") ; 778 SelfNode->TState = ObjectWaiter::TS_RUN ; 779 OrderAccess::fence() ; // see comments at the end of EnterI() 780 } 781 782 // after the thread acquires the lock in ::enter(). Equally, we could defer 783 // unlinking the thread until ::exit()-time. 784 785 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 786 { 787 assert (_owner == Self, "invariant") ; 788 assert (SelfNode->_thread == Self, "invariant") ; 789 790 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 791 // Normal case: remove Self from the DLL EntryList . 792 // This is a constant-time operation. 793 ObjectWaiter * nxt = SelfNode->_next ; 794 ObjectWaiter * prv = SelfNode->_prev ; 795 if (nxt != NULL) nxt->_prev = prv ; 796 if (prv != NULL) prv->_next = nxt ; 797 if (SelfNode == _EntryList ) _EntryList = nxt ; 798 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 799 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 800 TEVENT (Unlink from EntryList) ; 801 } else { 802 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 803 // Inopportune interleaving -- Self is still on the cxq. 804 // This usually means the enqueue of self raced an exiting thread. 805 // Normally we'll find Self near the front of the cxq, so 806 // dequeueing is typically fast. If needbe we can accelerate 807 // this with some MCS/CHL-like bidirectional list hints and advisory 808 // back-links so dequeueing from the interior will normally operate 809 // in constant-time. 810 // Dequeue Self from either the head (with CAS) or from the interior 811 // with a linear-time scan and normal non-atomic memory operations. 812 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 813 // and then unlink Self from EntryList. We have to drain eventually, 814 // so it might as well be now. 815 816 ObjectWaiter * v = _cxq ; 817 assert (v != NULL, "invariant") ; 818 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 819 // The CAS above can fail from interference IFF a "RAT" arrived. 820 // In that case Self must be in the interior and can no longer be 821 // at the head of cxq. 822 if (v == SelfNode) { 823 assert (_cxq != v, "invariant") ; 824 v = _cxq ; // CAS above failed - start scan at head of list 825 } 826 ObjectWaiter * p ; 827 ObjectWaiter * q = NULL ; 828 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 829 q = p ; 830 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 831 } 832 assert (v != SelfNode, "invariant") ; 833 assert (p == SelfNode, "Node not found on cxq") ; 834 assert (p != _cxq, "invariant") ; 835 assert (q != NULL, "invariant") ; 836 assert (q->_next == p, "invariant") ; 837 q->_next = p->_next ; 838 } 839 TEVENT (Unlink from cxq) ; 840 } 841 842 // Diagnostic hygiene ... 843 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 844 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 845 SelfNode->TState = ObjectWaiter::TS_RUN ; 846 } 847 848 // ----------------------------------------------------------------------------- 849 // Exit support 850 // 851 // exit() 852 // ~~~~~~ 853 // Note that the collector can't reclaim the objectMonitor or deflate 854 // the object out from underneath the thread calling ::exit() as the 855 // thread calling ::exit() never transitions to a stable state. 856 // This inhibits GC, which in turn inhibits asynchronous (and 857 // inopportune) reclamation of "this". 858 // 859 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 860 // There's one exception to the claim above, however. EnterI() can call 861 // exit() to drop a lock if the acquirer has been externally suspended. 862 // In that case exit() is called with _thread_state as _thread_blocked, 863 // but the monitor's _count field is > 0, which inhibits reclamation. 864 // 865 // 1-0 exit 866 // ~~~~~~~~ 867 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 868 // the fast-path operators have been optimized so the common ::exit() 869 // operation is 1-0. See i486.ad fast_unlock(), for instance. 870 // The code emitted by fast_unlock() elides the usual MEMBAR. This 871 // greatly improves latency -- MEMBAR and CAS having considerable local 872 // latency on modern processors -- but at the cost of "stranding". Absent the 873 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 874 // ::enter() path, resulting in the entering thread being stranding 875 // and a progress-liveness failure. Stranding is extremely rare. 876 // We use timers (timed park operations) & periodic polling to detect 877 // and recover from stranding. Potentially stranded threads periodically 878 // wake up and poll the lock. See the usage of the _Responsible variable. 879 // 880 // The CAS() in enter provides for safety and exclusion, while the CAS or 881 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 882 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 883 // We detect and recover from stranding with timers. 884 // 885 // If a thread transiently strands it'll park until (a) another 886 // thread acquires the lock and then drops the lock, at which time the 887 // exiting thread will notice and unpark the stranded thread, or, (b) 888 // the timer expires. If the lock is high traffic then the stranding latency 889 // will be low due to (a). If the lock is low traffic then the odds of 890 // stranding are lower, although the worst-case stranding latency 891 // is longer. Critically, we don't want to put excessive load in the 892 // platform's timer subsystem. We want to minimize both the timer injection 893 // rate (timers created/sec) as well as the number of timers active at 894 // any one time. (more precisely, we want to minimize timer-seconds, which is 895 // the integral of the # of active timers at any instant over time). 896 // Both impinge on OS scalability. Given that, at most one thread parked on 897 // a monitor will use a timer. 898 899 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 900 Thread * Self = THREAD ; 901 if (THREAD != _owner) { 902 if (THREAD->is_lock_owned((address) _owner)) { 903 // Transmute _owner from a BasicLock pointer to a Thread address. 904 // We don't need to hold _mutex for this transition. 905 // Non-null to Non-null is safe as long as all readers can 906 // tolerate either flavor. 907 assert (_recursions == 0, "invariant") ; 908 _owner = THREAD ; 909 _recursions = 0 ; 910 OwnerIsThread = 1 ; 911 } else { 912 // NOTE: we need to handle unbalanced monitor enter/exit 913 // in native code by throwing an exception. 914 // TODO: Throw an IllegalMonitorStateException ? 915 TEVENT (Exit - Throw IMSX) ; 916 assert(false, "Non-balanced monitor enter/exit!"); 917 if (false) { 918 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 919 } 920 return; 921 } 922 } 923 924 if (_recursions != 0) { 925 _recursions--; // this is simple recursive enter 926 TEVENT (Inflated exit - recursive) ; 927 return ; 928 } 929 930 // Invariant: after setting Responsible=null an thread must execute 931 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 932 if ((SyncFlags & 4) == 0) { 933 _Responsible = NULL ; 934 } 935 936 #if INCLUDE_TRACE 937 // get the owner's thread id for the MonitorEnter event 938 // if it is enabled and the thread isn't suspended 939 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 940 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 941 } 942 #endif 943 944 for (;;) { 945 assert (THREAD == _owner, "invariant") ; 946 947 948 if (Knob_ExitPolicy == 0) { 949 // release semantics: prior loads and stores from within the critical section 950 // must not float (reorder) past the following store that drops the lock. 951 // On SPARC that requires MEMBAR #loadstore|#storestore. 952 // But of course in TSO #loadstore|#storestore is not required. 953 // I'd like to write one of the following: 954 // A. OrderAccess::release() ; _owner = NULL 955 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 956 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 957 // store into a _dummy variable. That store is not needed, but can result 958 // in massive wasteful coherency traffic on classic SMP systems. 959 // Instead, I use release_store(), which is implemented as just a simple 960 // ST on x64, x86 and SPARC. 