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_SpinEarly = 1; 107 static int Knob_SuccEnabled = 1; // futile wake throttling 108 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 109 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 110 static int Knob_Bonus = 100; // spin success bonus 111 static int Knob_BonusB = 100; // spin success bonus 112 static int Knob_Penalty = 200; // spin failure penalty 113 static int Knob_Poverty = 1000; 114 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 115 static int Knob_FixedSpin = 0; 116 static int Knob_OState = 3; // Spinner checks thread state of _owner 117 static int Knob_UsePause = 1; 118 static int Knob_ExitPolicy = 0; 119 static int Knob_PreSpin = 10; // 20-100 likely better 120 static int Knob_ResetEvent = 0; 121 122 static int Knob_FastHSSEC = 0; 123 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 124 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 125 static volatile int InitDone = 0; 126 127 // ----------------------------------------------------------------------------- 128 // Theory of operations -- Monitors lists, thread residency, etc: 129 // 130 // * A thread acquires ownership of a monitor by successfully 131 // CAS()ing the _owner field from null to non-null. 132 // 133 // * Invariant: A thread appears on at most one monitor list -- 134 // cxq, EntryList or WaitSet -- at any one time. 135 // 136 // * Contending threads "push" themselves onto the cxq with CAS 137 // and then spin/park. 138 // 139 // * After a contending thread eventually acquires the lock it must 140 // dequeue itself from either the EntryList or the cxq. 141 // 142 // * The exiting thread identifies and unparks an "heir presumptive" 143 // tentative successor thread on the EntryList. Critically, the 144 // exiting thread doesn't unlink the successor thread from the EntryList. 145 // After having been unparked, the wakee will recontend for ownership of 146 // the monitor. The successor (wakee) will either acquire the lock or 147 // re-park itself. 148 // 149 // Succession is provided for by a policy of competitive handoff. 150 // The exiting thread does _not_ grant or pass ownership to the 151 // successor thread. (This is also referred to as "handoff" succession"). 152 // Instead the exiting thread releases ownership and possibly wakes 153 // a successor, so the successor can (re)compete for ownership of the lock. 154 // If the EntryList is empty but the cxq is populated the exiting 155 // thread will drain the cxq into the EntryList. It does so by 156 // by detaching the cxq (installing null with CAS) and folding 157 // the threads from the cxq into the EntryList. The EntryList is 158 // doubly linked, while the cxq is singly linked because of the 159 // CAS-based "push" used to enqueue recently arrived threads (RATs). 160 // 161 // * Concurrency invariants: 162 // 163 // -- only the monitor owner may access or mutate the EntryList. 164 // The mutex property of the monitor itself protects the EntryList 165 // from concurrent interference. 166 // -- Only the monitor owner may detach the cxq. 167 // 168 // * The monitor entry list operations avoid locks, but strictly speaking 169 // they're not lock-free. Enter is lock-free, exit is not. 170 // For a description of 'Methods and apparatus providing non-blocking access 171 // to a resource,' see U.S. Pat. No. 7844973. 172 // 173 // * The cxq can have multiple concurrent "pushers" but only one concurrent 174 // detaching thread. This mechanism is immune from the ABA corruption. 175 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 176 // 177 // * Taken together, the cxq and the EntryList constitute or form a 178 // single logical queue of threads stalled trying to acquire the lock. 179 // We use two distinct lists to improve the odds of a constant-time 180 // dequeue operation after acquisition (in the ::enter() epilogue) and 181 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 182 // A key desideratum is to minimize queue & monitor metadata manipulation 183 // that occurs while holding the monitor lock -- that is, we want to 184 // minimize monitor lock holds times. Note that even a small amount of 185 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 186 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 187 // locks and monitor metadata. 188 // 189 // Cxq points to the set of Recently Arrived Threads attempting entry. 190 // Because we push threads onto _cxq with CAS, the RATs must take the form of 191 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 192 // the unlocking thread notices that EntryList is null but _cxq is != null. 193 // 194 // The EntryList is ordered by the prevailing queue discipline and 195 // can be organized in any convenient fashion, such as a doubly-linked list or 196 // a circular doubly-linked list. Critically, we want insert and delete operations 197 // to operate in constant-time. If we need a priority queue then something akin 198 // to Solaris' sleepq would work nicely. Viz., 199 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 200 // Queue discipline is enforced at ::exit() time, when the unlocking thread 201 // drains the cxq into the EntryList, and orders or reorders the threads on the 202 // EntryList accordingly. 203 // 204 // Barring "lock barging", this mechanism provides fair cyclic ordering, 205 // somewhat similar to an elevator-scan. 206 // 207 // * The monitor synchronization subsystem avoids the use of native 208 // synchronization primitives except for the narrow platform-specific 209 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 210 // the semantics of park-unpark. Put another way, this monitor implementation 211 // depends only on atomic operations and park-unpark. The monitor subsystem 212 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 213 // underlying OS manages the READY<->RUN transitions. 214 // 215 // * Waiting threads reside on the WaitSet list -- wait() puts 216 // the caller onto the WaitSet. 217 // 218 // * notify() or notifyAll() simply transfers threads from the WaitSet to 219 // either the EntryList or cxq. Subsequent exit() operations will 220 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 221 // it's likely the notifyee would simply impale itself on the lock held 222 // by the notifier. 223 // 224 // * An interesting alternative is to encode cxq as (List,LockByte) where 225 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 226 // variable, like _recursions, in the scheme. The threads or Events that form 227 // the list would have to be aligned in 256-byte addresses. A thread would 228 // try to acquire the lock or enqueue itself with CAS, but exiting threads 229 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 230 // Note that is is *not* word-tearing, but it does presume that full-word 231 // CAS operations are coherent with intermix with STB operations. That's true 232 // on most common processors. 233 // 234 // * See also http://blogs.sun.com/dave 235 236 237 void* ObjectMonitor::operator new (size_t size) throw() { 238 return AllocateHeap(size, mtInternal); 239 } 240 void* ObjectMonitor::operator new[] (size_t size) throw() { 241 return operator new (size); 242 } 243 void ObjectMonitor::operator delete(void* p) { 244 FreeHeap(p); 245 } 246 void ObjectMonitor::operator delete[] (void *p) { 247 operator delete(p); 248 } 249 250 // ----------------------------------------------------------------------------- 251 // Enter support 252 253 void ObjectMonitor::enter(TRAPS) { 254 // The following code is ordered to check the most common cases first 255 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 256 Thread * const Self = THREAD; 257 258 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 259 if (cur == NULL) { 260 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 261 assert(_recursions == 0, "invariant"); 262 assert(_owner == Self, "invariant"); 263 return; 264 } 265 266 if (cur == Self) { 267 // TODO-FIXME: check for integer overflow! BUGID 6557169. 268 _recursions++; 269 return; 270 } 271 272 if (Self->is_lock_owned ((address)cur)) { 273 assert(_recursions == 0, "internal state error"); 274 _recursions = 1; 275 // Commute owner from a thread-specific on-stack BasicLockObject address to 276 // a full-fledged "Thread *". 277 _owner = Self; 278 return; 279 } 280 281 // We've encountered genuine contention. 282 assert(Self->_Stalled == 0, "invariant"); 283 Self->_Stalled = intptr_t(this); 284 285 // Try one round of spinning *before* enqueueing Self 286 // and before going through the awkward and expensive state 287 // transitions. The following spin is strictly optional ... 288 // Note that if we acquire the monitor from an initial spin 289 // we forgo posting JVMTI events and firing DTRACE probes. 290 if (Knob_SpinEarly && TrySpin (Self) > 0) { 291 assert(_owner == Self, "invariant"); 292 assert(_recursions == 0, "invariant"); 293 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 294 Self->_Stalled = 0; 295 return; 296 } 297 298 assert(_owner != Self, "invariant"); 299 assert(_succ != Self, "invariant"); 300 assert(Self->is_Java_thread(), "invariant"); 301 JavaThread * jt = (JavaThread *) Self; 302 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 303 assert(jt->thread_state() != _thread_blocked, "invariant"); 304 assert(this->object() != NULL, "invariant"); 305 assert(_count >= 0, "invariant"); 306 307 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 308 // Ensure the object-monitor relationship remains stable while there's contention. 309 Atomic::inc(&_count); 310 311 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 312 EventJavaMonitorEnter event; 313 if (event.should_commit()) { 314 event.set_monitorClass(((oop)this->object())->klass()); 315 event.set_address((uintptr_t)(this->object_addr())); 316 } 317 318 { // Change java thread status to indicate blocked on monitor enter. 319 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 320 321 Self->set_current_pending_monitor(this); 322 323 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 324 if (JvmtiExport::should_post_monitor_contended_enter()) { 325 JvmtiExport::post_monitor_contended_enter(jt, this); 326 327 // The current thread does not yet own the monitor and does not 328 // yet appear on any queues that would get it made the successor. 329 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 330 // handler cannot accidentally consume an unpark() meant for the 331 // ParkEvent associated with this ObjectMonitor. 