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