1 /* 2 * Copyright (c) 1998, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "oops/markOop.hpp" 30 #include "oops/oop.inline.hpp" 31 #include "runtime/atomic.hpp" 32 #include "runtime/handles.inline.hpp" 33 #include "runtime/interfaceSupport.hpp" 34 #include "runtime/mutexLocker.hpp" 35 #include "runtime/objectMonitor.hpp" 36 #include "runtime/objectMonitor.inline.hpp" 37 #include "runtime/orderAccess.inline.hpp" 38 #include "runtime/osThread.hpp" 39 #include "runtime/safepointMechanism.inline.hpp" 40 #include "runtime/stubRoutines.hpp" 41 #include "runtime/thread.inline.hpp" 42 #include "services/threadService.hpp" 43 #include "trace/tracing.hpp" 44 #include "trace/traceMacros.hpp" 45 #include "utilities/dtrace.hpp" 46 #include "utilities/macros.hpp" 47 #include "utilities/preserveException.hpp" 48 49 #ifdef DTRACE_ENABLED 50 51 // Only bother with this argument setup if dtrace is available 52 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 53 54 55 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 56 char* bytes = NULL; \ 57 int len = 0; \ 58 jlong jtid = SharedRuntime::get_java_tid(thread); \ 59 Symbol* klassname = ((oop)obj)->klass()->name(); \ 60 if (klassname != NULL) { \ 61 bytes = (char*)klassname->bytes(); \ 62 len = klassname->utf8_length(); \ 63 } 64 65 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 66 { \ 67 if (DTraceMonitorProbes) { \ 68 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 69 HOTSPOT_MONITOR_WAIT(jtid, \ 70 (monitor), bytes, len, (millis)); \ 71 } \ 72 } 73 74 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 75 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 76 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 77 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 78 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 79 80 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 81 { \ 82 if (DTraceMonitorProbes) { \ 83 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 84 HOTSPOT_MONITOR_##probe(jtid, \ 85 (uintptr_t)(monitor), bytes, len); \ 86 } \ 87 } 88 89 #else // ndef DTRACE_ENABLED 90 91 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 92 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 93 94 #endif // ndef DTRACE_ENABLED 95 96 // Tunables ... 97 // The knob* variables are effectively final. Once set they should 98 // never be modified hence. Consider using __read_mostly with GCC. 99 100 int ObjectMonitor::Knob_ExitRelease = 0; 101 int ObjectMonitor::Knob_Verbose = 0; 102 int ObjectMonitor::Knob_VerifyInUse = 0; 103 int ObjectMonitor::Knob_VerifyMatch = 0; 104 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 105 static int Knob_LogSpins = 0; // enable jvmstat tally for spins 106 static int Knob_HandOff = 0; 107 static int Knob_ReportSettings = 0; 108 109 static int Knob_SpinBase = 0; // Floor AKA SpinMin 110 static int Knob_SpinBackOff = 0; // spin-loop backoff 111 static int Knob_CASPenalty = -1; // Penalty for failed CAS 112 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 113 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 114 static int Knob_SpinEarly = 1; 115 static int Knob_SuccEnabled = 1; // futile wake throttling 116 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 117 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 118 static int Knob_Bonus = 100; // spin success bonus 119 static int Knob_BonusB = 100; // spin success bonus 120 static int Knob_Penalty = 200; // spin failure penalty 121 static int Knob_Poverty = 1000; 122 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 123 static int Knob_FixedSpin = 0; 124 static int Knob_OState = 3; // Spinner checks thread state of _owner 125 static int Knob_UsePause = 1; 126 static int Knob_ExitPolicy = 0; 127 static int Knob_PreSpin = 10; // 20-100 likely better 128 static int Knob_ResetEvent = 0; 129 static int BackOffMask = 0; 130 131 static int Knob_FastHSSEC = 0; 132 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 133 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 134 static volatile int InitDone = 0; 135 136 // ----------------------------------------------------------------------------- 137 // Theory of operations -- Monitors lists, thread residency, etc: 138 // 139 // * A thread acquires ownership of a monitor by successfully 140 // CAS()ing the _owner field from null to non-null. 141 // 142 // * Invariant: A thread appears on at most one monitor list -- 143 // cxq, EntryList or WaitSet -- at any one time. 144 // 145 // * Contending threads "push" themselves onto the cxq with CAS 146 // and then spin/park. 147 // 148 // * After a contending thread eventually acquires the lock it must 149 // dequeue itself from either the EntryList or the cxq. 150 // 151 // * The exiting thread identifies and unparks an "heir presumptive" 152 // tentative successor thread on the EntryList. Critically, the 153 // exiting thread doesn't unlink the successor thread from the EntryList. 154 // After having been unparked, the wakee will recontend for ownership of 155 // the monitor. The successor (wakee) will either acquire the lock or 156 // re-park itself. 157 // 158 // Succession is provided for by a policy of competitive handoff. 159 // The exiting thread does _not_ grant or pass ownership to the 160 // successor thread. (This is also referred to as "handoff" succession"). 161 // Instead the exiting thread releases ownership and possibly wakes 162 // a successor, so the successor can (re)compete for ownership of the lock. 163 // If the EntryList is empty but the cxq is populated the exiting 164 // thread will drain the cxq into the EntryList. It does so by 165 // by detaching the cxq (installing null with CAS) and folding 166 // the threads from the cxq into the EntryList. The EntryList is 167 // doubly linked, while the cxq is singly linked because of the 168 // CAS-based "push" used to enqueue recently arrived threads (RATs). 169 // 170 // * Concurrency invariants: 171 // 172 // -- only the monitor owner may access or mutate the EntryList. 173 // The mutex property of the monitor itself protects the EntryList 174 // from concurrent interference. 175 // -- Only the monitor owner may detach the cxq. 176 // 177 // * The monitor entry list operations avoid locks, but strictly speaking 178 // they're not lock-free. Enter is lock-free, exit is not. 179 // For a description of 'Methods and apparatus providing non-blocking access 180 // to a resource,' see U.S. Pat. No. 7844973. 181 // 182 // * The cxq can have multiple concurrent "pushers" but only one concurrent 183 // detaching thread. This mechanism is immune from the ABA corruption. 184 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 185 // 186 // * Taken together, the cxq and the EntryList constitute or form a 187 // single logical queue of threads stalled trying to acquire the lock. 188 // We use two distinct lists to improve the odds of a constant-time 189 // dequeue operation after acquisition (in the ::enter() epilogue) and 190 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 191 // A key desideratum is to minimize queue & monitor metadata manipulation 192 // that occurs while holding the monitor lock -- that is, we want to 193 // minimize monitor lock holds times. Note that even a small amount of 194 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 195 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 196 // locks and monitor metadata. 197 // 198 // Cxq points to the set of Recently Arrived Threads attempting entry. 199 // Because we push threads onto _cxq with CAS, the RATs must take the form of 200 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 201 // the unlocking thread notices that EntryList is null but _cxq is != null. 202 // 203 // The EntryList is ordered by the prevailing queue discipline and 204 // can be organized in any convenient fashion, such as a doubly-linked list or 205 // a circular doubly-linked list. Critically, we want insert and delete operations 206 // to operate in constant-time. If we need a priority queue then something akin 207 // to Solaris' sleepq would work nicely. Viz., 208 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 209 // Queue discipline is enforced at ::exit() time, when the unlocking thread 210 // drains the cxq into the EntryList, and orders or reorders the threads on the 211 // EntryList accordingly. 212 // 213 // Barring "lock barging", this mechanism provides fair cyclic ordering, 214 // somewhat similar to an elevator-scan. 215 // 216 // * The monitor synchronization subsystem avoids the use of native 217 // synchronization primitives except for the narrow platform-specific 218 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 219 // the semantics of park-unpark. Put another way, this monitor implementation 220 // depends only on atomic operations and park-unpark. The monitor subsystem 221 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 222 // underlying OS manages the READY<->RUN transitions. 223 // 224 // * Waiting threads reside on the WaitSet list -- wait() puts 225 // the caller onto the WaitSet. 226 // 227 // * notify() or notifyAll() simply transfers threads from the WaitSet to 228 // either the EntryList or cxq. Subsequent exit() operations will 229 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 230 // it's likely the notifyee would simply impale itself on the lock held 231 // by the notifier. 232 // 233 // * An interesting alternative is to encode cxq as (List,LockByte) where 234 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 235 // variable, like _recursions, in the scheme. The threads or Events that form 236 // the list would have to be aligned in 256-byte addresses. A thread would 237 // try to acquire the lock or enqueue itself with CAS, but exiting threads 238 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 239 // Note that is is *not* word-tearing, but it does presume that full-word 240 // CAS operations are coherent with intermix with STB operations. That's true 241 // on most common processors. 242 // 243 // * See also http://blogs.sun.com/dave 244 245 246 void* ObjectMonitor::operator new (size_t size) throw() { 247 return AllocateHeap(size, mtInternal); 248 } 249 void* ObjectMonitor::operator new[] (size_t size) throw() { 250 return operator new (size); 251 } 252 void ObjectMonitor::operator delete(void* p) { 253 FreeHeap(p); 254 } 255 void ObjectMonitor::operator delete[] (void *p) { 256 operator delete(p); 257 } 258 259 // ----------------------------------------------------------------------------- 260 // Enter support 261 262 void ObjectMonitor::enter(TRAPS) { 263 // The following code is ordered to check the most common cases first 264 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 265 Thread * const Self = THREAD; 266 267 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 268 if (cur == NULL) { 269 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 270 assert(_recursions == 0, "invariant"); 271 assert(_owner == Self, "invariant"); 272 return; 273 } 274 275 if (cur == Self) { 276 // TODO-FIXME: check for integer overflow! BUGID 6557169. 277 _recursions++; 278 return; 279 } 280 281 if (Self->is_lock_owned ((address)cur)) { 282 assert(_recursions == 0, "internal state error"); 283 _recursions = 1; 284 // Commute owner from a thread-specific on-stack BasicLockObject address to 285 // a full-fledged "Thread *". 286 _owner = Self; 287 return; 288 } 289 290 // We've encountered genuine contention. 291 assert(Self->_Stalled == 0, "invariant"); 292 Self->_Stalled = intptr_t(this); 293 294 // Try one round of spinning *before* enqueueing Self 295 // and before going through the awkward and expensive state 296 // transitions. The following spin is strictly optional ... 297 // Note that if we acquire the monitor from an initial spin 298 // we forgo posting JVMTI events and firing DTRACE probes. 