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_CASPenalty = -1; // Penalty for failed CAS 107 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 108 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 109 static int Knob_SpinEarly = 1; 110 static int Knob_SuccEnabled = 1; // futile wake throttling 111 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 112 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 113 static int Knob_Bonus = 100; // spin success bonus 114 static int Knob_BonusB = 100; // spin success bonus 115 static int Knob_Penalty = 200; // spin failure penalty 116 static int Knob_Poverty = 1000; 117 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 118 static int Knob_FixedSpin = 0; 119 static int Knob_OState = 3; // Spinner checks thread state of _owner 120 static int Knob_UsePause = 1; 121 static int Knob_ExitPolicy = 0; 122 static int Knob_PreSpin = 10; // 20-100 likely better 123 static int Knob_ResetEvent = 0; 124 125 static int Knob_FastHSSEC = 0; 126 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 127 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 128 static volatile int InitDone = 0; 129 130 // ----------------------------------------------------------------------------- 131 // Theory of operations -- Monitors lists, thread residency, etc: 132 // 133 // * A thread acquires ownership of a monitor by successfully 134 // CAS()ing the _owner field from null to non-null. 135 // 136 // * Invariant: A thread appears on at most one monitor list -- 137 // cxq, EntryList or WaitSet -- at any one time. 138 // 139 // * Contending threads "push" themselves onto the cxq with CAS 140 // and then spin/park. 141 // 142 // * After a contending thread eventually acquires the lock it must 143 // dequeue itself from either the EntryList or the cxq. 144 // 145 // * The exiting thread identifies and unparks an "heir presumptive" 146 // tentative successor thread on the EntryList. Critically, the 147 // exiting thread doesn't unlink the successor thread from the EntryList. 148 // After having been unparked, the wakee will recontend for ownership of 149 // the monitor. The successor (wakee) will either acquire the lock or 150 // re-park itself. 151 // 152 // Succession is provided for by a policy of competitive handoff. 153 // The exiting thread does _not_ grant or pass ownership to the 154 // successor thread. (This is also referred to as "handoff" succession"). 155 // Instead the exiting thread releases ownership and possibly wakes 156 // a successor, so the successor can (re)compete for ownership of the lock. 157 // If the EntryList is empty but the cxq is populated the exiting 158 // thread will drain the cxq into the EntryList. It does so by 159 // by detaching the cxq (installing null with CAS) and folding 160 // the threads from the cxq into the EntryList. The EntryList is 161 // doubly linked, while the cxq is singly linked because of the 162 // CAS-based "push" used to enqueue recently arrived threads (RATs). 163 // 164 // * Concurrency invariants: 165 // 166 // -- only the monitor owner may access or mutate the EntryList. 167 // The mutex property of the monitor itself protects the EntryList 168 // from concurrent interference. 169 // -- Only the monitor owner may detach the cxq. 170 // 171 // * The monitor entry list operations avoid locks, but strictly speaking 172 // they're not lock-free. Enter is lock-free, exit is not. 173 // For a description of 'Methods and apparatus providing non-blocking access 174 // to a resource,' see U.S. Pat. No. 7844973. 175 // 176 // * The cxq can have multiple concurrent "pushers" but only one concurrent 177 // detaching thread. This mechanism is immune from the ABA corruption. 178 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 179 // 180 // * Taken together, the cxq and the EntryList constitute or form a 181 // single logical queue of threads stalled trying to acquire the lock. 182 // We use two distinct lists to improve the odds of a constant-time 183 // dequeue operation after acquisition (in the ::enter() epilogue) and 184 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 185 // A key desideratum is to minimize queue & monitor metadata manipulation 186 // that occurs while holding the monitor lock -- that is, we want to 187 // minimize monitor lock holds times. Note that even a small amount of 188 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 189 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 190 // locks and monitor metadata. 191 // 192 // Cxq points to the set of Recently Arrived Threads attempting entry. 193 // Because we push threads onto _cxq with CAS, the RATs must take the form of 194 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 195 // the unlocking thread notices that EntryList is null but _cxq is != null. 196 // 197 // The EntryList is ordered by the prevailing queue discipline and 198 // can be organized in any convenient fashion, such as a doubly-linked list or 199 // a circular doubly-linked list. Critically, we want insert and delete operations 200 // to operate in constant-time. If we need a priority queue then something akin 201 // to Solaris' sleepq would work nicely. Viz., 202 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 203 // Queue discipline is enforced at ::exit() time, when the unlocking thread 204 // drains the cxq into the EntryList, and orders or reorders the threads on the 205 // EntryList accordingly. 206 // 207 // Barring "lock barging", this mechanism provides fair cyclic ordering, 208 // somewhat similar to an elevator-scan. 209 // 210 // * The monitor synchronization subsystem avoids the use of native 211 // synchronization primitives except for the narrow platform-specific 212 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 213 // the semantics of park-unpark. Put another way, this monitor implementation 214 // depends only on atomic operations and park-unpark. The monitor subsystem 215 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 216 // underlying OS manages the READY<->RUN transitions. 217 // 218 // * Waiting threads reside on the WaitSet list -- wait() puts 219 // the caller onto the WaitSet. 220 // 221 // * notify() or notifyAll() simply transfers threads from the WaitSet to 222 // either the EntryList or cxq. Subsequent exit() operations will 223 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 224 // it's likely the notifyee would simply impale itself on the lock held 225 // by the notifier. 226 // 227 // * An interesting alternative is to encode cxq as (List,LockByte) where 228 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 229 // variable, like _recursions, in the scheme. The threads or Events that form 230 // the list would have to be aligned in 256-byte addresses. A thread would 231 // try to acquire the lock or enqueue itself with CAS, but exiting threads 232 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 233 // Note that is is *not* word-tearing, but it does presume that full-word 234 // CAS operations are coherent with intermix with STB operations. That's true 235 // on most common processors. 236 // 237 // * See also http://blogs.sun.com/dave 238 239 240 void* ObjectMonitor::operator new (size_t size) throw() { 241 return AllocateHeap(size, mtInternal); 242 } 243 void* ObjectMonitor::operator new[] (size_t size) throw() { 244 return operator new (size); 245 } 246 void ObjectMonitor::operator delete(void* p) { 247 FreeHeap(p); 248 } 249 void ObjectMonitor::operator delete[] (void *p) { 250 operator delete(p); 251 } 252 253 // ----------------------------------------------------------------------------- 254 // Enter support 255 256 void ObjectMonitor::enter(TRAPS) { 257 // The following code is ordered to check the most common cases first 258 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 259 Thread * const Self = THREAD; 260 261 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 262 if (cur == NULL) { 263 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 264 assert(_recursions == 0, "invariant"); 265 assert(_owner == Self, "invariant"); 266 return; 267 } 268 269 if (cur == Self) { 270 // TODO-FIXME: check for integer overflow! BUGID 6557169. 271 _recursions++; 272 return; 273 } 274 275 if (Self->is_lock_owned ((address)cur)) { 276 assert(_recursions == 0, "internal state error"); 277 _recursions = 1; 278 // Commute owner from a thread-specific on-stack BasicLockObject address to 279 // a full-fledged "Thread *". 280 _owner = Self; 281 return; 282 } 283 284 // We've encountered genuine contention. 285 assert(Self->_Stalled == 0, "invariant"); 286 Self->_Stalled = intptr_t(this); 287 288 // Try one round of spinning *before* enqueueing Self 289 // and before going through the awkward and expensive state 290 // transitions. The following spin is strictly optional ... 291 // Note that if we acquire the monitor from an initial spin 292 // we forgo posting JVMTI events and firing DTRACE probes. 293 if (Knob_SpinEarly && TrySpin (Self) > 0) { 294 assert(_owner == Self, "invariant"); 295 assert(_recursions == 0, "invariant"); 296 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 297 Self->_Stalled = 0; 298 return; 299 } 300 301 assert(_owner != Self, "invariant"); 302 assert(_succ != Self, "invariant"); 303 assert(Self->is_Java_thread(), "invariant"); 304 JavaThread * jt = (JavaThread *) Self; 305 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 306 assert(jt->thread_state() != _thread_blocked, "invariant"); 307 assert(this->object() != NULL, "invariant"); 308 assert(_count >= 0, "invariant"); 309 310 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 311 // Ensure the object-monitor relationship remains stable while there's contention. 312 Atomic::inc(&_count); 313 314 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 315 EventJavaMonitorEnter event; 316 if (event.should_commit()) { 317 event.set_monitorClass(((oop)this->object())->klass()); 318 event.set_address((uintptr_t)(this->object_addr())); 319 } 320 321 { // Change java thread status to indicate blocked on monitor enter. 322 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 323 324 Self->set_current_pending_monitor(this); 325 326 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 327 if (JvmtiExport::should_post_monitor_contended_enter()) { 328 JvmtiExport::post_monitor_contended_enter(jt, this); 329 330 // The current thread does not yet own the monitor and does not 331 // yet appear on any queues that would get it made the successor. 