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