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_VerifyMatch = 0; 105 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 106 107 static int Knob_SpinBase = 0; // Floor AKA SpinMin 108 static int Knob_CASPenalty = -1; // Penalty for failed CAS 109 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 110 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 111 static int Knob_SpinEarly = 1; 112 static int Knob_SuccEnabled = 1; // futile wake throttling 113 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 114 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 115 static int Knob_Bonus = 100; // spin success bonus 116 static int Knob_BonusB = 100; // spin success bonus 117 static int Knob_Penalty = 200; // spin failure penalty 118 static int Knob_Poverty = 1000; 119 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 120 static int Knob_FixedSpin = 0; 121 static int Knob_OState = 3; // Spinner checks thread state of _owner 122 static int Knob_UsePause = 1; 123 static int Knob_ExitPolicy = 0; 124 static int Knob_PreSpin = 10; // 20-100 likely better 125 static int Knob_ResetEvent = 0; 126 127 static int Knob_FastHSSEC = 0; 128 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 129 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 130 static volatile int InitDone = 0; 131 132 // ----------------------------------------------------------------------------- 133 // Theory of operations -- Monitors lists, thread residency, etc: 134 // 135 // * A thread acquires ownership of a monitor by successfully 136 // CAS()ing the _owner field from null to non-null. 137 // 138 // * Invariant: A thread appears on at most one monitor list -- 139 // cxq, EntryList or WaitSet -- at any one time. 140 // 141 // * Contending threads "push" themselves onto the cxq with CAS 142 // and then spin/park. 143 // 144 // * After a contending thread eventually acquires the lock it must 145 // dequeue itself from either the EntryList or the cxq. 146 // 147 // * The exiting thread identifies and unparks an "heir presumptive" 148 // tentative successor thread on the EntryList. Critically, the 149 // exiting thread doesn't unlink the successor thread from the EntryList. 150 // After having been unparked, the wakee will recontend for ownership of 151 // the monitor. The successor (wakee) will either acquire the lock or 152 // re-park itself. 153 // 154 // Succession is provided for by a policy of competitive handoff. 155 // The exiting thread does _not_ grant or pass ownership to the 156 // successor thread. (This is also referred to as "handoff" succession"). 157 // Instead the exiting thread releases ownership and possibly wakes 158 // a successor, so the successor can (re)compete for ownership of the lock. 159 // If the EntryList is empty but the cxq is populated the exiting 160 // thread will drain the cxq into the EntryList. It does so by 161 // by detaching the cxq (installing null with CAS) and folding 162 // the threads from the cxq into the EntryList. The EntryList is 163 // doubly linked, while the cxq is singly linked because of the 164 // CAS-based "push" used to enqueue recently arrived threads (RATs). 165 // 166 // * Concurrency invariants: 167 // 168 // -- only the monitor owner may access or mutate the EntryList. 169 // The mutex property of the monitor itself protects the EntryList 170 // from concurrent interference. 171 // -- Only the monitor owner may detach the cxq. 172 // 173 // * The monitor entry list operations avoid locks, but strictly speaking 174 // they're not lock-free. Enter is lock-free, exit is not. 175 // For a description of 'Methods and apparatus providing non-blocking access 176 // to a resource,' see U.S. Pat. No. 7844973. 177 // 178 // * The cxq can have multiple concurrent "pushers" but only one concurrent 179 // detaching thread. This mechanism is immune from the ABA corruption. 180 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 181 // 182 // * Taken together, the cxq and the EntryList constitute or form a 183 // single logical queue of threads stalled trying to acquire the lock. 184 // We use two distinct lists to improve the odds of a constant-time 185 // dequeue operation after acquisition (in the ::enter() epilogue) and 186 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 187 // A key desideratum is to minimize queue & monitor metadata manipulation 188 // that occurs while holding the monitor lock -- that is, we want to 189 // minimize monitor lock holds times. Note that even a small amount of 190 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 191 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 192 // locks and monitor metadata. 193 // 194 // Cxq points to the set of Recently Arrived Threads attempting entry. 195 // Because we push threads onto _cxq with CAS, the RATs must take the form of 196 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 197 // the unlocking thread notices that EntryList is null but _cxq is != null. 198 // 199 // The EntryList is ordered by the prevailing queue discipline and 200 // can be organized in any convenient fashion, such as a doubly-linked list or 201 // a circular doubly-linked list. Critically, we want insert and delete operations 202 // to operate in constant-time. If we need a priority queue then something akin 203 // to Solaris' sleepq would work nicely. Viz., 204 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 205 // Queue discipline is enforced at ::exit() time, when the unlocking thread 206 // drains the cxq into the EntryList, and orders or reorders the threads on the 207 // EntryList accordingly. 208 // 209 // Barring "lock barging", this mechanism provides fair cyclic ordering, 210 // somewhat similar to an elevator-scan. 211 // 212 // * The monitor synchronization subsystem avoids the use of native 213 // synchronization primitives except for the narrow platform-specific 214 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 215 // the semantics of park-unpark. Put another way, this monitor implementation 216 // depends only on atomic operations and park-unpark. The monitor subsystem 217 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 218 // underlying OS manages the READY<->RUN transitions. 219 // 220 // * Waiting threads reside on the WaitSet list -- wait() puts 221 // the caller onto the WaitSet. 222 // 223 // * notify() or notifyAll() simply transfers threads from the WaitSet to 224 // either the EntryList or cxq. Subsequent exit() operations will 225 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 226 // it's likely the notifyee would simply impale itself on the lock held 227 // by the notifier. 228 // 229 // * An interesting alternative is to encode cxq as (List,LockByte) where 230 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 231 // variable, like _recursions, in the scheme. The threads or Events that form 232 // the list would have to be aligned in 256-byte addresses. A thread would 233 // try to acquire the lock or enqueue itself with CAS, but exiting threads 234 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 235 // Note that is is *not* word-tearing, but it does presume that full-word 236 // CAS operations are coherent with intermix with STB operations. That's true 237 // on most common processors. 