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