332 } 333 334 OSThreadContendState osts(Self->osthread()); 335 ThreadBlockInVM tbivm(jt); 336 337 // TODO-FIXME: change the following for(;;) loop to straight-line code. 338 for (;;) { 339 jt->set_suspend_equivalent(); 340 // cleared by handle_special_suspend_equivalent_condition() 341 // or java_suspend_self() 342 343 EnterI(THREAD); 344 345 if (!ExitSuspendEquivalent(jt)) break; 346 347 // We have acquired the contended monitor, but while we were 348 // waiting another thread suspended us. We don't want to enter 349 // the monitor while suspended because that would surprise the 350 // thread that suspended us. 351 // 352 _recursions = 0; 353 _succ = NULL; 354 exit(false, Self); 355 356 jt->java_suspend_self(); 357 } 358 Self->set_current_pending_monitor(NULL); 359 360 // We cleared the pending monitor info since we've just gotten past 361 // the enter-check-for-suspend dance and we now own the monitor free 362 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 363 // destructor can go to a safepoint at the end of this block. If we 364 // do a thread dump during that safepoint, then this thread will show 365 // as having "-locked" the monitor, but the OS and java.lang.Thread 366 // states will still report that the thread is blocked trying to 367 // acquire it. 368 } 369 370 Atomic::dec(&_count); 371 assert(_count >= 0, "invariant"); 372 Self->_Stalled = 0; 373 374 // Must either set _recursions = 0 or ASSERT _recursions == 0. 375 assert(_recursions == 0, "invariant"); 376 assert(_owner == Self, "invariant"); 377 assert(_succ != Self, "invariant"); 378 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 379 380 // The thread -- now the owner -- is back in vm mode. 381 // Report the glorious news via TI,DTrace and jvmstat. 382 // The probe effect is non-trivial. All the reportage occurs 383 // while we hold the monitor, increasing the length of the critical 384 // section. Amdahl's parallel speedup law comes vividly into play. 385 // 386 // Another option might be to aggregate the events (thread local or 387 // per-monitor aggregation) and defer reporting until a more opportune 388 // time -- such as next time some thread encounters contention but has 389 // yet to acquire the lock. While spinning that thread could 390 // spinning we could increment JVMStat counters, etc. 391 392 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 393 if (JvmtiExport::should_post_monitor_contended_entered()) { 394 JvmtiExport::post_monitor_contended_entered(jt, this); 395 396 // The current thread already owns the monitor and is not going to 397 // call park() for the remainder of the monitor enter protocol. So 398 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 399 // event handler consumed an unpark() issued by the thread that 400 // just exited the monitor. 401 } 402 if (event.should_commit()) { 403 event.set_previousOwner((uintptr_t)_previous_owner_tid); 404 event.commit(); 405 } 406 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 407 } 408 409 // Caveat: TryLock() is not necessarily serializing if it returns failure. 410 // Callers must compensate as needed. 411 412 int ObjectMonitor::TryLock(Thread * Self) { 413 void * own = _owner; 414 if (own != NULL) return 0; 415 if (Atomic::replace_if_null(Self, &_owner)) { 416 // Either guarantee _recursions == 0 or set _recursions = 0. 417 assert(_recursions == 0, "invariant"); 418 assert(_owner == Self, "invariant"); 419 return 1; 420 } 421 // The lock had been free momentarily, but we lost the race to the lock. 422 // Interference -- the CAS failed. 423 // We can either return -1 or retry. 424 // Retry doesn't make as much sense because the lock was just acquired. 425 return -1; 426 } 427 428 #define MAX_RECHECK_INTERVAL 1000 429 430 void ObjectMonitor::EnterI(TRAPS) { 431 Thread * const Self = THREAD; 432 assert(Self->is_Java_thread(), "invariant"); 433 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 434 435 // Try the lock - TATAS 436 if (TryLock (Self) > 0) { 437 assert(_succ != Self, "invariant"); 438 assert(_owner == Self, "invariant"); 439 assert(_Responsible != Self, "invariant"); 440 return; 441 } 442 443 DeferredInitialize(); 444 445 // We try one round of spinning *before* enqueueing Self. 446 // 447 // If the _owner is ready but OFFPROC we could use a YieldTo() 448 // operation to donate the remainder of this thread's quantum 449 // to the owner. This has subtle but beneficial affinity 450 // effects. 451 452 if (TrySpin (Self) > 0) { 453 assert(_owner == Self, "invariant"); 454 assert(_succ != Self, "invariant"); 455 assert(_Responsible != Self, "invariant"); 456 return; 457 } 458 459 // The Spin failed -- Enqueue and park the thread ... 460 assert(_succ != Self, "invariant"); 461 assert(_owner != Self, "invariant"); 462 assert(_Responsible != Self, "invariant"); 463 464 // Enqueue "Self" on ObjectMonitor's _cxq. 465 // 466 // Node acts as a proxy for Self. 467 // As an aside, if were to ever rewrite the synchronization code mostly 468 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 469 // Java objects. This would avoid awkward lifecycle and liveness issues, 470 // as well as eliminate a subset of ABA issues. 471 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 472 473 ObjectWaiter node(Self); 474 Self->_ParkEvent->reset(); 475 node._prev = (ObjectWaiter *) 0xBAD; 476 node.TState = ObjectWaiter::TS_CXQ; 477 478 // Push "Self" onto the front of the _cxq. 479 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 480 // Note that spinning tends to reduce the rate at which threads 481 // enqueue and dequeue on EntryList|cxq. 482 ObjectWaiter * nxt; 483 for (;;) { 484 node._next = nxt = _cxq; 485 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 486 487 // Interference - the CAS failed because _cxq changed. Just retry. 488 // As an optional optimization we retry the lock. 489 if (TryLock (Self) > 0) { 490 assert(_succ != Self, "invariant"); 491 assert(_owner == Self, "invariant"); 492 assert(_Responsible != Self, "invariant"); 493 return; 494 } 495 } 496 497 // Check for cxq|EntryList edge transition to non-null. This indicates 498 // the onset of contention. While contention persists exiting threads 499 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 500 // operations revert to the faster 1-0 mode. This enter operation may interleave 501 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 502 // arrange for one of the contending thread to use a timed park() operations 503 // to detect and recover from the race. (Stranding is form of progress failure 504 // where the monitor is unlocked but all the contending threads remain parked). 505 // That is, at least one of the contended threads will periodically poll _owner. 506 // One of the contending threads will become the designated "Responsible" thread. 507 // The Responsible thread uses a timed park instead of a normal indefinite park 508 // operation -- it periodically wakes and checks for and recovers from potential 509 // strandings admitted by 1-0 exit operations. We need at most one Responsible 510 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 511 // be responsible for a monitor. 512 // 513 // Currently, one of the contended threads takes on the added role of "Responsible". 514 // A viable alternative would be to use a dedicated "stranding checker" thread 515 // that periodically iterated over all the threads (or active monitors) and unparked 516 // successors where there was risk of stranding. This would help eliminate the 517 // timer scalability issues we see on some platforms as we'd only have one thread 518 // -- the checker -- parked on a timer. 519 520 if (nxt == NULL && _EntryList == NULL) { 521 // Try to assume the role of responsible thread for the monitor. 522 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 523 Atomic::replace_if_null(Self, &_Responsible); 524 } 525 526 // The lock might have been released while this thread was occupied queueing 527 // itself onto _cxq. To close the race and avoid "stranding" and 528 // progress-liveness failure we must resample-retry _owner before parking. 529 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 530 // In this case the ST-MEMBAR is accomplished with CAS(). 531 // 532 // TODO: Defer all thread state transitions until park-time. 533 // Since state transitions are heavy and inefficient we'd like 534 // to defer the state transitions until absolutely necessary, 535 // and in doing so avoid some transitions ... 536 537 int nWakeups = 0; 538 int recheckInterval = 1; 539 540 for (;;) { 541 542 if (TryLock(Self) > 0) break; 543 assert(_owner != Self, "invariant"); 544 545 // park self 546 if (_Responsible == Self) { 547 Self->_ParkEvent->park((jlong) recheckInterval); 548 // Increase the recheckInterval, but clamp the value. 549 recheckInterval *= 8; 550 if (recheckInterval > MAX_RECHECK_INTERVAL) { 551 recheckInterval = MAX_RECHECK_INTERVAL; 552 } 553 } else { 554 Self->_ParkEvent->park(); 555 } 556 557 if (TryLock(Self) > 0) break; 558 559 // The lock is still contested. 560 // Keep a tally of the # of futile wakeups. 561 // Note that the counter is not protected by a lock or updated by atomics. 562 // That is by design - we trade "lossy" counters which are exposed to 563 // races during updates for a lower probe effect. 564 565 // This PerfData object can be used in parallel with a safepoint. 566 // See the work around in PerfDataManager::destroy(). 567 OM_PERFDATA_OP(FutileWakeups, inc()); 568 ++nWakeups; 569 570 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 571 // We can defer clearing _succ until after the spin completes 572 // TrySpin() must tolerate being called with _succ == Self. 573 // Try yet another round of adaptive spinning. 574 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 575 576 // We can find that we were unpark()ed and redesignated _succ while 577 // we were spinning. That's harmless. If we iterate and call park(), 578 // park() will consume the event and return immediately and we'll 579 // just spin again. This pattern can repeat, leaving _succ to simply 580 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 581 // Alternately, we can sample fired() here, and if set, forgo spinning 582 // in the next iteration. 