238 // 239 // * See also http://blogs.sun.com/dave 240 241 242 void* ObjectMonitor::operator new (size_t size) throw() { 243 return AllocateHeap(size, mtInternal); 244 } 245 void* ObjectMonitor::operator new[] (size_t size) throw() { 246 return operator new (size); 247 } 248 void ObjectMonitor::operator delete(void* p) { 249 FreeHeap(p); 250 } 251 void ObjectMonitor::operator delete[] (void *p) { 252 operator delete(p); 253 } 254 255 // ----------------------------------------------------------------------------- 256 // Enter support 257 258 void ObjectMonitor::enter(TRAPS) { 259 // The following code is ordered to check the most common cases first 260 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 261 Thread * const Self = THREAD; 262 263 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 264 if (cur == NULL) { 265 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 266 assert(_recursions == 0, "invariant"); 267 assert(_owner == Self, "invariant"); 268 return; 269 } 270 271 if (cur == Self) { 272 // TODO-FIXME: check for integer overflow! BUGID 6557169. 273 _recursions++; 274 return; 275 } 276 277 if (Self->is_lock_owned ((address)cur)) { 278 assert(_recursions == 0, "internal state error"); 279 _recursions = 1; 280 // Commute owner from a thread-specific on-stack BasicLockObject address to 281 // a full-fledged "Thread *". 282 _owner = Self; 283 return; 284 } 285 286 // We've encountered genuine contention. 287 assert(Self->_Stalled == 0, "invariant"); 288 Self->_Stalled = intptr_t(this); 289 290 // Try one round of spinning *before* enqueueing Self 291 // and before going through the awkward and expensive state 292 // transitions. The following spin is strictly optional ... 293 // Note that if we acquire the monitor from an initial spin 294 // we forgo posting JVMTI events and firing DTRACE probes. 295 if (Knob_SpinEarly && TrySpin (Self) > 0) { 296 assert(_owner == Self, "invariant"); 297 assert(_recursions == 0, "invariant"); 298 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 299 Self->_Stalled = 0; 300 return; 301 } 302 303 assert(_owner != Self, "invariant"); 304 assert(_succ != Self, "invariant"); 305 assert(Self->is_Java_thread(), "invariant"); 306 JavaThread * jt = (JavaThread *) Self; 307 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 308 assert(jt->thread_state() != _thread_blocked, "invariant"); 309 assert(this->object() != NULL, "invariant"); 310 assert(_count >= 0, "invariant"); 311 312 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 313 // Ensure the object-monitor relationship remains stable while there's contention. 314 Atomic::inc(&_count); 315 316 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 317 EventJavaMonitorEnter event; 318 if (event.should_commit()) { 319 event.set_monitorClass(((oop)this->object())->klass()); 320 event.set_address((uintptr_t)(this->object_addr())); 321 } 322 323 { // Change java thread status to indicate blocked on monitor enter. 324 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 325 326 Self->set_current_pending_monitor(this); 327 328 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 329 if (JvmtiExport::should_post_monitor_contended_enter()) { 330 JvmtiExport::post_monitor_contended_enter(jt, this); 331 332 // The current thread does not yet own the monitor and does not 333 // yet appear on any queues that would get it made the successor. 334 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 335 // handler cannot accidentally consume an unpark() meant for the 336 // ParkEvent associated with this ObjectMonitor. 337 } 338 339 OSThreadContendState osts(Self->osthread()); 340 ThreadBlockInVM tbivm(jt); 341 342 // TODO-FIXME: change the following for(;;) loop to straight-line code. 343 for (;;) { 344 jt->set_suspend_equivalent(); 345 // cleared by handle_special_suspend_equivalent_condition() 346 // or java_suspend_self() 347 348 EnterI(THREAD); 349 350 if (!ExitSuspendEquivalent(jt)) break; 351 352 // We have acquired the contended monitor, but while we were 353 // waiting another thread suspended us. We don't want to enter 354 // the monitor while suspended because that would surprise the 355 // thread that suspended us. 356 // 357 _recursions = 0; 358 _succ = NULL; 359 exit(false, Self); 360 361 jt->java_suspend_self(); 362 } 363 Self->set_current_pending_monitor(NULL); 364 365 // We cleared the pending monitor info since we've just gotten past 366 // the enter-check-for-suspend dance and we now own the monitor free 367 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 368 // destructor can go to a safepoint at the end of this block. If we 369 // do a thread dump during that safepoint, then this thread will show 370 // as having "-locked" the monitor, but the OS and java.lang.Thread 371 // states will still report that the thread is blocked trying to 372 // acquire it. 373 } 374 375 Atomic::dec(&_count); 376 assert(_count >= 0, "invariant"); 377 Self->_Stalled = 0; 378 379 // Must either set _recursions = 0 or ASSERT _recursions == 0. 380 assert(_recursions == 0, "invariant"); 381 assert(_owner == Self, "invariant"); 382 assert(_succ != Self, "invariant"); 383 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 384 385 // The thread -- now the owner -- is back in vm mode. 386 // Report the glorious news via TI,DTrace and jvmstat. 387 // The probe effect is non-trivial. All the reportage occurs 388 // while we hold the monitor, increasing the length of the critical 389 // section. Amdahl's parallel speedup law comes vividly into play. 390 // 391 // Another option might be to aggregate the events (thread local or 392 // per-monitor aggregation) and defer reporting until a more opportune 393 // time -- such as next time some thread encounters contention but has 394 // yet to acquire the lock. While spinning that thread could 395 // spinning we could increment JVMStat counters, etc. 396 397 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 398 if (JvmtiExport::should_post_monitor_contended_entered()) { 399 JvmtiExport::post_monitor_contended_entered(jt, this); 400 401 // The current thread already owns the monitor and is not going to 402 // call park() for the remainder of the monitor enter protocol. So 403 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 404 // event handler consumed an unpark() issued by the thread that 405 // just exited the monitor. 406 } 407 if (event.should_commit()) { 408 event.set_previousOwner((uintptr_t)_previous_owner_tid); 409 event.commit(); 410 } 411 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 412 } 413 414 // Caveat: TryLock() is not necessarily serializing if it returns failure. 415 // Callers must compensate as needed. 416 417 int ObjectMonitor::TryLock(Thread * Self) { 418 void * own = _owner; 419 if (own != NULL) return 0; 420 if (Atomic::replace_if_null(Self, &_owner)) { 421 // Either guarantee _recursions == 0 or set _recursions = 0. 