961 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 962 OrderAccess::storeload() ; // See if we need to wake a successor 963 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 964 TEVENT (Inflated exit - simple egress) ; 965 return ; 966 } 967 TEVENT (Inflated exit - complex egress) ; 968 969 // Normally the exiting thread is responsible for ensuring succession, 970 // but if other successors are ready or other entering threads are spinning 971 // then this thread can simply store NULL into _owner and exit without 972 // waking a successor. The existence of spinners or ready successors 973 // guarantees proper succession (liveness). Responsibility passes to the 974 // ready or running successors. The exiting thread delegates the duty. 975 // More precisely, if a successor already exists this thread is absolved 976 // of the responsibility of waking (unparking) one. 977 // 978 // The _succ variable is critical to reducing futile wakeup frequency. 979 // _succ identifies the "heir presumptive" thread that has been made 980 // ready (unparked) but that has not yet run. We need only one such 981 // successor thread to guarantee progress. 982 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 983 // section 3.3 "Futile Wakeup Throttling" for details. 984 // 985 // Note that spinners in Enter() also set _succ non-null. 986 // In the current implementation spinners opportunistically set 987 // _succ so that exiting threads might avoid waking a successor. 988 // Another less appealing alternative would be for the exiting thread 989 // to drop the lock and then spin briefly to see if a spinner managed 990 // to acquire the lock. If so, the exiting thread could exit 991 // immediately without waking a successor, otherwise the exiting 992 // thread would need to dequeue and wake a successor. 993 // (Note that we'd need to make the post-drop spin short, but no 994 // shorter than the worst-case round-trip cache-line migration time. 995 // The dropped lock needs to become visible to the spinner, and then 996 // the acquisition of the lock by the spinner must become visible to 997 // the exiting thread). 998 // 999 1000 // It appears that an heir-presumptive (successor) must be made ready. 1001 // Only the current lock owner can manipulate the EntryList or 1002 // drain _cxq, so we need to reacquire the lock. If we fail 1003 // to reacquire the lock the responsibility for ensuring succession 1004 // falls to the new owner. 1005 // 1006 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1007 return ; 1008 } 1009 TEVENT (Exit - Reacquired) ; 1010 } else { 1011 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1012 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1013 OrderAccess::storeload() ; 1014 // Ratify the previously observed values. 1015 if (_cxq == NULL || _succ != NULL) { 1016 TEVENT (Inflated exit - simple egress) ; 1017 return ; 1018 } 1019 1020 // inopportune interleaving -- the exiting thread (this thread) 1021 // in the fast-exit path raced an entering thread in the slow-enter 1022 // path. 1023 // We have two choices: 1024 // A. Try to reacquire the lock. 1025 // If the CAS() fails return immediately, otherwise 1026 // we either restart/rerun the exit operation, or simply 1027 // fall-through into the code below which wakes a successor. 1028 // B. If the elements forming the EntryList|cxq are TSM 1029 // we could simply unpark() the lead thread and return 1030 // without having set _succ. 1031 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1032 TEVENT (Inflated exit - reacquired succeeded) ; 1033 return ; 1034 } 1035 TEVENT (Inflated exit - reacquired failed) ; 1036 } else { 1037 TEVENT (Inflated exit - complex egress) ; 1038 } 1039 } 1040 1041 guarantee (_owner == THREAD, "invariant") ; 1042 1043 ObjectWaiter * w = NULL ; 1044 int QMode = Knob_QMode ; 1045 1046 if (QMode == 2 && _cxq != NULL) { 1047 // QMode == 2 : cxq has precedence over EntryList. 1048 // Try to directly wake a successor from the cxq. 1049 // If successful, the successor will need to unlink itself from cxq. 1050 w = _cxq ; 1051 assert (w != NULL, "invariant") ; 1052 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1053 ExitEpilog (Self, w) ; 1054 return ; 1055 } 1056 1057 if (QMode == 3 && _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 // Drain _cxq into EntryList - bulk transfer. 1061 // First, detach _cxq. 1062 // The following loop is tantamount to: w = swap (&cxq, NULL) 1063 w = _cxq ; 1064 for (;;) { 1065 assert (w != NULL, "Invariant") ; 1066 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1067 if (u == w) break ; 1068 w = u ; 1069 } 1070 assert (w != NULL , "invariant") ; 1071 1072 ObjectWaiter * q = NULL ; 1073 ObjectWaiter * p ; 1074 for (p = w ; p != NULL ; p = p->_next) { 1075 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1076 p->TState = ObjectWaiter::TS_ENTER ; 1077 p->_prev = q ; 1078 q = p ; 1079 } 1080 1081 // Append the RATs to the EntryList 1082 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1083 ObjectWaiter * Tail ; 1084 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1085 if (Tail == NULL) { 1086 _EntryList = w ; 1087 } else { 1088 Tail->_next = w ; 1089 w->_prev = Tail ; 1090 } 1091 1092 // Fall thru into code that tries to wake a successor from EntryList 1093 } 1094 1095 if (QMode == 4 && _cxq != NULL) { 1096 // Aggressively drain cxq into EntryList at the first opportunity. 1097 // This policy ensure that recently-run threads live at the head of EntryList. 1098 1099 // Drain _cxq into EntryList - bulk transfer. 1100 // First, detach _cxq. 1101 // The following loop is tantamount to: w = swap (&cxq, NULL) 1102 w = _cxq ; 1103 for (;;) { 1104 assert (w != NULL, "Invariant") ; 1105 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1106 if (u == w) break ; 1107 w = u ; 1108 } 1109 assert (w != NULL , "invariant") ; 1110 1111 ObjectWaiter * q = NULL ; 1112 ObjectWaiter * p ; 1113 for (p = w ; p != NULL ; p = p->_next) { 1114 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1115 p->TState = ObjectWaiter::TS_ENTER ; 1116 p->_prev = q ; 1117 q = p ; 1118 } 1119 1120 // Prepend the RATs to the EntryList 1121 if (_EntryList != NULL) { 1122 q->_next = _EntryList ; 1123 _EntryList->_prev = q ; 1124 } 1125 _EntryList = w ; 1126 1127 // Fall thru into code that tries to wake a successor from EntryList 1128 } 1129 1130 w = _EntryList ; 1131 if (w != NULL) { 1132 // I'd like to write: guarantee (w->_thread != Self). 1133 // But in practice an exiting thread may find itself on the EntryList. 1134 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1135 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1136 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1137 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1138 // release the lock "O". T2 resumes immediately after the ST of null into 1139 // _owner, above. T2 notices that the EntryList is populated, so it 1140 // reacquires the lock and then finds itself on the EntryList. 1141 // Given all that, we have to tolerate the circumstance where "w" is 1142 // associated with Self. 1143 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1144 ExitEpilog (Self, w) ; 1145 return ; 1146 } 1147 1148 // If we find that both _cxq and EntryList are null then just 1149 // re-run the exit protocol from the top. 1150 w = _cxq ; 1151 if (w == NULL) continue ; 1152 1153 // Drain _cxq into EntryList - bulk transfer. 1154 // First, detach _cxq. 1155 // The following loop is tantamount to: w = swap (&cxq, NULL) 1156 for (;;) { 1157 assert (w != NULL, "Invariant") ; 1158 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1159 if (u == w) break ; 1160 w = u ; 1161 } 1162 TEVENT (Inflated exit - drain cxq into EntryList) ; 1163 1164 assert (w != NULL , "invariant") ; 1165 assert (_EntryList == NULL , "invariant") ; 1166 1167 // Convert the LIFO SLL anchored by _cxq into a DLL. 1168 // The list reorganization step operates in O(LENGTH(w)) time. 1169 // It's critical that this step operate quickly as 1170 // "Self" still holds the outer-lock, restricting parallelism 1171 // and effectively lengthening the critical section. 