583 584 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 585 Self->_ParkEvent->reset(); 586 OrderAccess::fence(); 587 } 588 if (_succ == Self) _succ = NULL; 589 590 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 591 OrderAccess::fence(); 592 } 593 594 // Egress : 595 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 596 // Normally we'll find Self on the EntryList . 597 // From the perspective of the lock owner (this thread), the 598 // EntryList is stable and cxq is prepend-only. 599 // The head of cxq is volatile but the interior is stable. 600 // In addition, Self.TState is stable. 601 602 assert(_owner == Self, "invariant"); 603 assert(object() != NULL, "invariant"); 604 // I'd like to write: 605 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 606 // but as we're at a safepoint that's not safe. 607 608 UnlinkAfterAcquire(Self, &node); 609 if (_succ == Self) _succ = NULL; 610 611 assert(_succ != Self, "invariant"); 612 if (_Responsible == Self) { 613 _Responsible = NULL; 614 OrderAccess::fence(); // Dekker pivot-point 615 616 // We may leave threads on cxq|EntryList without a designated 617 // "Responsible" thread. This is benign. When this thread subsequently 618 // exits the monitor it can "see" such preexisting "old" threads -- 619 // threads that arrived on the cxq|EntryList before the fence, above -- 620 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 621 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 622 // non-null and elect a new "Responsible" timer thread. 623 // 624 // This thread executes: 625 // ST Responsible=null; MEMBAR (in enter epilogue - here) 626 // LD cxq|EntryList (in subsequent exit) 627 // 628 // Entering threads in the slow/contended path execute: 629 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 630 // The (ST cxq; MEMBAR) is accomplished with CAS(). 631 // 632 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 633 // exit operation from floating above the ST Responsible=null. 634 } 635 636 // We've acquired ownership with CAS(). 637 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 638 // But since the CAS() this thread may have also stored into _succ, 639 // EntryList, cxq or Responsible. These meta-data updates must be 640 // visible __before this thread subsequently drops the lock. 641 // Consider what could occur if we didn't enforce this constraint -- 642 // STs to monitor meta-data and user-data could reorder with (become 643 // visible after) the ST in exit that drops ownership of the lock. 644 // Some other thread could then acquire the lock, but observe inconsistent 645 // or old monitor meta-data and heap data. That violates the JMM. 646 // To that end, the 1-0 exit() operation must have at least STST|LDST 647 // "release" barrier semantics. Specifically, there must be at least a 648 // STST|LDST barrier in exit() before the ST of null into _owner that drops 649 // the lock. The barrier ensures that changes to monitor meta-data and data 650 // protected by the lock will be visible before we release the lock, and 651 // therefore before some other thread (CPU) has a chance to acquire the lock. 652 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 653 // 654 // Critically, any prior STs to _succ or EntryList must be visible before 655 // the ST of null into _owner in the *subsequent* (following) corresponding 656 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 657 // execute a serializing instruction. 658 659 return; 660 } 661 662 // ReenterI() is a specialized inline form of the latter half of the 663 // contended slow-path from EnterI(). We use ReenterI() only for 664 // monitor reentry in wait(). 665 // 666 // In the future we should reconcile EnterI() and ReenterI(), adding 667 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 668 // loop accordingly. 669 670 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 671 assert(Self != NULL, "invariant"); 672 assert(SelfNode != NULL, "invariant"); 673 assert(SelfNode->_thread == Self, "invariant"); 674 assert(_waiters > 0, "invariant"); 675 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 676 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 677 JavaThread * jt = (JavaThread *) Self; 678 679 int nWakeups = 0; 680 for (;;) { 681 ObjectWaiter::TStates v = SelfNode->TState; 682 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 683 assert(_owner != Self, "invariant"); 684 685 if (TryLock(Self) > 0) break; 686 if (TrySpin(Self) > 0) break; 687 688 // State transition wrappers around park() ... 689 // ReenterI() wisely defers state transitions until 690 // it's clear we must park the thread. 691 { 692 OSThreadContendState osts(Self->osthread()); 693 ThreadBlockInVM tbivm(jt); 694 695 // cleared by handle_special_suspend_equivalent_condition() 696 // or java_suspend_self() 697 jt->set_suspend_equivalent(); 698 Self->_ParkEvent->park(); 699 700 // were we externally suspended while we were waiting? 701 for (;;) { 702 if (!ExitSuspendEquivalent(jt)) break; 703 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 704 jt->java_suspend_self(); 705 jt->set_suspend_equivalent(); 706 } 707 } 708 709 // Try again, but just so we distinguish between futile wakeups and 710 // successful wakeups. The following test isn't algorithmically 711 // necessary, but it helps us maintain sensible statistics. 712 if (TryLock(Self) > 0) break; 713 714 // The lock is still contested. 715 // Keep a tally of the # of futile wakeups. 716 // Note that the counter is not protected by a lock or updated by atomics. 717 // That is by design - we trade "lossy" counters which are exposed to 718 // races during updates for a lower probe effect. 719 ++nWakeups; 720 721 // Assuming this is not a spurious wakeup we'll normally 722 // find that _succ == Self. 723 if (_succ == Self) _succ = NULL; 724 725 // Invariant: after clearing _succ a contending thread 726 // *must* retry _owner before parking. 727 OrderAccess::fence(); 728 729 // This PerfData object can be used in parallel with a safepoint. 730 // See the work around in PerfDataManager::destroy(). 731 OM_PERFDATA_OP(FutileWakeups, inc()); 732 } 733 734 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 735 // Normally we'll find Self on the EntryList. 736 // Unlinking from the EntryList is constant-time and atomic-free. 737 // From the perspective of the lock owner (this thread), the 738 // EntryList is stable and cxq is prepend-only. 739 // The head of cxq is volatile but the interior is stable. 740 // In addition, Self.TState is stable. 741 742 assert(_owner == Self, "invariant"); 743 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 744 UnlinkAfterAcquire(Self, SelfNode); 745 if (_succ == Self) _succ = NULL; 746 assert(_succ != Self, "invariant"); 747 SelfNode->TState = ObjectWaiter::TS_RUN; 748 OrderAccess::fence(); // see comments at the end of EnterI() 749 } 750 751 // By convention we unlink a contending thread from EntryList|cxq immediately 752 // after the thread acquires the lock in ::enter(). Equally, we could defer 753 // unlinking the thread until ::exit()-time. 754 755 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 756 assert(_owner == Self, "invariant"); 757 assert(SelfNode->_thread == Self, "invariant"); 758 759 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 760 // Normal case: remove Self from the DLL EntryList . 761 // This is a constant-time operation. 762 ObjectWaiter * nxt = SelfNode->_next; 763 ObjectWaiter * prv = SelfNode->_prev; 764 if (nxt != NULL) nxt->_prev = prv; 765 if (prv != NULL) prv->_next = nxt; 766 if (SelfNode == _EntryList) _EntryList = nxt; 767 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 768 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 769 } else { 770 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 771 // Inopportune interleaving -- Self is still on the cxq. 772 // This usually means the enqueue of self raced an exiting thread. 773 // Normally we'll find Self near the front of the cxq, so 774 // dequeueing is typically fast. If needbe we can accelerate 775 // this with some MCS/CHL-like bidirectional list hints and advisory 776 // back-links so dequeueing from the interior will normally operate 777 // in constant-time. 778 // Dequeue Self from either the head (with CAS) or from the interior 779 // with a linear-time scan and normal non-atomic memory operations. 780 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 781 // and then unlink Self from EntryList. We have to drain eventually, 782 // so it might as well be now. 783 784 ObjectWaiter * v = _cxq; 785 assert(v != NULL, "invariant"); 786 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 787 // The CAS above can fail from interference IFF a "RAT" arrived. 788 // In that case Self must be in the interior and can no longer be 789 // at the head of cxq. 790 if (v == SelfNode) { 791 assert(_cxq != v, "invariant"); 792 v = _cxq; // CAS above failed - start scan at head of list 793 } 794 ObjectWaiter * p; 795 ObjectWaiter * q = NULL; 796 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 797 q = p; 798 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 799 } 800 assert(v != SelfNode, "invariant"); 801 assert(p == SelfNode, "Node not found on cxq"); 802 assert(p != _cxq, "invariant"); 803 assert(q != NULL, "invariant"); 804 assert(q->_next == p, "invariant"); 805 q->_next = p->_next; 806 } 807 } 808 809 #ifdef ASSERT 810 // Diagnostic hygiene ... 811 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 812 SelfNode->_next = (ObjectWaiter *) 0xBAD; 813 SelfNode->TState = ObjectWaiter::TS_RUN; 814 #endif 815 } 816 817 // ----------------------------------------------------------------------------- 818 // Exit support 819 // 820 // exit() 821 // ~~~~~~ 822 // Note that the collector can't reclaim the objectMonitor or deflate 823 // the object out from underneath the thread calling ::exit() as the 824 // thread calling ::exit() never transitions to a stable state. 825 // This inhibits GC, which in turn inhibits asynchronous (and 826 // inopportune) reclamation of "this". 827 // 828 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 829 // There's one exception to the claim above, however. EnterI() can call 830 // exit() to drop a lock if the acquirer has been externally suspended. 831 // In that case exit() is called with _thread_state as _thread_blocked, 832 // but the monitor's _count field is > 0, which inhibits reclamation. 