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