1172 // Invariant: s chases t chases u. 1173 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1174 // we have faster access to the tail. 1175 1176 if (QMode == 1) { 1177 // QMode == 1 : drain cxq to EntryList, reversing order 1178 // We also reverse the order of the list. 1179 ObjectWaiter * s = NULL ; 1180 ObjectWaiter * t = w ; 1181 ObjectWaiter * u = NULL ; 1182 while (t != NULL) { 1183 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1184 t->TState = ObjectWaiter::TS_ENTER ; 1185 u = t->_next ; 1186 t->_prev = u ; 1187 t->_next = s ; 1188 s = t; 1189 t = u ; 1190 } 1191 _EntryList = s ; 1192 assert (s != NULL, "invariant") ; 1193 } else { 1194 // QMode == 0 or QMode == 2 1195 _EntryList = w ; 1196 ObjectWaiter * q = NULL ; 1197 ObjectWaiter * p ; 1198 for (p = w ; p != NULL ; p = p->_next) { 1199 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1200 p->TState = ObjectWaiter::TS_ENTER ; 1201 p->_prev = q ; 1202 q = p ; 1203 } 1204 } 1205 1206 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1207 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1208 1209 // See if we can abdicate to a spinner instead of waking a thread. 1210 // A primary goal of the implementation is to reduce the 1211 // context-switch rate. 1212 if (_succ != NULL) continue; 1213 1214 w = _EntryList ; 1215 if (w != NULL) { 1216 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1217 ExitEpilog (Self, w) ; 1218 return ; 1219 } 1220 } 1221 } 1222 1223 // ExitSuspendEquivalent: 1224 // A faster alternate to handle_special_suspend_equivalent_condition() 1225 // 1226 // handle_special_suspend_equivalent_condition() unconditionally 1227 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1228 // operations have high latency. Note that in ::enter() we call HSSEC 1229 // while holding the monitor, so we effectively lengthen the critical sections. 1230 // 1231 // There are a number of possible solutions: 1232 // 1233 // A. To ameliorate the problem we might also defer state transitions 1234 // to as late as possible -- just prior to parking. 1235 // Given that, we'd call HSSEC after having returned from park(), 1236 // but before attempting to acquire the monitor. This is only a 1237 // partial solution. It avoids calling HSSEC while holding the 1238 // monitor (good), but it still increases successor reacquisition latency -- 1239 // the interval between unparking a successor and the time the successor 1240 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1241 // If we use this technique we can also avoid EnterI()-exit() loop 1242 // in ::enter() where we iteratively drop the lock and then attempt 1243 // to reacquire it after suspending. 1244 // 1245 // B. In the future we might fold all the suspend bits into a 1246 // composite per-thread suspend flag and then update it with CAS(). 1247 // Alternately, a Dekker-like mechanism with multiple variables 1248 // would suffice: 1249 // ST Self->_suspend_equivalent = false 1250 // MEMBAR 1251 // LD Self_>_suspend_flags 1252 // 1253 1254 1255 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1256 int Mode = Knob_FastHSSEC ; 1257 if (Mode && !jSelf->is_external_suspend()) { 1258 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1259 jSelf->clear_suspend_equivalent() ; 1260 if (2 == Mode) OrderAccess::storeload() ; 1261 if (!jSelf->is_external_suspend()) return false ; 1262 // We raced a suspension -- fall thru into the slow path 1263 TEVENT (ExitSuspendEquivalent - raced) ; 1264 jSelf->set_suspend_equivalent() ; 1265 } 1266 return jSelf->handle_special_suspend_equivalent_condition() ; 1267 } 1268 1269 1270 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1271 assert (_owner == Self, "invariant") ; 1272 1273 // Exit protocol: 1274 // 1. ST _succ = wakee 1275 // 2. membar #loadstore|#storestore; 1276 // 2. ST _owner = NULL 1277 // 3. unpark(wakee) 1278 1279 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1280 ParkEvent * Trigger = Wakee->_event ; 1281 1282 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1283 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1284 // out-of-scope (non-extant). 1285 Wakee = NULL ; 1286 1287 // Drop the lock 1288 OrderAccess::release_store_ptr (&_owner, NULL) ; 1289 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1290 1291 if (SafepointSynchronize::do_call_back()) { 1292 TEVENT (unpark before SAFEPOINT) ; 1293 } 1294 1295 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1296 Trigger->unpark() ; 1297 1298 // Maintain stats and report events to JVMTI 1299 if (ObjectMonitor::_sync_Parks != NULL) { 1300 ObjectMonitor::_sync_Parks->inc() ; 1301 } 1302 } 1303 1304 1305 // ----------------------------------------------------------------------------- 1306 // Class Loader deadlock handling. 1307 // 1308 // complete_exit exits a lock returning recursion count 1309 // complete_exit/reenter operate as a wait without waiting 1310 // complete_exit requires an inflated monitor 1311 // The _owner field is not always the Thread addr even with an 1312 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1313 // thread due to contention. 1314 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1315 Thread * const Self = THREAD; 1316 assert(Self->is_Java_thread(), "Must be Java thread!"); 1317 JavaThread *jt = (JavaThread *)THREAD; 1318 1319 DeferredInitialize(); 1320 1321 if (THREAD != _owner) { 1322 if (THREAD->is_lock_owned ((address)_owner)) { 1323 assert(_recursions == 0, "internal state error"); 1324 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1325 _recursions = 0 ; 1326 OwnerIsThread = 1 ; 1327 } 1328 } 1329 1330 guarantee(Self == _owner, "complete_exit not owner"); 1331 intptr_t save = _recursions; // record the old recursion count 1332 _recursions = 0; // set the recursion level to be 0 1333 exit (true, Self) ; // exit the monitor 1334 guarantee (_owner != Self, "invariant"); 1335 return save; 1336 } 1337 1338 // reenter() enters a lock and sets recursion count 1339 // complete_exit/reenter operate as a wait without waiting 1340 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1341 Thread * const Self = THREAD; 1342 assert(Self->is_Java_thread(), "Must be Java thread!"); 1343 JavaThread *jt = (JavaThread *)THREAD; 1344 1345 guarantee(_owner != Self, "reenter already owner"); 1346 enter (THREAD); // enter the monitor 1347 guarantee (_recursions == 0, "reenter recursion"); 1348 _recursions = recursions; 1349 return; 1350 } 1351 1352 1353 // ----------------------------------------------------------------------------- 1354 // A macro is used below because there may already be a pending 1355 // exception which should not abort the execution of the routines 1356 // which use this (which is why we don't put this into check_slow and 1357 // call it with a CHECK argument). 1358 1359 #define CHECK_OWNER() \ 1360 do { \ 1361 if (THREAD != _owner) { \ 1362 if (THREAD->is_lock_owned((address) _owner)) { \ 1363 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1364 _recursions = 0; \ 1365 OwnerIsThread = 1 ; \ 1366 } else { \ 1367 TEVENT (Throw IMSX) ; \ 1368 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1369 } \ 1370 } \ 1371 } while (false) 1372 1373 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1374 // TODO-FIXME: remove check_slow() -- it's likely dead. 1375 1376 void ObjectMonitor::check_slow(TRAPS) { 1377 TEVENT (check_slow - throw IMSX) ; 1378 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1379 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1380 } 1381 1382 static int Adjust (volatile int * adr, int dx) { 1383 int v ; 1384 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1385 return v ; 1386 } 1387 1388 // helper method for posting a monitor wait event 1389 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1390 jlong notifier_tid, 1391 jlong timeout, 1392 bool timedout) { 1393 event->set_klass(((oop)this->object())->klass()); 1394 event->set_timeout((TYPE_ULONG)timeout); 1395 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1396 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1397 event->set_timedOut((TYPE_BOOLEAN)timedout); 1398 event->commit(); 1399 } 1400 1401 // ----------------------------------------------------------------------------- 1402 // Wait/Notify/NotifyAll 1403 // 1404 // Note: a subset of changes to ObjectMonitor::wait() 1405 // will need to be replicated in complete_exit above 1406 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1407 Thread * const Self = THREAD ; 1408 assert(Self->is_Java_thread(), "Must be Java thread!"); 1409 JavaThread *jt = (JavaThread *)THREAD; 1410 1411 DeferredInitialize () ; 1412 1413 // Throw IMSX or IEX. 1414 CHECK_OWNER(); 1415 1416 EventJavaMonitorWait event; 1417 1418 // check for a pending interrupt 1419 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1420 // post monitor waited event. Note that this is past-tense, we are done waiting. 