299 if (Knob_SpinEarly && TrySpin (Self) > 0) { 300 assert(_owner == Self, "invariant"); 301 assert(_recursions == 0, "invariant"); 302 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 303 Self->_Stalled = 0; 304 return; 305 } 306 307 assert(_owner != Self, "invariant"); 308 assert(_succ != Self, "invariant"); 309 assert(Self->is_Java_thread(), "invariant"); 310 JavaThread * jt = (JavaThread *) Self; 311 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 312 assert(jt->thread_state() != _thread_blocked, "invariant"); 313 assert(this->object() != NULL, "invariant"); 314 assert(_count >= 0, "invariant"); 315 316 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 317 // Ensure the object-monitor relationship remains stable while there's contention. 318 Atomic::inc(&_count); 319 320 EventJavaMonitorEnter event; 321 322 { // Change java thread status to indicate blocked on monitor enter. 323 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 324 325 Self->set_current_pending_monitor(this); 326 327 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 328 if (JvmtiExport::should_post_monitor_contended_enter()) { 329 JvmtiExport::post_monitor_contended_enter(jt, this); 330 331 // The current thread does not yet own the monitor and does not 332 // yet appear on any queues that would get it made the successor. 333 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 334 // handler cannot accidentally consume an unpark() meant for the 335 // ParkEvent associated with this ObjectMonitor. 336 } 337 338 OSThreadContendState osts(Self->osthread()); 339 ThreadBlockInVM tbivm(jt); 340 341 // TODO-FIXME: change the following for(;;) loop to straight-line code. 342 for (;;) { 343 jt->set_suspend_equivalent(); 344 // cleared by handle_special_suspend_equivalent_condition() 345 // or java_suspend_self() 346 347 EnterI(THREAD); 348 349 if (!ExitSuspendEquivalent(jt)) break; 350 351 // We have acquired the contended monitor, but while we were 352 // waiting another thread suspended us. We don't want to enter 353 // the monitor while suspended because that would surprise the 354 // thread that suspended us. 355 // 356 _recursions = 0; 357 _succ = NULL; 358 exit(false, Self); 359 360 jt->java_suspend_self(); 361 } 362 Self->set_current_pending_monitor(NULL); 363 364 // We cleared the pending monitor info since we've just gotten past 365 // the enter-check-for-suspend dance and we now own the monitor free 366 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 367 // destructor can go to a safepoint at the end of this block. If we 368 // do a thread dump during that safepoint, then this thread will show 369 // as having "-locked" the monitor, but the OS and java.lang.Thread 370 // states will still report that the thread is blocked trying to 371 // acquire it. 372 } 373 374 Atomic::dec(&_count); 375 assert(_count >= 0, "invariant"); 376 Self->_Stalled = 0; 377 378 // Must either set _recursions = 0 or ASSERT _recursions == 0. 379 assert(_recursions == 0, "invariant"); 380 assert(_owner == Self, "invariant"); 381 assert(_succ != Self, "invariant"); 382 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 383 384 // The thread -- now the owner -- is back in vm mode. 385 // Report the glorious news via TI,DTrace and jvmstat. 386 // The probe effect is non-trivial. All the reportage occurs 387 // while we hold the monitor, increasing the length of the critical 388 // section. Amdahl's parallel speedup law comes vividly into play. 389 // 390 // Another option might be to aggregate the events (thread local or 391 // per-monitor aggregation) and defer reporting until a more opportune 392 // time -- such as next time some thread encounters contention but has 393 // yet to acquire the lock. While spinning that thread could 394 // spinning we could increment JVMStat counters, etc. 395 396 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 397 if (JvmtiExport::should_post_monitor_contended_entered()) { 398 JvmtiExport::post_monitor_contended_entered(jt, this); 399 400 // The current thread already owns the monitor and is not going to 401 // call park() for the remainder of the monitor enter protocol. So 402 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 403 // event handler consumed an unpark() issued by the thread that 404 // just exited the monitor. 405 } 406 407 if (event.should_commit()) { 408 event.set_monitorClass(((oop)this->object())->klass()); 409 event.set_previousOwner((TYPE_THREAD)_previous_owner_tid); 410 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 411 event.commit(); 412 } 413 414 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 415 } 416 417 418 // Caveat: TryLock() is not necessarily serializing if it returns failure. 419 // Callers must compensate as needed. 420 421 int ObjectMonitor::TryLock(Thread * Self) { 422 void * own = _owner; 423 if (own != NULL) return 0; 424 if (Atomic::cmpxchg(Self, &_owner, (void*)NULL) == NULL) { 425 // Either guarantee _recursions == 0 or set _recursions = 0. 426 assert(_recursions == 0, "invariant"); 427 assert(_owner == Self, "invariant"); 428 return 1; 429 } 430 // The lock had been free momentarily, but we lost the race to the lock. 431 // Interference -- the CAS failed. 432 // We can either return -1 or retry. 433 // Retry doesn't make as much sense because the lock was just acquired. 434 return -1; 435 } 436 437 #define MAX_RECHECK_INTERVAL 1000 438 439 void ObjectMonitor::EnterI(TRAPS) { 440 Thread * const Self = THREAD; 441 assert(Self->is_Java_thread(), "invariant"); 442 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 443 444 // Try the lock - TATAS 445 if (TryLock (Self) > 0) { 446 assert(_succ != Self, "invariant"); 447 assert(_owner == Self, "invariant"); 448 assert(_Responsible != Self, "invariant"); 449 return; 450 } 451 452 DeferredInitialize(); 453 454 // We try one round of spinning *before* enqueueing Self. 455 // 456 // If the _owner is ready but OFFPROC we could use a YieldTo() 457 // operation to donate the remainder of this thread's quantum 458 // to the owner. This has subtle but beneficial affinity 459 // effects. 460 461 if (TrySpin (Self) > 0) { 462 assert(_owner == Self, "invariant"); 463 assert(_succ != Self, "invariant"); 464 assert(_Responsible != Self, "invariant"); 465 return; 466 } 467 468 // The Spin failed -- Enqueue and park the thread ... 469 assert(_succ != Self, "invariant"); 470 assert(_owner != Self, "invariant"); 471 assert(_Responsible != Self, "invariant"); 472 473 // Enqueue "Self" on ObjectMonitor's _cxq. 474 // 475 // Node acts as a proxy for Self. 476 // As an aside, if were to ever rewrite the synchronization code mostly 477 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 478 // Java objects. This would avoid awkward lifecycle and liveness issues, 479 // as well as eliminate a subset of ABA issues. 480 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 481 482 ObjectWaiter node(Self); 483 Self->_ParkEvent->reset(); 484 node._prev = (ObjectWaiter *) 0xBAD; 485 node.TState = ObjectWaiter::TS_CXQ; 486 487 // Push "Self" onto the front of the _cxq. 488 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 489 // Note that spinning tends to reduce the rate at which threads 490 // enqueue and dequeue on EntryList|cxq. 491 ObjectWaiter * nxt; 492 for (;;) { 493 node._next = nxt = _cxq; 494 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 495 496 // Interference - the CAS failed because _cxq changed. Just retry. 497 // As an optional optimization we retry the lock. 498 if (TryLock (Self) > 0) { 499 assert(_succ != Self, "invariant"); 500 assert(_owner == Self, "invariant"); 501 assert(_Responsible != Self, "invariant"); 502 return; 503 } 504 } 505 506 // Check for cxq|EntryList edge transition to non-null. This indicates 507 // the onset of contention. While contention persists exiting threads 508 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 509 // operations revert to the faster 1-0 mode. This enter operation may interleave 510 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 511 // arrange for one of the contending thread to use a timed park() operations 512 // to detect and recover from the race. (Stranding is form of progress failure 513 // where the monitor is unlocked but all the contending threads remain parked). 514 // That is, at least one of the contended threads will periodically poll _owner. 515 // One of the contending threads will become the designated "Responsible" thread. 516 // The Responsible thread uses a timed park instead of a normal indefinite park 517 // operation -- it periodically wakes and checks for and recovers from potential 518 // strandings admitted by 1-0 exit operations. We need at most one Responsible 519 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 520 // be responsible for a monitor. 521 // 522 // Currently, one of the contended threads takes on the added role of "Responsible". 523 // A viable alternative would be to use a dedicated "stranding checker" thread 524 // that periodically iterated over all the threads (or active monitors) and unparked 525 // successors where there was risk of stranding. This would help eliminate the 526 // timer scalability issues we see on some platforms as we'd only have one thread 527 // -- the checker -- parked on a timer. 528 529 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 530 // Try to assume the role of responsible thread for the monitor. 531 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 532 Atomic::cmpxchg(Self, &_Responsible, (Thread*)NULL); 533 } 534 535 // The lock might have been released while this thread was occupied queueing 536 // itself onto _cxq. To close the race and avoid "stranding" and 537 // progress-liveness failure we must resample-retry _owner before parking. 538 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 539 // In this case the ST-MEMBAR is accomplished with CAS(). 540 // 541 // TODO: Defer all thread state transitions until park-time. 542 // Since state transitions are heavy and inefficient we'd like 543 // to defer the state transitions until absolutely necessary, 544 // and in doing so avoid some transitions ... 545 546 TEVENT(Inflated enter - Contention); 547 int nWakeups = 0; 548 int recheckInterval = 1; 549 550 for (;;) { 551 552 if (TryLock(Self) > 0) break; 553 assert(_owner != Self, "invariant"); 554 555 if ((SyncFlags & 2) && _Responsible == NULL) { 556 Atomic::cmpxchg(Self, &_Responsible, (Thread*)NULL); 557 } 558 559 // park self 560 if (_Responsible == Self || (SyncFlags & 1)) { 561 TEVENT(Inflated enter - park TIMED); 562 Self->_ParkEvent->park((jlong) recheckInterval); 563 // Increase the recheckInterval, but clamp the value. 564 recheckInterval *= 8; 565 if (recheckInterval > MAX_RECHECK_INTERVAL) { 566 recheckInterval = MAX_RECHECK_INTERVAL; 567 } 568 } else { 569 TEVENT(Inflated enter - park UNTIMED); 570 Self->_ParkEvent->park(); 571 } 572 573 if (TryLock(Self) > 0) break; 574 575 // The lock is still contested. 576 // Keep a tally of the # of futile wakeups. 577 // Note that the counter is not protected by a lock or updated by atomics. 578 // That is by design - we trade "lossy" counters which are exposed to 579 // races during updates for a lower probe effect. 580 TEVENT(Inflated enter - Futile wakeup); 581 // This PerfData object can be used in parallel with a safepoint. 582 // See the work around in PerfDataManager::destroy(). 583 OM_PERFDATA_OP(FutileWakeups, inc()); 584 ++nWakeups; 585 586 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 587 // We can defer clearing _succ until after the spin completes 588 // TrySpin() must tolerate being called with _succ == Self. 589 // Try yet another round of adaptive spinning. 590 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 591 592 // We can find that we were unpark()ed and redesignated _succ while 593 // we were spinning. That's harmless. If we iterate and call park(), 594 // park() will consume the event and return immediately and we'll 595 // just spin again. This pattern can repeat, leaving _succ to simply 596 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 597 // Alternately, we can sample fired() here, and if set, forgo spinning 598 // in the next iteration. 