332 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 333 // handler cannot accidentally consume an unpark() meant for the 334 // ParkEvent associated with this ObjectMonitor. 335 } 336 337 OSThreadContendState osts(Self->osthread()); 338 ThreadBlockInVM tbivm(jt); 339 340 // TODO-FIXME: change the following for(;;) loop to straight-line code. 341 for (;;) { 342 jt->set_suspend_equivalent(); 343 // cleared by handle_special_suspend_equivalent_condition() 344 // or java_suspend_self() 345 346 EnterI(THREAD); 347 348 if (!ExitSuspendEquivalent(jt)) break; 349 350 // We have acquired the contended monitor, but while we were 351 // waiting another thread suspended us. We don't want to enter 352 // the monitor while suspended because that would surprise the 353 // thread that suspended us. 354 // 355 _recursions = 0; 356 _succ = NULL; 357 exit(false, Self); 358 359 jt->java_suspend_self(); 360 } 361 Self->set_current_pending_monitor(NULL); 362 363 // We cleared the pending monitor info since we've just gotten past 364 // the enter-check-for-suspend dance and we now own the monitor free 365 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 366 // destructor can go to a safepoint at the end of this block. If we 367 // do a thread dump during that safepoint, then this thread will show 368 // as having "-locked" the monitor, but the OS and java.lang.Thread 369 // states will still report that the thread is blocked trying to 370 // acquire it. 371 } 372 373 Atomic::dec(&_count); 374 assert(_count >= 0, "invariant"); 375 Self->_Stalled = 0; 376 377 // Must either set _recursions = 0 or ASSERT _recursions == 0. 378 assert(_recursions == 0, "invariant"); 379 assert(_owner == Self, "invariant"); 380 assert(_succ != Self, "invariant"); 381 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 382 383 // The thread -- now the owner -- is back in vm mode. 384 // Report the glorious news via TI,DTrace and jvmstat. 385 // The probe effect is non-trivial. All the reportage occurs 386 // while we hold the monitor, increasing the length of the critical 387 // section. Amdahl's parallel speedup law comes vividly into play. 388 // 389 // Another option might be to aggregate the events (thread local or 390 // per-monitor aggregation) and defer reporting until a more opportune 391 // time -- such as next time some thread encounters contention but has 392 // yet to acquire the lock. While spinning that thread could 393 // spinning we could increment JVMStat counters, etc. 394 395 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 396 if (JvmtiExport::should_post_monitor_contended_entered()) { 397 JvmtiExport::post_monitor_contended_entered(jt, this); 398 399 // The current thread already owns the monitor and is not going to 400 // call park() for the remainder of the monitor enter protocol. So 401 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 402 // event handler consumed an unpark() issued by the thread that 403 // just exited the monitor. 404 } 405 if (event.should_commit()) { 406 event.set_previousOwner((uintptr_t)_previous_owner_tid); 407 event.commit(); 408 } 409 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 410 } 411 412 // Caveat: TryLock() is not necessarily serializing if it returns failure. 413 // Callers must compensate as needed. 414 415 int ObjectMonitor::TryLock(Thread * Self) { 416 void * own = _owner; 417 if (own != NULL) return 0; 418 if (Atomic::replace_if_null(Self, &_owner)) { 419 // Either guarantee _recursions == 0 or set _recursions = 0. 420 assert(_recursions == 0, "invariant"); 421 assert(_owner == Self, "invariant"); 422 return 1; 423 } 424 // The lock had been free momentarily, but we lost the race to the lock. 425 // Interference -- the CAS failed. 426 // We can either return -1 or retry. 427 // Retry doesn't make as much sense because the lock was just acquired. 428 return -1; 429 } 430 431 #define MAX_RECHECK_INTERVAL 1000 432 433 void ObjectMonitor::EnterI(TRAPS) { 434 Thread * const Self = THREAD; 435 assert(Self->is_Java_thread(), "invariant"); 436 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 437 438 // Try the lock - TATAS 439 if (TryLock (Self) > 0) { 440 assert(_succ != Self, "invariant"); 441 assert(_owner == Self, "invariant"); 442 assert(_Responsible != Self, "invariant"); 443 return; 444 } 445 446 DeferredInitialize(); 447 448 // We try one round of spinning *before* enqueueing Self. 449 // 450 // If the _owner is ready but OFFPROC we could use a YieldTo() 451 // operation to donate the remainder of this thread's quantum 452 // to the owner. This has subtle but beneficial affinity 453 // effects. 454 455 if (TrySpin (Self) > 0) { 456 assert(_owner == Self, "invariant"); 457 assert(_succ != Self, "invariant"); 458 assert(_Responsible != Self, "invariant"); 459 return; 460 } 461 462 // The Spin failed -- Enqueue and park the thread ... 463 assert(_succ != Self, "invariant"); 464 assert(_owner != Self, "invariant"); 465 assert(_Responsible != Self, "invariant"); 466 467 // Enqueue "Self" on ObjectMonitor's _cxq. 468 // 469 // Node acts as a proxy for Self. 470 // As an aside, if were to ever rewrite the synchronization code mostly 471 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 472 // Java objects. This would avoid awkward lifecycle and liveness issues, 473 // as well as eliminate a subset of ABA issues. 474 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 475 476 ObjectWaiter node(Self); 477 Self->_ParkEvent->reset(); 478 node._prev = (ObjectWaiter *) 0xBAD; 479 node.TState = ObjectWaiter::TS_CXQ; 480 481 // Push "Self" onto the front of the _cxq. 482 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 483 // Note that spinning tends to reduce the rate at which threads 484 // enqueue and dequeue on EntryList|cxq. 485 ObjectWaiter * nxt; 486 for (;;) { 487 node._next = nxt = _cxq; 488 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 489 490 // Interference - the CAS failed because _cxq changed. Just retry. 491 // As an optional optimization we retry the lock. 492 if (TryLock (Self) > 0) { 493 assert(_succ != Self, "invariant"); 494 assert(_owner == Self, "invariant"); 495 assert(_Responsible != Self, "invariant"); 496 return; 497 } 498 } 499 500 // Check for cxq|EntryList edge transition to non-null. This indicates 501 // the onset of contention. While contention persists exiting threads 502 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 503 // operations revert to the faster 1-0 mode. This enter operation may interleave 504 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 505 // arrange for one of the contending thread to use a timed park() operations 506 // to detect and recover from the race. (Stranding is form of progress failure 507 // where the monitor is unlocked but all the contending threads remain parked). 508 // That is, at least one of the contended threads will periodically poll _owner. 509 // One of the contending threads will become the designated "Responsible" thread. 510 // The Responsible thread uses a timed park instead of a normal indefinite park 511 // operation -- it periodically wakes and checks for and recovers from potential 512 // strandings admitted by 1-0 exit operations. We need at most one Responsible 513 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 514 // be responsible for a monitor. 515 // 516 // Currently, one of the contended threads takes on the added role of "Responsible". 517 // A viable alternative would be to use a dedicated "stranding checker" thread 518 // that periodically iterated over all the threads (or active monitors) and unparked 519 // successors where there was risk of stranding. This would help eliminate the 520 // timer scalability issues we see on some platforms as we'd only have one thread 521 // -- the checker -- parked on a timer. 522 523 if (nxt == NULL && _EntryList == NULL) { 524 // Try to assume the role of responsible thread for the monitor. 525 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 526 Atomic::replace_if_null(Self, &_Responsible); 527 } 528 529 // The lock might have been released while this thread was occupied queueing 530 // itself onto _cxq. To close the race and avoid "stranding" and 531 // progress-liveness failure we must resample-retry _owner before parking. 532 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 533 // In this case the ST-MEMBAR is accomplished with CAS(). 534 // 535 // TODO: Defer all thread state transitions until park-time. 536 // Since state transitions are heavy and inefficient we'd like 537 // to defer the state transitions until absolutely necessary, 538 // and in doing so avoid some transitions ... 539 540 int nWakeups = 0; 541 int recheckInterval = 1; 542 543 for (;;) { 544 545 if (TryLock(Self) > 0) break; 546 assert(_owner != Self, "invariant"); 547 548 // park self 549 if (_Responsible == Self) { 550 Self->_ParkEvent->park((jlong) recheckInterval); 551 // Increase the recheckInterval, but clamp the value. 552 recheckInterval *= 8; 553 if (recheckInterval > MAX_RECHECK_INTERVAL) { 554 recheckInterval = MAX_RECHECK_INTERVAL; 555 } 556 } else { 557 Self->_ParkEvent->park(); 558 } 559 560 if (TryLock(Self) > 0) break; 561 562 // The lock is still contested. 563 // Keep a tally of the # of futile wakeups. 564 // Note that the counter is not protected by a lock or updated by atomics. 565 // That is by design - we trade "lossy" counters which are exposed to 566 // races during updates for a lower probe effect. 567 568 // This PerfData object can be used in parallel with a safepoint. 569 // See the work around in PerfDataManager::destroy(). 570 OM_PERFDATA_OP(FutileWakeups, inc()); 571 ++nWakeups; 572 573 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 574 // We can defer clearing _succ until after the spin completes 575 // TrySpin() must tolerate being called with _succ == Self. 576 // Try yet another round of adaptive spinning. 577 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 578 579 // We can find that we were unpark()ed and redesignated _succ while 580 // we were spinning. That's harmless. If we iterate and call park(), 581 // park() will consume the event and return immediately and we'll 582 // just spin again. This pattern can repeat, leaving _succ to simply 583 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 584 // Alternately, we can sample fired() here, and if set, forgo spinning 585 // in the next iteration. 