833 // 834 // 1-0 exit 835 // ~~~~~~~~ 836 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 837 // the fast-path operators have been optimized so the common ::exit() 838 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 839 // The code emitted by fast_unlock() elides the usual MEMBAR. This 840 // greatly improves latency -- MEMBAR and CAS having considerable local 841 // latency on modern processors -- but at the cost of "stranding". Absent the 842 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 843 // ::enter() path, resulting in the entering thread being stranding 844 // and a progress-liveness failure. Stranding is extremely rare. 845 // We use timers (timed park operations) & periodic polling to detect 846 // and recover from stranding. Potentially stranded threads periodically 847 // wake up and poll the lock. See the usage of the _Responsible variable. 848 // 849 // The CAS() in enter provides for safety and exclusion, while the CAS or 850 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 851 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 852 // We detect and recover from stranding with timers. 853 // 854 // If a thread transiently strands it'll park until (a) another 855 // thread acquires the lock and then drops the lock, at which time the 856 // exiting thread will notice and unpark the stranded thread, or, (b) 857 // the timer expires. If the lock is high traffic then the stranding latency 858 // will be low due to (a). If the lock is low traffic then the odds of 859 // stranding are lower, although the worst-case stranding latency 860 // is longer. Critically, we don't want to put excessive load in the 861 // platform's timer subsystem. We want to minimize both the timer injection 862 // rate (timers created/sec) as well as the number of timers active at 863 // any one time. (more precisely, we want to minimize timer-seconds, which is 864 // the integral of the # of active timers at any instant over time). 865 // Both impinge on OS scalability. Given that, at most one thread parked on 866 // a monitor will use a timer. 867 // 868 // There is also the risk of a futile wake-up. If we drop the lock 869 // another thread can reacquire the lock immediately, and we can 870 // then wake a thread unnecessarily. This is benign, and we've 871 // structured the code so the windows are short and the frequency 872 // of such futile wakups is low. 873 874 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 875 Thread * const Self = THREAD; 876 if (THREAD != _owner) { 877 if (THREAD->is_lock_owned((address) _owner)) { 878 // Transmute _owner from a BasicLock pointer to a Thread address. 879 // We don't need to hold _mutex for this transition. 880 // Non-null to Non-null is safe as long as all readers can 881 // tolerate either flavor. 882 assert(_recursions == 0, "invariant"); 883 _owner = THREAD; 884 _recursions = 0; 885 } else { 886 // Apparent unbalanced locking ... 887 // Naively we'd like to throw IllegalMonitorStateException. 888 // As a practical matter we can neither allocate nor throw an 889 // exception as ::exit() can be called from leaf routines. 890 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 891 // Upon deeper reflection, however, in a properly run JVM the only 892 // way we should encounter this situation is in the presence of 893 // unbalanced JNI locking. TODO: CheckJNICalls. 894 // See also: CR4414101 895 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 896 return; 897 } 898 } 899 900 if (_recursions != 0) { 901 _recursions--; // this is simple recursive enter 902 return; 903 } 904 905 // Invariant: after setting Responsible=null an thread must execute 906 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 907 _Responsible = NULL; 908 909 #if INCLUDE_JFR 910 // get the owner's thread id for the MonitorEnter event 911 // if it is enabled and the thread isn't suspended 912 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 913 _previous_owner_tid = JFR_THREAD_ID(Self); 914 } 915 #endif 916 917 for (;;) { 918 assert(THREAD == _owner, "invariant"); 919 920 if (Knob_ExitPolicy == 0) { 921 // release semantics: prior loads and stores from within the critical section 922 // must not float (reorder) past the following store that drops the lock. 923 // On SPARC that requires MEMBAR #loadstore|#storestore. 924 // But of course in TSO #loadstore|#storestore is not required. 925 // I'd like to write one of the following: 926 // A. OrderAccess::release() ; _owner = NULL 927 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 928 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 929 // store into a _dummy variable. That store is not needed, but can result 930 // in massive wasteful coherency traffic on classic SMP systems. 931 // Instead, I use release_store(), which is implemented as just a simple 932 // ST on x64, x86 and SPARC. 933 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 934 OrderAccess::storeload(); // See if we need to wake a successor 935 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 936 return; 937 } 938 // Other threads are blocked trying to acquire the lock. 939 940 // Normally the exiting thread is responsible for ensuring succession, 941 // but if other successors are ready or other entering threads are spinning 942 // then this thread can simply store NULL into _owner and exit without 943 // waking a successor. The existence of spinners or ready successors 944 // guarantees proper succession (liveness). Responsibility passes to the 945 // ready or running successors. The exiting thread delegates the duty. 946 // More precisely, if a successor already exists this thread is absolved 947 // of the responsibility of waking (unparking) one. 948 // 949 // The _succ variable is critical to reducing futile wakeup frequency. 950 // _succ identifies the "heir presumptive" thread that has been made 951 // ready (unparked) but that has not yet run. We need only one such 952 // successor thread to guarantee progress. 953 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 954 // section 3.3 "Futile Wakeup Throttling" for details. 955 // 956 // Note that spinners in Enter() also set _succ non-null. 957 // In the current implementation spinners opportunistically set 958 // _succ so that exiting threads might avoid waking a successor. 959 // Another less appealing alternative would be for the exiting thread 960 // to drop the lock and then spin briefly to see if a spinner managed 961 // to acquire the lock. If so, the exiting thread could exit 962 // immediately without waking a successor, otherwise the exiting 963 // thread would need to dequeue and wake a successor. 964 // (Note that we'd need to make the post-drop spin short, but no 965 // shorter than the worst-case round-trip cache-line migration time. 966 // The dropped lock needs to become visible to the spinner, and then 967 // the acquisition of the lock by the spinner must become visible to 968 // the exiting thread). 969 970 // It appears that an heir-presumptive (successor) must be made ready. 971 // Only the current lock owner can manipulate the EntryList or 972 // drain _cxq, so we need to reacquire the lock. If we fail 973 // to reacquire the lock the responsibility for ensuring succession 974 // falls to the new owner. 975 // 976 if (!Atomic::replace_if_null(THREAD, &_owner)) { 977 return; 978 } 979 } else { 980 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 981 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 982 OrderAccess::storeload(); 983 // Ratify the previously observed values. 984 if (_cxq == NULL || _succ != NULL) { 985 return; 986 } 987 988 // inopportune interleaving -- the exiting thread (this thread) 989 // in the fast-exit path raced an entering thread in the slow-enter 990 // path. 991 // We have two choices: 992 // A. Try to reacquire the lock. 993 // If the CAS() fails return immediately, otherwise 994 // we either restart/rerun the exit operation, or simply 995 // fall-through into the code below which wakes a successor. 996 // B. If the elements forming the EntryList|cxq are TSM 997 // we could simply unpark() the lead thread and return 998 // without having set _succ. 999 if (!Atomic::replace_if_null(THREAD, &_owner)) { 1000 return; 1001 } 1002 } 1003 } 1004 1005 guarantee(_owner == THREAD, "invariant"); 1006 1007 ObjectWaiter * w = NULL; 1008 int QMode = Knob_QMode; 1009 1010 if (QMode == 2 && _cxq != NULL) { 1011 // QMode == 2 : cxq has precedence over EntryList. 1012 // Try to directly wake a successor from the cxq. 1013 // If successful, the successor will need to unlink itself from cxq. 1014 w = _cxq; 1015 assert(w != NULL, "invariant"); 1016 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1017 ExitEpilog(Self, w); 1018 return; 1019 } 1020 1021 if (QMode == 3 && _cxq != NULL) { 1022 // Aggressively drain cxq into EntryList at the first opportunity. 1023 // This policy ensure that recently-run threads live at the head of EntryList. 1024 // Drain _cxq into EntryList - bulk transfer. 1025 // First, detach _cxq. 1026 // The following loop is tantamount to: w = swap(&cxq, NULL) 1027 w = _cxq; 1028 for (;;) { 1029 assert(w != NULL, "Invariant"); 1030 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1031 if (u == w) break; 1032 w = u; 1033 } 1034 assert(w != NULL, "invariant"); 1035 1036 ObjectWaiter * q = NULL; 1037 ObjectWaiter * p; 1038 for (p = w; p != NULL; p = p->_next) { 1039 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1040 p->TState = ObjectWaiter::TS_ENTER; 1041 p->_prev = q; 1042 q = p; 1043 } 1044 1045 // Append the RATs to the EntryList 1046 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1047 ObjectWaiter * Tail; 1048 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1049 Tail = Tail->_next) 1050 /* empty */; 1051 if (Tail == NULL) { 1052 _EntryList = w; 1053 } else { 1054 Tail->_next = w; 1055 w->_prev = Tail; 1056 } 1057 1058 // Fall thru into code that tries to wake a successor from EntryList 1059 } 1060 1061 if (QMode == 4 && _cxq != NULL) { 1062 // Aggressively drain cxq into EntryList at the first opportunity. 1063 // This policy ensure that recently-run threads live at the head of EntryList. 1064 1065 // Drain _cxq into EntryList - bulk transfer. 