1421 if (JvmtiExport::should_post_monitor_waited()) { 1422 // Note: 'false' parameter is passed here because the 1423 // wait was not timed out due to thread interrupt. 1424 JvmtiExport::post_monitor_waited(jt, this, false); 1425 } 1426 if (event.should_commit()) { 1427 post_monitor_wait_event(&event, 0, millis, false); 1428 } 1429 TEVENT (Wait - Throw IEX) ; 1430 THROW(vmSymbols::java_lang_InterruptedException()); 1431 return ; 1432 } 1433 1434 TEVENT (Wait) ; 1435 1436 assert (Self->_Stalled == 0, "invariant") ; 1437 Self->_Stalled = intptr_t(this) ; 1438 jt->set_current_waiting_monitor(this); 1439 1440 // create a node to be put into the queue 1441 // Critically, after we reset() the event but prior to park(), we must check 1442 // for a pending interrupt. 1443 ObjectWaiter node(Self); 1444 node.TState = ObjectWaiter::TS_WAIT ; 1445 Self->_ParkEvent->reset() ; 1446 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1447 1448 // Enter the waiting queue, which is a circular doubly linked list in this case 1449 // but it could be a priority queue or any data structure. 1450 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1451 // by the the owner of the monitor *except* in the case where park() 1452 // returns because of a timeout of interrupt. Contention is exceptionally rare 1453 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1454 1455 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1456 AddWaiter (&node) ; 1457 Thread::SpinRelease (&_WaitSetLock) ; 1458 1459 if ((SyncFlags & 4) == 0) { 1460 _Responsible = NULL ; 1461 } 1462 intptr_t save = _recursions; // record the old recursion count 1463 _waiters++; // increment the number of waiters 1464 _recursions = 0; // set the recursion level to be 1 1465 exit (true, Self) ; // exit the monitor 1466 guarantee (_owner != Self, "invariant") ; 1467 1468 // As soon as the ObjectMonitor's ownership is dropped in the exit() 1469 // call above, another thread can enter() the ObjectMonitor, do the 1470 // notify(), and exit() the ObjectMonitor. If the other thread's 1471 // exit() call chooses this thread as the successor and the unpark() 1472 // call happens to occur while this thread is posting a 1473 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event 1474 // handler using RawMonitors and consuming the unpark(). 1475 // 1476 // To avoid the problem, we re-post the event. This does no harm 1477 // even if the original unpark() was not consumed because we are the 1478 // chosen successor for this monitor. 1479 if (node._notified != 0 && _succ == Self) { 1480 node._event->unpark(); 1481 } 1482 1483 // The thread is on the WaitSet list - now park() it. 1484 // On MP systems it's conceivable that a brief spin before we park 1485 // could be profitable. 1486 // 1487 // TODO-FIXME: change the following logic to a loop of the form 1488 // while (!timeout && !interrupted && _notified == 0) park() 1489 1490 int ret = OS_OK ; 1491 int WasNotified = 0 ; 1492 { // State transition wrappers 1493 OSThread* osthread = Self->osthread(); 1494 OSThreadWaitState osts(osthread, true); 1495 { 1496 ThreadBlockInVM tbivm(jt); 1497 // Thread is in thread_blocked state and oop access is unsafe. 1498 jt->set_suspend_equivalent(); 1499 1500 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1501 // Intentionally empty 1502 } else 1503 if (node._notified == 0) { 1504 if (millis <= 0) { 1505 Self->_ParkEvent->park () ; 1506 } else { 1507 ret = Self->_ParkEvent->park (millis) ; 1508 } 1509 } 1510 1511 // were we externally suspended while we were waiting? 1512 if (ExitSuspendEquivalent (jt)) { 1513 // TODO-FIXME: add -- if succ == Self then succ = null. 1514 jt->java_suspend_self(); 1515 } 1516 1517 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1518 1519 1520 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1521 // from the WaitSet to the EntryList. 1522 // See if we need to remove Node from the WaitSet. 1523 // We use double-checked locking to avoid grabbing _WaitSetLock 1524 // if the thread is not on the wait queue. 1525 // 1526 // Note that we don't need a fence before the fetch of TState. 1527 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1528 // written by the is thread. (perhaps the fetch might even be satisfied 1529 // by a look-aside into the processor's own store buffer, although given 1530 // the length of the code path between the prior ST and this load that's 1531 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1532 // then we'll acquire the lock and then re-fetch a fresh TState value. 1533 // That is, we fail toward safety. 1534 1535 if (node.TState == ObjectWaiter::TS_WAIT) { 1536 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1537 if (node.TState == ObjectWaiter::TS_WAIT) { 1538 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1539 assert(node._notified == 0, "invariant"); 1540 node.TState = ObjectWaiter::TS_RUN ; 1541 } 1542 Thread::SpinRelease (&_WaitSetLock) ; 1543 } 1544 1545 // The thread is now either on off-list (TS_RUN), 1546 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1547 // The Node's TState variable is stable from the perspective of this thread. 1548 // No other threads will asynchronously modify TState. 1549 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1550 OrderAccess::loadload() ; 1551 if (_succ == Self) _succ = NULL ; 1552 WasNotified = node._notified ; 1553 1554 // Reentry phase -- reacquire the monitor. 1555 // re-enter contended monitor after object.wait(). 1556 // retain OBJECT_WAIT state until re-enter successfully completes 1557 // Thread state is thread_in_vm and oop access is again safe, 1558 // although the raw address of the object may have changed. 1559 // (Don't cache naked oops over safepoints, of course). 1560 1561 // post monitor waited event. Note that this is past-tense, we are done waiting. 1562 if (JvmtiExport::should_post_monitor_waited()) { 1563 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1564 } 1565 1566 if (event.should_commit()) { 1567 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1568 } 1569 1570 OrderAccess::fence() ; 1571 1572 assert (Self->_Stalled != 0, "invariant") ; 1573 Self->_Stalled = 0 ; 1574 1575 assert (_owner != Self, "invariant") ; 1576 ObjectWaiter::TStates v = node.TState ; 1577 if (v == ObjectWaiter::TS_RUN) { 1578 enter (Self) ; 1579 } else { 1580 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1581 ReenterI (Self, &node) ; 1582 node.wait_reenter_end(this); 1583 } 1584 1585 // Self has reacquired the lock. 1586 // Lifecycle - the node representing Self must not appear on any queues. 1587 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1588 // want residual elements associated with this thread left on any lists. 