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