586 587 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 588 Self->_ParkEvent->reset(); 589 OrderAccess::fence(); 590 } 591 if (_succ == Self) _succ = NULL; 592 593 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 594 OrderAccess::fence(); 595 } 596 597 // Egress : 598 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 599 // Normally we'll find Self on the EntryList . 600 // From the perspective of the lock owner (this thread), the 601 // EntryList is stable and cxq is prepend-only. 602 // The head of cxq is volatile but the interior is stable. 603 // In addition, Self.TState is stable. 604 605 assert(_owner == Self, "invariant"); 606 assert(object() != NULL, "invariant"); 607 // I'd like to write: 608 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 609 // but as we're at a safepoint that's not safe. 610 611 UnlinkAfterAcquire(Self, &node); 612 if (_succ == Self) _succ = NULL; 613 614 assert(_succ != Self, "invariant"); 615 if (_Responsible == Self) { 616 _Responsible = NULL; 617 OrderAccess::fence(); // Dekker pivot-point 618 619 // We may leave threads on cxq|EntryList without a designated 620 // "Responsible" thread. This is benign. When this thread subsequently 621 // exits the monitor it can "see" such preexisting "old" threads -- 622 // threads that arrived on the cxq|EntryList before the fence, above -- 623 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 624 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 625 // non-null and elect a new "Responsible" timer thread. 626 // 627 // This thread executes: 628 // ST Responsible=null; MEMBAR (in enter epilogue - here) 629 // LD cxq|EntryList (in subsequent exit) 630 // 631 // Entering threads in the slow/contended path execute: 632 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 633 // The (ST cxq; MEMBAR) is accomplished with CAS(). 634 // 635 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 636 // exit operation from floating above the ST Responsible=null. 637 } 638 639 // We've acquired ownership with CAS(). 640 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 641 // But since the CAS() this thread may have also stored into _succ, 642 // EntryList, cxq or Responsible. These meta-data updates must be 643 // visible __before this thread subsequently drops the lock. 644 // Consider what could occur if we didn't enforce this constraint -- 645 // STs to monitor meta-data and user-data could reorder with (become 646 // visible after) the ST in exit that drops ownership of the lock. 647 // Some other thread could then acquire the lock, but observe inconsistent 648 // or old monitor meta-data and heap data. That violates the JMM. 649 // To that end, the 1-0 exit() operation must have at least STST|LDST 650 // "release" barrier semantics. Specifically, there must be at least a 651 // STST|LDST barrier in exit() before the ST of null into _owner that drops 652 // the lock. The barrier ensures that changes to monitor meta-data and data 653 // protected by the lock will be visible before we release the lock, and 654 // therefore before some other thread (CPU) has a chance to acquire the lock. 655 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 656 // 657 // Critically, any prior STs to _succ or EntryList must be visible before 658 // the ST of null into _owner in the *subsequent* (following) corresponding 659 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 660 // execute a serializing instruction. 661 662 return; 663 } 664 665 // ReenterI() is a specialized inline form of the latter half of the 666 // contended slow-path from EnterI(). We use ReenterI() only for 667 // monitor reentry in wait(). 668 // 669 // In the future we should reconcile EnterI() and ReenterI(), adding 670 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 671 // loop accordingly. 672 673 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 674 assert(Self != NULL, "invariant"); 675 assert(SelfNode != NULL, "invariant"); 676 assert(SelfNode->_thread == Self, "invariant"); 677 assert(_waiters > 0, "invariant"); 678 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 679 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 680 JavaThread * jt = (JavaThread *) Self; 681 682 int nWakeups = 0; 683 for (;;) { 684 ObjectWaiter::TStates v = SelfNode->TState; 685 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 686 assert(_owner != Self, "invariant"); 687 688 if (TryLock(Self) > 0) break; 689 if (TrySpin(Self) > 0) break; 690 691 // State transition wrappers around park() ... 692 // ReenterI() wisely defers state transitions until 693 // it's clear we must park the thread. 694 { 695 OSThreadContendState osts(Self->osthread()); 696 ThreadBlockInVM tbivm(jt); 697 698 // cleared by handle_special_suspend_equivalent_condition() 699 // or java_suspend_self() 700 jt->set_suspend_equivalent(); 701 Self->_ParkEvent->park(); 702 703 // were we externally suspended while we were waiting? 704 for (;;) { 705 if (!ExitSuspendEquivalent(jt)) break; 706 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 707 jt->java_suspend_self(); 708 jt->set_suspend_equivalent(); 709 } 710 } 711 712 // Try again, but just so we distinguish between futile wakeups and 713 // successful wakeups. The following test isn't algorithmically 714 // necessary, but it helps us maintain sensible statistics. 715 if (TryLock(Self) > 0) break; 716 717 // The lock is still contested. 718 // Keep a tally of the # of futile wakeups. 719 // Note that the counter is not protected by a lock or updated by atomics. 720 // That is by design - we trade "lossy" counters which are exposed to 721 // races during updates for a lower probe effect. 722 ++nWakeups; 723 724 // Assuming this is not a spurious wakeup we'll normally 725 // find that _succ == Self. 726 if (_succ == Self) _succ = NULL; 727 728 // Invariant: after clearing _succ a contending thread 729 // *must* retry _owner before parking. 730 OrderAccess::fence(); 731 732 // This PerfData object can be used in parallel with a safepoint. 733 // See the work around in PerfDataManager::destroy(). 734 OM_PERFDATA_OP(FutileWakeups, inc()); 735 } 736 737 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 738 // Normally we'll find Self on the EntryList. 739 // Unlinking from the EntryList is constant-time and atomic-free. 740 // From the perspective of the lock owner (this thread), the 741 // EntryList is stable and cxq is prepend-only. 742 // The head of cxq is volatile but the interior is stable. 743 // In addition, Self.TState is stable. 744 745 assert(_owner == Self, "invariant"); 746 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 747 UnlinkAfterAcquire(Self, SelfNode); 748 if (_succ == Self) _succ = NULL; 749 assert(_succ != Self, "invariant"); 750 SelfNode->TState = ObjectWaiter::TS_RUN; 751 OrderAccess::fence(); // see comments at the end of EnterI() 752 } 753 754 // By convention we unlink a contending thread from EntryList|cxq immediately 755 // after the thread acquires the lock in ::enter(). Equally, we could defer 756 // unlinking the thread until ::exit()-time. 757 758 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 759 assert(_owner == Self, "invariant"); 760 assert(SelfNode->_thread == Self, "invariant"); 761 762 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 763 // Normal case: remove Self from the DLL EntryList . 764 // This is a constant-time operation. 765 ObjectWaiter * nxt = SelfNode->_next; 766 ObjectWaiter * prv = SelfNode->_prev; 767 if (nxt != NULL) nxt->_prev = prv; 768 if (prv != NULL) prv->_next = nxt; 769 if (SelfNode == _EntryList) _EntryList = nxt; 770 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 771 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 772 } else { 773 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 774 // Inopportune interleaving -- Self is still on the cxq. 775 // This usually means the enqueue of self raced an exiting thread. 776 // Normally we'll find Self near the front of the cxq, so 777 // dequeueing is typically fast. If needbe we can accelerate 778 // this with some MCS/CHL-like bidirectional list hints and advisory 779 // back-links so dequeueing from the interior will normally operate 780 // in constant-time. 781 // Dequeue Self from either the head (with CAS) or from the interior 782 // with a linear-time scan and normal non-atomic memory operations. 783 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 784 // and then unlink Self from EntryList. We have to drain eventually, 785 // so it might as well be now. 786 787 ObjectWaiter * v = _cxq; 788 assert(v != NULL, "invariant"); 789 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 790 // The CAS above can fail from interference IFF a "RAT" arrived. 791 // In that case Self must be in the interior and can no longer be 792 // at the head of cxq. 793 if (v == SelfNode) { 794 assert(_cxq != v, "invariant"); 795 v = _cxq; // CAS above failed - start scan at head of list 796 } 797 ObjectWaiter * p; 798 ObjectWaiter * q = NULL; 799 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 800 q = p; 801 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 802 } 803 assert(v != SelfNode, "invariant"); 804 assert(p == SelfNode, "Node not found on cxq"); 805 assert(p != _cxq, "invariant"); 806 assert(q != NULL, "invariant"); 807 assert(q->_next == p, "invariant"); 808 q->_next = p->_next; 809 } 810 } 811 812 #ifdef ASSERT 813 // Diagnostic hygiene ... 814 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 815 SelfNode->_next = (ObjectWaiter *) 0xBAD; 816 SelfNode->TState = ObjectWaiter::TS_RUN; 817 #endif 818 } 819 820 // ----------------------------------------------------------------------------- 821 // Exit support 822 // 823 // exit() 824 // ~~~~~~ 825 // Note that the collector can't reclaim the objectMonitor or deflate 826 // the object out from underneath the thread calling ::exit() as the 827 // thread calling ::exit() never transitions to a stable state. 828 // This inhibits GC, which in turn inhibits asynchronous (and 829 // inopportune) reclamation of "this". 830 // 831 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 832 // There's one exception to the claim above, however. EnterI() can call 833 // exit() to drop a lock if the acquirer has been externally suspended. 834 // In that case exit() is called with _thread_state as _thread_blocked, 835 // but the monitor's _count field is > 0, which inhibits reclamation. 