422 assert(_recursions == 0, "invariant"); 423 assert(_owner == Self, "invariant"); 424 return 1; 425 } 426 // The lock had been free momentarily, but we lost the race to the lock. 427 // Interference -- the CAS failed. 428 // We can either return -1 or retry. 429 // Retry doesn't make as much sense because the lock was just acquired. 430 return -1; 431 } 432 433 #define MAX_RECHECK_INTERVAL 1000 434 435 void ObjectMonitor::EnterI(TRAPS) { 436 Thread * const Self = THREAD; 437 assert(Self->is_Java_thread(), "invariant"); 438 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 439 440 // Try the lock - TATAS 441 if (TryLock (Self) > 0) { 442 assert(_succ != Self, "invariant"); 443 assert(_owner == Self, "invariant"); 444 assert(_Responsible != Self, "invariant"); 445 return; 446 } 447 448 DeferredInitialize(); 449 450 // We try one round of spinning *before* enqueueing Self. 451 // 452 // If the _owner is ready but OFFPROC we could use a YieldTo() 453 // operation to donate the remainder of this thread's quantum 454 // to the owner. This has subtle but beneficial affinity 455 // effects. 456 457 if (TrySpin (Self) > 0) { 458 assert(_owner == Self, "invariant"); 459 assert(_succ != Self, "invariant"); 460 assert(_Responsible != Self, "invariant"); 461 return; 462 } 463 464 // The Spin failed -- Enqueue and park the thread ... 465 assert(_succ != Self, "invariant"); 466 assert(_owner != Self, "invariant"); 467 assert(_Responsible != Self, "invariant"); 468 469 // Enqueue "Self" on ObjectMonitor's _cxq. 470 // 471 // Node acts as a proxy for Self. 472 // As an aside, if were to ever rewrite the synchronization code mostly 473 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 474 // Java objects. This would avoid awkward lifecycle and liveness issues, 475 // as well as eliminate a subset of ABA issues. 476 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 477 478 ObjectWaiter node(Self); 479 Self->_ParkEvent->reset(); 480 node._prev = (ObjectWaiter *) 0xBAD; 481 node.TState = ObjectWaiter::TS_CXQ; 482 483 // Push "Self" onto the front of the _cxq. 484 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 485 // Note that spinning tends to reduce the rate at which threads 486 // enqueue and dequeue on EntryList|cxq. 487 ObjectWaiter * nxt; 488 for (;;) { 489 node._next = nxt = _cxq; 490 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 491 492 // Interference - the CAS failed because _cxq changed. Just retry. 493 // As an optional optimization we retry the lock. 494 if (TryLock (Self) > 0) { 495 assert(_succ != Self, "invariant"); 496 assert(_owner == Self, "invariant"); 497 assert(_Responsible != Self, "invariant"); 498 return; 499 } 500 } 501 502 // Check for cxq|EntryList edge transition to non-null. This indicates 503 // the onset of contention. While contention persists exiting threads 504 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 505 // operations revert to the faster 1-0 mode. This enter operation may interleave 506 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 507 // arrange for one of the contending thread to use a timed park() operations 508 // to detect and recover from the race. (Stranding is form of progress failure 509 // where the monitor is unlocked but all the contending threads remain parked). 510 // That is, at least one of the contended threads will periodically poll _owner. 511 // One of the contending threads will become the designated "Responsible" thread. 512 // The Responsible thread uses a timed park instead of a normal indefinite park 513 // operation -- it periodically wakes and checks for and recovers from potential 514 // strandings admitted by 1-0 exit operations. We need at most one Responsible 515 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 516 // be responsible for a monitor. 517 // 518 // Currently, one of the contended threads takes on the added role of "Responsible". 519 // A viable alternative would be to use a dedicated "stranding checker" thread 520 // that periodically iterated over all the threads (or active monitors) and unparked 521 // successors where there was risk of stranding. This would help eliminate the 522 // timer scalability issues we see on some platforms as we'd only have one thread 523 // -- the checker -- parked on a timer. 524 525 if (nxt == NULL && _EntryList == NULL) { 526 // Try to assume the role of responsible thread for the monitor. 527 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 528 Atomic::replace_if_null(Self, &_Responsible); 529 } 530 531 // The lock might have been released while this thread was occupied queueing 532 // itself onto _cxq. To close the race and avoid "stranding" and 533 // progress-liveness failure we must resample-retry _owner before parking. 534 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 535 // In this case the ST-MEMBAR is accomplished with CAS(). 536 // 537 // TODO: Defer all thread state transitions until park-time. 538 // Since state transitions are heavy and inefficient we'd like 539 // to defer the state transitions until absolutely necessary, 540 // and in doing so avoid some transitions ... 541 542 int nWakeups = 0; 543 int recheckInterval = 1; 544 545 for (;;) { 546 547 if (TryLock(Self) > 0) break; 548 assert(_owner != Self, "invariant"); 549 550 // park self 551 if (_Responsible == Self) { 552 Self->_ParkEvent->park((jlong) recheckInterval); 553 // Increase the recheckInterval, but clamp the value. 554 recheckInterval *= 8; 555 if (recheckInterval > MAX_RECHECK_INTERVAL) { 556 recheckInterval = MAX_RECHECK_INTERVAL; 557 } 558 } else { 559 Self->_ParkEvent->park(); 560 } 561 562 if (TryLock(Self) > 0) break; 563 564 // The lock is still contested. 565 // Keep a tally of the # of futile wakeups. 566 // Note that the counter is not protected by a lock or updated by atomics. 567 // That is by design - we trade "lossy" counters which are exposed to 568 // races during updates for a lower probe effect. 569 570 // This PerfData object can be used in parallel with a safepoint. 571 // See the work around in PerfDataManager::destroy(). 572 OM_PERFDATA_OP(FutileWakeups, inc()); 573 ++nWakeups; 574 575 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 576 // We can defer clearing _succ until after the spin completes 577 // TrySpin() must tolerate being called with _succ == Self. 578 // Try yet another round of adaptive spinning. 579 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 580 581 // We can find that we were unpark()ed and redesignated _succ while 582 // we were spinning. That's harmless. If we iterate and call park(), 583 // park() will consume the event and return immediately and we'll 584 // just spin again. This pattern can repeat, leaving _succ to simply 585 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 586 // Alternately, we can sample fired() here, and if set, forgo spinning 587 // in the next iteration. 588 589 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 590 Self->_ParkEvent->reset(); 591 OrderAccess::fence(); 592 } 593 if (_succ == Self) _succ = NULL; 594 595 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 596 OrderAccess::fence(); 597 } 598 599 // Egress : 600 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 601 // Normally we'll find Self on the EntryList . 