599 600 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 601 Self->_ParkEvent->reset(); 602 OrderAccess::fence(); 603 } 604 if (_succ == Self) _succ = NULL; 605 606 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 607 OrderAccess::fence(); 608 } 609 610 // Egress : 611 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 612 // Normally we'll find Self on the EntryList . 613 // From the perspective of the lock owner (this thread), the 614 // EntryList is stable and cxq is prepend-only. 615 // The head of cxq is volatile but the interior is stable. 616 // In addition, Self.TState is stable. 617 618 assert(_owner == Self, "invariant"); 619 assert(object() != NULL, "invariant"); 620 // I'd like to write: 621 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 622 // but as we're at a safepoint that's not safe. 623 624 UnlinkAfterAcquire(Self, &node); 625 if (_succ == Self) _succ = NULL; 626 627 assert(_succ != Self, "invariant"); 628 if (_Responsible == Self) { 629 _Responsible = NULL; 630 OrderAccess::fence(); // Dekker pivot-point 631 632 // We may leave threads on cxq|EntryList without a designated 633 // "Responsible" thread. This is benign. When this thread subsequently 634 // exits the monitor it can "see" such preexisting "old" threads -- 635 // threads that arrived on the cxq|EntryList before the fence, above -- 636 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 637 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 638 // non-null and elect a new "Responsible" timer thread. 639 // 640 // This thread executes: 641 // ST Responsible=null; MEMBAR (in enter epilogue - here) 642 // LD cxq|EntryList (in subsequent exit) 643 // 644 // Entering threads in the slow/contended path execute: 645 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 646 // The (ST cxq; MEMBAR) is accomplished with CAS(). 647 // 648 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 649 // exit operation from floating above the ST Responsible=null. 650 } 651 652 // We've acquired ownership with CAS(). 653 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 654 // But since the CAS() this thread may have also stored into _succ, 655 // EntryList, cxq or Responsible. These meta-data updates must be 656 // visible __before this thread subsequently drops the lock. 657 // Consider what could occur if we didn't enforce this constraint -- 658 // STs to monitor meta-data and user-data could reorder with (become 659 // visible after) the ST in exit that drops ownership of the lock. 660 // Some other thread could then acquire the lock, but observe inconsistent 661 // or old monitor meta-data and heap data. That violates the JMM. 662 // To that end, the 1-0 exit() operation must have at least STST|LDST 663 // "release" barrier semantics. Specifically, there must be at least a 664 // STST|LDST barrier in exit() before the ST of null into _owner that drops 665 // the lock. The barrier ensures that changes to monitor meta-data and data 666 // protected by the lock will be visible before we release the lock, and 667 // therefore before some other thread (CPU) has a chance to acquire the lock. 668 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 669 // 670 // Critically, any prior STs to _succ or EntryList must be visible before 671 // the ST of null into _owner in the *subsequent* (following) corresponding 672 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 673 // execute a serializing instruction. 674 675 if (SyncFlags & 8) { 676 OrderAccess::fence(); 677 } 678 return; 679 } 680 681 // ReenterI() is a specialized inline form of the latter half of the 682 // contended slow-path from EnterI(). We use ReenterI() only for 683 // monitor reentry in wait(). 684 // 685 // In the future we should reconcile EnterI() and ReenterI(), adding 686 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 687 // loop accordingly. 688 689 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 690 assert(Self != NULL, "invariant"); 691 assert(SelfNode != NULL, "invariant"); 692 assert(SelfNode->_thread == Self, "invariant"); 693 assert(_waiters > 0, "invariant"); 694 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 695 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 696 JavaThread * jt = (JavaThread *) Self; 697 698 int nWakeups = 0; 699 for (;;) { 700 ObjectWaiter::TStates v = SelfNode->TState; 701 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 702 assert(_owner != Self, "invariant"); 703 704 if (TryLock(Self) > 0) break; 705 if (TrySpin(Self) > 0) break; 706 707 TEVENT(Wait Reentry - parking); 708 709 // State transition wrappers around park() ... 710 // ReenterI() wisely defers state transitions until 711 // it's clear we must park the thread. 712 { 713 OSThreadContendState osts(Self->osthread()); 714 ThreadBlockInVM tbivm(jt); 715 716 // cleared by handle_special_suspend_equivalent_condition() 717 // or java_suspend_self() 718 jt->set_suspend_equivalent(); 719 if (SyncFlags & 1) { 720 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL); 721 } else { 722 Self->_ParkEvent->park(); 723 } 724 725 // were we externally suspended while we were waiting? 726 for (;;) { 727 if (!ExitSuspendEquivalent(jt)) break; 728 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 729 jt->java_suspend_self(); 730 jt->set_suspend_equivalent(); 731 } 732 } 733 734 // Try again, but just so we distinguish between futile wakeups and 735 // successful wakeups. The following test isn't algorithmically 736 // necessary, but it helps us maintain sensible statistics. 737 if (TryLock(Self) > 0) break; 738 739 // The lock is still contested. 740 // Keep a tally of the # of futile wakeups. 741 // Note that the counter is not protected by a lock or updated by atomics. 742 // That is by design - we trade "lossy" counters which are exposed to 743 // races during updates for a lower probe effect. 744 TEVENT(Wait Reentry - futile wakeup); 745 ++nWakeups; 746 747 // Assuming this is not a spurious wakeup we'll normally 748 // find that _succ == Self. 749 if (_succ == Self) _succ = NULL; 750 751 // Invariant: after clearing _succ a contending thread 752 // *must* retry _owner before parking. 753 OrderAccess::fence(); 754 755 // This PerfData object can be used in parallel with a safepoint. 756 // See the work around in PerfDataManager::destroy(). 757 OM_PERFDATA_OP(FutileWakeups, inc()); 758 } 759 760 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 761 // Normally we'll find Self on the EntryList. 762 // Unlinking from the EntryList is constant-time and atomic-free. 763 // From the perspective of the lock owner (this thread), the 764 // EntryList is stable and cxq is prepend-only. 765 // The head of cxq is volatile but the interior is stable. 766 // In addition, Self.TState is stable. 767 768 assert(_owner == Self, "invariant"); 769 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 770 UnlinkAfterAcquire(Self, SelfNode); 771 if (_succ == Self) _succ = NULL; 772 assert(_succ != Self, "invariant"); 773 SelfNode->TState = ObjectWaiter::TS_RUN; 774 OrderAccess::fence(); // see comments at the end of EnterI() 775 } 776 777 // By convention we unlink a contending thread from EntryList|cxq immediately 778 // after the thread acquires the lock in ::enter(). Equally, we could defer 779 // unlinking the thread until ::exit()-time. 780 781 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 782 assert(_owner == Self, "invariant"); 783 assert(SelfNode->_thread == Self, "invariant"); 784 785 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 786 // Normal case: remove Self from the DLL EntryList . 787 // This is a constant-time operation. 788 ObjectWaiter * nxt = SelfNode->_next; 789 ObjectWaiter * prv = SelfNode->_prev; 790 if (nxt != NULL) nxt->_prev = prv; 791 if (prv != NULL) prv->_next = nxt; 792 if (SelfNode == _EntryList) _EntryList = nxt; 793 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 794 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 795 TEVENT(Unlink from EntryList); 796 } else { 797 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 798 // Inopportune interleaving -- Self is still on the cxq. 799 // This usually means the enqueue of self raced an exiting thread. 800 // Normally we'll find Self near the front of the cxq, so 801 // dequeueing is typically fast. If needbe we can accelerate 802 // this with some MCS/CHL-like bidirectional list hints and advisory 803 // back-links so dequeueing from the interior will normally operate 804 // in constant-time. 805 // Dequeue Self from either the head (with CAS) or from the interior 806 // with a linear-time scan and normal non-atomic memory operations. 807 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 808 // and then unlink Self from EntryList. We have to drain eventually, 809 // so it might as well be now. 810 811 ObjectWaiter * v = _cxq; 812 assert(v != NULL, "invariant"); 813 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 814 // The CAS above can fail from interference IFF a "RAT" arrived. 815 // In that case Self must be in the interior and can no longer be 816 // at the head of cxq. 817 if (v == SelfNode) { 818 assert(_cxq != v, "invariant"); 819 v = _cxq; // CAS above failed - start scan at head of list 820 } 821 ObjectWaiter * p; 822 ObjectWaiter * q = NULL; 823 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 824 q = p; 825 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 826 } 827 assert(v != SelfNode, "invariant"); 828 assert(p == SelfNode, "Node not found on cxq"); 829 assert(p != _cxq, "invariant"); 830 assert(q != NULL, "invariant"); 831 assert(q->_next == p, "invariant"); 832 q->_next = p->_next; 833 } 834 TEVENT(Unlink from cxq); 835 } 836 837 #ifdef ASSERT 838 // Diagnostic hygiene ... 839 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 840 SelfNode->_next = (ObjectWaiter *) 0xBAD; 841 SelfNode->TState = ObjectWaiter::TS_RUN; 842 #endif 843 } 844 845 // ----------------------------------------------------------------------------- 846 // Exit support 847 // 848 // exit() 849 // ~~~~~~ 850 // Note that the collector can't reclaim the objectMonitor or deflate 851 // the object out from underneath the thread calling ::exit() as the 852 // thread calling ::exit() never transitions to a stable state. 853 // This inhibits GC, which in turn inhibits asynchronous (and 854 // inopportune) reclamation of "this". 855 // 856 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 857 // There's one exception to the claim above, however. EnterI() can call 858 // exit() to drop a lock if the acquirer has been externally suspended. 859 // In that case exit() is called with _thread_state as _thread_blocked, 860 // but the monitor's _count field is > 0, which inhibits reclamation. 861 // 862 // 1-0 exit 863 // ~~~~~~~~ 864 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 865 // the fast-path operators have been optimized so the common ::exit() 866 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 867 // The code emitted by fast_unlock() elides the usual MEMBAR. This 868 // greatly improves latency -- MEMBAR and CAS having considerable local 869 // latency on modern processors -- but at the cost of "stranding". Absent the 870 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 871 // ::enter() path, resulting in the entering thread being stranding 872 // and a progress-liveness failure. Stranding is extremely rare. 873 // We use timers (timed park operations) & periodic polling to detect 874 // and recover from stranding. Potentially stranded threads periodically 875 // wake up and poll the lock. See the usage of the _Responsible variable. 876 // 877 // The CAS() in enter provides for safety and exclusion, while the CAS or 878 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 879 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 880 // We detect and recover from stranding with timers. 