1066 // First, detach _cxq. 1067 // The following loop is tantamount to: w = swap(&cxq, NULL) 1068 w = _cxq; 1069 for (;;) { 1070 assert(w != NULL, "Invariant"); 1071 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1072 if (u == w) break; 1073 w = u; 1074 } 1075 assert(w != NULL, "invariant"); 1076 1077 ObjectWaiter * q = NULL; 1078 ObjectWaiter * p; 1079 for (p = w; p != NULL; p = p->_next) { 1080 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1081 p->TState = ObjectWaiter::TS_ENTER; 1082 p->_prev = q; 1083 q = p; 1084 } 1085 1086 // Prepend the RATs to the EntryList 1087 if (_EntryList != NULL) { 1088 q->_next = _EntryList; 1089 _EntryList->_prev = q; 1090 } 1091 _EntryList = w; 1092 1093 // Fall thru into code that tries to wake a successor from EntryList 1094 } 1095 1096 w = _EntryList; 1097 if (w != NULL) { 1098 // I'd like to write: guarantee (w->_thread != Self). 1099 // But in practice an exiting thread may find itself on the EntryList. 1100 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1101 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1102 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1103 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1104 // release the lock "O". T2 resumes immediately after the ST of null into 1105 // _owner, above. T2 notices that the EntryList is populated, so it 1106 // reacquires the lock and then finds itself on the EntryList. 1107 // Given all that, we have to tolerate the circumstance where "w" is 1108 // associated with Self. 1109 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1110 ExitEpilog(Self, w); 1111 return; 1112 } 1113 1114 // If we find that both _cxq and EntryList are null then just 1115 // re-run the exit protocol from the top. 1116 w = _cxq; 1117 if (w == NULL) continue; 1118 1119 // Drain _cxq into EntryList - bulk transfer. 1120 // First, detach _cxq. 1121 // The following loop is tantamount to: w = swap(&cxq, NULL) 1122 for (;;) { 1123 assert(w != NULL, "Invariant"); 1124 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1125 if (u == w) break; 1126 w = u; 1127 } 1128 1129 assert(w != NULL, "invariant"); 1130 assert(_EntryList == NULL, "invariant"); 1131 1132 // Convert the LIFO SLL anchored by _cxq into a DLL. 1133 // The list reorganization step operates in O(LENGTH(w)) time. 1134 // It's critical that this step operate quickly as 1135 // "Self" still holds the outer-lock, restricting parallelism 1136 // and effectively lengthening the critical section. 1137 // Invariant: s chases t chases u. 1138 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1139 // we have faster access to the tail. 1140 1141 if (QMode == 1) { 1142 // QMode == 1 : drain cxq to EntryList, reversing order 1143 // We also reverse the order of the list. 1144 ObjectWaiter * s = NULL; 1145 ObjectWaiter * t = w; 1146 ObjectWaiter * u = NULL; 1147 while (t != NULL) { 1148 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1149 t->TState = ObjectWaiter::TS_ENTER; 1150 u = t->_next; 1151 t->_prev = u; 1152 t->_next = s; 1153 s = t; 1154 t = u; 1155 } 1156 _EntryList = s; 1157 assert(s != NULL, "invariant"); 1158 } else { 1159 // QMode == 0 or QMode == 2 1160 _EntryList = w; 1161 ObjectWaiter * q = NULL; 1162 ObjectWaiter * p; 1163 for (p = w; p != NULL; p = p->_next) { 1164 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1165 p->TState = ObjectWaiter::TS_ENTER; 1166 p->_prev = q; 1167 q = p; 1168 } 1169 } 1170 1171 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1172 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1173 1174 // See if we can abdicate to a spinner instead of waking a thread. 1175 // A primary goal of the implementation is to reduce the 1176 // context-switch rate. 1177 if (_succ != NULL) continue; 1178 1179 w = _EntryList; 1180 if (w != NULL) { 1181 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1182 ExitEpilog(Self, w); 1183 return; 1184 } 1185 } 1186 } 1187 1188 // ExitSuspendEquivalent: 1189 // A faster alternate to handle_special_suspend_equivalent_condition() 1190 // 1191 // handle_special_suspend_equivalent_condition() unconditionally 1192 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1193 // operations have high latency. Note that in ::enter() we call HSSEC 1194 // while holding the monitor, so we effectively lengthen the critical sections. 1195 // 1196 // There are a number of possible solutions: 1197 // 1198 // A. To ameliorate the problem we might also defer state transitions 1199 // to as late as possible -- just prior to parking. 1200 // Given that, we'd call HSSEC after having returned from park(), 1201 // but before attempting to acquire the monitor. This is only a 1202 // partial solution. It avoids calling HSSEC while holding the 1203 // monitor (good), but it still increases successor reacquisition latency -- 1204 // the interval between unparking a successor and the time the successor 1205 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1206 // If we use this technique we can also avoid EnterI()-exit() loop 1207 // in ::enter() where we iteratively drop the lock and then attempt 1208 // to reacquire it after suspending. 1209 // 1210 // B. In the future we might fold all the suspend bits into a 1211 // composite per-thread suspend flag and then update it with CAS(). 1212 // Alternately, a Dekker-like mechanism with multiple variables 1213 // would suffice: 1214 // ST Self->_suspend_equivalent = false 1215 // MEMBAR 1216 // LD Self_>_suspend_flags 1217 // 1218 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1219 // with a more efficient implementation, the need to use "FastHSSEC" has 1220 // decreased. - Dave 1221 1222 1223 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1224 const int Mode = Knob_FastHSSEC; 1225 if (Mode && !jSelf->is_external_suspend()) { 1226 assert(jSelf->is_suspend_equivalent(), "invariant"); 1227 jSelf->clear_suspend_equivalent(); 1228 if (2 == Mode) OrderAccess::storeload(); 1229 if (!jSelf->is_external_suspend()) return false; 1230 // We raced a suspension -- fall thru into the slow path 1231 jSelf->set_suspend_equivalent(); 1232 } 1233 return jSelf->handle_special_suspend_equivalent_condition(); 1234 } 1235 1236 1237 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1238 assert(_owner == Self, "invariant"); 1239 1240 // Exit protocol: 1241 // 1. ST _succ = wakee 1242 // 2. membar #loadstore|#storestore; 1243 // 2. ST _owner = NULL 1244 // 3. unpark(wakee) 1245 1246 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1247 ParkEvent * Trigger = Wakee->_event; 1248 1249 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1250 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1251 // out-of-scope (non-extant). 1252 Wakee = NULL; 1253 1254 // Drop the lock 1255 OrderAccess::release_store(&_owner, (void*)NULL); 1256 OrderAccess::fence(); // ST _owner vs LD in unpark() 1257 1258 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1259 Trigger->unpark(); 1260 1261 // Maintain stats and report events to JVMTI 1262 OM_PERFDATA_OP(Parks, inc()); 1263 } 1264 1265 1266 // ----------------------------------------------------------------------------- 1267 // Class Loader deadlock handling. 1268 // 1269 // complete_exit exits a lock returning recursion count 1270 // complete_exit/reenter operate as a wait without waiting 1271 // complete_exit requires an inflated monitor 1272 // The _owner field is not always the Thread addr even with an 1273 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1274 // thread due to contention. 1275 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1276 Thread * const Self = THREAD; 1277 assert(Self->is_Java_thread(), "Must be Java thread!"); 1278 JavaThread *jt = (JavaThread *)THREAD; 1279 1280 DeferredInitialize(); 1281 1282 if (THREAD != _owner) { 1283 if (THREAD->is_lock_owned ((address)_owner)) { 1284 assert(_recursions == 0, "internal state error"); 1285 _owner = THREAD; // Convert from basiclock addr to Thread addr 1286 _recursions = 0; 1287 } 1288 } 1289 1290 guarantee(Self == _owner, "complete_exit not owner"); 1291 intptr_t save = _recursions; // record the old recursion count 1292 _recursions = 0; // set the recursion level to be 0 1293 exit(true, Self); // exit the monitor 1294 guarantee(_owner != Self, "invariant"); 1295 return save; 1296 } 1297 1298 // reenter() enters a lock and sets recursion count 1299 // complete_exit/reenter operate as a wait without waiting 1300 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1301 Thread * const Self = THREAD; 1302 assert(Self->is_Java_thread(), "Must be Java thread!"); 1303 JavaThread *jt = (JavaThread *)THREAD; 1304 1305 guarantee(_owner != Self, "reenter already owner"); 1306 enter(THREAD); // enter the monitor 1307 guarantee(_recursions == 0, "reenter recursion"); 1308 _recursions = recursions; 1309 return; 1310 } 1311 1312 1313 // ----------------------------------------------------------------------------- 1314 // A macro is used below because there may already be a pending 1315 // exception which should not abort the execution of the routines 1316 // which use this (which is why we don't put this into check_slow and 1317 // call it with a CHECK argument). 1318 1319 #define CHECK_OWNER() \ 1320 do { \ 1321 if (THREAD != _owner) { \ 1322 if (THREAD->is_lock_owned((address) _owner)) { \ 1323 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1324 _recursions = 0; \ 1325 } else { \ 1326 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1327 } \ 1328 } \ 1329 } while (false) 1330 1331 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1332 // TODO-FIXME: remove check_slow() -- it's likely dead. 1333 1334 void ObjectMonitor::check_slow(TRAPS) { 1335 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1336 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1337 } 1338 1339 static int Adjust(volatile int * adr, int dx) { 1340 int v; 1341 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1342 return v; 1343 } 1344 1345 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1346 ObjectMonitor* monitor, 1347 jlong notifier_tid, 1348 jlong timeout, 1349 bool timedout) { 1350 assert(event != NULL, "invariant"); 1351 assert(monitor != NULL, "invariant"); 1352 event->set_monitorClass(((oop)monitor->object())->klass()); 1353 event->set_timeout(timeout); 1354 event->set_address((uintptr_t)monitor->object_addr()); 1355 event->set_notifier(notifier_tid); 1356 event->set_timedOut(timedout); 1357 event->commit(); 1358 } 1359 1360 // ----------------------------------------------------------------------------- 1361 // Wait/Notify/NotifyAll 1362 // 1363 // Note: a subset of changes to ObjectMonitor::wait() 1364 // will need to be replicated in complete_exit 1365 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1366 Thread * const Self = THREAD; 1367 assert(Self->is_Java_thread(), "Must be Java thread!"); 1368 JavaThread *jt = (JavaThread *)THREAD; 1369 1370 DeferredInitialize(); 1371 1372 // Throw IMSX or IEX. 1373 CHECK_OWNER(); 1374 1375 EventJavaMonitorWait event; 1376 1377 // check for a pending interrupt 1378 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1379 // post monitor waited event. Note that this is past-tense, we are done waiting. 