602 // From the perspective of the lock owner (this thread), the 603 // EntryList is stable and cxq is prepend-only. 604 // The head of cxq is volatile but the interior is stable. 605 // In addition, Self.TState is stable. 606 607 assert(_owner == Self, "invariant"); 608 assert(object() != NULL, "invariant"); 609 // I'd like to write: 610 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 611 // but as we're at a safepoint that's not safe. 612 613 UnlinkAfterAcquire(Self, &node); 614 if (_succ == Self) _succ = NULL; 615 616 assert(_succ != Self, "invariant"); 617 if (_Responsible == Self) { 618 _Responsible = NULL; 619 OrderAccess::fence(); // Dekker pivot-point 620 621 // We may leave threads on cxq|EntryList without a designated 622 // "Responsible" thread. This is benign. When this thread subsequently 623 // exits the monitor it can "see" such preexisting "old" threads -- 624 // threads that arrived on the cxq|EntryList before the fence, above -- 625 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 626 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 627 // non-null and elect a new "Responsible" timer thread. 628 // 629 // This thread executes: 630 // ST Responsible=null; MEMBAR (in enter epilogue - here) 631 // LD cxq|EntryList (in subsequent exit) 632 // 633 // Entering threads in the slow/contended path execute: 634 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 635 // The (ST cxq; MEMBAR) is accomplished with CAS(). 636 // 637 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 638 // exit operation from floating above the ST Responsible=null. 639 } 640 641 // We've acquired ownership with CAS(). 642 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 643 // But since the CAS() this thread may have also stored into _succ, 644 // EntryList, cxq or Responsible. These meta-data updates must be 645 // visible __before this thread subsequently drops the lock. 646 // Consider what could occur if we didn't enforce this constraint -- 647 // STs to monitor meta-data and user-data could reorder with (become 648 // visible after) the ST in exit that drops ownership of the lock. 649 // Some other thread could then acquire the lock, but observe inconsistent 650 // or old monitor meta-data and heap data. That violates the JMM. 651 // To that end, the 1-0 exit() operation must have at least STST|LDST 652 // "release" barrier semantics. Specifically, there must be at least a 653 // STST|LDST barrier in exit() before the ST of null into _owner that drops 654 // the lock. The barrier ensures that changes to monitor meta-data and data 655 // protected by the lock will be visible before we release the lock, and 656 // therefore before some other thread (CPU) has a chance to acquire the lock. 657 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 658 // 659 // Critically, any prior STs to _succ or EntryList must be visible before 660 // the ST of null into _owner in the *subsequent* (following) corresponding 661 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 662 // execute a serializing instruction. 663 664 return; 665 } 666 667 // ReenterI() is a specialized inline form of the latter half of the 668 // contended slow-path from EnterI(). We use ReenterI() only for 669 // monitor reentry in wait(). 670 // 671 // In the future we should reconcile EnterI() and ReenterI(), adding 672 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 673 // loop accordingly. 674 675 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 676 assert(Self != NULL, "invariant"); 677 assert(SelfNode != NULL, "invariant"); 678 assert(SelfNode->_thread == Self, "invariant"); 679 assert(_waiters > 0, "invariant"); 680 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 681 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 682 JavaThread * jt = (JavaThread *) Self; 683 684 int nWakeups = 0; 685 for (;;) { 686 ObjectWaiter::TStates v = SelfNode->TState; 687 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 688 assert(_owner != Self, "invariant"); 689 690 if (TryLock(Self) > 0) break; 691 if (TrySpin(Self) > 0) break; 692 693 // State transition wrappers around park() ... 694 // ReenterI() wisely defers state transitions until 695 // it's clear we must park the thread. 696 { 697 OSThreadContendState osts(Self->osthread()); 698 ThreadBlockInVM tbivm(jt); 699 700 // cleared by handle_special_suspend_equivalent_condition() 701 // or java_suspend_self() 702 jt->set_suspend_equivalent(); 703 Self->_ParkEvent->park(); 704 705 // were we externally suspended while we were waiting? 706 for (;;) { 707 if (!ExitSuspendEquivalent(jt)) break; 708 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 709 jt->java_suspend_self(); 710 jt->set_suspend_equivalent(); 711 } 712 } 713 714 // Try again, but just so we distinguish between futile wakeups and 715 // successful wakeups. The following test isn't algorithmically 716 // necessary, but it helps us maintain sensible statistics. 717 if (TryLock(Self) > 0) break; 718 719 // The lock is still contested. 720 // Keep a tally of the # of futile wakeups. 721 // Note that the counter is not protected by a lock or updated by atomics. 722 // That is by design - we trade "lossy" counters which are exposed to 723 // races during updates for a lower probe effect. 724 ++nWakeups; 725 726 // Assuming this is not a spurious wakeup we'll normally 727 // find that _succ == Self. 728 if (_succ == Self) _succ = NULL; 729 730 // Invariant: after clearing _succ a contending thread 731 // *must* retry _owner before parking. 732 OrderAccess::fence(); 733 734 // This PerfData object can be used in parallel with a safepoint. 735 // See the work around in PerfDataManager::destroy(). 736 OM_PERFDATA_OP(FutileWakeups, inc()); 737 } 738 739 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 740 // Normally we'll find Self on the EntryList. 741 // Unlinking from the EntryList is constant-time and atomic-free. 742 // From the perspective of the lock owner (this thread), the 743 // EntryList is stable and cxq is prepend-only. 744 // The head of cxq is volatile but the interior is stable. 745 // In addition, Self.TState is stable. 746 747 assert(_owner == Self, "invariant"); 748 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 749 UnlinkAfterAcquire(Self, SelfNode); 750 if (_succ == Self) _succ = NULL; 751 assert(_succ != Self, "invariant"); 752 SelfNode->TState = ObjectWaiter::TS_RUN; 753 OrderAccess::fence(); // see comments at the end of EnterI() 754 } 755 756 // By convention we unlink a contending thread from EntryList|cxq immediately 757 // after the thread acquires the lock in ::enter(). Equally, we could defer 758 // unlinking the thread until ::exit()-time. 759 760 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 761 assert(_owner == Self, "invariant"); 762 assert(SelfNode->_thread == Self, "invariant"); 763 764 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 765 // Normal case: remove Self from the DLL EntryList . 766 // This is a constant-time operation. 