881 // 882 // If a thread transiently strands it'll park until (a) another 883 // thread acquires the lock and then drops the lock, at which time the 884 // exiting thread will notice and unpark the stranded thread, or, (b) 885 // the timer expires. If the lock is high traffic then the stranding latency 886 // will be low due to (a). If the lock is low traffic then the odds of 887 // stranding are lower, although the worst-case stranding latency 888 // is longer. Critically, we don't want to put excessive load in the 889 // platform's timer subsystem. We want to minimize both the timer injection 890 // rate (timers created/sec) as well as the number of timers active at 891 // any one time. (more precisely, we want to minimize timer-seconds, which is 892 // the integral of the # of active timers at any instant over time). 893 // Both impinge on OS scalability. Given that, at most one thread parked on 894 // a monitor will use a timer. 895 // 896 // There is also the risk of a futile wake-up. If we drop the lock 897 // another thread can reacquire the lock immediately, and we can 898 // then wake a thread unnecessarily. This is benign, and we've 899 // structured the code so the windows are short and the frequency 900 // of such futile wakups is low. 901 902 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 903 Thread * const Self = THREAD; 904 if (THREAD != _owner) { 905 if (THREAD->is_lock_owned((address) _owner)) { 906 // Transmute _owner from a BasicLock pointer to a Thread address. 907 // We don't need to hold _mutex for this transition. 908 // Non-null to Non-null is safe as long as all readers can 909 // tolerate either flavor. 910 assert(_recursions == 0, "invariant"); 911 _owner = THREAD; 912 _recursions = 0; 913 } else { 914 // Apparent unbalanced locking ... 915 // Naively we'd like to throw IllegalMonitorStateException. 916 // As a practical matter we can neither allocate nor throw an 917 // exception as ::exit() can be called from leaf routines. 918 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 919 // Upon deeper reflection, however, in a properly run JVM the only 920 // way we should encounter this situation is in the presence of 921 // unbalanced JNI locking. TODO: CheckJNICalls. 922 // See also: CR4414101 923 TEVENT(Exit - Throw IMSX); 924 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 925 return; 926 } 927 } 928 929 if (_recursions != 0) { 930 _recursions--; // this is simple recursive enter 931 TEVENT(Inflated exit - recursive); 932 return; 933 } 934 935 // Invariant: after setting Responsible=null an thread must execute 936 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 937 if ((SyncFlags & 4) == 0) { 938 _Responsible = NULL; 939 } 940 941 #if INCLUDE_TRACE 942 // get the owner's thread id for the MonitorEnter event 943 // if it is enabled and the thread isn't suspended 944 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 945 _previous_owner_tid = THREAD_TRACE_ID(Self); 946 } 947 #endif 948 949 for (;;) { 950 assert(THREAD == _owner, "invariant"); 951 952 if (Knob_ExitPolicy == 0) { 953 // release semantics: prior loads and stores from within the critical section 954 // must not float (reorder) past the following store that drops the lock. 955 // On SPARC that requires MEMBAR #loadstore|#storestore. 956 // But of course in TSO #loadstore|#storestore is not required. 957 // I'd like to write one of the following: 958 // A. OrderAccess::release() ; _owner = NULL 959 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 960 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 961 // store into a _dummy variable. That store is not needed, but can result 962 // in massive wasteful coherency traffic on classic SMP systems. 963 // Instead, I use release_store(), which is implemented as just a simple 964 // ST on x64, x86 and SPARC. 965 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 966 OrderAccess::storeload(); // See if we need to wake a successor 967 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 968 TEVENT(Inflated exit - simple egress); 969 return; 970 } 971 TEVENT(Inflated exit - complex egress); 972 // Other threads are blocked trying to acquire the lock. 973 974 // Normally the exiting thread is responsible for ensuring succession, 975 // but if other successors are ready or other entering threads are spinning 976 // then this thread can simply store NULL into _owner and exit without 977 // waking a successor. The existence of spinners or ready successors 978 // guarantees proper succession (liveness). Responsibility passes to the 979 // ready or running successors. The exiting thread delegates the duty. 980 // More precisely, if a successor already exists this thread is absolved 981 // of the responsibility of waking (unparking) one. 982 // 983 // The _succ variable is critical to reducing futile wakeup frequency. 984 // _succ identifies the "heir presumptive" thread that has been made 985 // ready (unparked) but that has not yet run. We need only one such 986 // successor thread to guarantee progress. 987 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 988 // section 3.3 "Futile Wakeup Throttling" for details. 989 // 990 // Note that spinners in Enter() also set _succ non-null. 991 // In the current implementation spinners opportunistically set 992 // _succ so that exiting threads might avoid waking a successor. 993 // Another less appealing alternative would be for the exiting thread 994 // to drop the lock and then spin briefly to see if a spinner managed 995 // to acquire the lock. If so, the exiting thread could exit 996 // immediately without waking a successor, otherwise the exiting 997 // thread would need to dequeue and wake a successor. 998 // (Note that we'd need to make the post-drop spin short, but no 999 // shorter than the worst-case round-trip cache-line migration time. 1000 // The dropped lock needs to become visible to the spinner, and then 1001 // the acquisition of the lock by the spinner must become visible to 1002 // the exiting thread). 1003 1004 // It appears that an heir-presumptive (successor) must be made ready. 1005 // Only the current lock owner can manipulate the EntryList or 1006 // drain _cxq, so we need to reacquire the lock. If we fail 1007 // to reacquire the lock the responsibility for ensuring succession 1008 // falls to the new owner. 1009 // 1010 if (Atomic::cmpxchg(THREAD, &_owner, (void*)NULL) != NULL) { 1011 return; 1012 } 1013 TEVENT(Exit - Reacquired); 1014 } else { 1015 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1016 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 1017 OrderAccess::storeload(); 1018 // Ratify the previously observed values. 1019 if (_cxq == NULL || _succ != NULL) { 1020 TEVENT(Inflated exit - simple egress); 1021 return; 1022 } 1023 1024 // inopportune interleaving -- the exiting thread (this thread) 1025 // in the fast-exit path raced an entering thread in the slow-enter 1026 // path. 1027 // We have two choices: 1028 // A. Try to reacquire the lock. 1029 // If the CAS() fails return immediately, otherwise 1030 // we either restart/rerun the exit operation, or simply 1031 // fall-through into the code below which wakes a successor. 1032 // B. If the elements forming the EntryList|cxq are TSM 1033 // we could simply unpark() the lead thread and return 1034 // without having set _succ. 1035 if (Atomic::cmpxchg(THREAD, &_owner, (void*)NULL) != NULL) { 1036 TEVENT(Inflated exit - reacquired succeeded); 1037 return; 1038 } 1039 TEVENT(Inflated exit - reacquired failed); 1040 } else { 1041 TEVENT(Inflated exit - complex egress); 1042 } 1043 } 1044 1045 guarantee(_owner == THREAD, "invariant"); 1046 1047 ObjectWaiter * w = NULL; 1048 int QMode = Knob_QMode; 1049 1050 if (QMode == 2 && _cxq != NULL) { 1051 // QMode == 2 : cxq has precedence over EntryList. 1052 // Try to directly wake a successor from the cxq. 1053 // If successful, the successor will need to unlink itself from cxq. 1054 w = _cxq; 1055 assert(w != NULL, "invariant"); 1056 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1057 ExitEpilog(Self, w); 1058 return; 1059 } 1060 1061 if (QMode == 3 && _cxq != NULL) { 1062 // Aggressively drain cxq into EntryList at the first opportunity. 1063 // This policy ensure that recently-run threads live at the head of EntryList. 1064 // Drain _cxq into EntryList - bulk transfer. 1065 // First, detach _cxq. 1066 // The following loop is tantamount to: w = swap(&cxq, NULL) 1067 w = _cxq; 1068 for (;;) { 1069 assert(w != NULL, "Invariant"); 1070 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1071 if (u == w) break; 1072 w = u; 1073 } 1074 assert(w != NULL, "invariant"); 1075 1076 ObjectWaiter * q = NULL; 1077 ObjectWaiter * p; 1078 for (p = w; p != NULL; p = p->_next) { 1079 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1080 p->TState = ObjectWaiter::TS_ENTER; 1081 p->_prev = q; 1082 q = p; 1083 } 1084 1085 // Append the RATs to the EntryList 1086 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1087 ObjectWaiter * Tail; 1088 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1089 Tail = Tail->_next) 1090 /* empty */; 1091 if (Tail == NULL) { 1092 _EntryList = w; 1093 } else { 1094 Tail->_next = w; 1095 w->_prev = Tail; 1096 } 1097 1098 // Fall thru into code that tries to wake a successor from EntryList 1099 } 1100 1101 if (QMode == 4 && _cxq != NULL) { 1102 // Aggressively drain cxq into EntryList at the first opportunity. 1103 // This policy ensure that recently-run threads live at the head of EntryList. 1104 1105 // Drain _cxq into EntryList - bulk transfer. 1106 // First, detach _cxq. 1107 // The following loop is tantamount to: w = swap(&cxq, NULL) 1108 w = _cxq; 1109 for (;;) { 1110 assert(w != NULL, "Invariant"); 1111 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1112 if (u == w) break; 1113 w = u; 1114 } 1115 assert(w != NULL, "invariant"); 1116 1117 ObjectWaiter * q = NULL; 1118 ObjectWaiter * p; 1119 for (p = w; p != NULL; p = p->_next) { 1120 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1121 p->TState = ObjectWaiter::TS_ENTER; 1122 p->_prev = q; 1123 q = p; 1124 } 1125 1126 // Prepend the RATs to the EntryList 1127 if (_EntryList != NULL) { 1128 q->_next = _EntryList; 1129 _EntryList->_prev = q; 1130 } 1131 _EntryList = w; 1132 1133 // Fall thru into code that tries to wake a successor from EntryList 1134 } 1135 1136 w = _EntryList; 1137 if (w != NULL) { 1138 // I'd like to write: guarantee (w->_thread != Self). 1139 // But in practice an exiting thread may find itself on the EntryList. 1140 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1141 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1142 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1143 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1144 // release the lock "O". T2 resumes immediately after the ST of null into 1145 // _owner, above. T2 notices that the EntryList is populated, so it 1146 // reacquires the lock and then finds itself on the EntryList. 1147 // Given all that, we have to tolerate the circumstance where "w" is 1148 // associated with Self. 1149 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1150 ExitEpilog(Self, w); 1151 return; 1152 } 1153 1154 // If we find that both _cxq and EntryList are null then just 1155 // re-run the exit protocol from the top. 1156 w = _cxq; 1157 if (w == NULL) continue; 1158 1159 // Drain _cxq into EntryList - bulk transfer. 1160 // First, detach _cxq. 1161 // The following loop is tantamount to: w = swap(&cxq, NULL) 1162 for (;;) { 1163 assert(w != NULL, "Invariant"); 1164 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1165 if (u == w) break; 1166 w = u; 1167 } 1168 TEVENT(Inflated exit - drain cxq into EntryList); 1169 1170 assert(w != NULL, "invariant"); 1171 assert(_EntryList == NULL, "invariant"); 1172 1173 // Convert the LIFO SLL anchored by _cxq into a DLL. 1174 // The list reorganization step operates in O(LENGTH(w)) time. 1175 // It's critical that this step operate quickly as 1176 // "Self" still holds the outer-lock, restricting parallelism 1177 // and effectively lengthening the critical section. 