1589 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1590 assert (_owner == Self, "invariant") ; 1591 assert (_succ != Self , "invariant") ; 1592 } // OSThreadWaitState() 1593 1594 jt->set_current_waiting_monitor(NULL); 1595 1596 guarantee (_recursions == 0, "invariant") ; 1597 _recursions = save; // restore the old recursion count 1598 _waiters--; // decrement the number of waiters 1599 1600 // Verify a few postconditions 1601 assert (_owner == Self , "invariant") ; 1602 assert (_succ != Self , "invariant") ; 1603 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1604 1605 if (SyncFlags & 32) { 1606 OrderAccess::fence() ; 1607 } 1608 1609 // check if the notification happened 1610 if (!WasNotified) { 1611 // no, it could be timeout or Thread.interrupt() or both 1612 // check for interrupt event, otherwise it is timeout 1613 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1614 TEVENT (Wait - throw IEX from epilog) ; 1615 THROW(vmSymbols::java_lang_InterruptedException()); 1616 } 1617 } 1618 1619 // NOTE: Spurious wake up will be consider as timeout. 1620 // Monitor notify has precedence over thread interrupt. 1621 } 1622 1623 1624 // Consider: 1625 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1626 // then instead of transferring a thread from the WaitSet to the EntryList 1627 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1628 1629 void ObjectMonitor::notify(TRAPS) { 1630 CHECK_OWNER(); 1631 if (_WaitSet == NULL) { 1632 TEVENT (Empty-Notify) ; 1633 return ; 1634 } 1635 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1636 1637 int Policy = Knob_MoveNotifyee ; 1638 1639 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1640 ObjectWaiter * iterator = DequeueWaiter() ; 1641 if (iterator != NULL) { 1642 TEVENT (Notify1 - Transfer) ; 1643 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1644 guarantee (iterator->_notified == 0, "invariant") ; 1645 if (Policy != 4) { 1646 iterator->TState = ObjectWaiter::TS_ENTER ; 1647 } 1648 iterator->_notified = 1 ; 1649 Thread * Self = THREAD; 1650 iterator->_notifier_tid = Self->osthread()->thread_id(); 1651 1652 ObjectWaiter * List = _EntryList ; 1653 if (List != NULL) { 1654 assert (List->_prev == NULL, "invariant") ; 1655 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1656 assert (List != iterator, "invariant") ; 1657 } 1658 1659 if (Policy == 0) { // prepend to EntryList 1660 if (List == NULL) { 1661 iterator->_next = iterator->_prev = NULL ; 1662 _EntryList = iterator ; 1663 } else { 1664 List->_prev = iterator ; 1665 iterator->_next = List ; 1666 iterator->_prev = NULL ; 1667 _EntryList = iterator ; 1668 } 1669 } else 1670 if (Policy == 1) { // append to EntryList 1671 if (List == NULL) { 1672 iterator->_next = iterator->_prev = NULL ; 1673 _EntryList = iterator ; 1674 } else { 1675 // CONSIDER: finding the tail currently requires a linear-time walk of 1676 // the EntryList. We can make tail access constant-time by converting to 1677 // a CDLL instead of using our current DLL. 1678 ObjectWaiter * Tail ; 1679 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1680 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1681 Tail->_next = iterator ; 1682 iterator->_prev = Tail ; 1683 iterator->_next = NULL ; 1684 } 1685 } else 1686 if (Policy == 2) { // prepend to cxq 1687 // prepend to cxq 1688 if (List == NULL) { 1689 iterator->_next = iterator->_prev = NULL ; 1690 _EntryList = iterator ; 1691 } else { 1692 iterator->TState = ObjectWaiter::TS_CXQ ; 1693 for (;;) { 1694 ObjectWaiter * Front = _cxq ; 1695 iterator->_next = Front ; 1696 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1697 break ; 1698 } 1699 } 1700 } 1701 } else 1702 if (Policy == 3) { // append to cxq 1703 iterator->TState = ObjectWaiter::TS_CXQ ; 1704 for (;;) { 1705 ObjectWaiter * Tail ; 1706 Tail = _cxq ; 1707 if (Tail == NULL) { 1708 iterator->_next = NULL ; 1709 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1710 break ; 1711 } 1712 } else { 1713 while (Tail->_next != NULL) Tail = Tail->_next ; 1714 Tail->_next = iterator ; 1715 iterator->_prev = Tail ; 1716 iterator->_next = NULL ; 1717 break ; 1718 } 1719 } 1720 } else { 1721 ParkEvent * ev = iterator->_event ; 1722 iterator->TState = ObjectWaiter::TS_RUN ; 1723 OrderAccess::fence() ; 1724 ev->unpark() ; 1725 } 1726 1727 if (Policy < 4) { 1728 iterator->wait_reenter_begin(this); 1729 } 1730 1731 // _WaitSetLock protects the wait queue, not the EntryList. We could 1732 // move the add-to-EntryList operation, above, outside the critical section 1733 // protected by _WaitSetLock. In practice that's not useful. With the 1734 // exception of wait() timeouts and interrupts the monitor owner 1735 // is the only thread that grabs _WaitSetLock. There's almost no contention 1736 // on _WaitSetLock so it's not profitable to reduce the length of the 1737 // critical section. 1738 } 1739 1740 Thread::SpinRelease (&_WaitSetLock) ; 1741 1742 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1743 ObjectMonitor::_sync_Notifications->inc() ; 1744 } 1745 } 1746 1747 1748 void ObjectMonitor::notifyAll(TRAPS) { 1749 CHECK_OWNER(); 1750 ObjectWaiter* iterator; 1751 if (_WaitSet == NULL) { 1752 TEVENT (Empty-NotifyAll) ; 1753 return ; 1754 } 1755 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1756 1757 int Policy = Knob_MoveNotifyee ; 1758 int Tally = 0 ; 1759 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1760 1761 for (;;) { 1762 iterator = DequeueWaiter () ; 1763 if (iterator == NULL) break ; 1764 TEVENT (NotifyAll - Transfer1) ; 1765 ++Tally ; 1766 1767 // Disposition - what might we do with iterator ? 1768 // a. add it directly to the EntryList - either tail or head. 1769 // b. push it onto the front of the _cxq. 1770 // For now we use (a). 1771 1772 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1773 guarantee (iterator->_notified == 0, "invariant") ; 1774 iterator->_notified = 1 ; 1775 Thread * Self = THREAD; 1776 iterator->_notifier_tid = Self->osthread()->thread_id(); 1777 if (Policy != 4) { 1778 iterator->TState = ObjectWaiter::TS_ENTER ; 1779 } 1780 1781 ObjectWaiter * List = _EntryList ; 1782 if (List != NULL) { 1783 assert (List->_prev == NULL, "invariant") ; 1784 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1785 assert (List != iterator, "invariant") ; 1786 } 1787 1788 if (Policy == 0) { // prepend to EntryList 1789 if (List == NULL) { 1790 iterator->_next = iterator->_prev = NULL ; 1791 _EntryList = iterator ; 1792 } else { 1793 List->_prev = iterator ; 1794 iterator->_next = List ; 1795 iterator->_prev = NULL ; 1796 _EntryList = iterator ; 1797 } 1798 } else 1799 if (Policy == 1) { // append to EntryList 1800 if (List == NULL) { 1801 iterator->_next = iterator->_prev = NULL ; 1802 _EntryList = iterator ; 1803 } else { 1804 // CONSIDER: finding the tail currently requires a linear-time walk of 1805 // the EntryList. We can make tail access constant-time by converting to 1806 // a CDLL instead of using our current DLL. 1807 ObjectWaiter * Tail ; 1808 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1809 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1810 Tail->_next = iterator ; 1811 iterator->_prev = Tail ; 1812 iterator->_next = NULL ; 1813 } 1814 } else 1815 if (Policy == 2) { // prepend to cxq 1816 // prepend to cxq 1817 iterator->TState = ObjectWaiter::TS_CXQ ; 1818 for (;;) { 1819 ObjectWaiter * Front = _cxq ; 1820 iterator->_next = Front ; 1821 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1822 break ; 1823 } 1824 } 1825 } else 1826 if (Policy == 3) { // append to cxq 1827 iterator->TState = ObjectWaiter::TS_CXQ ; 1828 for (;;) { 1829 ObjectWaiter * Tail ; 1830 Tail = _cxq ; 1831 if (Tail == NULL) { 1832 iterator->_next = NULL ; 1833 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1834 break ; 1835 } 1836 } else { 1837 while (Tail->_next != NULL) Tail = Tail->_next ; 1838 Tail->_next = iterator ; 1839 iterator->_prev = Tail ; 1840 iterator->_next = NULL ; 1841 break ; 1842 } 1843 } 1844 } else { 1845 ParkEvent * ev = iterator->_event ; 1846 iterator->TState = ObjectWaiter::TS_RUN ; 1847 OrderAccess::fence() ; 1848 ev->unpark() ; 1849 } 1850 1851 if (Policy < 4) { 1852 iterator->wait_reenter_begin(this); 1853 } 1854 1855 // _WaitSetLock protects the wait queue, not the EntryList. We could 1856 // move the add-to-EntryList operation, above, outside the critical section 1857 // protected by _WaitSetLock. In practice that's not useful. With the 1858 // exception of wait() timeouts and interrupts the monitor owner 1859 // is the only thread that grabs _WaitSetLock. There's almost no contention 1860 // on _WaitSetLock so it's not profitable to reduce the length of the 1861 // critical section. 