767 ObjectWaiter * nxt = SelfNode->_next; 768 ObjectWaiter * prv = SelfNode->_prev; 769 if (nxt != NULL) nxt->_prev = prv; 770 if (prv != NULL) prv->_next = nxt; 771 if (SelfNode == _EntryList) _EntryList = nxt; 772 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 773 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 774 } else { 775 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 776 // Inopportune interleaving -- Self is still on the cxq. 777 // This usually means the enqueue of self raced an exiting thread. 778 // Normally we'll find Self near the front of the cxq, so 779 // dequeueing is typically fast. If needbe we can accelerate 780 // this with some MCS/CHL-like bidirectional list hints and advisory 781 // back-links so dequeueing from the interior will normally operate 782 // in constant-time. 783 // Dequeue Self from either the head (with CAS) or from the interior 784 // with a linear-time scan and normal non-atomic memory operations. 785 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 786 // and then unlink Self from EntryList. We have to drain eventually, 787 // so it might as well be now. 788 789 ObjectWaiter * v = _cxq; 790 assert(v != NULL, "invariant"); 791 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 792 // The CAS above can fail from interference IFF a "RAT" arrived. 793 // In that case Self must be in the interior and can no longer be 794 // at the head of cxq. 795 if (v == SelfNode) { 796 assert(_cxq != v, "invariant"); 797 v = _cxq; // CAS above failed - start scan at head of list 798 } 799 ObjectWaiter * p; 800 ObjectWaiter * q = NULL; 801 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 802 q = p; 803 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 804 } 805 assert(v != SelfNode, "invariant"); 806 assert(p == SelfNode, "Node not found on cxq"); 807 assert(p != _cxq, "invariant"); 808 assert(q != NULL, "invariant"); 809 assert(q->_next == p, "invariant"); 810 q->_next = p->_next; 811 } 812 } 813 814 #ifdef ASSERT 815 // Diagnostic hygiene ... 816 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 817 SelfNode->_next = (ObjectWaiter *) 0xBAD; 818 SelfNode->TState = ObjectWaiter::TS_RUN; 819 #endif 820 } 821 822 // ----------------------------------------------------------------------------- 823 // Exit support 824 // 825 // exit() 826 // ~~~~~~ 827 // Note that the collector can't reclaim the objectMonitor or deflate 828 // the object out from underneath the thread calling ::exit() as the 829 // thread calling ::exit() never transitions to a stable state. 830 // This inhibits GC, which in turn inhibits asynchronous (and 831 // inopportune) reclamation of "this". 832 // 833 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 834 // There's one exception to the claim above, however. EnterI() can call 835 // exit() to drop a lock if the acquirer has been externally suspended. 836 // In that case exit() is called with _thread_state as _thread_blocked, 837 // but the monitor's _count field is > 0, which inhibits reclamation. 838 // 839 // 1-0 exit 840 // ~~~~~~~~ 841 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 842 // the fast-path operators have been optimized so the common ::exit() 843 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 844 // The code emitted by fast_unlock() elides the usual MEMBAR. This 845 // greatly improves latency -- MEMBAR and CAS having considerable local 846 // latency on modern processors -- but at the cost of "stranding". Absent the 847 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 848 // ::enter() path, resulting in the entering thread being stranding 849 // and a progress-liveness failure. Stranding is extremely rare. 850 // We use timers (timed park operations) & periodic polling to detect 851 // and recover from stranding. Potentially stranded threads periodically 852 // wake up and poll the lock. See the usage of the _Responsible variable. 853 // 854 // The CAS() in enter provides for safety and exclusion, while the CAS or 855 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 856 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 857 // We detect and recover from stranding with timers. 858 // 859 // If a thread transiently strands it'll park until (a) another 860 // thread acquires the lock and then drops the lock, at which time the 861 // exiting thread will notice and unpark the stranded thread, or, (b) 862 // the timer expires. If the lock is high traffic then the stranding latency 863 // will be low due to (a). If the lock is low traffic then the odds of 864 // stranding are lower, although the worst-case stranding latency 865 // is longer. Critically, we don't want to put excessive load in the 866 // platform's timer subsystem. We want to minimize both the timer injection 867 // rate (timers created/sec) as well as the number of timers active at 868 // any one time. (more precisely, we want to minimize timer-seconds, which is 869 // the integral of the # of active timers at any instant over time). 870 // Both impinge on OS scalability. Given that, at most one thread parked on 871 // a monitor will use a timer. 872 // 873 // There is also the risk of a futile wake-up. If we drop the lock 874 // another thread can reacquire the lock immediately, and we can 875 // then wake a thread unnecessarily. This is benign, and we've 876 // structured the code so the windows are short and the frequency 877 // of such futile wakups is low. 878 879 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 880 Thread * const Self = THREAD; 881 if (THREAD != _owner) { 882 if (THREAD->is_lock_owned((address) _owner)) { 883 // Transmute _owner from a BasicLock pointer to a Thread address. 884 // We don't need to hold _mutex for this transition. 885 // Non-null to Non-null is safe as long as all readers can 886 // tolerate either flavor. 887 assert(_recursions == 0, "invariant"); 888 _owner = THREAD; 889 _recursions = 0; 890 } else { 891 // Apparent unbalanced locking ... 892 // Naively we'd like to throw IllegalMonitorStateException. 893 // As a practical matter we can neither allocate nor throw an 894 // exception as ::exit() can be called from leaf routines. 895 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 896 // Upon deeper reflection, however, in a properly run JVM the only 897 // way we should encounter this situation is in the presence of 898 // unbalanced JNI locking. TODO: CheckJNICalls. 899 // See also: CR4414101 900 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 901 return; 902 } 903 } 904 905 if (_recursions != 0) { 906 _recursions--; // this is simple recursive enter 907 return; 908 } 909 910 // Invariant: after setting Responsible=null an thread must execute 911 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 912 _Responsible = NULL; 913 914 #if INCLUDE_JFR 915 // get the owner's thread id for the MonitorEnter event 916 // if it is enabled and the thread isn't suspended 917 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 918 _previous_owner_tid = JFR_THREAD_ID(Self); 919 } 920 #endif 921 922 for (;;) { 923 assert(THREAD == _owner, "invariant"); 924 925 if (Knob_ExitPolicy == 0) { 926 // release semantics: prior loads and stores from within the critical section 927 // must not float (reorder) past the following store that drops the lock. 928 // On SPARC that requires MEMBAR #loadstore|#storestore. 