1178 // Invariant: s chases t chases u. 1179 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1180 // we have faster access to the tail. 1181 1182 if (QMode == 1) { 1183 // QMode == 1 : drain cxq to EntryList, reversing order 1184 // We also reverse the order of the list. 1185 ObjectWaiter * s = NULL; 1186 ObjectWaiter * t = w; 1187 ObjectWaiter * u = NULL; 1188 while (t != NULL) { 1189 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1190 t->TState = ObjectWaiter::TS_ENTER; 1191 u = t->_next; 1192 t->_prev = u; 1193 t->_next = s; 1194 s = t; 1195 t = u; 1196 } 1197 _EntryList = s; 1198 assert(s != NULL, "invariant"); 1199 } else { 1200 // QMode == 0 or QMode == 2 1201 _EntryList = w; 1202 ObjectWaiter * q = NULL; 1203 ObjectWaiter * p; 1204 for (p = w; p != NULL; p = p->_next) { 1205 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1206 p->TState = ObjectWaiter::TS_ENTER; 1207 p->_prev = q; 1208 q = p; 1209 } 1210 } 1211 1212 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1213 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1214 1215 // See if we can abdicate to a spinner instead of waking a thread. 1216 // A primary goal of the implementation is to reduce the 1217 // context-switch rate. 1218 if (_succ != NULL) continue; 1219 1220 w = _EntryList; 1221 if (w != NULL) { 1222 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1223 ExitEpilog(Self, w); 1224 return; 1225 } 1226 } 1227 } 1228 1229 // ExitSuspendEquivalent: 1230 // A faster alternate to handle_special_suspend_equivalent_condition() 1231 // 1232 // handle_special_suspend_equivalent_condition() unconditionally 1233 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1234 // operations have high latency. Note that in ::enter() we call HSSEC 1235 // while holding the monitor, so we effectively lengthen the critical sections. 1236 // 1237 // There are a number of possible solutions: 1238 // 1239 // A. To ameliorate the problem we might also defer state transitions 1240 // to as late as possible -- just prior to parking. 1241 // Given that, we'd call HSSEC after having returned from park(), 1242 // but before attempting to acquire the monitor. This is only a 1243 // partial solution. It avoids calling HSSEC while holding the 1244 // monitor (good), but it still increases successor reacquisition latency -- 1245 // the interval between unparking a successor and the time the successor 1246 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1247 // If we use this technique we can also avoid EnterI()-exit() loop 1248 // in ::enter() where we iteratively drop the lock and then attempt 1249 // to reacquire it after suspending. 1250 // 1251 // B. In the future we might fold all the suspend bits into a 1252 // composite per-thread suspend flag and then update it with CAS(). 1253 // Alternately, a Dekker-like mechanism with multiple variables 1254 // would suffice: 1255 // ST Self->_suspend_equivalent = false 1256 // MEMBAR 1257 // LD Self_>_suspend_flags 1258 // 1259 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1260 // with a more efficient implementation, the need to use "FastHSSEC" has 1261 // decreased. - Dave 1262 1263 1264 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1265 const int Mode = Knob_FastHSSEC; 1266 if (Mode && !jSelf->is_external_suspend()) { 1267 assert(jSelf->is_suspend_equivalent(), "invariant"); 1268 jSelf->clear_suspend_equivalent(); 1269 if (2 == Mode) OrderAccess::storeload(); 1270 if (!jSelf->is_external_suspend()) return false; 1271 // We raced a suspension -- fall thru into the slow path 1272 TEVENT(ExitSuspendEquivalent - raced); 1273 jSelf->set_suspend_equivalent(); 1274 } 1275 return jSelf->handle_special_suspend_equivalent_condition(); 1276 } 1277 1278 1279 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1280 assert(_owner == Self, "invariant"); 1281 1282 // Exit protocol: 1283 // 1. ST _succ = wakee 1284 // 2. membar #loadstore|#storestore; 1285 // 2. ST _owner = NULL 1286 // 3. unpark(wakee) 1287 1288 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1289 ParkEvent * Trigger = Wakee->_event; 1290 1291 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1292 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1293 // out-of-scope (non-extant). 1294 Wakee = NULL; 1295 1296 // Drop the lock 1297 OrderAccess::release_store(&_owner, (void*)NULL); 1298 OrderAccess::fence(); // ST _owner vs LD in unpark() 1299 1300 if (SafepointMechanism::poll(Self)) { 1301 TEVENT(unpark before SAFEPOINT); 1302 } 1303 1304 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1305 Trigger->unpark(); 1306 1307 // Maintain stats and report events to JVMTI 1308 OM_PERFDATA_OP(Parks, inc()); 1309 } 1310 1311 1312 // ----------------------------------------------------------------------------- 1313 // Class Loader deadlock handling. 1314 // 1315 // complete_exit exits a lock returning recursion count 1316 // complete_exit/reenter operate as a wait without waiting 1317 // complete_exit requires an inflated monitor 1318 // The _owner field is not always the Thread addr even with an 1319 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1320 // thread due to contention. 1321 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1322 Thread * const Self = THREAD; 1323 assert(Self->is_Java_thread(), "Must be Java thread!"); 1324 JavaThread *jt = (JavaThread *)THREAD; 1325 1326 DeferredInitialize(); 1327 1328 if (THREAD != _owner) { 1329 if (THREAD->is_lock_owned ((address)_owner)) { 1330 assert(_recursions == 0, "internal state error"); 1331 _owner = THREAD; // Convert from basiclock addr to Thread addr 1332 _recursions = 0; 1333 } 1334 } 1335 1336 guarantee(Self == _owner, "complete_exit not owner"); 1337 intptr_t save = _recursions; // record the old recursion count 1338 _recursions = 0; // set the recursion level to be 0 1339 exit(true, Self); // exit the monitor 1340 guarantee(_owner != Self, "invariant"); 1341 return save; 1342 } 1343 1344 // reenter() enters a lock and sets recursion count 1345 // complete_exit/reenter operate as a wait without waiting 1346 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1347 Thread * const Self = THREAD; 1348 assert(Self->is_Java_thread(), "Must be Java thread!"); 1349 JavaThread *jt = (JavaThread *)THREAD; 1350 1351 guarantee(_owner != Self, "reenter already owner"); 1352 enter(THREAD); // enter the monitor 1353 guarantee(_recursions == 0, "reenter recursion"); 1354 _recursions = recursions; 1355 return; 1356 } 1357 1358 1359 // ----------------------------------------------------------------------------- 1360 // A macro is used below because there may already be a pending 1361 // exception which should not abort the execution of the routines 1362 // which use this (which is why we don't put this into check_slow and 1363 // call it with a CHECK argument). 1364 1365 #define CHECK_OWNER() \ 1366 do { \ 1367 if (THREAD != _owner) { \ 1368 if (THREAD->is_lock_owned((address) _owner)) { \ 1369 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1370 _recursions = 0; \ 1371 } else { \ 1372 TEVENT(Throw IMSX); \ 1373 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1374 } \ 1375 } \ 1376 } while (false) 1377 1378 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1379 // TODO-FIXME: remove check_slow() -- it's likely dead. 1380 1381 void ObjectMonitor::check_slow(TRAPS) { 1382 TEVENT(check_slow - throw IMSX); 1383 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1384 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1385 } 1386 1387 static int Adjust(volatile int * adr, int dx) { 1388 int v; 1389 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1390 return v; 1391 } 1392 1393 // helper method for posting a monitor wait event 1394 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1395 jlong notifier_tid, 1396 jlong timeout, 1397 bool timedout) { 1398 assert(event != NULL, "invariant"); 1399 event->set_monitorClass(((oop)this->object())->klass()); 1400 event->set_timeout(timeout); 1401 event->set_address((TYPE_ADDRESS)this->object_addr()); 1402 event->set_notifier(notifier_tid); 1403 event->set_timedOut(timedout); 1404 event->commit(); 1405 } 1406 1407 // ----------------------------------------------------------------------------- 1408 // Wait/Notify/NotifyAll 1409 // 1410 // Note: a subset of changes to ObjectMonitor::wait() 1411 // will need to be replicated in complete_exit 1412 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1413 Thread * const Self = THREAD; 1414 assert(Self->is_Java_thread(), "Must be Java thread!"); 1415 JavaThread *jt = (JavaThread *)THREAD; 1416 1417 DeferredInitialize(); 1418 1419 // Throw IMSX or IEX. 1420 CHECK_OWNER(); 1421 1422 EventJavaMonitorWait event; 1423 1424 // check for a pending interrupt 1425 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1426 // post monitor waited event. Note that this is past-tense, we are done waiting. 1427 if (JvmtiExport::should_post_monitor_waited()) { 1428 // Note: 'false' parameter is passed here because the 1429 // wait was not timed out due to thread interrupt. 1430 JvmtiExport::post_monitor_waited(jt, this, false); 1431 1432 // In this short circuit of the monitor wait protocol, the 1433 // current thread never drops ownership of the monitor and 1434 // never gets added to the wait queue so the current thread 1435 // cannot be made the successor. This means that the 1436 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1437 // consume an unpark() meant for the ParkEvent associated with 1438 // this ObjectMonitor. 1439 } 1440 if (event.should_commit()) { 1441 post_monitor_wait_event(&event, 0, millis, false); 1442 } 1443 TEVENT(Wait - Throw IEX); 1444 THROW(vmSymbols::java_lang_InterruptedException()); 1445 return; 1446 } 1447 1448 TEVENT(Wait); 1449 1450 assert(Self->_Stalled == 0, "invariant"); 1451 Self->_Stalled = intptr_t(this); 1452 jt->set_current_waiting_monitor(this); 1453 1454 // create a node to be put into the queue 1455 // Critically, after we reset() the event but prior to park(), we must check 1456 // for a pending interrupt. 1457 ObjectWaiter node(Self); 1458 node.TState = ObjectWaiter::TS_WAIT; 1459 Self->_ParkEvent->reset(); 1460 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1461 1462 // Enter the waiting queue, which is a circular doubly linked list in this case 1463 // but it could be a priority queue or any data structure. 1464 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1465 // by the the owner of the monitor *except* in the case where park() 1466 // returns because of a timeout of interrupt. Contention is exceptionally rare 1467 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1468 1469 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1470 AddWaiter(&node); 1471 Thread::SpinRelease(&_WaitSetLock); 1472 1473 if ((SyncFlags & 4) == 0) { 1474 _Responsible = NULL; 1475 } 1476 intptr_t save = _recursions; // record the old recursion count 1477 _waiters++; // increment the number of waiters 1478 _recursions = 0; // set the recursion level to be 1 1479 exit(true, Self); // exit the monitor 1480 guarantee(_owner != Self, "invariant"); 1481 1482 // The thread is on the WaitSet list - now park() it. 1483 // On MP systems it's conceivable that a brief spin before we park 1484 // could be profitable. 1485 // 1486 // TODO-FIXME: change the following logic to a loop of the form 1487 // while (!timeout && !interrupted && _notified == 0) park() 1488 1489 int ret = OS_OK; 1490 int WasNotified = 0; 1491 { // State transition wrappers 1492 OSThread* osthread = Self->osthread(); 1493 OSThreadWaitState osts(osthread, true); 1494 { 1495 ThreadBlockInVM tbivm(jt); 1496 // Thread is in thread_blocked state and oop access is unsafe. 1497 jt->set_suspend_equivalent(); 1498 1499 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1500 // Intentionally empty 1501 } else if (node._notified == 0) { 1502 if (millis <= 0) { 1503 Self->_ParkEvent->park(); 1504 } else { 1505 ret = Self->_ParkEvent->park(millis); 1506 } 1507 } 1508 1509 // were we externally suspended while we were waiting? 