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