929 // But of course in TSO #loadstore|#storestore is not required. 930 // I'd like to write one of the following: 931 // A. OrderAccess::release() ; _owner = NULL 932 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 933 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 934 // store into a _dummy variable. That store is not needed, but can result 935 // in massive wasteful coherency traffic on classic SMP systems. 936 // Instead, I use release_store(), which is implemented as just a simple 937 // ST on x64, x86 and SPARC. 938 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 939 OrderAccess::storeload(); // See if we need to wake a successor 940 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 941 return; 942 } 943 // Other threads are blocked trying to acquire the lock. 944 945 // Normally the exiting thread is responsible for ensuring succession, 946 // but if other successors are ready or other entering threads are spinning 947 // then this thread can simply store NULL into _owner and exit without 948 // waking a successor. The existence of spinners or ready successors 949 // guarantees proper succession (liveness). Responsibility passes to the 950 // ready or running successors. The exiting thread delegates the duty. 951 // More precisely, if a successor already exists this thread is absolved 952 // of the responsibility of waking (unparking) one. 953 // 954 // The _succ variable is critical to reducing futile wakeup frequency. 955 // _succ identifies the "heir presumptive" thread that has been made 956 // ready (unparked) but that has not yet run. We need only one such 957 // successor thread to guarantee progress. 958 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 959 // section 3.3 "Futile Wakeup Throttling" for details. 960 // 961 // Note that spinners in Enter() also set _succ non-null. 962 // In the current implementation spinners opportunistically set 963 // _succ so that exiting threads might avoid waking a successor. 964 // Another less appealing alternative would be for the exiting thread 965 // to drop the lock and then spin briefly to see if a spinner managed 966 // to acquire the lock. If so, the exiting thread could exit 967 // immediately without waking a successor, otherwise the exiting 968 // thread would need to dequeue and wake a successor. 969 // (Note that we'd need to make the post-drop spin short, but no 970 // shorter than the worst-case round-trip cache-line migration time. 971 // The dropped lock needs to become visible to the spinner, and then 972 // the acquisition of the lock by the spinner must become visible to 973 // the exiting thread). 974 975 // It appears that an heir-presumptive (successor) must be made ready. 976 // Only the current lock owner can manipulate the EntryList or 977 // drain _cxq, so we need to reacquire the lock. If we fail 978 // to reacquire the lock the responsibility for ensuring succession 979 // falls to the new owner. 980 // 981 if (!Atomic::replace_if_null(THREAD, &_owner)) { 982 return; 983 } 984 } else { 985 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 986 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 987 OrderAccess::storeload(); 988 // Ratify the previously observed values. 989 if (_cxq == NULL || _succ != NULL) { 990 return; 991 } 992 993 // inopportune interleaving -- the exiting thread (this thread) 994 // in the fast-exit path raced an entering thread in the slow-enter 995 // path. 996 // We have two choices: 997 // A. Try to reacquire the lock. 998 // If the CAS() fails return immediately, otherwise 999 // we either restart/rerun the exit operation, or simply 1000 // fall-through into the code below which wakes a successor. 1001 // B. If the elements forming the EntryList|cxq are TSM 1002 // we could simply unpark() the lead thread and return 1003 // without having set _succ. 1004 if (!Atomic::replace_if_null(THREAD, &_owner)) { 1005 return; 1006 } 1007 } 1008 } 1009 1010 guarantee(_owner == THREAD, "invariant"); 1011 1012 ObjectWaiter * w = NULL; 1013 int QMode = Knob_QMode; 1014 1015 if (QMode == 2 && _cxq != NULL) { 1016 // QMode == 2 : cxq has precedence over EntryList. 1017 // Try to directly wake a successor from the cxq. 1018 // If successful, the successor will need to unlink itself from cxq. 1019 w = _cxq; 1020 assert(w != NULL, "invariant"); 1021 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1022 ExitEpilog(Self, w); 1023 return; 1024 } 1025 1026 if (QMode == 3 && _cxq != NULL) { 1027 // Aggressively drain cxq into EntryList at the first opportunity. 1028 // This policy ensure that recently-run threads live at the head of EntryList. 1029 // Drain _cxq into EntryList - bulk transfer. 1030 // First, detach _cxq. 1031 // The following loop is tantamount to: w = swap(&cxq, NULL) 1032 w = _cxq; 1033 for (;;) { 1034 assert(w != NULL, "Invariant"); 1035 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1036 if (u == w) break; 1037 w = u; 1038 } 1039 assert(w != NULL, "invariant"); 1040 1041 ObjectWaiter * q = NULL; 1042 ObjectWaiter * p; 1043 for (p = w; p != NULL; p = p->_next) { 1044 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1045 p->TState = ObjectWaiter::TS_ENTER; 1046 p->_prev = q; 1047 q = p; 1048 } 1049 1050 // Append the RATs to the EntryList 1051 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1052 ObjectWaiter * Tail; 1053 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1054 Tail = Tail->_next) 1055 /* empty */; 1056 if (Tail == NULL) { 1057 _EntryList = w; 1058 } else { 1059 Tail->_next = w; 1060 w->_prev = Tail; 1061 } 1062 1063 // Fall thru into code that tries to wake a successor from EntryList 1064 } 1065 1066 if (QMode == 4 && _cxq != NULL) { 1067 // Aggressively drain cxq into EntryList at the first opportunity. 1068 // This policy ensure that recently-run threads live at the head of EntryList. 1069 1070 // Drain _cxq into EntryList - bulk transfer. 1071 // First, detach _cxq. 1072 // The following loop is tantamount to: w = swap(&cxq, NULL) 1073 w = _cxq; 1074 for (;;) { 1075 assert(w != NULL, "Invariant"); 1076 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1077 if (u == w) break; 1078 w = u; 1079 } 1080 assert(w != NULL, "invariant"); 1081 1082 ObjectWaiter * q = NULL; 1083 ObjectWaiter * p; 1084 for (p = w; p != NULL; p = p->_next) { 1085 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1086 p->TState = ObjectWaiter::TS_ENTER; 1087 p->_prev = q; 1088 q = p; 1089 } 1090 1091 // Prepend the RATs to the EntryList 1092 if (_EntryList != NULL) { 1093 q->_next = _EntryList; 1094 _EntryList->_prev = q; 1095 } 1096 _EntryList = w; 1097 1098 // Fall thru into code that tries to wake a successor from EntryList 1099 } 1100 1101 w = _EntryList; 1102 if (w != NULL) { 1103 // I'd like to write: guarantee (w->_thread != Self). 1104 // But in practice an exiting thread may find itself on the EntryList. 