836 // 837 // 1-0 exit 838 // ~~~~~~~~ 839 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 840 // the fast-path operators have been optimized so the common ::exit() 841 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 842 // The code emitted by fast_unlock() elides the usual MEMBAR. This 843 // greatly improves latency -- MEMBAR and CAS having considerable local 844 // latency on modern processors -- but at the cost of "stranding". Absent the 845 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 846 // ::enter() path, resulting in the entering thread being stranding 847 // and a progress-liveness failure. Stranding is extremely rare. 848 // We use timers (timed park operations) & periodic polling to detect 849 // and recover from stranding. Potentially stranded threads periodically 850 // wake up and poll the lock. See the usage of the _Responsible variable. 851 // 852 // The CAS() in enter provides for safety and exclusion, while the CAS or 853 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 854 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 855 // We detect and recover from stranding with timers. 856 // 857 // If a thread transiently strands it'll park until (a) another 858 // thread acquires the lock and then drops the lock, at which time the 859 // exiting thread will notice and unpark the stranded thread, or, (b) 860 // the timer expires. If the lock is high traffic then the stranding latency 861 // will be low due to (a). If the lock is low traffic then the odds of 862 // stranding are lower, although the worst-case stranding latency 863 // is longer. Critically, we don't want to put excessive load in the 864 // platform's timer subsystem. We want to minimize both the timer injection 865 // rate (timers created/sec) as well as the number of timers active at 866 // any one time. (more precisely, we want to minimize timer-seconds, which is 867 // the integral of the # of active timers at any instant over time). 868 // Both impinge on OS scalability. Given that, at most one thread parked on 869 // a monitor will use a timer. 870 // 871 // There is also the risk of a futile wake-up. If we drop the lock 872 // another thread can reacquire the lock immediately, and we can 873 // then wake a thread unnecessarily. This is benign, and we've 874 // structured the code so the windows are short and the frequency 875 // of such futile wakups is low. 876 877 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 878 Thread * const Self = THREAD; 879 if (THREAD != _owner) { 880 if (THREAD->is_lock_owned((address) _owner)) { 881 // Transmute _owner from a BasicLock pointer to a Thread address. 882 // We don't need to hold _mutex for this transition. 883 // Non-null to Non-null is safe as long as all readers can 884 // tolerate either flavor. 885 assert(_recursions == 0, "invariant"); 886 _owner = THREAD; 887 _recursions = 0; 888 } else { 889 // Apparent unbalanced locking ... 890 // Naively we'd like to throw IllegalMonitorStateException. 891 // As a practical matter we can neither allocate nor throw an 892 // exception as ::exit() can be called from leaf routines. 893 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 894 // Upon deeper reflection, however, in a properly run JVM the only 895 // way we should encounter this situation is in the presence of 896 // unbalanced JNI locking. TODO: CheckJNICalls. 897 // See also: CR4414101 898 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 899 return; 900 } 901 } 902 903 if (_recursions != 0) { 904 _recursions--; // this is simple recursive enter 905 return; 906 } 907 908 // Invariant: after setting Responsible=null an thread must execute 909 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 910 _Responsible = NULL; 911 912 #if INCLUDE_JFR 913 // get the owner's thread id for the MonitorEnter event 914 // if it is enabled and the thread isn't suspended 915 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 916 _previous_owner_tid = JFR_THREAD_ID(Self); 917 } 918 #endif 919 920 for (;;) { 921 assert(THREAD == _owner, "invariant"); 922 923 if (Knob_ExitPolicy == 0) { 924 // release semantics: prior loads and stores from within the critical section 925 // must not float (reorder) past the following store that drops the lock. 926 // On SPARC that requires MEMBAR #loadstore|#storestore. 927 // But of course in TSO #loadstore|#storestore is not required. 928 // I'd like to write one of the following: 929 // A. OrderAccess::release() ; _owner = NULL 930 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 931 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 932 // store into a _dummy variable. That store is not needed, but can result 933 // in massive wasteful coherency traffic on classic SMP systems. 934 // Instead, I use release_store(), which is implemented as just a simple 935 // ST on x64, x86 and SPARC. 936 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 937 OrderAccess::storeload(); // See if we need to wake a successor 938 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 939 return; 940 } 941 // Other threads are blocked trying to acquire the lock. 942 943 // Normally the exiting thread is responsible for ensuring succession, 944 // but if other successors are ready or other entering threads are spinning 945 // then this thread can simply store NULL into _owner and exit without 946 // waking a successor. The existence of spinners or ready successors 947 // guarantees proper succession (liveness). Responsibility passes to the 948 // ready or running successors. The exiting thread delegates the duty. 949 // More precisely, if a successor already exists this thread is absolved 950 // of the responsibility of waking (unparking) one. 951 // 952 // The _succ variable is critical to reducing futile wakeup frequency. 953 // _succ identifies the "heir presumptive" thread that has been made 954 // ready (unparked) but that has not yet run. We need only one such 955 // successor thread to guarantee progress. 956 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 957 // section 3.3 "Futile Wakeup Throttling" for details. 958 // 959 // Note that spinners in Enter() also set _succ non-null. 960 // In the current implementation spinners opportunistically set 961 // _succ so that exiting threads might avoid waking a successor. 962 // Another less appealing alternative would be for the exiting thread 963 // to drop the lock and then spin briefly to see if a spinner managed 964 // to acquire the lock. If so, the exiting thread could exit 965 // immediately without waking a successor, otherwise the exiting 966 // thread would need to dequeue and wake a successor. 967 // (Note that we'd need to make the post-drop spin short, but no 968 // shorter than the worst-case round-trip cache-line migration time. 969 // The dropped lock needs to become visible to the spinner, and then 970 // the acquisition of the lock by the spinner must become visible to 971 // the exiting thread). 972 973 // It appears that an heir-presumptive (successor) must be made ready. 974 // Only the current lock owner can manipulate the EntryList or 975 // drain _cxq, so we need to reacquire the lock. If we fail 976 // to reacquire the lock the responsibility for ensuring succession 977 // falls to the new owner. 978 // 979 if (!Atomic::replace_if_null(THREAD, &_owner)) { 980 return; 981 } 982 } else { 983 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 984 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 985 OrderAccess::storeload(); 986 // Ratify the previously observed values. 987 if (_cxq == NULL || _succ != NULL) { 988 return; 989 } 990 991 // inopportune interleaving -- the exiting thread (this thread) 992 // in the fast-exit path raced an entering thread in the slow-enter 993 // path. 994 // We have two choices: 995 // A. Try to reacquire the lock. 996 // If the CAS() fails return immediately, otherwise 997 // we either restart/rerun the exit operation, or simply 998 // fall-through into the code below which wakes a successor. 999 // B. If the elements forming the EntryList|cxq are TSM 1000 // we could simply unpark() the lead thread and return 1001 // without having set _succ. 1002 if (!Atomic::replace_if_null(THREAD, &_owner)) { 1003 return; 1004 } 1005 } 1006 } 1007 1008 guarantee(_owner == THREAD, "invariant"); 1009 1010 ObjectWaiter * w = NULL; 1011 int QMode = Knob_QMode; 1012 1013 if (QMode == 2 && _cxq != NULL) { 1014 // QMode == 2 : cxq has precedence over EntryList. 1015 // Try to directly wake a successor from the cxq. 1016 // If successful, the successor will need to unlink itself from cxq. 1017 w = _cxq; 1018 assert(w != NULL, "invariant"); 1019 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1020 ExitEpilog(Self, w); 1021 return; 1022 } 1023 1024 if (QMode == 3 && _cxq != NULL) { 1025 // Aggressively drain cxq into EntryList at the first opportunity. 1026 // This policy ensure that recently-run threads live at the head of EntryList. 1027 // Drain _cxq into EntryList - bulk transfer. 1028 // First, detach _cxq. 1029 // The following loop is tantamount to: w = swap(&cxq, NULL) 1030 w = _cxq; 1031 for (;;) { 1032 assert(w != NULL, "Invariant"); 1033 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1034 if (u == w) break; 1035 w = u; 1036 } 1037 assert(w != NULL, "invariant"); 1038 1039 ObjectWaiter * q = NULL; 1040 ObjectWaiter * p; 1041 for (p = w; p != NULL; p = p->_next) { 1042 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1043 p->TState = ObjectWaiter::TS_ENTER; 1044 p->_prev = q; 1045 q = p; 1046 } 1047 1048 // Append the RATs to the EntryList 1049 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1050 ObjectWaiter * Tail; 1051 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1052 Tail = Tail->_next) 1053 /* empty */; 1054 if (Tail == NULL) { 1055 _EntryList = w; 1056 } else { 1057 Tail->_next = w; 1058 w->_prev = Tail; 1059 } 1060 1061 // Fall thru into code that tries to wake a successor from EntryList 1062 } 1063 1064 if (QMode == 4 && _cxq != NULL) { 1065 // Aggressively drain cxq into EntryList at the first opportunity. 1066 // This policy ensure that recently-run threads live at the head of EntryList. 1067 1068 // Drain _cxq into EntryList - bulk transfer. 1069 // First, detach _cxq. 1070 // The following loop is tantamount to: w = swap(&cxq, NULL) 1071 w = _cxq; 1072 for (;;) { 1073 assert(w != NULL, "Invariant"); 1074 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1075 if (u == w) break; 1076 w = u; 1077 } 1078 assert(w != NULL, "invariant"); 1079 1080 ObjectWaiter * q = NULL; 1081 ObjectWaiter * p; 1082 for (p = w; p != NULL; p = p->_next) { 1083 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1084 p->TState = ObjectWaiter::TS_ENTER; 1085 p->_prev = q; 1086 q = p; 1087 } 1088 1089 // Prepend the RATs to the EntryList 1090 if (_EntryList != NULL) { 1091 q->_next = _EntryList; 1092 _EntryList->_prev = q; 1093 } 1094 _EntryList = w; 1095 1096 // Fall thru into code that tries to wake a successor from EntryList 1097 } 1098 1099 w = _EntryList; 1100 if (w != NULL) { 1101 // I'd like to write: guarantee (w->_thread != Self). 