1380 if (JvmtiExport::should_post_monitor_waited()) { 1381 // Note: 'false' parameter is passed here because the 1382 // wait was not timed out due to thread interrupt. 1383 JvmtiExport::post_monitor_waited(jt, this, false); 1384 1385 // In this short circuit of the monitor wait protocol, the 1386 // current thread never drops ownership of the monitor and 1387 // never gets added to the wait queue so the current thread 1388 // cannot be made the successor. This means that the 1389 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1390 // consume an unpark() meant for the ParkEvent associated with 1391 // this ObjectMonitor. 1392 } 1393 if (event.should_commit()) { 1394 post_monitor_wait_event(&event, this, 0, millis, false); 1395 } 1396 THROW(vmSymbols::java_lang_InterruptedException()); 1397 return; 1398 } 1399 1400 assert(Self->_Stalled == 0, "invariant"); 1401 Self->_Stalled = intptr_t(this); 1402 jt->set_current_waiting_monitor(this); 1403 1404 // create a node to be put into the queue 1405 // Critically, after we reset() the event but prior to park(), we must check 1406 // for a pending interrupt. 1407 ObjectWaiter node(Self); 1408 node.TState = ObjectWaiter::TS_WAIT; 1409 Self->_ParkEvent->reset(); 1410 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1411 1412 // Enter the waiting queue, which is a circular doubly linked list in this case 1413 // but it could be a priority queue or any data structure. 1414 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1415 // by the the owner of the monitor *except* in the case where park() 1416 // returns because of a timeout of interrupt. Contention is exceptionally rare 1417 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1418 1419 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1420 AddWaiter(&node); 1421 Thread::SpinRelease(&_WaitSetLock); 1422 1423 _Responsible = NULL; 1424 1425 intptr_t save = _recursions; // record the old recursion count 1426 _waiters++; // increment the number of waiters 1427 _recursions = 0; // set the recursion level to be 1 1428 exit(true, Self); // exit the monitor 1429 guarantee(_owner != Self, "invariant"); 1430 1431 // The thread is on the WaitSet list - now park() it. 1432 // On MP systems it's conceivable that a brief spin before we park 1433 // could be profitable. 1434 // 1435 // TODO-FIXME: change the following logic to a loop of the form 1436 // while (!timeout && !interrupted && _notified == 0) park() 1437 1438 int ret = OS_OK; 1439 int WasNotified = 0; 1440 { // State transition wrappers 1441 OSThread* osthread = Self->osthread(); 1442 OSThreadWaitState osts(osthread, true); 1443 { 1444 ThreadBlockInVM tbivm(jt); 1445 // Thread is in thread_blocked state and oop access is unsafe. 1446 jt->set_suspend_equivalent(); 1447 1448 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1449 // Intentionally empty 1450 } else if (node._notified == 0) { 1451 if (millis <= 0) { 1452 Self->_ParkEvent->park(); 1453 } else { 1454 ret = Self->_ParkEvent->park(millis); 1455 } 1456 } 1457 1458 // were we externally suspended while we were waiting? 1459 if (ExitSuspendEquivalent (jt)) { 1460 // TODO-FIXME: add -- if succ == Self then succ = null. 1461 jt->java_suspend_self(); 1462 } 1463 1464 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1465 1466 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1467 // from the WaitSet to the EntryList. 1468 // See if we need to remove Node from the WaitSet. 1469 // We use double-checked locking to avoid grabbing _WaitSetLock 1470 // if the thread is not on the wait queue. 1471 // 1472 // Note that we don't need a fence before the fetch of TState. 1473 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1474 // written by the is thread. (perhaps the fetch might even be satisfied 1475 // by a look-aside into the processor's own store buffer, although given 1476 // the length of the code path between the prior ST and this load that's 1477 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1478 // then we'll acquire the lock and then re-fetch a fresh TState value. 1479 // That is, we fail toward safety. 1480 1481 if (node.TState == ObjectWaiter::TS_WAIT) { 1482 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1483 if (node.TState == ObjectWaiter::TS_WAIT) { 1484 DequeueSpecificWaiter(&node); // unlink from WaitSet 1485 assert(node._notified == 0, "invariant"); 1486 node.TState = ObjectWaiter::TS_RUN; 1487 } 1488 Thread::SpinRelease(&_WaitSetLock); 1489 } 1490 1491 // The thread is now either on off-list (TS_RUN), 1492 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1493 // The Node's TState variable is stable from the perspective of this thread. 1494 // No other threads will asynchronously modify TState. 1495 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1496 OrderAccess::loadload(); 1497 if (_succ == Self) _succ = NULL; 1498 WasNotified = node._notified; 1499 1500 // Reentry phase -- reacquire the monitor. 1501 // re-enter contended monitor after object.wait(). 1502 // retain OBJECT_WAIT state until re-enter successfully completes 1503 // Thread state is thread_in_vm and oop access is again safe, 1504 // although the raw address of the object may have changed. 1505 // (Don't cache naked oops over safepoints, of course). 1506 1507 // post monitor waited event. Note that this is past-tense, we are done waiting. 1508 if (JvmtiExport::should_post_monitor_waited()) { 1509 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1510 1511 if (node._notified != 0 && _succ == Self) { 1512 // In this part of the monitor wait-notify-reenter protocol it 1513 // is possible (and normal) for another thread to do a fastpath 1514 // monitor enter-exit while this thread is still trying to get 1515 // to the reenter portion of the protocol. 1516 // 1517 // The ObjectMonitor was notified and the current thread is 1518 // the successor which also means that an unpark() has already 1519 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1520 // consume the unpark() that was done when the successor was 1521 // set because the same ParkEvent is shared between Java 1522 // monitors and JVM/TI RawMonitors (for now). 1523 // 1524 // We redo the unpark() to ensure forward progress, i.e., we 1525 // don't want all pending threads hanging (parked) with none 1526 // entering the unlocked monitor. 1527 node._event->unpark(); 1528 } 1529 } 1530 1531 if (event.should_commit()) { 1532 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1533 } 1534 1535 OrderAccess::fence(); 1536 1537 assert(Self->_Stalled != 0, "invariant"); 1538 Self->_Stalled = 0; 1539 1540 assert(_owner != Self, "invariant"); 1541 ObjectWaiter::TStates v = node.TState; 1542 if (v == ObjectWaiter::TS_RUN) { 1543 enter(Self); 1544 } else { 1545 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1546 ReenterI(Self, &node); 1547 node.wait_reenter_end(this); 1548 } 1549 1550 // Self has reacquired the lock. 1551 // Lifecycle - the node representing Self must not appear on any queues. 1552 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1553 // want residual elements associated with this thread left on any lists. 1554 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1555 assert(_owner == Self, "invariant"); 1556 assert(_succ != Self, "invariant"); 1557 } // OSThreadWaitState() 1558 1559 jt->set_current_waiting_monitor(NULL); 1560 1561 guarantee(_recursions == 0, "invariant"); 1562 _recursions = save; // restore the old recursion count 1563 _waiters--; // decrement the number of waiters 1564 1565 // Verify a few postconditions 1566 assert(_owner == Self, "invariant"); 1567 assert(_succ != Self, "invariant"); 1568 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1569 1570 // check if the notification happened 1571 if (!WasNotified) { 1572 // no, it could be timeout or Thread.interrupt() or both 1573 // check for interrupt event, otherwise it is timeout 1574 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1575 THROW(vmSymbols::java_lang_InterruptedException()); 1576 } 1577 } 1578 1579 // NOTE: Spurious wake up will be consider as timeout. 1580 // Monitor notify has precedence over thread interrupt. 1581 } 1582 1583 1584 // Consider: 1585 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1586 // then instead of transferring a thread from the WaitSet to the EntryList 1587 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1588 1589 void ObjectMonitor::INotify(Thread * Self) { 1590 const int policy = Knob_MoveNotifyee; 1591 1592 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1593 ObjectWaiter * iterator = DequeueWaiter(); 1594 if (iterator != NULL) { 1595 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1596 guarantee(iterator->_notified == 0, "invariant"); 1597 // Disposition - what might we do with iterator ? 1598 // a. add it directly to the EntryList - either tail (policy == 1) 1599 // or head (policy == 0). 1600 // b. push it onto the front of the _cxq (policy == 2). 1601 // For now we use (b). 