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