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