1102 // But in practice an exiting thread may find itself on the EntryList. 1103 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1104 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1105 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1106 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1107 // release the lock "O". T2 resumes immediately after the ST of null into 1108 // _owner, above. T2 notices that the EntryList is populated, so it 1109 // reacquires the lock and then finds itself on the EntryList. 1110 // Given all that, we have to tolerate the circumstance where "w" is 1111 // associated with Self. 1112 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1113 ExitEpilog(Self, w); 1114 return; 1115 } 1116 1117 // If we find that both _cxq and EntryList are null then just 1118 // re-run the exit protocol from the top. 1119 w = _cxq; 1120 if (w == NULL) continue; 1121 1122 // Drain _cxq into EntryList - bulk transfer. 1123 // First, detach _cxq. 1124 // The following loop is tantamount to: w = swap(&cxq, NULL) 1125 for (;;) { 1126 assert(w != NULL, "Invariant"); 1127 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1128 if (u == w) break; 1129 w = u; 1130 } 1131 1132 assert(w != NULL, "invariant"); 1133 assert(_EntryList == NULL, "invariant"); 1134 1135 // Convert the LIFO SLL anchored by _cxq into a DLL. 1136 // The list reorganization step operates in O(LENGTH(w)) time. 1137 // It's critical that this step operate quickly as 1138 // "Self" still holds the outer-lock, restricting parallelism 1139 // and effectively lengthening the critical section. 1140 // Invariant: s chases t chases u. 1141 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1142 // we have faster access to the tail. 1143 1144 if (QMode == 1) { 1145 // QMode == 1 : drain cxq to EntryList, reversing order 1146 // We also reverse the order of the list. 1147 ObjectWaiter * s = NULL; 1148 ObjectWaiter * t = w; 1149 ObjectWaiter * u = NULL; 1150 while (t != NULL) { 1151 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1152 t->TState = ObjectWaiter::TS_ENTER; 1153 u = t->_next; 1154 t->_prev = u; 1155 t->_next = s; 1156 s = t; 1157 t = u; 1158 } 1159 _EntryList = s; 1160 assert(s != NULL, "invariant"); 1161 } else { 1162 // QMode == 0 or QMode == 2 1163 _EntryList = w; 1164 ObjectWaiter * q = NULL; 1165 ObjectWaiter * p; 1166 for (p = w; p != NULL; p = p->_next) { 1167 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1168 p->TState = ObjectWaiter::TS_ENTER; 1169 p->_prev = q; 1170 q = p; 1171 } 1172 } 1173 1174 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1175 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1176 1177 // See if we can abdicate to a spinner instead of waking a thread. 1178 // A primary goal of the implementation is to reduce the 1179 // context-switch rate. 1180 if (_succ != NULL) continue; 1181 1182 w = _EntryList; 1183 if (w != NULL) { 1184 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1185 ExitEpilog(Self, w); 1186 return; 1187 } 1188 } 1189 } 1190 1191 // ExitSuspendEquivalent: 1192 // A faster alternate to handle_special_suspend_equivalent_condition() 1193 // 1194 // handle_special_suspend_equivalent_condition() unconditionally 1195 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1196 // operations have high latency. Note that in ::enter() we call HSSEC 1197 // while holding the monitor, so we effectively lengthen the critical sections. 1198 // 1199 // There are a number of possible solutions: 1200 // 1201 // A. To ameliorate the problem we might also defer state transitions 1202 // to as late as possible -- just prior to parking. 1203 // Given that, we'd call HSSEC after having returned from park(), 1204 // but before attempting to acquire the monitor. This is only a 1205 // partial solution. It avoids calling HSSEC while holding the 1206 // monitor (good), but it still increases successor reacquisition latency -- 1207 // the interval between unparking a successor and the time the successor 1208 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1209 // If we use this technique we can also avoid EnterI()-exit() loop 1210 // in ::enter() where we iteratively drop the lock and then attempt 1211 // to reacquire it after suspending. 1212 // 1213 // B. In the future we might fold all the suspend bits into a 1214 // composite per-thread suspend flag and then update it with CAS(). 1215 // Alternately, a Dekker-like mechanism with multiple variables 1216 // would suffice: 1217 // ST Self->_suspend_equivalent = false 1218 // MEMBAR 1219 // LD Self_>_suspend_flags 1220 // 1221 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1222 // with a more efficient implementation, the need to use "FastHSSEC" has 1223 // decreased. - Dave 1224 1225 1226 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1227 const int Mode = Knob_FastHSSEC; 1228 if (Mode && !jSelf->is_external_suspend()) { 1229 assert(jSelf->is_suspend_equivalent(), "invariant"); 1230 jSelf->clear_suspend_equivalent(); 1231 if (2 == Mode) OrderAccess::storeload(); 1232 if (!jSelf->is_external_suspend()) return false; 1233 // We raced a suspension -- fall thru into the slow path 1234 jSelf->set_suspend_equivalent(); 1235 } 1236 return jSelf->handle_special_suspend_equivalent_condition(); 1237 } 1238 1239 1240 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1241 assert(_owner == Self, "invariant"); 1242 1243 // Exit protocol: 1244 // 1. ST _succ = wakee 1245 // 2. membar #loadstore|#storestore; 1246 // 2. ST _owner = NULL 1247 // 3. unpark(wakee) 1248 1249 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1250 ParkEvent * Trigger = Wakee->_event; 1251 1252 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1253 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1254 // out-of-scope (non-extant). 1255 Wakee = NULL; 1256 1257 // Drop the lock 1258 OrderAccess::release_store(&_owner, (void*)NULL); 1259 OrderAccess::fence(); // ST _owner vs LD in unpark() 1260 1261 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1262 Trigger->unpark(); 1263 1264 // Maintain stats and report events to JVMTI 1265 OM_PERFDATA_OP(Parks, inc()); 1266 } 1267 1268 1269 // ----------------------------------------------------------------------------- 1270 // Class Loader deadlock handling. 1271 // 1272 // complete_exit exits a lock returning recursion count 1273 // complete_exit/reenter operate as a wait without waiting 1274 // complete_exit requires an inflated monitor 1275 // The _owner field is not always the Thread addr even with an 1276 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1277 // thread due to contention. 1278 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1279 Thread * const Self = THREAD; 1280 assert(Self->is_Java_thread(), "Must be Java thread!"); 1281 JavaThread *jt = (JavaThread *)THREAD; 1282 1283 DeferredInitialize(); 1284 1285 if (THREAD != _owner) { 1286 if (THREAD->is_lock_owned ((address)_owner)) { 1287 assert(_recursions == 0, "internal state error"); 1288 _owner = THREAD; // Convert from basiclock addr to Thread addr 1289 _recursions = 0; 1290 } 1291 } 1292 1293 guarantee(Self == _owner, "complete_exit not owner"); 1294 intptr_t save = _recursions; // record the old recursion count 1295 _recursions = 0; // set the recursion level to be 0 1296 exit(true, Self); // exit the monitor 1297 guarantee(_owner != Self, "invariant"); 1298 return save; 1299 } 1300 1301 // reenter() enters a lock and sets recursion count 1302 // complete_exit/reenter operate as a wait without waiting 1303 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1304 Thread * const Self = THREAD; 1305 assert(Self->is_Java_thread(), "Must be Java thread!"); 1306 JavaThread *jt = (JavaThread *)THREAD; 1307 1308 guarantee(_owner != Self, "reenter already owner"); 1309 enter(THREAD); // enter the monitor 1310 guarantee(_recursions == 0, "reenter recursion"); 1311 _recursions = recursions; 1312 return; 1313 } 1314 1315 1316 // ----------------------------------------------------------------------------- 1317 // A macro is used below because there may already be a pending 1318 // exception which should not abort the execution of the routines 1319 // which use this (which is why we don't put this into check_slow and 1320 // call it with a CHECK argument). 1321 1322 #define CHECK_OWNER() \ 1323 do { \ 1324 if (THREAD != _owner) { \ 1325 if (THREAD->is_lock_owned((address) _owner)) { \ 1326 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1327 _recursions = 0; \ 1328 } else { \ 1329 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1330 } \ 1331 } \ 1332 } while (false) 1333 1334 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1335 // TODO-FIXME: remove check_slow() -- it's likely dead. 1336 1337 void ObjectMonitor::check_slow(TRAPS) { 1338 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1339 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1340 } 1341 1342 static int Adjust(volatile int * adr, int dx) { 1343 int v; 1344 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1345 return v; 1346 } 1347 1348 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1349 ObjectMonitor* monitor, 1350 jlong notifier_tid, 1351 jlong timeout, 1352 bool timedout) { 1353 assert(event != NULL, "invariant"); 1354 assert(monitor != NULL, "invariant"); 1355 event->set_monitorClass(((oop)monitor->object())->klass()); 1356 event->set_timeout(timeout); 1357 event->set_address((uintptr_t)monitor->object_addr()); 1358 event->set_notifier(notifier_tid); 1359 event->set_timedOut(timedout); 1360 event->commit(); 1361 } 1362 1363 // ----------------------------------------------------------------------------- 1364 // Wait/Notify/NotifyAll 1365 // 1366 // Note: a subset of changes to ObjectMonitor::wait() 1367 // will need to be replicated in complete_exit 1368 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1369 Thread * const Self = THREAD; 1370 assert(Self->is_Java_thread(), "Must be Java thread!"); 1371 JavaThread *jt = (JavaThread *)THREAD; 1372 1373 DeferredInitialize(); 1374 1375 // Throw IMSX or IEX. 1376 CHECK_OWNER(); 1377 1378 EventJavaMonitorWait event; 1379 1380 // check for a pending interrupt 1381 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1382 // post monitor waited event. Note that this is past-tense, we are done waiting. 