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