1510 if (ExitSuspendEquivalent (jt)) { 1511 // TODO-FIXME: add -- if succ == Self then succ = null. 1512 jt->java_suspend_self(); 1513 } 1514 1515 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1516 1517 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1518 // from the WaitSet to the EntryList. 1519 // See if we need to remove Node from the WaitSet. 1520 // We use double-checked locking to avoid grabbing _WaitSetLock 1521 // if the thread is not on the wait queue. 1522 // 1523 // Note that we don't need a fence before the fetch of TState. 1524 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1525 // written by the is thread. (perhaps the fetch might even be satisfied 1526 // by a look-aside into the processor's own store buffer, although given 1527 // the length of the code path between the prior ST and this load that's 1528 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1529 // then we'll acquire the lock and then re-fetch a fresh TState value. 1530 // That is, we fail toward safety. 1531 1532 if (node.TState == ObjectWaiter::TS_WAIT) { 1533 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1534 if (node.TState == ObjectWaiter::TS_WAIT) { 1535 DequeueSpecificWaiter(&node); // unlink from WaitSet 1536 assert(node._notified == 0, "invariant"); 1537 node.TState = ObjectWaiter::TS_RUN; 1538 } 1539 Thread::SpinRelease(&_WaitSetLock); 1540 } 1541 1542 // The thread is now either on off-list (TS_RUN), 1543 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1544 // The Node's TState variable is stable from the perspective of this thread. 1545 // No other threads will asynchronously modify TState. 1546 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1547 OrderAccess::loadload(); 1548 if (_succ == Self) _succ = NULL; 1549 WasNotified = node._notified; 1550 1551 // Reentry phase -- reacquire the monitor. 1552 // re-enter contended monitor after object.wait(). 1553 // retain OBJECT_WAIT state until re-enter successfully completes 1554 // Thread state is thread_in_vm and oop access is again safe, 1555 // although the raw address of the object may have changed. 1556 // (Don't cache naked oops over safepoints, of course). 1557 1558 // post monitor waited event. Note that this is past-tense, we are done waiting. 1559 if (JvmtiExport::should_post_monitor_waited()) { 1560 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1561 1562 if (node._notified != 0 && _succ == Self) { 1563 // In this part of the monitor wait-notify-reenter protocol it 1564 // is possible (and normal) for another thread to do a fastpath 1565 // monitor enter-exit while this thread is still trying to get 1566 // to the reenter portion of the protocol. 1567 // 1568 // The ObjectMonitor was notified and the current thread is 1569 // the successor which also means that an unpark() has already 1570 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1571 // consume the unpark() that was done when the successor was 1572 // set because the same ParkEvent is shared between Java 1573 // monitors and JVM/TI RawMonitors (for now). 1574 // 1575 // We redo the unpark() to ensure forward progress, i.e., we 1576 // don't want all pending threads hanging (parked) with none 1577 // entering the unlocked monitor. 1578 node._event->unpark(); 1579 } 1580 } 1581 1582 if (event.should_commit()) { 1583 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1584 } 1585 1586 OrderAccess::fence(); 1587 1588 assert(Self->_Stalled != 0, "invariant"); 1589 Self->_Stalled = 0; 1590 1591 assert(_owner != Self, "invariant"); 1592 ObjectWaiter::TStates v = node.TState; 1593 if (v == ObjectWaiter::TS_RUN) { 1594 enter(Self); 1595 } else { 1596 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1597 ReenterI(Self, &node); 1598 node.wait_reenter_end(this); 1599 } 1600 1601 // Self has reacquired the lock. 1602 // Lifecycle - the node representing Self must not appear on any queues. 1603 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1604 // want residual elements associated with this thread left on any lists. 1605 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1606 assert(_owner == Self, "invariant"); 1607 assert(_succ != Self, "invariant"); 1608 } // OSThreadWaitState() 1609 1610 jt->set_current_waiting_monitor(NULL); 1611 1612 guarantee(_recursions == 0, "invariant"); 1613 _recursions = save; // restore the old recursion count 1614 _waiters--; // decrement the number of waiters 1615 1616 // Verify a few postconditions 1617 assert(_owner == Self, "invariant"); 1618 assert(_succ != Self, "invariant"); 1619 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1620 1621 if (SyncFlags & 32) { 1622 OrderAccess::fence(); 1623 } 1624 1625 // check if the notification happened 1626 if (!WasNotified) { 1627 // no, it could be timeout or Thread.interrupt() or both 1628 // check for interrupt event, otherwise it is timeout 1629 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1630 TEVENT(Wait - throw IEX from epilog); 1631 THROW(vmSymbols::java_lang_InterruptedException()); 1632 } 1633 } 1634 1635 // NOTE: Spurious wake up will be consider as timeout. 1636 // Monitor notify has precedence over thread interrupt. 1637 } 1638 1639 1640 // Consider: 1641 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1642 // then instead of transferring a thread from the WaitSet to the EntryList 1643 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1644 1645 void ObjectMonitor::INotify(Thread * Self) { 1646 const int policy = Knob_MoveNotifyee; 1647 1648 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1649 ObjectWaiter * iterator = DequeueWaiter(); 1650 if (iterator != NULL) { 1651 TEVENT(Notify1 - Transfer); 1652 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1653 guarantee(iterator->_notified == 0, "invariant"); 1654 // Disposition - what might we do with iterator ? 1655 // a. add it directly to the EntryList - either tail (policy == 1) 1656 // or head (policy == 0). 1657 // b. push it onto the front of the _cxq (policy == 2). 1658 // For now we use (b). 1659 if (policy != 4) { 1660 iterator->TState = ObjectWaiter::TS_ENTER; 1661 } 1662 iterator->_notified = 1; 1663 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1664 1665 ObjectWaiter * list = _EntryList; 1666 if (list != NULL) { 1667 assert(list->_prev == NULL, "invariant"); 1668 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1669 assert(list != iterator, "invariant"); 1670 } 1671 1672 if (policy == 0) { // prepend to EntryList 1673 if (list == NULL) { 1674 iterator->_next = iterator->_prev = NULL; 1675 _EntryList = iterator; 1676 } else { 1677 list->_prev = iterator; 1678 iterator->_next = list; 1679 iterator->_prev = NULL; 1680 _EntryList = iterator; 1681 } 1682 } else if (policy == 1) { // append to EntryList 1683 if (list == NULL) { 1684 iterator->_next = iterator->_prev = NULL; 1685 _EntryList = iterator; 1686 } else { 1687 // CONSIDER: finding the tail currently requires a linear-time walk of 1688 // the EntryList. We can make tail access constant-time by converting to 1689 // a CDLL instead of using our current DLL. 1690 ObjectWaiter * tail; 1691 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1692 assert(tail != NULL && tail->_next == NULL, "invariant"); 1693 tail->_next = iterator; 1694 iterator->_prev = tail; 1695 iterator->_next = NULL; 1696 } 1697 } else if (policy == 2) { // prepend to cxq 1698 if (list == NULL) { 1699 iterator->_next = iterator->_prev = NULL; 1700 _EntryList = iterator; 1701 } else { 1702 iterator->TState = ObjectWaiter::TS_CXQ; 1703 for (;;) { 1704 ObjectWaiter * front = _cxq; 1705 iterator->_next = front; 1706 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1707 break; 1708 } 1709 } 1710 } 1711 } else if (policy == 3) { // append to cxq 1712 iterator->TState = ObjectWaiter::TS_CXQ; 1713 for (;;) { 1714 ObjectWaiter * tail = _cxq; 1715 if (tail == NULL) { 1716 iterator->_next = NULL; 1717 if (Atomic::cmpxchg(iterator, &_cxq, (ObjectWaiter*)NULL) == NULL) { 1718 break; 1719 } 1720 } else { 1721 while (tail->_next != NULL) tail = tail->_next; 1722 tail->_next = iterator; 1723 iterator->_prev = tail; 1724 iterator->_next = NULL; 1725 break; 1726 } 1727 } 1728 } else { 1729 ParkEvent * ev = iterator->_event; 1730 iterator->TState = ObjectWaiter::TS_RUN; 1731 OrderAccess::fence(); 1732 ev->unpark(); 1733 } 1734 1735 // _WaitSetLock protects the wait queue, not the EntryList. We could 1736 // move the add-to-EntryList operation, above, outside the critical section 1737 // protected by _WaitSetLock. In practice that's not useful. With the 1738 // exception of wait() timeouts and interrupts the monitor owner 1739 // is the only thread that grabs _WaitSetLock. There's almost no contention 1740 // on _WaitSetLock so it's not profitable to reduce the length of the 1741 // critical section. 1742 1743 if (policy < 4) { 1744 iterator->wait_reenter_begin(this); 1745 } 1746 } 1747 Thread::SpinRelease(&_WaitSetLock); 1748 } 1749 1750 // Consider: a not-uncommon synchronization bug is to use notify() when 1751 // notifyAll() is more appropriate, potentially resulting in stranded 1752 // threads; this is one example of a lost wakeup. A useful diagnostic 1753 // option is to force all notify() operations to behave as notifyAll(). 1754 // 1755 // Note: We can also detect many such problems with a "minimum wait". 1756 // When the "minimum wait" is set to a small non-zero timeout value 1757 // and the program does not hang whereas it did absent "minimum wait", 1758 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses 1759 // a "minimum wait" for all park() operations; see the recheckInterval 1760 // variable and MAX_RECHECK_INTERVAL. 1761 1762 void ObjectMonitor::notify(TRAPS) { 1763 CHECK_OWNER(); 1764 if (_WaitSet == NULL) { 1765 TEVENT(Empty-Notify); 1766 return; 1767 } 1768 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1769 INotify(THREAD); 1770 OM_PERFDATA_OP(Notifications, inc(1)); 1771 } 1772 1773 1774 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1775 // from the waitset to the EntryList. This could be done more efficiently with a 1776 // single bulk transfer but in practice it's not time-critical. Beware too, 1777 // that in prepend-mode we invert the order of the waiters. Let's say that the 1778 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1779 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1780 1781 void ObjectMonitor::notifyAll(TRAPS) { 1782 CHECK_OWNER(); 1783 if (_WaitSet == NULL) { 1784 TEVENT(Empty-NotifyAll); 1785 return; 1786 } 1787 1788 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1789 int tally = 0; 1790 while (_WaitSet != NULL) { 1791 tally++; 1792 INotify(THREAD); 1793 } 1794 1795 OM_PERFDATA_OP(Notifications, inc(tally)); 1796 } 1797 1798 // ----------------------------------------------------------------------------- 1799 // Adaptive Spinning Support 1800 // 1801 // Adaptive spin-then-block - rational spinning 1802 // 1803 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1804 // algorithm. On high order SMP systems it would be better to start with 1805 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1806 // a contending thread could enqueue itself on the cxq and then spin locally 1807 // on a thread-specific variable such as its ParkEvent._Event flag. 1808 // That's left as an exercise for the reader. Note that global spinning is 1809 // not problematic on Niagara, as the L2 cache serves the interconnect and 1810 // has both low latency and massive bandwidth. 1811 // 1812 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1813 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1814 // (duration) or we can fix the count at approximately the duration of 1815 // a context switch and vary the frequency. Of course we could also 1816 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1817 // For a description of 'Adaptive spin-then-block mutual exclusion in 1818 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1819 // 1820 // This implementation varies the duration "D", where D varies with 1821 // the success rate of recent spin attempts. (D is capped at approximately 1822 // length of a round-trip context switch). The success rate for recent 1823 // spin attempts is a good predictor of the success rate of future spin 1824 // attempts. The mechanism adapts automatically to varying critical 1825 // section length (lock modality), system load and degree of parallelism. 