1862 } 1863 1864 Thread::SpinRelease (&_WaitSetLock) ; 1865 1866 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1867 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1868 } 1869 } 1870 1871 // ----------------------------------------------------------------------------- 1872 // Adaptive Spinning Support 1873 // 1874 // Adaptive spin-then-block - rational spinning 1875 // 1876 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1877 // algorithm. On high order SMP systems it would be better to start with 1878 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1879 // a contending thread could enqueue itself on the cxq and then spin locally 1880 // on a thread-specific variable such as its ParkEvent._Event flag. 1881 // That's left as an exercise for the reader. Note that global spinning is 1882 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1883 // low latency and massive bandwidth. 1884 // 1885 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1886 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1887 // (duration) or we can fix the count at approximately the duration of 1888 // a context switch and vary the frequency. Of course we could also 1889 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1890 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1891 // 1892 // This implementation varies the duration "D", where D varies with 1893 // the success rate of recent spin attempts. (D is capped at approximately 1894 // length of a round-trip context switch). The success rate for recent 1895 // spin attempts is a good predictor of the success rate of future spin 1896 // attempts. The mechanism adapts automatically to varying critical 1897 // section length (lock modality), system load and degree of parallelism. 1898 // D is maintained per-monitor in _SpinDuration and is initialized 1899 // optimistically. Spin frequency is fixed at 100%. 1900 // 1901 // Note that _SpinDuration is volatile, but we update it without locks 1902 // or atomics. The code is designed so that _SpinDuration stays within 1903 // a reasonable range even in the presence of races. The arithmetic 1904 // operations on _SpinDuration are closed over the domain of legal values, 1905 // so at worst a race will install and older but still legal value. 1906 // At the very worst this introduces some apparent non-determinism. 1907 // We might spin when we shouldn't or vice-versa, but since the spin 1908 // count are relatively short, even in the worst case, the effect is harmless. 1909 // 1910 // Care must be taken that a low "D" value does not become an 1911 // an absorbing state. Transient spinning failures -- when spinning 1912 // is overall profitable -- should not cause the system to converge 1913 // on low "D" values. We want spinning to be stable and predictable 1914 // and fairly responsive to change and at the same time we don't want 1915 // it to oscillate, become metastable, be "too" non-deterministic, 1916 // or converge on or enter undesirable stable absorbing states. 1917 // 1918 // We implement a feedback-based control system -- using past behavior 1919 // to predict future behavior. We face two issues: (a) if the 1920 // input signal is random then the spin predictor won't provide optimal 1921 // results, and (b) if the signal frequency is too high then the control 1922 // system, which has some natural response lag, will "chase" the signal. 1923 // (b) can arise from multimodal lock hold times. Transient preemption 1924 // can also result in apparent bimodal lock hold times. 1925 // Although sub-optimal, neither condition is particularly harmful, as 1926 // in the worst-case we'll spin when we shouldn't or vice-versa. 1927 // The maximum spin duration is rather short so the failure modes aren't bad. 1928 // To be conservative, I've tuned the gain in system to bias toward 1929 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1930 // "rings" or oscillates between spinning and not spinning. This happens 1931 // when spinning is just on the cusp of profitability, however, so the 1932 // situation is not dire. The state is benign -- there's no need to add 1933 // hysteresis control to damp the transition rate between spinning and 1934 // not spinning. 1935 // 1936 1937 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1938 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1939 1940 // Spinning: Fixed frequency (100%), vary duration 1941 1942 1943 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1944 1945 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1946 int ctr = Knob_FixedSpin ; 1947 if (ctr != 0) { 1948 while (--ctr >= 0) { 1949 if (TryLock (Self) > 0) return 1 ; 1950 SpinPause () ; 1951 } 1952 return 0 ; 1953 } 1954 1955 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 1956 if (TryLock(Self) > 0) { 1957 // Increase _SpinDuration ... 1958 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1959 // Raising _SpurDuration to the poverty line is key. 1960 int x = _SpinDuration ; 1961 if (x < Knob_SpinLimit) { 1962 if (x < Knob_Poverty) x = Knob_Poverty ; 1963 _SpinDuration = x + Knob_BonusB ; 1964 } 1965 return 1 ; 1966 } 1967 SpinPause () ; 1968 } 1969 1970 // Admission control - verify preconditions for spinning 1971 // 1972 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1973 // becoming an absorbing state. Put another way, we spin briefly to 1974 // sample, just in case the system load, parallelism, contention, or lock 1975 // modality changed. 1976 // 1977 // Consider the following alternative: 1978 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1979 // spin attempt. "Periodically" might mean after a tally of 1980 // the # of failed spin attempts (or iterations) reaches some threshold. 1981 // This takes us into the realm of 1-out-of-N spinning, where we 1982 // hold the duration constant but vary the frequency. 1983 1984 ctr = _SpinDuration ; 1985 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 1986 if (ctr <= 0) return 0 ; 1987 1988 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 1989 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1990 TEVENT (Spin abort - notrunnable [TOP]); 1991 return 0 ; 1992 } 1993 1994 int MaxSpin = Knob_MaxSpinners ; 1995 if (MaxSpin >= 0) { 1996 if (_Spinner > MaxSpin) { 1997 TEVENT (Spin abort -- too many spinners) ; 1998 return 0 ; 1999 } 2000 // Slightly racy, but benign ... 2001 Adjust (&_Spinner, 1) ; 2002 } 2003 2004 // We're good to spin ... spin ingress. 2005 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2006 // when preparing to LD...CAS _owner, etc and the CAS is likely 2007 // to succeed. 2008 int hits = 0 ; 2009 int msk = 0 ; 2010 int caspty = Knob_CASPenalty ; 2011 int oxpty = Knob_OXPenalty ; 2012 int sss = Knob_SpinSetSucc ; 2013 if (sss && _succ == NULL ) _succ = Self ; 2014 Thread * prv = NULL ; 2015 2016 // There are three ways to exit the following loop: 2017 // 1. A successful spin where this thread has acquired the lock. 2018 // 2. Spin failure with prejudice 2019 // 3. Spin failure without prejudice 2020 2021 while (--ctr >= 0) { 2022 2023 // Periodic polling -- Check for pending GC 2024 // Threads may spin while they're unsafe. 2025 // We don't want spinning threads to delay the JVM from reaching 2026 // a stop-the-world safepoint or to steal cycles from GC. 2027 // If we detect a pending safepoint we abort in order that 2028 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2029 // this thread, if safe, doesn't steal cycles from GC. 2030 // This is in keeping with the "no loitering in runtime" rule. 2031 // We periodically check to see if there's a safepoint pending. 2032 if ((ctr & 0xFF) == 0) { 2033 if (SafepointSynchronize::do_call_back()) { 2034 TEVENT (Spin: safepoint) ; 2035 goto Abort ; // abrupt spin egress 2036 } 2037 if (Knob_UsePause & 1) SpinPause () ; 2038 2039 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2040 if (hits > 50 && scb != NULL) { 2041 int abend = (*scb)(SpinCallbackArgument, 0) ; 2042 } 2043 } 2044 2045 if (Knob_UsePause & 2) SpinPause() ; 2046 2047 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2048 // This is useful on classic SMP systems, but is of less utility on 2049 // N1-style CMT platforms. 