1602 if (policy != 4) { 1603 iterator->TState = ObjectWaiter::TS_ENTER; 1604 } 1605 iterator->_notified = 1; 1606 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1607 1608 ObjectWaiter * list = _EntryList; 1609 if (list != NULL) { 1610 assert(list->_prev == NULL, "invariant"); 1611 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1612 assert(list != iterator, "invariant"); 1613 } 1614 1615 if (policy == 0) { // prepend to EntryList 1616 if (list == NULL) { 1617 iterator->_next = iterator->_prev = NULL; 1618 _EntryList = iterator; 1619 } else { 1620 list->_prev = iterator; 1621 iterator->_next = list; 1622 iterator->_prev = NULL; 1623 _EntryList = iterator; 1624 } 1625 } else if (policy == 1) { // append to EntryList 1626 if (list == NULL) { 1627 iterator->_next = iterator->_prev = NULL; 1628 _EntryList = iterator; 1629 } else { 1630 // CONSIDER: finding the tail currently requires a linear-time walk of 1631 // the EntryList. We can make tail access constant-time by converting to 1632 // a CDLL instead of using our current DLL. 1633 ObjectWaiter * tail; 1634 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1635 assert(tail != NULL && tail->_next == NULL, "invariant"); 1636 tail->_next = iterator; 1637 iterator->_prev = tail; 1638 iterator->_next = NULL; 1639 } 1640 } else if (policy == 2) { // prepend to cxq 1641 if (list == NULL) { 1642 iterator->_next = iterator->_prev = NULL; 1643 _EntryList = iterator; 1644 } else { 1645 iterator->TState = ObjectWaiter::TS_CXQ; 1646 for (;;) { 1647 ObjectWaiter * front = _cxq; 1648 iterator->_next = front; 1649 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1650 break; 1651 } 1652 } 1653 } 1654 } else if (policy == 3) { // append to cxq 1655 iterator->TState = ObjectWaiter::TS_CXQ; 1656 for (;;) { 1657 ObjectWaiter * tail = _cxq; 1658 if (tail == NULL) { 1659 iterator->_next = NULL; 1660 if (Atomic::replace_if_null(iterator, &_cxq)) { 1661 break; 1662 } 1663 } else { 1664 while (tail->_next != NULL) tail = tail->_next; 1665 tail->_next = iterator; 1666 iterator->_prev = tail; 1667 iterator->_next = NULL; 1668 break; 1669 } 1670 } 1671 } else { 1672 ParkEvent * ev = iterator->_event; 1673 iterator->TState = ObjectWaiter::TS_RUN; 1674 OrderAccess::fence(); 1675 ev->unpark(); 1676 } 1677 1678 // _WaitSetLock protects the wait queue, not the EntryList. We could 1679 // move the add-to-EntryList operation, above, outside the critical section 1680 // protected by _WaitSetLock. In practice that's not useful. With the 1681 // exception of wait() timeouts and interrupts the monitor owner 1682 // is the only thread that grabs _WaitSetLock. There's almost no contention 1683 // on _WaitSetLock so it's not profitable to reduce the length of the 1684 // critical section. 1685 1686 if (policy < 4) { 1687 iterator->wait_reenter_begin(this); 1688 } 1689 } 1690 Thread::SpinRelease(&_WaitSetLock); 1691 } 1692 1693 // Consider: a not-uncommon synchronization bug is to use notify() when 1694 // notifyAll() is more appropriate, potentially resulting in stranded 1695 // threads; this is one example of a lost wakeup. A useful diagnostic 1696 // option is to force all notify() operations to behave as notifyAll(). 1697 // 1698 // Note: We can also detect many such problems with a "minimum wait". 1699 // When the "minimum wait" is set to a small non-zero timeout value 1700 // and the program does not hang whereas it did absent "minimum wait", 1701 // that suggests a lost wakeup bug. 1702 1703 void ObjectMonitor::notify(TRAPS) { 1704 CHECK_OWNER(); 1705 if (_WaitSet == NULL) { 1706 return; 1707 } 1708 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1709 INotify(THREAD); 1710 OM_PERFDATA_OP(Notifications, inc(1)); 1711 } 1712 1713 1714 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1715 // from the waitset to the EntryList. This could be done more efficiently with a 1716 // single bulk transfer but in practice it's not time-critical. Beware too, 1717 // that in prepend-mode we invert the order of the waiters. Let's say that the 1718 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1719 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1720 1721 void ObjectMonitor::notifyAll(TRAPS) { 1722 CHECK_OWNER(); 1723 if (_WaitSet == NULL) { 1724 return; 1725 } 1726 1727 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1728 int tally = 0; 1729 while (_WaitSet != NULL) { 1730 tally++; 1731 INotify(THREAD); 1732 } 1733 1734 OM_PERFDATA_OP(Notifications, inc(tally)); 1735 } 1736 1737 // ----------------------------------------------------------------------------- 1738 // Adaptive Spinning Support 1739 // 1740 // Adaptive spin-then-block - rational spinning 1741 // 1742 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1743 // algorithm. On high order SMP systems it would be better to start with 1744 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1745 // a contending thread could enqueue itself on the cxq and then spin locally 1746 // on a thread-specific variable such as its ParkEvent._Event flag. 1747 // That's left as an exercise for the reader. Note that global spinning is 1748 // not problematic on Niagara, as the L2 cache serves the interconnect and 1749 // has both low latency and massive bandwidth. 1750 // 1751 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1752 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1753 // (duration) or we can fix the count at approximately the duration of 1754 // a context switch and vary the frequency. Of course we could also 1755 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1756 // For a description of 'Adaptive spin-then-block mutual exclusion in 1757 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1758 // 1759 // This implementation varies the duration "D", where D varies with 1760 // the success rate of recent spin attempts. (D is capped at approximately 1761 // length of a round-trip context switch). The success rate for recent 1762 // spin attempts is a good predictor of the success rate of future spin 1763 // attempts. The mechanism adapts automatically to varying critical 1764 // section length (lock modality), system load and degree of parallelism. 1765 // D is maintained per-monitor in _SpinDuration and is initialized 1766 // optimistically. Spin frequency is fixed at 100%. 1767 // 1768 // Note that _SpinDuration is volatile, but we update it without locks 1769 // or atomics. The code is designed so that _SpinDuration stays within 1770 // a reasonable range even in the presence of races. The arithmetic 1771 // operations on _SpinDuration are closed over the domain of legal values, 1772 // so at worst a race will install and older but still legal value. 1773 // At the very worst this introduces some apparent non-determinism. 1774 // We might spin when we shouldn't or vice-versa, but since the spin 1775 // count are relatively short, even in the worst case, the effect is harmless. 1776 // 1777 // Care must be taken that a low "D" value does not become an 1778 // an absorbing state. Transient spinning failures -- when spinning 1779 // is overall profitable -- should not cause the system to converge 1780 // on low "D" values. We want spinning to be stable and predictable 1781 // and fairly responsive to change and at the same time we don't want 1782 // it to oscillate, become metastable, be "too" non-deterministic, 1783 // or converge on or enter undesirable stable absorbing states. 1784 // 1785 // We implement a feedback-based control system -- using past behavior 1786 // to predict future behavior. We face two issues: (a) if the 1787 // input signal is random then the spin predictor won't provide optimal 1788 // results, and (b) if the signal frequency is too high then the control 1789 // system, which has some natural response lag, will "chase" the signal. 1790 // (b) can arise from multimodal lock hold times. Transient preemption 1791 // can also result in apparent bimodal lock hold times. 1792 // Although sub-optimal, neither condition is particularly harmful, as 1793 // in the worst-case we'll spin when we shouldn't or vice-versa. 1794 // The maximum spin duration is rather short so the failure modes aren't bad. 1795 // To be conservative, I've tuned the gain in system to bias toward 1796 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1797 // "rings" or oscillates between spinning and not spinning. This happens 1798 // when spinning is just on the cusp of profitability, however, so the 1799 // situation is not dire. The state is benign -- there's no need to add 1800 // hysteresis control to damp the transition rate between spinning and 1801 // not spinning. 1802 1803 // Spinning: Fixed frequency (100%), vary duration 1804 int ObjectMonitor::TrySpin(Thread * Self) { 1805 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1806 int ctr = Knob_FixedSpin; 1807 if (ctr != 0) { 1808 while (--ctr >= 0) { 1809 if (TryLock(Self) > 0) return 1; 1810 SpinPause(); 1811 } 1812 return 0; 1813 } 1814 1815 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1816 if (TryLock(Self) > 0) { 1817 // Increase _SpinDuration ... 1818 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1819 // Raising _SpurDuration to the poverty line is key. 1820 int x = _SpinDuration; 1821 if (x < Knob_SpinLimit) { 1822 if (x < Knob_Poverty) x = Knob_Poverty; 1823 _SpinDuration = x + Knob_BonusB; 1824 } 1825 return 1; 1826 } 1827 SpinPause(); 1828 } 1829 1830 // Admission control - verify preconditions for spinning 1831 // 1832 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1833 // becoming an absorbing state. Put another way, we spin briefly to 1834 // sample, just in case the system load, parallelism, contention, or lock 1835 // modality changed. 1836 // 1837 // Consider the following alternative: 1838 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1839 // spin attempt. "Periodically" might mean after a tally of 1840 // the # of failed spin attempts (or iterations) reaches some threshold. 1841 // This takes us into the realm of 1-out-of-N spinning, where we 1842 // hold the duration constant but vary the frequency. 1843 1844 ctr = _SpinDuration; 1845 if (ctr <= 0) return 0; 1846 1847 if (Knob_SuccRestrict && _succ != NULL) return 0; 1848 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1849 return 0; 1850 } 1851 1852 int MaxSpin = Knob_MaxSpinners; 1853 if (MaxSpin >= 0) { 1854 if (_Spinner > MaxSpin) { 1855 return 0; 1856 } 1857 // Slightly racy, but benign ... 1858 Adjust(&_Spinner, 1); 1859 } 1860 1861 // We're good to spin ... spin ingress. 1862 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1863 // when preparing to LD...CAS _owner, etc and the CAS is likely 1864 // to succeed. 1865 if (_succ == NULL) { 1866 _succ = Self; 1867 } 1868 Thread * prv = NULL; 1869 1870 // There are three ways to exit the following loop: 1871 // 1. A successful spin where this thread has acquired the lock. 1872 // 2. Spin failure with prejudice 1873 // 3. Spin failure without prejudice 1874 1875 while (--ctr >= 0) { 1876 1877 // Periodic polling -- Check for pending GC 1878 // Threads may spin while they're unsafe. 1879 // We don't want spinning threads to delay the JVM from reaching 1880 // a stop-the-world safepoint or to steal cycles from GC. 1881 // If we detect a pending safepoint we abort in order that 1882 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1883 // this thread, if safe, doesn't steal cycles from GC. 1884 // This is in keeping with the "no loitering in runtime" rule. 1885 // We periodically check to see if there's a safepoint pending. 