1826 // D is maintained per-monitor in _SpinDuration and is initialized 1827 // optimistically. Spin frequency is fixed at 100%. 1828 // 1829 // Note that _SpinDuration is volatile, but we update it without locks 1830 // or atomics. The code is designed so that _SpinDuration stays within 1831 // a reasonable range even in the presence of races. The arithmetic 1832 // operations on _SpinDuration are closed over the domain of legal values, 1833 // so at worst a race will install and older but still legal value. 1834 // At the very worst this introduces some apparent non-determinism. 1835 // We might spin when we shouldn't or vice-versa, but since the spin 1836 // count are relatively short, even in the worst case, the effect is harmless. 1837 // 1838 // Care must be taken that a low "D" value does not become an 1839 // an absorbing state. Transient spinning failures -- when spinning 1840 // is overall profitable -- should not cause the system to converge 1841 // on low "D" values. We want spinning to be stable and predictable 1842 // and fairly responsive to change and at the same time we don't want 1843 // it to oscillate, become metastable, be "too" non-deterministic, 1844 // or converge on or enter undesirable stable absorbing states. 1845 // 1846 // We implement a feedback-based control system -- using past behavior 1847 // to predict future behavior. We face two issues: (a) if the 1848 // input signal is random then the spin predictor won't provide optimal 1849 // results, and (b) if the signal frequency is too high then the control 1850 // system, which has some natural response lag, will "chase" the signal. 1851 // (b) can arise from multimodal lock hold times. Transient preemption 1852 // can also result in apparent bimodal lock hold times. 1853 // Although sub-optimal, neither condition is particularly harmful, as 1854 // in the worst-case we'll spin when we shouldn't or vice-versa. 1855 // The maximum spin duration is rather short so the failure modes aren't bad. 1856 // To be conservative, I've tuned the gain in system to bias toward 1857 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1858 // "rings" or oscillates between spinning and not spinning. This happens 1859 // when spinning is just on the cusp of profitability, however, so the 1860 // situation is not dire. The state is benign -- there's no need to add 1861 // hysteresis control to damp the transition rate between spinning and 1862 // not spinning. 1863 1864 // Spinning: Fixed frequency (100%), vary duration 1865 int ObjectMonitor::TrySpin(Thread * Self) { 1866 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1867 int ctr = Knob_FixedSpin; 1868 if (ctr != 0) { 1869 while (--ctr >= 0) { 1870 if (TryLock(Self) > 0) return 1; 1871 SpinPause(); 1872 } 1873 return 0; 1874 } 1875 1876 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1877 if (TryLock(Self) > 0) { 1878 // Increase _SpinDuration ... 1879 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1880 // Raising _SpurDuration to the poverty line is key. 1881 int x = _SpinDuration; 1882 if (x < Knob_SpinLimit) { 1883 if (x < Knob_Poverty) x = Knob_Poverty; 1884 _SpinDuration = x + Knob_BonusB; 1885 } 1886 return 1; 1887 } 1888 SpinPause(); 1889 } 1890 1891 // Admission control - verify preconditions for spinning 1892 // 1893 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1894 // becoming an absorbing state. Put another way, we spin briefly to 1895 // sample, just in case the system load, parallelism, contention, or lock 1896 // modality changed. 1897 // 1898 // Consider the following alternative: 1899 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1900 // spin attempt. "Periodically" might mean after a tally of 1901 // the # of failed spin attempts (or iterations) reaches some threshold. 1902 // This takes us into the realm of 1-out-of-N spinning, where we 1903 // hold the duration constant but vary the frequency. 1904 1905 ctr = _SpinDuration; 1906 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1907 if (ctr <= 0) return 0; 1908 1909 if (Knob_SuccRestrict && _succ != NULL) return 0; 1910 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1911 TEVENT(Spin abort - notrunnable [TOP]); 1912 return 0; 1913 } 1914 1915 int MaxSpin = Knob_MaxSpinners; 1916 if (MaxSpin >= 0) { 1917 if (_Spinner > MaxSpin) { 1918 TEVENT(Spin abort -- too many spinners); 1919 return 0; 1920 } 1921 // Slightly racy, but benign ... 1922 Adjust(&_Spinner, 1); 1923 } 1924 1925 // We're good to spin ... spin ingress. 1926 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1927 // when preparing to LD...CAS _owner, etc and the CAS is likely 1928 // to succeed. 1929 int hits = 0; 1930 int msk = 0; 1931 int caspty = Knob_CASPenalty; 1932 int oxpty = Knob_OXPenalty; 1933 int sss = Knob_SpinSetSucc; 1934 if (sss && _succ == NULL) _succ = Self; 1935 Thread * prv = NULL; 1936 1937 // There are three ways to exit the following loop: 1938 // 1. A successful spin where this thread has acquired the lock. 1939 // 2. Spin failure with prejudice 1940 // 3. Spin failure without prejudice 1941 1942 while (--ctr >= 0) { 1943 1944 // Periodic polling -- Check for pending GC 1945 // Threads may spin while they're unsafe. 1946 // We don't want spinning threads to delay the JVM from reaching 1947 // a stop-the-world safepoint or to steal cycles from GC. 1948 // If we detect a pending safepoint we abort in order that 1949 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1950 // this thread, if safe, doesn't steal cycles from GC. 1951 // This is in keeping with the "no loitering in runtime" rule. 1952 // We periodically check to see if there's a safepoint pending. 1953 if ((ctr & 0xFF) == 0) { 1954 if (SafepointMechanism::poll(Self)) { 1955 TEVENT(Spin: safepoint); 1956 goto Abort; // abrupt spin egress 1957 } 1958 if (Knob_UsePause & 1) SpinPause(); 1959 } 1960 1961 if (Knob_UsePause & 2) SpinPause(); 1962 1963 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1964 // This is useful on classic SMP systems, but is of less utility on 1965 // N1-style CMT platforms. 1966 // 1967 // Trade-off: lock acquisition latency vs coherency bandwidth. 1968 // Lock hold times are typically short. A histogram 1969 // of successful spin attempts shows that we usually acquire 1970 // the lock early in the spin. That suggests we want to 1971 // sample _owner frequently in the early phase of the spin, 1972 // but then back-off and sample less frequently as the spin 1973 // progresses. The back-off makes a good citizen on SMP big 1974 // SMP systems. Oversampling _owner can consume excessive 1975 // coherency bandwidth. Relatedly, if we _oversample _owner we 1976 // can inadvertently interfere with the the ST m->owner=null. 1977 // executed by the lock owner. 1978 if (ctr & msk) continue; 1979 ++hits; 1980 if ((hits & 0xF) == 0) { 1981 // The 0xF, above, corresponds to the exponent. 1982 // Consider: (msk+1)|msk 1983 msk = ((msk << 2)|3) & BackOffMask; 1984 } 1985 1986 // Probe _owner with TATAS 1987 // If this thread observes the monitor transition or flicker 1988 // from locked to unlocked to locked, then the odds that this 1989 // thread will acquire the lock in this spin attempt go down 1990 // considerably. The same argument applies if the CAS fails 1991 // or if we observe _owner change from one non-null value to 1992 // another non-null value. In such cases we might abort 1993 // the spin without prejudice or apply a "penalty" to the 1994 // spin count-down variable "ctr", reducing it by 100, say. 1995 1996 Thread * ox = (Thread *) _owner; 1997 if (ox == NULL) { 1998 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1999 if (ox == NULL) { 2000 // The CAS succeeded -- this thread acquired ownership 2001 // Take care of some bookkeeping to exit spin state. 2002 if (sss && _succ == Self) { 2003 _succ = NULL; 2004 } 2005 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2006 2007 // Increase _SpinDuration : 2008 // The spin was successful (profitable) so we tend toward 2009 // longer spin attempts in the future. 2010 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2011 // If we acquired the lock early in the spin cycle it 2012 // makes sense to increase _SpinDuration proportionally. 2013 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2014 int x = _SpinDuration; 2015 if (x < Knob_SpinLimit) { 2016 if (x < Knob_Poverty) x = Knob_Poverty; 2017 _SpinDuration = x + Knob_Bonus; 2018 } 2019 return 1; 2020 } 2021 2022 // The CAS failed ... we can take any of the following actions: 2023 // * penalize: ctr -= Knob_CASPenalty 2024 // * exit spin with prejudice -- goto Abort; 2025 // * exit spin without prejudice. 2026 // * Since CAS is high-latency, retry again immediately. 2027 prv = ox; 2028 TEVENT(Spin: cas failed); 2029 if (caspty == -2) break; 2030 if (caspty == -1) goto Abort; 2031 ctr -= caspty; 2032 continue; 2033 } 2034 2035 // Did lock ownership change hands ? 2036 if (ox != prv && prv != NULL) { 2037 TEVENT(spin: Owner changed) 2038 if (oxpty == -2) break; 2039 if (oxpty == -1) goto Abort; 2040 ctr -= oxpty; 2041 } 2042 prv = ox; 2043 2044 // Abort the spin if the owner is not executing. 2045 // The owner must be executing in order to drop the lock. 2046 // Spinning while the owner is OFFPROC is idiocy. 2047 // Consider: ctr -= RunnablePenalty ; 2048 if (Knob_OState && NotRunnable (Self, ox)) { 2049 TEVENT(Spin abort - notrunnable); 2050 goto Abort; 2051 } 2052 if (sss && _succ == NULL) _succ = Self; 2053 } 2054 2055 // Spin failed with prejudice -- reduce _SpinDuration. 2056 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2057 // AIMD is globally stable. 2058 TEVENT(Spin failure); 2059 { 2060 int x = _SpinDuration; 2061 if (x > 0) { 2062 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2063 // This is globally sample and tends to damp the response. 2064 x -= Knob_Penalty; 2065 if (x < 0) x = 0; 2066 _SpinDuration = x; 2067 } 2068 } 2069 2070 Abort: 2071 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2072 if (sss && _succ == Self) { 2073 _succ = NULL; 2074 // Invariant: after setting succ=null a contending thread 2075 // must recheck-retry _owner before parking. This usually happens 2076 // in the normal usage of TrySpin(), but it's safest 2077 // to make TrySpin() as foolproof as possible. 2078 OrderAccess::fence(); 2079 if (TryLock(Self) > 0) return 1; 2080 } 2081 return 0; 2082 } 2083 2084 // NotRunnable() -- informed spinning 2085 // 2086 // Don't bother spinning if the owner is not eligible to drop the lock. 2087 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2088 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2089 // The thread must be runnable in order to drop the lock in timely fashion. 2090 // If the _owner is not runnable then spinning will not likely be 2091 // successful (profitable). 2092 // 2093 // Beware -- the thread referenced by _owner could have died 2094 // so a simply fetch from _owner->_thread_state might trap. 2095 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2096 // Because of the lifecycle issues the schedctl and _thread_state values 2097 // observed by NotRunnable() might be garbage. NotRunnable must 2098 // tolerate this and consider the observed _thread_state value 2099 // as advisory. 2100 // 2101 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2102 // a thread pointer. 2103 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2104 // with the LSB of _owner. Another option would be to probablistically probe 2105 // the putative _owner->TypeTag value. 2106 // 2107 // Checking _thread_state isn't perfect. Even if the thread is 2108 // in_java it might be blocked on a page-fault or have been preempted 2109 // and sitting on a ready/dispatch queue. _thread state in conjunction 2110 // with schedctl.sc_state gives us a good picture of what the 2111 // thread is doing, however. 