1383 if (JvmtiExport::should_post_monitor_waited()) { 1384 // Note: 'false' parameter is passed here because the 1385 // wait was not timed out due to thread interrupt. 1386 JvmtiExport::post_monitor_waited(jt, this, false); 1387 1388 // In this short circuit of the monitor wait protocol, the 1389 // current thread never drops ownership of the monitor and 1390 // never gets added to the wait queue so the current thread 1391 // cannot be made the successor. This means that the 1392 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1393 // consume an unpark() meant for the ParkEvent associated with 1394 // this ObjectMonitor. 1395 } 1396 if (event.should_commit()) { 1397 post_monitor_wait_event(&event, this, 0, millis, false); 1398 } 1399 THROW(vmSymbols::java_lang_InterruptedException()); 1400 return; 1401 } 1402 1403 assert(Self->_Stalled == 0, "invariant"); 1404 Self->_Stalled = intptr_t(this); 1405 jt->set_current_waiting_monitor(this); 1406 1407 // create a node to be put into the queue 1408 // Critically, after we reset() the event but prior to park(), we must check 1409 // for a pending interrupt. 1410 ObjectWaiter node(Self); 1411 node.TState = ObjectWaiter::TS_WAIT; 1412 Self->_ParkEvent->reset(); 1413 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1414 1415 // Enter the waiting queue, which is a circular doubly linked list in this case 1416 // but it could be a priority queue or any data structure. 1417 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1418 // by the the owner of the monitor *except* in the case where park() 1419 // returns because of a timeout of interrupt. Contention is exceptionally rare 1420 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1421 1422 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1423 AddWaiter(&node); 1424 Thread::SpinRelease(&_WaitSetLock); 1425 1426 _Responsible = NULL; 1427 1428 intptr_t save = _recursions; // record the old recursion count 1429 _waiters++; // increment the number of waiters 1430 _recursions = 0; // set the recursion level to be 1 1431 exit(true, Self); // exit the monitor 1432 guarantee(_owner != Self, "invariant"); 1433 1434 // The thread is on the WaitSet list - now park() it. 1435 // On MP systems it's conceivable that a brief spin before we park 1436 // could be profitable. 1437 // 1438 // TODO-FIXME: change the following logic to a loop of the form 1439 // while (!timeout && !interrupted && _notified == 0) park() 1440 1441 int ret = OS_OK; 1442 int WasNotified = 0; 1443 { // State transition wrappers 1444 OSThread* osthread = Self->osthread(); 1445 OSThreadWaitState osts(osthread, true); 1446 { 1447 ThreadBlockInVM tbivm(jt); 1448 // Thread is in thread_blocked state and oop access is unsafe. 1449 jt->set_suspend_equivalent(); 1450 1451 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1452 // Intentionally empty 1453 } else if (node._notified == 0) { 1454 if (millis <= 0) { 1455 Self->_ParkEvent->park(); 1456 } else { 1457 ret = Self->_ParkEvent->park(millis); 1458 } 1459 } 1460 1461 // were we externally suspended while we were waiting? 1462 if (ExitSuspendEquivalent (jt)) { 1463 // TODO-FIXME: add -- if succ == Self then succ = null. 1464 jt->java_suspend_self(); 1465 } 1466 1467 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1468 1469 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1470 // from the WaitSet to the EntryList. 1471 // See if we need to remove Node from the WaitSet. 1472 // We use double-checked locking to avoid grabbing _WaitSetLock 1473 // if the thread is not on the wait queue. 1474 // 1475 // Note that we don't need a fence before the fetch of TState. 1476 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1477 // written by the is thread. (perhaps the fetch might even be satisfied 1478 // by a look-aside into the processor's own store buffer, although given 1479 // the length of the code path between the prior ST and this load that's 1480 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1481 // then we'll acquire the lock and then re-fetch a fresh TState value. 1482 // That is, we fail toward safety. 1483 1484 if (node.TState == ObjectWaiter::TS_WAIT) { 1485 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1486 if (node.TState == ObjectWaiter::TS_WAIT) { 1487 DequeueSpecificWaiter(&node); // unlink from WaitSet 1488 assert(node._notified == 0, "invariant"); 1489 node.TState = ObjectWaiter::TS_RUN; 1490 } 1491 Thread::SpinRelease(&_WaitSetLock); 1492 } 1493 1494 // The thread is now either on off-list (TS_RUN), 1495 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1496 // The Node's TState variable is stable from the perspective of this thread. 1497 // No other threads will asynchronously modify TState. 1498 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1499 OrderAccess::loadload(); 1500 if (_succ == Self) _succ = NULL; 1501 WasNotified = node._notified; 1502 1503 // Reentry phase -- reacquire the monitor. 1504 // re-enter contended monitor after object.wait(). 1505 // retain OBJECT_WAIT state until re-enter successfully completes 1506 // Thread state is thread_in_vm and oop access is again safe, 1507 // although the raw address of the object may have changed. 1508 // (Don't cache naked oops over safepoints, of course). 1509 1510 // post monitor waited event. Note that this is past-tense, we are done waiting. 1511 if (JvmtiExport::should_post_monitor_waited()) { 1512 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1513 1514 if (node._notified != 0 && _succ == Self) { 1515 // In this part of the monitor wait-notify-reenter protocol it 1516 // is possible (and normal) for another thread to do a fastpath 1517 // monitor enter-exit while this thread is still trying to get 1518 // to the reenter portion of the protocol. 1519 // 1520 // The ObjectMonitor was notified and the current thread is 1521 // the successor which also means that an unpark() has already 1522 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1523 // consume the unpark() that was done when the successor was 1524 // set because the same ParkEvent is shared between Java 1525 // monitors and JVM/TI RawMonitors (for now). 1526 // 1527 // We redo the unpark() to ensure forward progress, i.e., we 1528 // don't want all pending threads hanging (parked) with none 1529 // entering the unlocked monitor. 1530 node._event->unpark(); 1531 } 1532 } 1533 1534 if (event.should_commit()) { 1535 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1536 } 1537 1538 OrderAccess::fence(); 1539 1540 assert(Self->_Stalled != 0, "invariant"); 1541 Self->_Stalled = 0; 1542 1543 assert(_owner != Self, "invariant"); 1544 ObjectWaiter::TStates v = node.TState; 1545 if (v == ObjectWaiter::TS_RUN) { 1546 enter(Self); 1547 } else { 1548 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1549 ReenterI(Self, &node); 1550 node.wait_reenter_end(this); 1551 } 1552 1553 // Self has reacquired the lock. 1554 // Lifecycle - the node representing Self must not appear on any queues. 1555 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1556 // want residual elements associated with this thread left on any lists. 1557 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1558 assert(_owner == Self, "invariant"); 1559 assert(_succ != Self, "invariant"); 1560 } // OSThreadWaitState() 1561 1562 jt->set_current_waiting_monitor(NULL); 1563 1564 guarantee(_recursions == 0, "invariant"); 1565 _recursions = save; // restore the old recursion count 1566 _waiters--; // decrement the number of waiters 1567 1568 // Verify a few postconditions 1569 assert(_owner == Self, "invariant"); 1570 assert(_succ != Self, "invariant"); 1571 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1572 1573 // check if the notification happened 1574 if (!WasNotified) { 1575 // no, it could be timeout or Thread.interrupt() or both 1576 // check for interrupt event, otherwise it is timeout 1577 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1578 THROW(vmSymbols::java_lang_InterruptedException()); 1579 } 1580 } 1581 1582 // NOTE: Spurious wake up will be consider as timeout. 1583 // Monitor notify has precedence over thread interrupt. 1584 } 1585 1586 1587 // Consider: 1588 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1589 // then instead of transferring a thread from the WaitSet to the EntryList 1590 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1591 1592 void ObjectMonitor::INotify(Thread * Self) { 1593 const int policy = Knob_MoveNotifyee; 1594 1595 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1596 ObjectWaiter * iterator = DequeueWaiter(); 1597 if (iterator != NULL) { 1598 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1599 guarantee(iterator->_notified == 0, "invariant"); 1600 // Disposition - what might we do with iterator ? 1601 // a. add it directly to the EntryList - either tail (policy == 1) 1602 // or head (policy == 0). 1603 // b. push it onto the front of the _cxq (policy == 2). 1604 // For now we use (b). 