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