2112 // 2113 // TODO: check schedctl.sc_state. 2114 // We'll need to use SafeFetch32() to read from the schedctl block. 2115 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2116 // 2117 // The return value from NotRunnable() is *advisory* -- the 2118 // result is based on sampling and is not necessarily coherent. 2119 // The caller must tolerate false-negative and false-positive errors. 2120 // Spinning, in general, is probabilistic anyway. 2121 2122 2123 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2124 // Check ox->TypeTag == 2BAD. 2125 if (ox == NULL) return 0; 2126 2127 // Avoid transitive spinning ... 2128 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2129 // Immediately after T1 acquires L it's possible that T2, also 2130 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2131 // This occurs transiently after T1 acquired L but before 2132 // T1 managed to clear T1.Stalled. T2 does not need to abort 2133 // its spin in this circumstance. 2134 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2135 2136 if (BlockedOn == 1) return 1; 2137 if (BlockedOn != 0) { 2138 return BlockedOn != intptr_t(this) && _owner == ox; 2139 } 2140 2141 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2142 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2143 // consider also: jst != _thread_in_Java -- but that's overspecific. 2144 return jst == _thread_blocked || jst == _thread_in_native; 2145 } 2146 2147 2148 // ----------------------------------------------------------------------------- 2149 // WaitSet management ... 2150 2151 ObjectWaiter::ObjectWaiter(Thread* thread) { 2152 _next = NULL; 2153 _prev = NULL; 2154 _notified = 0; 2155 _notifier_tid = 0; 2156 TState = TS_RUN; 2157 _thread = thread; 2158 _event = thread->_ParkEvent; 2159 _active = false; 2160 assert(_event != NULL, "invariant"); 2161 } 2162 2163 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2164 JavaThread *jt = (JavaThread *)this->_thread; 2165 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2166 } 2167 2168 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2169 JavaThread *jt = (JavaThread *)this->_thread; 2170 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2171 } 2172 2173 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2174 assert(node != NULL, "should not add NULL node"); 2175 assert(node->_prev == NULL, "node already in list"); 2176 assert(node->_next == NULL, "node already in list"); 2177 // put node at end of queue (circular doubly linked list) 2178 if (_WaitSet == NULL) { 2179 _WaitSet = node; 2180 node->_prev = node; 2181 node->_next = node; 2182 } else { 2183 ObjectWaiter* head = _WaitSet; 2184 ObjectWaiter* tail = head->_prev; 2185 assert(tail->_next == head, "invariant check"); 2186 tail->_next = node; 2187 head->_prev = node; 2188 node->_next = head; 2189 node->_prev = tail; 2190 } 2191 } 2192 2193 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2194 // dequeue the very first waiter 2195 ObjectWaiter* waiter = _WaitSet; 2196 if (waiter) { 2197 DequeueSpecificWaiter(waiter); 2198 } 2199 return waiter; 2200 } 2201 2202 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2203 assert(node != NULL, "should not dequeue NULL node"); 2204 assert(node->_prev != NULL, "node already removed from list"); 2205 assert(node->_next != NULL, "node already removed from list"); 2206 // when the waiter has woken up because of interrupt, 2207 // timeout or other spurious wake-up, dequeue the 2208 // waiter from waiting list 2209 ObjectWaiter* next = node->_next; 2210 if (next == node) { 2211 assert(node->_prev == node, "invariant check"); 2212 _WaitSet = NULL; 2213 } else { 2214 ObjectWaiter* prev = node->_prev; 2215 assert(prev->_next == node, "invariant check"); 2216 assert(next->_prev == node, "invariant check"); 2217 next->_prev = prev; 2218 prev->_next = next; 2219 if (_WaitSet == node) { 2220 _WaitSet = next; 2221 } 2222 } 2223 node->_next = NULL; 2224 node->_prev = NULL; 2225 } 2226 2227 // ----------------------------------------------------------------------------- 2228 // PerfData support 2229 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2230 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2231 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2232 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2233 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2234 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2235 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2236 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2237 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2238 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2239 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2240 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2241 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2242 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2243 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2244 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2245 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2246 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2247 2248 // One-shot global initialization for the sync subsystem. 2249 // We could also defer initialization and initialize on-demand 2250 // the first time we call inflate(). Initialization would 2251 // be protected - like so many things - by the MonitorCache_lock. 2252 2253 void ObjectMonitor::Initialize() { 2254 static int InitializationCompleted = 0; 2255 assert(InitializationCompleted == 0, "invariant"); 2256 InitializationCompleted = 1; 2257 if (UsePerfData) { 2258 EXCEPTION_MARK; 2259 #define NEWPERFCOUNTER(n) \ 2260 { \ 2261 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2262 CHECK); \ 2263 } 2264 #define NEWPERFVARIABLE(n) \ 2265 { \ 2266 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2267 CHECK); \ 2268 } 2269 NEWPERFCOUNTER(_sync_Inflations); 2270 NEWPERFCOUNTER(_sync_Deflations); 2271 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2272 NEWPERFCOUNTER(_sync_FutileWakeups); 2273 NEWPERFCOUNTER(_sync_Parks); 2274 NEWPERFCOUNTER(_sync_EmptyNotifications); 2275 NEWPERFCOUNTER(_sync_Notifications); 2276 NEWPERFCOUNTER(_sync_SlowEnter); 2277 NEWPERFCOUNTER(_sync_SlowExit); 2278 NEWPERFCOUNTER(_sync_SlowNotify); 2279 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2280 NEWPERFCOUNTER(_sync_FailedSpins); 2281 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2282 NEWPERFCOUNTER(_sync_PrivateA); 2283 NEWPERFCOUNTER(_sync_PrivateB); 2284 NEWPERFCOUNTER(_sync_MonInCirculation); 2285 NEWPERFCOUNTER(_sync_MonScavenged); 2286 NEWPERFVARIABLE(_sync_MonExtant); 2287 #undef NEWPERFCOUNTER 2288 #undef NEWPERFVARIABLE 2289 } 2290 } 2291 2292 static char * kvGet(char * kvList, const char * Key) { 2293 if (kvList == NULL) return NULL; 2294 size_t n = strlen(Key); 2295 char * Search; 2296 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2297 if (strncmp (Search, Key, n) == 0) { 2298 if (Search[n] == '=') return Search + n + 1; 2299 if (Search[n] == 0) return(char *) "1"; 2300 } 2301 } 2302 return NULL; 2303 } 2304 2305 static int kvGetInt(char * kvList, const char * Key, int Default) { 2306 char * v = kvGet(kvList, Key); 2307 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2308 if (Knob_ReportSettings && v != NULL) { 2309 tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ; 2310 tty->flush(); 2311 } 2312 return rslt; 2313 } 2314 2315 void ObjectMonitor::DeferredInitialize() { 2316 if (InitDone > 0) return; 2317 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2318 while (InitDone != 1) /* empty */; 2319 return; 2320 } 2321 2322 // One-shot global initialization ... 2323 // The initialization is idempotent, so we don't need locks. 2324 // In the future consider doing this via os::init_2(). 2325 // SyncKnobs consist of <Key>=<Value> pairs in the style 2326 // of environment variables. Start by converting ':' to NUL. 2327 2328 if (SyncKnobs == NULL) SyncKnobs = ""; 2329 2330 size_t sz = strlen(SyncKnobs); 2331 char * knobs = (char *) malloc(sz + 2); 2332 if (knobs == NULL) { 2333 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2334 guarantee(0, "invariant"); 2335 } 2336 strcpy(knobs, SyncKnobs); 2337 knobs[sz+1] = 0; 2338 for (char * p = knobs; *p; p++) { 2339 if (*p == ':') *p = 0; 2340 } 2341 2342 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2343 SETKNOB(ReportSettings); 2344 SETKNOB(ExitRelease); 2345 SETKNOB(Verbose); 2346 SETKNOB(VerifyInUse); 2347 SETKNOB(VerifyMatch); 2348 SETKNOB(FixedSpin); 2349 SETKNOB(SpinLimit); 2350 SETKNOB(SpinBase); 2351 SETKNOB(SpinBackOff); 2352 SETKNOB(CASPenalty); 2353 SETKNOB(OXPenalty); 2354 SETKNOB(LogSpins); 2355 SETKNOB(SpinSetSucc); 2356 SETKNOB(SuccEnabled); 2357 SETKNOB(SuccRestrict); 2358 SETKNOB(Penalty); 2359 SETKNOB(Bonus); 2360 SETKNOB(BonusB); 2361 SETKNOB(Poverty); 2362 SETKNOB(SpinAfterFutile); 2363 SETKNOB(UsePause); 2364 SETKNOB(SpinEarly); 2365 SETKNOB(OState); 2366 SETKNOB(MaxSpinners); 2367 SETKNOB(PreSpin); 2368 SETKNOB(ExitPolicy); 2369 SETKNOB(QMode); 2370 SETKNOB(ResetEvent); 2371 SETKNOB(MoveNotifyee); 2372 SETKNOB(FastHSSEC); 2373 #undef SETKNOB 2374 2375 if (Knob_Verbose) { 2376 sanity_checks(); 2377 } 2378 2379 if (os::is_MP()) { 2380 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2381 if (Knob_ReportSettings) { 2382 tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask); 2383 } 2384 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2385 } else { 2386 Knob_SpinLimit = 0; 2387 Knob_SpinBase = 0; 2388 Knob_PreSpin = 0; 2389 Knob_FixedSpin = -1; 2390 } 2391 2392 if (Knob_LogSpins == 0) { 2393 ObjectMonitor::_sync_FailedSpins = NULL; 2394 } 2395 2396 free(knobs); 2397 OrderAccess::fence(); 2398 InitDone = 1; 2399 } 2400 2401 void ObjectMonitor::sanity_checks() { 2402 int error_cnt = 0; 2403 int warning_cnt = 0; 2404 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2405 2406 if (verbose) { 2407 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2408 sizeof(ObjectMonitor)); 2409 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT, 2410 sizeof(PaddedEnd<ObjectMonitor>)); 2411 } 2412 2413 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2414 if (verbose) { 2415 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2416 } 2417 2418 ObjectMonitor dummy; 2419 u_char *addr_begin = (u_char*)&dummy; 2420 u_char *addr_header = (u_char*)&dummy._header; 2421 u_char *addr_owner = (u_char*)&dummy._owner; 2422 2423 uint offset_header = (uint)(addr_header - addr_begin); 2424 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2425 2426 uint offset_owner = (uint)(addr_owner - addr_begin); 2427 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2428 2429 if ((uint)(addr_header - addr_begin) != 0) { 2430 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2431 error_cnt++; 2432 } 2433 2434 if (cache_line_size != 0) { 2435 // We were able to determine the L1 data cache line size so 2436 // do some cache line specific sanity checks 2437 2438 if ((offset_owner - offset_header) < cache_line_size) { 2439 tty->print_cr("WARNING: the _header and _owner fields are closer " 2440 "than a cache line which permits false sharing."); 2441 warning_cnt++; 2442 } 2443 2444 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) { 2445 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a " 2446 "multiple of a cache line which permits false sharing."); 2447 warning_cnt++; 2448 } 2449 } 2450 2451 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2452 &warning_cnt); 2453 2454 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2455 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2456 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2457 } 2458 2459 guarantee(error_cnt == 0, 2460 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2461 } 2462 2463 #ifndef PRODUCT 2464 void ObjectMonitor_test() { 2465 ObjectMonitor::sanity_checks(); 2466 } 2467 #endif