1605 if (policy != 4) { 1606 iterator->TState = ObjectWaiter::TS_ENTER; 1607 } 1608 iterator->_notified = 1; 1609 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1610 1611 ObjectWaiter * list = _EntryList; 1612 if (list != NULL) { 1613 assert(list->_prev == NULL, "invariant"); 1614 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1615 assert(list != iterator, "invariant"); 1616 } 1617 1618 if (policy == 0) { // prepend to EntryList 1619 if (list == NULL) { 1620 iterator->_next = iterator->_prev = NULL; 1621 _EntryList = iterator; 1622 } else { 1623 list->_prev = iterator; 1624 iterator->_next = list; 1625 iterator->_prev = NULL; 1626 _EntryList = iterator; 1627 } 1628 } else if (policy == 1) { // append to EntryList 1629 if (list == NULL) { 1630 iterator->_next = iterator->_prev = NULL; 1631 _EntryList = iterator; 1632 } else { 1633 // CONSIDER: finding the tail currently requires a linear-time walk of 1634 // the EntryList. We can make tail access constant-time by converting to 1635 // a CDLL instead of using our current DLL. 1636 ObjectWaiter * tail; 1637 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1638 assert(tail != NULL && tail->_next == NULL, "invariant"); 1639 tail->_next = iterator; 1640 iterator->_prev = tail; 1641 iterator->_next = NULL; 1642 } 1643 } else if (policy == 2) { // prepend to cxq 1644 if (list == NULL) { 1645 iterator->_next = iterator->_prev = NULL; 1646 _EntryList = iterator; 1647 } else { 1648 iterator->TState = ObjectWaiter::TS_CXQ; 1649 for (;;) { 1650 ObjectWaiter * front = _cxq; 1651 iterator->_next = front; 1652 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1653 break; 1654 } 1655 } 1656 } 1657 } else if (policy == 3) { // append to cxq 1658 iterator->TState = ObjectWaiter::TS_CXQ; 1659 for (;;) { 1660 ObjectWaiter * tail = _cxq; 1661 if (tail == NULL) { 1662 iterator->_next = NULL; 1663 if (Atomic::replace_if_null(iterator, &_cxq)) { 1664 break; 1665 } 1666 } else { 1667 while (tail->_next != NULL) tail = tail->_next; 1668 tail->_next = iterator; 1669 iterator->_prev = tail; 1670 iterator->_next = NULL; 1671 break; 1672 } 1673 } 1674 } else { 1675 ParkEvent * ev = iterator->_event; 1676 iterator->TState = ObjectWaiter::TS_RUN; 1677 OrderAccess::fence(); 1678 ev->unpark(); 1679 } 1680 1681 // _WaitSetLock protects the wait queue, not the EntryList. We could 1682 // move the add-to-EntryList operation, above, outside the critical section 1683 // protected by _WaitSetLock. In practice that's not useful. With the 1684 // exception of wait() timeouts and interrupts the monitor owner 1685 // is the only thread that grabs _WaitSetLock. There's almost no contention 1686 // on _WaitSetLock so it's not profitable to reduce the length of the 1687 // critical section. 1688 1689 if (policy < 4) { 1690 iterator->wait_reenter_begin(this); 1691 } 1692 } 1693 Thread::SpinRelease(&_WaitSetLock); 1694 } 1695 1696 // Consider: a not-uncommon synchronization bug is to use notify() when 1697 // notifyAll() is more appropriate, potentially resulting in stranded 1698 // threads; this is one example of a lost wakeup. A useful diagnostic 1699 // option is to force all notify() operations to behave as notifyAll(). 1700 // 1701 // Note: We can also detect many such problems with a "minimum wait". 1702 // When the "minimum wait" is set to a small non-zero timeout value 1703 // and the program does not hang whereas it did absent "minimum wait", 1704 // that suggests a lost wakeup bug. 1705 1706 void ObjectMonitor::notify(TRAPS) { 1707 CHECK_OWNER(); 1708 if (_WaitSet == NULL) { 1709 return; 1710 } 1711 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1712 INotify(THREAD); 1713 OM_PERFDATA_OP(Notifications, inc(1)); 1714 } 1715 1716 1717 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1718 // from the waitset to the EntryList. This could be done more efficiently with a 1719 // single bulk transfer but in practice it's not time-critical. Beware too, 1720 // that in prepend-mode we invert the order of the waiters. Let's say that the 1721 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1722 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1723 1724 void ObjectMonitor::notifyAll(TRAPS) { 1725 CHECK_OWNER(); 1726 if (_WaitSet == NULL) { 1727 return; 1728 } 1729 1730 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1731 int tally = 0; 1732 while (_WaitSet != NULL) { 1733 tally++; 1734 INotify(THREAD); 1735 } 1736 1737 OM_PERFDATA_OP(Notifications, inc(tally)); 1738 } 1739 1740 // ----------------------------------------------------------------------------- 1741 // Adaptive Spinning Support 1742 // 1743 // Adaptive spin-then-block - rational spinning 1744 // 1745 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1746 // algorithm. On high order SMP systems it would be better to start with 1747 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1748 // a contending thread could enqueue itself on the cxq and then spin locally 1749 // on a thread-specific variable such as its ParkEvent._Event flag. 1750 // That's left as an exercise for the reader. Note that global spinning is 1751 // not problematic on Niagara, as the L2 cache serves the interconnect and 1752 // has both low latency and massive bandwidth. 1753 // 1754 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1755 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1756 // (duration) or we can fix the count at approximately the duration of 1757 // a context switch and vary the frequency. Of course we could also 1758 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1759 // For a description of 'Adaptive spin-then-block mutual exclusion in 1760 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1761 // 1762 // This implementation varies the duration "D", where D varies with 1763 // the success rate of recent spin attempts. (D is capped at approximately 1764 // length of a round-trip context switch). The success rate for recent 1765 // spin attempts is a good predictor of the success rate of future spin 1766 // attempts. The mechanism adapts automatically to varying critical 1767 // section length (lock modality), system load and degree of parallelism. 1768 // D is maintained per-monitor in _SpinDuration and is initialized 1769 // optimistically. Spin frequency is fixed at 100%. 1770 // 1771 // Note that _SpinDuration is volatile, but we update it without locks 1772 // or atomics. The code is designed so that _SpinDuration stays within 1773 // a reasonable range even in the presence of races. The arithmetic 1774 // operations on _SpinDuration are closed over the domain of legal values, 1775 // so at worst a race will install and older but still legal value. 1776 // At the very worst this introduces some apparent non-determinism. 1777 // We might spin when we shouldn't or vice-versa, but since the spin 1778 // count are relatively short, even in the worst case, the effect is harmless. 1779 // 1780 // Care must be taken that a low "D" value does not become an 1781 // an absorbing state. Transient spinning failures -- when spinning 1782 // is overall profitable -- should not cause the system to converge 1783 // on low "D" values. We want spinning to be stable and predictable 1784 // and fairly responsive to change and at the same time we don't want 1785 // it to oscillate, become metastable, be "too" non-deterministic, 1786 // or converge on or enter undesirable stable absorbing states. 1787 // 1788 // We implement a feedback-based control system -- using past behavior 1789 // to predict future behavior. We face two issues: (a) if the 1790 // input signal is random then the spin predictor won't provide optimal 1791 // results, and (b) if the signal frequency is too high then the control 1792 // system, which has some natural response lag, will "chase" the signal. 1793 // (b) can arise from multimodal lock hold times. Transient preemption 1794 // can also result in apparent bimodal lock hold times. 1795 // Although sub-optimal, neither condition is particularly harmful, as 1796 // in the worst-case we'll spin when we shouldn't or vice-versa. 1797 // The maximum spin duration is rather short so the failure modes aren't bad. 1798 // To be conservative, I've tuned the gain in system to bias toward 1799 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1800 // "rings" or oscillates between spinning and not spinning. This happens 1801 // when spinning is just on the cusp of profitability, however, so the 1802 // situation is not dire. The state is benign -- there's no need to add 1803 // hysteresis control to damp the transition rate between spinning and 1804 // not spinning. 1805 1806 // Spinning: Fixed frequency (100%), vary duration 1807 int ObjectMonitor::TrySpin(Thread * Self) { 1808 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1809 int ctr = Knob_FixedSpin; 1810 if (ctr != 0) { 1811 while (--ctr >= 0) { 1812 if (TryLock(Self) > 0) return 1; 1813 SpinPause(); 1814 } 1815 return 0; 1816 } 1817 1818 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1819 if (TryLock(Self) > 0) { 1820 // Increase _SpinDuration ... 1821 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1822 // Raising _SpurDuration to the poverty line is key. 1823 int x = _SpinDuration; 1824 if (x < Knob_SpinLimit) { 1825 if (x < Knob_Poverty) x = Knob_Poverty; 1826 _SpinDuration = x + Knob_BonusB; 1827 } 1828 return 1; 1829 } 1830 SpinPause(); 1831 } 1832 1833 // Admission control - verify preconditions for spinning 1834 // 1835 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1836 // becoming an absorbing state. Put another way, we spin briefly to 1837 // sample, just in case the system load, parallelism, contention, or lock 1838 // modality changed. 1839 // 1840 // Consider the following alternative: 1841 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1842 // spin attempt. "Periodically" might mean after a tally of 1843 // the # of failed spin attempts (or iterations) reaches some threshold. 1844 // This takes us into the realm of 1-out-of-N spinning, where we 1845 // hold the duration constant but vary the frequency. 1846 1847 ctr = _SpinDuration; 1848 if (ctr <= 0) return 0; 1849 1850 if (Knob_SuccRestrict && _succ != NULL) return 0; 1851 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1852 return 0; 1853 } 1854 1855 int MaxSpin = Knob_MaxSpinners; 1856 if (MaxSpin >= 0) { 1857 if (_Spinner > MaxSpin) { 1858 return 0; 1859 } 1860 // Slightly racy, but benign ... 1861 Adjust(&_Spinner, 1); 1862 } 1863 1864 // We're good to spin ... spin ingress. 1865 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1866 // when preparing to LD...CAS _owner, etc and the CAS is likely 1867 // to succeed. 1868 int caspty = Knob_CASPenalty; 1869 int oxpty = Knob_OXPenalty; 1870 int sss = Knob_SpinSetSucc; 1871 if (sss && _succ == NULL) _succ = Self; 1872 Thread * prv = NULL; 1873 1874 // There are three ways to exit the following loop: 1875 // 1. A successful spin where this thread has acquired the lock. 1876 // 2. Spin failure with prejudice 1877 // 3. Spin failure without prejudice 1878 1879 while (--ctr >= 0) { 1880 1881 // Periodic polling -- Check for pending GC 1882 // Threads may spin while they're unsafe. 1883 // We don't want spinning threads to delay the JVM from reaching 1884 // a stop-the-world safepoint or to steal cycles from GC. 1885 // If we detect a pending safepoint we abort in order that 1886 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1887 // this thread, if safe, doesn't steal cycles from GC. 1888 // This is in keeping with the "no loitering in runtime" rule. 1889 // We periodically check to see if there's a safepoint pending. 