2050 // 2051 // Trade-off: lock acquisition latency vs coherency bandwidth. 2052 // Lock hold times are typically short. A histogram 2053 // of successful spin attempts shows that we usually acquire 2054 // the lock early in the spin. That suggests we want to 2055 // sample _owner frequently in the early phase of the spin, 2056 // but then back-off and sample less frequently as the spin 2057 // progresses. The back-off makes a good citizen on SMP big 2058 // SMP systems. Oversampling _owner can consume excessive 2059 // coherency bandwidth. Relatedly, if we _oversample _owner we 2060 // can inadvertently interfere with the the ST m->owner=null. 2061 // executed by the lock owner. 2062 if (ctr & msk) continue ; 2063 ++hits ; 2064 if ((hits & 0xF) == 0) { 2065 // The 0xF, above, corresponds to the exponent. 2066 // Consider: (msk+1)|msk 2067 msk = ((msk << 2)|3) & BackOffMask ; 2068 } 2069 2070 // Probe _owner with TATAS 2071 // If this thread observes the monitor transition or flicker 2072 // from locked to unlocked to locked, then the odds that this 2073 // thread will acquire the lock in this spin attempt go down 2074 // considerably. The same argument applies if the CAS fails 2075 // or if we observe _owner change from one non-null value to 2076 // another non-null value. In such cases we might abort 2077 // the spin without prejudice or apply a "penalty" to the 2078 // spin count-down variable "ctr", reducing it by 100, say. 2079 2080 Thread * ox = (Thread *) _owner ; 2081 if (ox == NULL) { 2082 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2083 if (ox == NULL) { 2084 // The CAS succeeded -- this thread acquired ownership 2085 // Take care of some bookkeeping to exit spin state. 2086 if (sss && _succ == Self) { 2087 _succ = NULL ; 2088 } 2089 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2090 2091 // Increase _SpinDuration : 2092 // The spin was successful (profitable) so we tend toward 2093 // longer spin attempts in the future. 2094 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2095 // If we acquired the lock early in the spin cycle it 2096 // makes sense to increase _SpinDuration proportionally. 2097 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2098 int x = _SpinDuration ; 2099 if (x < Knob_SpinLimit) { 2100 if (x < Knob_Poverty) x = Knob_Poverty ; 2101 _SpinDuration = x + Knob_Bonus ; 2102 } 2103 return 1 ; 2104 } 2105 2106 // The CAS failed ... we can take any of the following actions: 2107 // * penalize: ctr -= Knob_CASPenalty 2108 // * exit spin with prejudice -- goto Abort; 2109 // * exit spin without prejudice. 2110 // * Since CAS is high-latency, retry again immediately. 2111 prv = ox ; 2112 TEVENT (Spin: cas failed) ; 2113 if (caspty == -2) break ; 2114 if (caspty == -1) goto Abort ; 2115 ctr -= caspty ; 2116 continue ; 2117 } 2118 2119 // Did lock ownership change hands ? 2120 if (ox != prv && prv != NULL ) { 2121 TEVENT (spin: Owner changed) 2122 if (oxpty == -2) break ; 2123 if (oxpty == -1) goto Abort ; 2124 ctr -= oxpty ; 2125 } 2126 prv = ox ; 2127 2128 // Abort the spin if the owner is not executing. 2129 // The owner must be executing in order to drop the lock. 2130 // Spinning while the owner is OFFPROC is idiocy. 2131 // Consider: ctr -= RunnablePenalty ; 2132 if (Knob_OState && NotRunnable (Self, ox)) { 2133 TEVENT (Spin abort - notrunnable); 2134 goto Abort ; 2135 } 2136 if (sss && _succ == NULL ) _succ = Self ; 2137 } 2138 2139 // Spin failed with prejudice -- reduce _SpinDuration. 2140 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2141 // AIMD is globally stable. 2142 TEVENT (Spin failure) ; 2143 { 2144 int x = _SpinDuration ; 2145 if (x > 0) { 2146 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2147 // This is globally sample and tends to damp the response. 2148 x -= Knob_Penalty ; 2149 if (x < 0) x = 0 ; 2150 _SpinDuration = x ; 2151 } 2152 } 2153 2154 Abort: 2155 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2156 if (sss && _succ == Self) { 2157 _succ = NULL ; 2158 // Invariant: after setting succ=null a contending thread 2159 // must recheck-retry _owner before parking. This usually happens 2160 // in the normal usage of TrySpin(), but it's safest 2161 // to make TrySpin() as foolproof as possible. 2162 OrderAccess::fence() ; 2163 if (TryLock(Self) > 0) return 1 ; 2164 } 2165 return 0 ; 2166 } 2167 2168 // NotRunnable() -- informed spinning 2169 // 2170 // Don't bother spinning if the owner is not eligible to drop the lock. 2171 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2172 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2173 // The thread must be runnable in order to drop the lock in timely fashion. 2174 // If the _owner is not runnable then spinning will not likely be 2175 // successful (profitable). 2176 // 2177 // Beware -- the thread referenced by _owner could have died 2178 // so a simply fetch from _owner->_thread_state might trap. 2179 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2180 // Because of the lifecycle issues the schedctl and _thread_state values 2181 // observed by NotRunnable() might be garbage. NotRunnable must 2182 // tolerate this and consider the observed _thread_state value 2183 // as advisory. 2184 // 2185 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2186 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2187 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2188 // with the LSB of _owner. Another option would be to probablistically probe 2189 // the putative _owner->TypeTag value. 2190 // 2191 // Checking _thread_state isn't perfect. Even if the thread is 2192 // in_java it might be blocked on a page-fault or have been preempted 2193 // and sitting on a ready/dispatch queue. _thread state in conjunction 2194 // with schedctl.sc_state gives us a good picture of what the 2195 // thread is doing, however. 2196 // 2197 // TODO: check schedctl.sc_state. 2198 // We'll need to use SafeFetch32() to read from the schedctl block. 2199 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2200 // 2201 // The return value from NotRunnable() is *advisory* -- the 2202 // result is based on sampling and is not necessarily coherent. 2203 // The caller must tolerate false-negative and false-positive errors. 2204 // Spinning, in general, is probabilistic anyway. 2205 2206 2207 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2208 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2209 if (!OwnerIsThread) return 0 ; 2210 2211 if (ox == NULL) return 0 ; 2212 2213 // Avoid transitive spinning ... 2214 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2215 // Immediately after T1 acquires L it's possible that T2, also 2216 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2217 // This occurs transiently after T1 acquired L but before 2218 // T1 managed to clear T1.Stalled. T2 does not need to abort 2219 // its spin in this circumstance. 2220 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2221 2222 if (BlockedOn == 1) return 1 ; 2223 if (BlockedOn != 0) { 2224 return BlockedOn != intptr_t(this) && _owner == ox ; 2225 } 2226 2227 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2228 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2229 // consider also: jst != _thread_in_Java -- but that's overspecific. 2230 return jst == _thread_blocked || jst == _thread_in_native ; 2231 } 2232 2233 2234 // ----------------------------------------------------------------------------- 2235 // WaitSet management ... 