1105 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1106 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1107 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1108 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1109 // release the lock "O". T2 resumes immediately after the ST of null into 1110 // _owner, above. T2 notices that the EntryList is populated, so it 1111 // reacquires the lock and then finds itself on the EntryList. 1112 // Given all that, we have to tolerate the circumstance where "w" is 1113 // associated with Self. 1114 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1115 ExitEpilog(Self, w); 1116 return; 1117 } 1118 1119 // If we find that both _cxq and EntryList are null then just 1120 // re-run the exit protocol from the top. 1121 w = _cxq; 1122 if (w == NULL) continue; 1123 1124 // Drain _cxq into EntryList - bulk transfer. 1125 // First, detach _cxq. 1126 // The following loop is tantamount to: w = swap(&cxq, NULL) 1127 for (;;) { 1128 assert(w != NULL, "Invariant"); 1129 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1130 if (u == w) break; 1131 w = u; 1132 } 1133 1134 assert(w != NULL, "invariant"); 1135 assert(_EntryList == NULL, "invariant"); 1136 1137 // Convert the LIFO SLL anchored by _cxq into a DLL. 1138 // The list reorganization step operates in O(LENGTH(w)) time. 1139 // It's critical that this step operate quickly as 1140 // "Self" still holds the outer-lock, restricting parallelism 1141 // and effectively lengthening the critical section. 1142 // Invariant: s chases t chases u. 1143 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1144 // we have faster access to the tail. 1145 1146 if (QMode == 1) { 1147 // QMode == 1 : drain cxq to EntryList, reversing order 1148 // We also reverse the order of the list. 1149 ObjectWaiter * s = NULL; 1150 ObjectWaiter * t = w; 1151 ObjectWaiter * u = NULL; 1152 while (t != NULL) { 1153 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1154 t->TState = ObjectWaiter::TS_ENTER; 1155 u = t->_next; 1156 t->_prev = u; 1157 t->_next = s; 1158 s = t; 1159 t = u; 1160 } 1161 _EntryList = s; 1162 assert(s != NULL, "invariant"); 1163 } else { 1164 // QMode == 0 or QMode == 2 1165 _EntryList = w; 1166 ObjectWaiter * q = NULL; 1167 ObjectWaiter * p; 1168 for (p = w; p != NULL; p = p->_next) { 1169 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1170 p->TState = ObjectWaiter::TS_ENTER; 1171 p->_prev = q; 1172 q = p; 1173 } 1174 } 1175 1176 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1177 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1178 1179 // See if we can abdicate to a spinner instead of waking a thread. 1180 // A primary goal of the implementation is to reduce the 1181 // context-switch rate. 1182 if (_succ != NULL) continue; 1183 1184 w = _EntryList; 1185 if (w != NULL) { 1186 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1187 ExitEpilog(Self, w); 1188 return; 1189 } 1190 } 1191 } 1192 1193 // ExitSuspendEquivalent: 1194 // A faster alternate to handle_special_suspend_equivalent_condition() 1195 // 1196 // handle_special_suspend_equivalent_condition() unconditionally 1197 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1198 // operations have high latency. Note that in ::enter() we call HSSEC 1199 // while holding the monitor, so we effectively lengthen the critical sections. 1200 // 1201 // There are a number of possible solutions: 1202 // 1203 // A. To ameliorate the problem we might also defer state transitions 1204 // to as late as possible -- just prior to parking. 1205 // Given that, we'd call HSSEC after having returned from park(), 1206 // but before attempting to acquire the monitor. This is only a 1207 // partial solution. It avoids calling HSSEC while holding the 1208 // monitor (good), but it still increases successor reacquisition latency -- 1209 // the interval between unparking a successor and the time the successor 1210 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1211 // If we use this technique we can also avoid EnterI()-exit() loop 1212 // in ::enter() where we iteratively drop the lock and then attempt 1213 // to reacquire it after suspending. 1214 // 1215 // B. In the future we might fold all the suspend bits into a 1216 // composite per-thread suspend flag and then update it with CAS(). 1217 // Alternately, a Dekker-like mechanism with multiple variables 1218 // would suffice: 1219 // ST Self->_suspend_equivalent = false 1220 // MEMBAR 1221 // LD Self_>_suspend_flags 1222 // 1223 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1224 // with a more efficient implementation, the need to use "FastHSSEC" has 1225 // decreased. - Dave 1226 1227 1228 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1229 const int Mode = Knob_FastHSSEC; 1230 if (Mode && !jSelf->is_external_suspend()) { 1231 assert(jSelf->is_suspend_equivalent(), "invariant"); 1232 jSelf->clear_suspend_equivalent(); 1233 if (2 == Mode) OrderAccess::storeload(); 1234 if (!jSelf->is_external_suspend()) return false; 1235 // We raced a suspension -- fall thru into the slow path 1236 jSelf->set_suspend_equivalent(); 1237 } 1238 return jSelf->handle_special_suspend_equivalent_condition(); 1239 } 1240 1241 1242 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1243 assert(_owner == Self, "invariant"); 1244 1245 // Exit protocol: 1246 // 1. ST _succ = wakee 1247 // 2. membar #loadstore|#storestore; 1248 // 2. ST _owner = NULL 1249 // 3. unpark(wakee) 1250 1251 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1252 ParkEvent * Trigger = Wakee->_event; 1253 1254 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1255 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1256 // out-of-scope (non-extant). 1257 Wakee = NULL; 1258 1259 // Drop the lock 1260 OrderAccess::release_store(&_owner, (void*)NULL); 1261 OrderAccess::fence(); // ST _owner vs LD in unpark() 1262 1263 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1264 Trigger->unpark(); 1265 1266 // Maintain stats and report events to JVMTI 1267 OM_PERFDATA_OP(Parks, inc()); 1268 } 1269 1270 1271 // ----------------------------------------------------------------------------- 1272 // Class Loader deadlock handling. 1273 // 1274 // complete_exit exits a lock returning recursion count 1275 // complete_exit/reenter operate as a wait without waiting 1276 // complete_exit requires an inflated monitor 1277 // The _owner field is not always the Thread addr even with an 1278 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1279 // thread due to contention. 