1890 if ((ctr & 0xFF) == 0) { 1891 if (SafepointMechanism::poll(Self)) { 1892 goto Abort; // abrupt spin egress 1893 } 1894 if (Knob_UsePause & 1) SpinPause(); 1895 } 1896 1897 if (Knob_UsePause & 2) SpinPause(); 1898 1899 // Probe _owner with TATAS 1900 // If this thread observes the monitor transition or flicker 1901 // from locked to unlocked to locked, then the odds that this 1902 // thread will acquire the lock in this spin attempt go down 1903 // considerably. The same argument applies if the CAS fails 1904 // or if we observe _owner change from one non-null value to 1905 // another non-null value. In such cases we might abort 1906 // the spin without prejudice or apply a "penalty" to the 1907 // spin count-down variable "ctr", reducing it by 100, say. 1908 1909 Thread * ox = (Thread *) _owner; 1910 if (ox == NULL) { 1911 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1912 if (ox == NULL) { 1913 // The CAS succeeded -- this thread acquired ownership 1914 // Take care of some bookkeeping to exit spin state. 1915 if (sss && _succ == Self) { 1916 _succ = NULL; 1917 } 1918 if (MaxSpin > 0) Adjust(&_Spinner, -1); 1919 1920 // Increase _SpinDuration : 1921 // The spin was successful (profitable) so we tend toward 1922 // longer spin attempts in the future. 1923 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1924 // If we acquired the lock early in the spin cycle it 1925 // makes sense to increase _SpinDuration proportionally. 1926 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1927 int x = _SpinDuration; 1928 if (x < Knob_SpinLimit) { 1929 if (x < Knob_Poverty) x = Knob_Poverty; 1930 _SpinDuration = x + Knob_Bonus; 1931 } 1932 return 1; 1933 } 1934 1935 // The CAS failed ... we can take any of the following actions: 1936 // * penalize: ctr -= Knob_CASPenalty 1937 // * exit spin with prejudice -- goto Abort; 1938 // * exit spin without prejudice. 1939 // * Since CAS is high-latency, retry again immediately. 1940 prv = ox; 1941 if (caspty == -2) break; 1942 if (caspty == -1) goto Abort; 1943 ctr -= caspty; 1944 continue; 1945 } 1946 1947 // Did lock ownership change hands ? 1948 if (ox != prv && prv != NULL) { 1949 if (oxpty == -2) break; 1950 if (oxpty == -1) goto Abort; 1951 ctr -= oxpty; 1952 } 1953 prv = ox; 1954 1955 // Abort the spin if the owner is not executing. 1956 // The owner must be executing in order to drop the lock. 1957 // Spinning while the owner is OFFPROC is idiocy. 1958 // Consider: ctr -= RunnablePenalty ; 1959 if (Knob_OState && NotRunnable (Self, ox)) { 1960 goto Abort; 1961 } 1962 if (sss && _succ == NULL) _succ = Self; 1963 } 1964 1965 // Spin failed with prejudice -- reduce _SpinDuration. 1966 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1967 // AIMD is globally stable. 1968 { 1969 int x = _SpinDuration; 1970 if (x > 0) { 1971 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1972 // This is globally sample and tends to damp the response. 1973 x -= Knob_Penalty; 1974 if (x < 0) x = 0; 1975 _SpinDuration = x; 1976 } 1977 } 1978 1979 Abort: 1980 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 1981 if (sss && _succ == Self) { 1982 _succ = NULL; 1983 // Invariant: after setting succ=null a contending thread 1984 // must recheck-retry _owner before parking. This usually happens 1985 // in the normal usage of TrySpin(), but it's safest 1986 // to make TrySpin() as foolproof as possible. 1987 OrderAccess::fence(); 1988 if (TryLock(Self) > 0) return 1; 1989 } 1990 return 0; 1991 } 1992 1993 // NotRunnable() -- informed spinning 1994 // 1995 // Don't bother spinning if the owner is not eligible to drop the lock. 1996 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1997 // The thread must be runnable in order to drop the lock in timely fashion. 1998 // If the _owner is not runnable then spinning will not likely be 1999 // successful (profitable). 2000 // 2001 // Beware -- the thread referenced by _owner could have died 2002 // so a simply fetch from _owner->_thread_state might trap. 2003 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2004 // Because of the lifecycle issues, the _thread_state values 2005 // observed by NotRunnable() might be garbage. NotRunnable must 2006 // tolerate this and consider the observed _thread_state value 2007 // as advisory. 2008 // 2009 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2010 // a thread pointer. 2011 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2012 // with the LSB of _owner. Another option would be to probabilistically probe 2013 // the putative _owner->TypeTag value. 2014 // 2015 // Checking _thread_state isn't perfect. Even if the thread is 2016 // in_java it might be blocked on a page-fault or have been preempted 2017 // and sitting on a ready/dispatch queue. 2018 // 2019 // The return value from NotRunnable() is *advisory* -- the 2020 // result is based on sampling and is not necessarily coherent. 2021 // The caller must tolerate false-negative and false-positive errors. 2022 // Spinning, in general, is probabilistic anyway. 2023 2024 2025 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2026 // Check ox->TypeTag == 2BAD. 2027 if (ox == NULL) return 0; 2028 2029 // Avoid transitive spinning ... 2030 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2031 // Immediately after T1 acquires L it's possible that T2, also 2032 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2033 // This occurs transiently after T1 acquired L but before 2034 // T1 managed to clear T1.Stalled. T2 does not need to abort 2035 // its spin in this circumstance. 2036 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2037 2038 if (BlockedOn == 1) return 1; 2039 if (BlockedOn != 0) { 2040 return BlockedOn != intptr_t(this) && _owner == ox; 2041 } 2042 2043 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2044 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2045 // consider also: jst != _thread_in_Java -- but that's overspecific. 2046 return jst == _thread_blocked || jst == _thread_in_native; 2047 } 2048 2049 2050 // ----------------------------------------------------------------------------- 2051 // WaitSet management ... 2052 2053 ObjectWaiter::ObjectWaiter(Thread* thread) { 2054 _next = NULL; 2055 _prev = NULL; 2056 _notified = 0; 2057 _notifier_tid = 0; 2058 TState = TS_RUN; 2059 _thread = thread; 2060 _event = thread->_ParkEvent; 2061 _active = false; 2062 assert(_event != NULL, "invariant"); 2063 } 2064 2065 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2066 JavaThread *jt = (JavaThread *)this->_thread; 2067 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2068 } 2069 2070 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2071 JavaThread *jt = (JavaThread *)this->_thread; 2072 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2073 } 2074 2075 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2076 assert(node != NULL, "should not add NULL node"); 2077 assert(node->_prev == NULL, "node already in list"); 2078 assert(node->_next == NULL, "node already in list"); 2079 // put node at end of queue (circular doubly linked list) 2080 if (_WaitSet == NULL) { 2081 _WaitSet = node; 2082 node->_prev = node; 2083 node->_next = node; 2084 } else { 2085 ObjectWaiter* head = _WaitSet; 2086 ObjectWaiter* tail = head->_prev; 2087 assert(tail->_next == head, "invariant check"); 2088 tail->_next = node; 2089 head->_prev = node; 2090 node->_next = head; 2091 node->_prev = tail; 2092 } 2093 } 2094 2095 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2096 // dequeue the very first waiter 2097 ObjectWaiter* waiter = _WaitSet; 2098 if (waiter) { 2099 DequeueSpecificWaiter(waiter); 2100 } 2101 return waiter; 2102 } 2103 2104 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2105 assert(node != NULL, "should not dequeue NULL node"); 2106 assert(node->_prev != NULL, "node already removed from list"); 2107 assert(node->_next != NULL, "node already removed from list"); 2108 // when the waiter has woken up because of interrupt, 2109 // timeout or other spurious wake-up, dequeue the 2110 // waiter from waiting list 2111 ObjectWaiter* next = node->_next; 2112 if (next == node) { 2113 assert(node->_prev == node, "invariant check"); 2114 _WaitSet = NULL; 2115 } else { 2116 ObjectWaiter* prev = node->_prev; 2117 assert(prev->_next == node, "invariant check"); 2118 assert(next->_prev == node, "invariant check"); 2119 next->_prev = prev; 2120 prev->_next = next; 2121 if (_WaitSet == node) { 2122 _WaitSet = next; 2123 } 2124 } 2125 node->_next = NULL; 2126 node->_prev = NULL; 2127 } 2128 2129 // ----------------------------------------------------------------------------- 2130 // PerfData support 2131 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2132 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2133 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2134 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2135 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2136 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2137 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2138 2139 // One-shot global initialization for the sync subsystem. 2140 // We could also defer initialization and initialize on-demand 2141 // the first time we call inflate(). Initialization would 2142 // be protected - like so many things - by the MonitorCache_lock. 2143 2144 void ObjectMonitor::Initialize() { 2145 static int InitializationCompleted = 0; 2146 assert(InitializationCompleted == 0, "invariant"); 2147 InitializationCompleted = 1; 2148 if (UsePerfData) { 2149 EXCEPTION_MARK; 2150 #define NEWPERFCOUNTER(n) \ 2151 { \ 2152 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2153 CHECK); \ 2154 } 2155 #define NEWPERFVARIABLE(n) \ 2156 { \ 2157 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2158 CHECK); \ 2159 } 2160 NEWPERFCOUNTER(_sync_Inflations); 2161 NEWPERFCOUNTER(_sync_Deflations); 2162 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2163 NEWPERFCOUNTER(_sync_FutileWakeups); 2164 NEWPERFCOUNTER(_sync_Parks); 2165 NEWPERFCOUNTER(_sync_Notifications); 2166 NEWPERFVARIABLE(_sync_MonExtant); 2167 #undef NEWPERFCOUNTER 2168 #undef NEWPERFVARIABLE 2169 } 2170 } 2171 2172 void ObjectMonitor::DeferredInitialize() { 2173 if (InitDone > 0) return; 2174 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2175 while (InitDone != 1) /* empty */; 2176 return; 2177 } 2178 2179 // One-shot global initialization ... 2180 // The initialization is idempotent, so we don't need locks. 2181 // In the future consider doing this via os::init_2(). 2182 2183 if (!os::is_MP()) { 2184 Knob_SpinLimit = 0; 2185 Knob_PreSpin = 0; 2186 Knob_FixedSpin = -1; 2187 } 2188 2189 OrderAccess::fence(); 2190 InitDone = 1; 2191 } 2192