2236 2237 ObjectWaiter::ObjectWaiter(Thread* thread) { 2238 _next = NULL; 2239 _prev = NULL; 2240 _notified = 0; 2241 TState = TS_RUN ; 2242 _thread = thread; 2243 _event = thread->_ParkEvent ; 2244 _active = false; 2245 assert (_event != NULL, "invariant") ; 2246 } 2247 2248 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2249 JavaThread *jt = (JavaThread *)this->_thread; 2250 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2251 } 2252 2253 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2254 JavaThread *jt = (JavaThread *)this->_thread; 2255 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2256 } 2257 2258 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2259 assert(node != NULL, "should not dequeue NULL node"); 2260 assert(node->_prev == NULL, "node already in list"); 2261 assert(node->_next == NULL, "node already in list"); 2262 // put node at end of queue (circular doubly linked list) 2263 if (_WaitSet == NULL) { 2264 _WaitSet = node; 2265 node->_prev = node; 2266 node->_next = node; 2267 } else { 2268 ObjectWaiter* head = _WaitSet ; 2269 ObjectWaiter* tail = head->_prev; 2270 assert(tail->_next == head, "invariant check"); 2271 tail->_next = node; 2272 head->_prev = node; 2273 node->_next = head; 2274 node->_prev = tail; 2275 } 2276 } 2277 2278 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2279 // dequeue the very first waiter 2280 ObjectWaiter* waiter = _WaitSet; 2281 if (waiter) { 2282 DequeueSpecificWaiter(waiter); 2283 } 2284 return waiter; 2285 } 2286 2287 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2288 assert(node != NULL, "should not dequeue NULL node"); 2289 assert(node->_prev != NULL, "node already removed from list"); 2290 assert(node->_next != NULL, "node already removed from list"); 2291 // when the waiter has woken up because of interrupt, 2292 // timeout or other spurious wake-up, dequeue the 2293 // waiter from waiting list 2294 ObjectWaiter* next = node->_next; 2295 if (next == node) { 2296 assert(node->_prev == node, "invariant check"); 2297 _WaitSet = NULL; 2298 } else { 2299 ObjectWaiter* prev = node->_prev; 2300 assert(prev->_next == node, "invariant check"); 2301 assert(next->_prev == node, "invariant check"); 2302 next->_prev = prev; 2303 prev->_next = next; 2304 if (_WaitSet == node) { 2305 _WaitSet = next; 2306 } 2307 } 2308 node->_next = NULL; 2309 node->_prev = NULL; 2310 } 2311 2312 // ----------------------------------------------------------------------------- 2313 // PerfData support 2314 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2315 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2316 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2317 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2318 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2319 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2320 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2321 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2322 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2323 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2324 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2325 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2326 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2327 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2328 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2329 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2330 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2331 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2332 2333 // One-shot global initialization for the sync subsystem. 2334 // We could also defer initialization and initialize on-demand 2335 // the first time we call inflate(). Initialization would 2336 // be protected - like so many things - by the MonitorCache_lock. 2337 2338 void ObjectMonitor::Initialize () { 2339 static int InitializationCompleted = 0 ; 2340 assert (InitializationCompleted == 0, "invariant") ; 2341 InitializationCompleted = 1 ; 2342 if (UsePerfData) { 2343 EXCEPTION_MARK ; 2344 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2345 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2346 NEWPERFCOUNTER(_sync_Inflations) ; 2347 NEWPERFCOUNTER(_sync_Deflations) ; 2348 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2349 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2350 NEWPERFCOUNTER(_sync_Parks) ; 2351 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2352 NEWPERFCOUNTER(_sync_Notifications) ; 2353 NEWPERFCOUNTER(_sync_SlowEnter) ; 2354 NEWPERFCOUNTER(_sync_SlowExit) ; 2355 NEWPERFCOUNTER(_sync_SlowNotify) ; 2356 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2357 NEWPERFCOUNTER(_sync_FailedSpins) ; 2358 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2359 NEWPERFCOUNTER(_sync_PrivateA) ; 2360 NEWPERFCOUNTER(_sync_PrivateB) ; 2361 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2362 NEWPERFCOUNTER(_sync_MonScavenged) ; 2363 NEWPERFVARIABLE(_sync_MonExtant) ; 2364 #undef NEWPERFCOUNTER 2365 } 2366 } 2367 2368 2369 // Compile-time asserts 2370 // When possible, it's better to catch errors deterministically at 2371 // compile-time than at runtime. The down-side to using compile-time 2372 // asserts is that error message -- often something about negative array 2373 // indices -- is opaque. 2374 2375 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2376 2377 void ObjectMonitor::ctAsserts() { 2378 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2379 } 2380 2381 2382 static char * kvGet (char * kvList, const char * Key) { 2383 if (kvList == NULL) return NULL ; 2384 size_t n = strlen (Key) ; 2385 char * Search ; 2386 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2387 if (strncmp (Search, Key, n) == 0) { 2388 if (Search[n] == '=') return Search + n + 1 ; 2389 if (Search[n] == 0) return (char *) "1" ; 2390 } 2391 } 2392 return NULL ; 2393 } 2394 2395 static int kvGetInt (char * kvList, const char * Key, int Default) { 2396 char * v = kvGet (kvList, Key) ; 2397 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2398 if (Knob_ReportSettings && v != NULL) { 2399 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2400 ::fflush (stdout) ; 2401 } 2402 return rslt ; 2403 } 2404 2405 void ObjectMonitor::DeferredInitialize () { 2406 if (InitDone > 0) return ; 2407 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2408 while (InitDone != 1) ; 2409 return ; 2410 } 2411 2412 // One-shot global initialization ... 2413 // The initialization is idempotent, so we don't need locks. 2414 // In the future consider doing this via os::init_2(). 2415 // SyncKnobs consist of <Key>=<Value> pairs in the style 2416 // of environment variables. Start by converting ':' to NUL. 2417 2418 if (SyncKnobs == NULL) SyncKnobs = "" ; 2419 2420 size_t sz = strlen (SyncKnobs) ; 2421 char * knobs = (char *) malloc (sz + 2) ; 2422 if (knobs == NULL) { 2423 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; 2424 guarantee (0, "invariant") ; 2425 } 2426 strcpy (knobs, SyncKnobs) ; 2427 knobs[sz+1] = 0 ; 2428 for (char * p = knobs ; *p ; p++) { 2429 if (*p == ':') *p = 0 ; 2430 } 2431 2432 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2433 SETKNOB(ReportSettings) ; 2434 SETKNOB(Verbose) ; 2435 SETKNOB(FixedSpin) ; 2436 SETKNOB(SpinLimit) ; 2437 SETKNOB(SpinBase) ; 2438 SETKNOB(SpinBackOff); 2439 SETKNOB(CASPenalty) ; 2440 SETKNOB(OXPenalty) ; 2441 SETKNOB(LogSpins) ; 2442 SETKNOB(SpinSetSucc) ; 2443 SETKNOB(SuccEnabled) ; 2444 SETKNOB(SuccRestrict) ; 2445 SETKNOB(Penalty) ; 2446 SETKNOB(Bonus) ; 2447 SETKNOB(BonusB) ; 2448 SETKNOB(Poverty) ; 2449 SETKNOB(SpinAfterFutile) ; 2450 SETKNOB(UsePause) ; 2451 SETKNOB(SpinEarly) ; 2452 SETKNOB(OState) ; 2453 SETKNOB(MaxSpinners) ; 2454 SETKNOB(PreSpin) ; 2455 SETKNOB(ExitPolicy) ; 2456 SETKNOB(QMode); 2457 SETKNOB(ResetEvent) ; 2458 SETKNOB(MoveNotifyee) ; 2459 SETKNOB(FastHSSEC) ; 2460 #undef SETKNOB 2461 2462 if (os::is_MP()) { 2463 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2464 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2465 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2466 } else { 2467 Knob_SpinLimit = 0 ; 2468 Knob_SpinBase = 0 ; 2469 Knob_PreSpin = 0 ; 2470 Knob_FixedSpin = -1 ; 2471 } 2472 2473 if (Knob_LogSpins == 0) { 2474 ObjectMonitor::_sync_FailedSpins = NULL ; 2475 } 2476 2477 free (knobs) ; 2478 OrderAccess::fence() ; 2479 InitDone = 1 ; 2480 } 2481 2482 #ifndef PRODUCT 2483 void ObjectMonitor::verify() { 2484 } 2485 2486 void ObjectMonitor::print() { 2487 } 2488 #endif