1886 if ((ctr & 0xFF) == 0) { 1887 if (SafepointMechanism::poll(Self)) { 1888 goto Abort; // abrupt spin egress 1889 } 1890 if (Knob_UsePause & 1) SpinPause(); 1891 } 1892 1893 if (Knob_UsePause & 2) SpinPause(); 1894 1895 // Probe _owner with TATAS 1896 // If this thread observes the monitor transition or flicker 1897 // from locked to unlocked to locked, then the odds that this 1898 // thread will acquire the lock in this spin attempt go down 1899 // considerably. The same argument applies if the CAS fails 1900 // or if we observe _owner change from one non-null value to 1901 // another non-null value. In such cases we might abort 1902 // the spin without prejudice or apply a "penalty" to the 1903 // spin count-down variable "ctr", reducing it by 100, say. 1904 1905 Thread * ox = (Thread *) _owner; 1906 if (ox == NULL) { 1907 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1908 if (ox == NULL) { 1909 // The CAS succeeded -- this thread acquired ownership 1910 // Take care of some bookkeeping to exit spin state. 1911 if (_succ == Self) { 1912 _succ = NULL; 1913 } 1914 if (MaxSpin > 0) Adjust(&_Spinner, -1); 1915 1916 // Increase _SpinDuration : 1917 // The spin was successful (profitable) so we tend toward 1918 // longer spin attempts in the future. 1919 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1920 // If we acquired the lock early in the spin cycle it 1921 // makes sense to increase _SpinDuration proportionally. 1922 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1923 int x = _SpinDuration; 1924 if (x < Knob_SpinLimit) { 1925 if (x < Knob_Poverty) x = Knob_Poverty; 1926 _SpinDuration = x + Knob_Bonus; 1927 } 1928 return 1; 1929 } 1930 1931 // The CAS failed ... we can take any of the following actions: 1932 // * penalize: ctr -= CASPenalty 1933 // * exit spin with prejudice -- goto Abort; 1934 // * exit spin without prejudice. 1935 // * Since CAS is high-latency, retry again immediately. 1936 prv = ox; 1937 goto Abort; 1938 } 1939 1940 // Did lock ownership change hands ? 1941 if (ox != prv && prv != NULL) { 1942 goto Abort; 1943 } 1944 prv = ox; 1945 1946 // Abort the spin if the owner is not executing. 1947 // The owner must be executing in order to drop the lock. 1948 // Spinning while the owner is OFFPROC is idiocy. 1949 // Consider: ctr -= RunnablePenalty ; 1950 if (Knob_OState && NotRunnable (Self, ox)) { 1951 goto Abort; 1952 } 1953 if (_succ == NULL) { 1954 _succ = Self; 1955 } 1956 } 1957 1958 // Spin failed with prejudice -- reduce _SpinDuration. 1959 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1960 // AIMD is globally stable. 1961 { 1962 int x = _SpinDuration; 1963 if (x > 0) { 1964 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1965 // This is globally sample and tends to damp the response. 1966 x -= Knob_Penalty; 1967 if (x < 0) x = 0; 1968 _SpinDuration = x; 1969 } 1970 } 1971 1972 Abort: 1973 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 1974 if (_succ == Self) { 1975 _succ = NULL; 1976 // Invariant: after setting succ=null a contending thread 1977 // must recheck-retry _owner before parking. This usually happens 1978 // in the normal usage of TrySpin(), but it's safest 1979 // to make TrySpin() as foolproof as possible. 1980 OrderAccess::fence(); 1981 if (TryLock(Self) > 0) return 1; 1982 } 1983 return 0; 1984 } 1985 1986 // NotRunnable() -- informed spinning 1987 // 1988 // Don't bother spinning if the owner is not eligible to drop the lock. 1989 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1990 // The thread must be runnable in order to drop the lock in timely fashion. 1991 // If the _owner is not runnable then spinning will not likely be 1992 // successful (profitable). 1993 // 1994 // Beware -- the thread referenced by _owner could have died 1995 // so a simply fetch from _owner->_thread_state might trap. 1996 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1997 // Because of the lifecycle issues, the _thread_state values 1998 // observed by NotRunnable() might be garbage. NotRunnable must 1999 // tolerate this and consider the observed _thread_state value 2000 // as advisory. 2001 // 2002 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2003 // a thread pointer. 2004 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2005 // with the LSB of _owner. Another option would be to probabilistically probe 2006 // the putative _owner->TypeTag value. 2007 // 2008 // Checking _thread_state isn't perfect. Even if the thread is 2009 // in_java it might be blocked on a page-fault or have been preempted 2010 // and sitting on a ready/dispatch queue. 2011 // 2012 // The return value from NotRunnable() is *advisory* -- the 2013 // result is based on sampling and is not necessarily coherent. 2014 // The caller must tolerate false-negative and false-positive errors. 2015 // Spinning, in general, is probabilistic anyway. 2016 2017 2018 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2019 // Check ox->TypeTag == 2BAD. 2020 if (ox == NULL) return 0; 2021 2022 // Avoid transitive spinning ... 2023 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2024 // Immediately after T1 acquires L it's possible that T2, also 2025 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2026 // This occurs transiently after T1 acquired L but before 2027 // T1 managed to clear T1.Stalled. T2 does not need to abort 2028 // its spin in this circumstance. 2029 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2030 2031 if (BlockedOn == 1) return 1; 2032 if (BlockedOn != 0) { 2033 return BlockedOn != intptr_t(this) && _owner == ox; 2034 } 2035 2036 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2037 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2038 // consider also: jst != _thread_in_Java -- but that's overspecific. 2039 return jst == _thread_blocked || jst == _thread_in_native; 2040 } 2041 2042 2043 // ----------------------------------------------------------------------------- 2044 // WaitSet management ... 2045 2046 ObjectWaiter::ObjectWaiter(Thread* thread) { 2047 _next = NULL; 2048 _prev = NULL; 2049 _notified = 0; 2050 _notifier_tid = 0; 2051 TState = TS_RUN; 2052 _thread = thread; 2053 _event = thread->_ParkEvent; 2054 _active = false; 2055 assert(_event != NULL, "invariant"); 2056 } 2057 2058 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2059 JavaThread *jt = (JavaThread *)this->_thread; 2060 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2061 } 2062 2063 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2064 JavaThread *jt = (JavaThread *)this->_thread; 2065 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2066 } 2067 2068 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2069 assert(node != NULL, "should not add NULL node"); 2070 assert(node->_prev == NULL, "node already in list"); 2071 assert(node->_next == NULL, "node already in list"); 2072 // put node at end of queue (circular doubly linked list) 2073 if (_WaitSet == NULL) { 2074 _WaitSet = node; 2075 node->_prev = node; 2076 node->_next = node; 2077 } else { 2078 ObjectWaiter* head = _WaitSet; 2079 ObjectWaiter* tail = head->_prev; 2080 assert(tail->_next == head, "invariant check"); 2081 tail->_next = node; 2082 head->_prev = node; 2083 node->_next = head; 2084 node->_prev = tail; 2085 } 2086 } 2087 2088 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2089 // dequeue the very first waiter 2090 ObjectWaiter* waiter = _WaitSet; 2091 if (waiter) { 2092 DequeueSpecificWaiter(waiter); 2093 } 2094 return waiter; 2095 } 2096 2097 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2098 assert(node != NULL, "should not dequeue NULL node"); 2099 assert(node->_prev != NULL, "node already removed from list"); 2100 assert(node->_next != NULL, "node already removed from list"); 2101 // when the waiter has woken up because of interrupt, 2102 // timeout or other spurious wake-up, dequeue the 2103 // waiter from waiting list 2104 ObjectWaiter* next = node->_next; 2105 if (next == node) { 2106 assert(node->_prev == node, "invariant check"); 2107 _WaitSet = NULL; 2108 } else { 2109 ObjectWaiter* prev = node->_prev; 2110 assert(prev->_next == node, "invariant check"); 2111 assert(next->_prev == node, "invariant check"); 2112 next->_prev = prev; 2113 prev->_next = next; 2114 if (_WaitSet == node) { 2115 _WaitSet = next; 2116 } 2117 } 2118 node->_next = NULL; 2119 node->_prev = NULL; 2120 } 2121 2122 // ----------------------------------------------------------------------------- 2123 // PerfData support 2124 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2125 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2126 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2127 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2128 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2129 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2130 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2131 2132 // One-shot global initialization for the sync subsystem. 2133 // We could also defer initialization and initialize on-demand 2134 // the first time we call inflate(). Initialization would 2135 // be protected - like so many things - by the MonitorCache_lock. 2136 2137 void ObjectMonitor::Initialize() { 2138 static int InitializationCompleted = 0; 2139 assert(InitializationCompleted == 0, "invariant"); 2140 InitializationCompleted = 1; 2141 if (UsePerfData) { 2142 EXCEPTION_MARK; 2143 #define NEWPERFCOUNTER(n) \ 2144 { \ 2145 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2146 CHECK); \ 2147 } 2148 #define NEWPERFVARIABLE(n) \ 2149 { \ 2150 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2151 CHECK); \ 2152 } 2153 NEWPERFCOUNTER(_sync_Inflations); 2154 NEWPERFCOUNTER(_sync_Deflations); 2155 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2156 NEWPERFCOUNTER(_sync_FutileWakeups); 2157 NEWPERFCOUNTER(_sync_Parks); 2158 NEWPERFCOUNTER(_sync_Notifications); 2159 NEWPERFVARIABLE(_sync_MonExtant); 2160 #undef NEWPERFCOUNTER 2161 #undef NEWPERFVARIABLE 2162 } 2163 } 2164 2165 void ObjectMonitor::DeferredInitialize() { 2166 if (InitDone > 0) return; 2167 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2168 while (InitDone != 1) /* empty */; 2169 return; 2170 } 2171 2172 // One-shot global initialization ... 2173 // The initialization is idempotent, so we don't need locks. 2174 // In the future consider doing this via os::init_2(). 2175 2176 if (!os::is_MP()) { 2177 Knob_SpinLimit = 0; 2178 Knob_PreSpin = 0; 2179 Knob_FixedSpin = -1; 2180 } 2181 2182 OrderAccess::fence(); 2183 InitDone = 1; 2184 } 2185