1280 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1281 Thread * const Self = THREAD; 1282 assert(Self->is_Java_thread(), "Must be Java thread!"); 1283 JavaThread *jt = (JavaThread *)THREAD; 1284 1285 DeferredInitialize(); 1286 1287 if (THREAD != _owner) { 1288 if (THREAD->is_lock_owned ((address)_owner)) { 1289 assert(_recursions == 0, "internal state error"); 1290 _owner = THREAD; // Convert from basiclock addr to Thread addr 1291 _recursions = 0; 1292 } 1293 } 1294 1295 guarantee(Self == _owner, "complete_exit not owner"); 1296 intptr_t save = _recursions; // record the old recursion count 1297 _recursions = 0; // set the recursion level to be 0 1298 exit(true, Self); // exit the monitor 1299 guarantee(_owner != Self, "invariant"); 1300 return save; 1301 } 1302 1303 // reenter() enters a lock and sets recursion count 1304 // complete_exit/reenter operate as a wait without waiting 1305 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1306 Thread * const Self = THREAD; 1307 assert(Self->is_Java_thread(), "Must be Java thread!"); 1308 JavaThread *jt = (JavaThread *)THREAD; 1309 1310 guarantee(_owner != Self, "reenter already owner"); 1311 enter(THREAD); // enter the monitor 1312 guarantee(_recursions == 0, "reenter recursion"); 1313 _recursions = recursions; 1314 return; 1315 } 1316 1317 1318 // ----------------------------------------------------------------------------- 1319 // A macro is used below because there may already be a pending 1320 // exception which should not abort the execution of the routines 1321 // which use this (which is why we don't put this into check_slow and 1322 // call it with a CHECK argument). 1323 1324 #define CHECK_OWNER() \ 1325 do { \ 1326 if (THREAD != _owner) { \ 1327 if (THREAD->is_lock_owned((address) _owner)) { \ 1328 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1329 _recursions = 0; \ 1330 } else { \ 1331 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1332 } \ 1333 } \ 1334 } while (false) 1335 1336 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1337 // TODO-FIXME: remove check_slow() -- it's likely dead. 1338 1339 void ObjectMonitor::check_slow(TRAPS) { 1340 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1341 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1342 } 1343 1344 static int Adjust(volatile int * adr, int dx) { 1345 int v; 1346 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1347 return v; 1348 } 1349 1350 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1351 ObjectMonitor* monitor, 1352 jlong notifier_tid, 1353 jlong timeout, 1354 bool timedout) { 1355 assert(event != NULL, "invariant"); 1356 assert(monitor != NULL, "invariant"); 1357 event->set_monitorClass(((oop)monitor->object())->klass()); 1358 event->set_timeout(timeout); 1359 event->set_address((uintptr_t)monitor->object_addr()); 1360 event->set_notifier(notifier_tid); 1361 event->set_timedOut(timedout); 1362 event->commit(); 1363 } 1364 1365 // ----------------------------------------------------------------------------- 1366 // Wait/Notify/NotifyAll 1367 // 1368 // Note: a subset of changes to ObjectMonitor::wait() 1369 // will need to be replicated in complete_exit 1370 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1371 Thread * const Self = THREAD; 1372 assert(Self->is_Java_thread(), "Must be Java thread!"); 1373 JavaThread *jt = (JavaThread *)THREAD; 1374 1375 DeferredInitialize(); 1376 1377 // Throw IMSX or IEX. 1378 CHECK_OWNER(); 1379 1380 EventJavaMonitorWait event; 1381 1382 // check for a pending interrupt 1383 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1384 // post monitor waited event. Note that this is past-tense, we are done waiting. 1385 if (JvmtiExport::should_post_monitor_waited()) { 1386 // Note: 'false' parameter is passed here because the 1387 // wait was not timed out due to thread interrupt. 1388 JvmtiExport::post_monitor_waited(jt, this, false); 1389 1390 // In this short circuit of the monitor wait protocol, the 1391 // current thread never drops ownership of the monitor and 1392 // never gets added to the wait queue so the current thread 1393 // cannot be made the successor. This means that the 1394 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1395 // consume an unpark() meant for the ParkEvent associated with 1396 // this ObjectMonitor. 1397 } 1398 if (event.should_commit()) { 1399 post_monitor_wait_event(&event, this, 0, millis, false); 1400 } 1401 THROW(vmSymbols::java_lang_InterruptedException()); 1402 return; 1403 } 1404 1405 assert(Self->_Stalled == 0, "invariant"); 1406 Self->_Stalled = intptr_t(this); 1407 jt->set_current_waiting_monitor(this); 1408 1409 // create a node to be put into the queue 1410 // Critically, after we reset() the event but prior to park(), we must check 1411 // for a pending interrupt. 1412 ObjectWaiter node(Self); 1413 node.TState = ObjectWaiter::TS_WAIT; 1414 Self->_ParkEvent->reset(); 1415 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1416 1417 // Enter the waiting queue, which is a circular doubly linked list in this case 1418 // but it could be a priority queue or any data structure. 1419 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1420 // by the the owner of the monitor *except* in the case where park() 1421 // returns because of a timeout of interrupt. Contention is exceptionally rare 1422 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1423 1424 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1425 AddWaiter(&node); 1426 Thread::SpinRelease(&_WaitSetLock); 1427 1428 _Responsible = NULL; 1429 1430 intptr_t save = _recursions; // record the old recursion count 1431 _waiters++; // increment the number of waiters 1432 _recursions = 0; // set the recursion level to be 1 1433 exit(true, Self); // exit the monitor 1434 guarantee(_owner != Self, "invariant"); 1435 1436 // The thread is on the WaitSet list - now park() it. 1437 // On MP systems it's conceivable that a brief spin before we park 1438 // could be profitable. 1439 // 1440 // TODO-FIXME: change the following logic to a loop of the form 1441 // while (!timeout && !interrupted && _notified == 0) park() 1442 1443 int ret = OS_OK; 1444 int WasNotified = 0; 1445 { // State transition wrappers 1446 OSThread* osthread = Self->osthread(); 1447 OSThreadWaitState osts(osthread, true); 1448 { 1449 ThreadBlockInVM tbivm(jt); 1450 // Thread is in thread_blocked state and oop access is unsafe. 1451 jt->set_suspend_equivalent(); 1452 1453 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1454 // Intentionally empty 1455 } else if (node._notified == 0) { 1456 if (millis <= 0) { 1457 Self->_ParkEvent->park(); 1458 } else { 1459 ret = Self->_ParkEvent->park(millis); 1460 } 1461 } 1462 1463 // were we externally suspended while we were waiting? 1464 if (ExitSuspendEquivalent (jt)) { 1465 // TODO-FIXME: add -- if succ == Self then succ = null. 1466 jt->java_suspend_self(); 1467 } 1468 1469 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1470 1471 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1472 // from the WaitSet to the EntryList. 1473 // See if we need to remove Node from the WaitSet. 1474 // We use double-checked locking to avoid grabbing _WaitSetLock 1475 // if the thread is not on the wait queue. 1476 // 1477 // Note that we don't need a fence before the fetch of TState. 1478 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1479 // written by the is thread. (perhaps the fetch might even be satisfied 1480 // by a look-aside into the processor's own store buffer, although given 1481 // the length of the code path between the prior ST and this load that's 1482 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1483 // then we'll acquire the lock and then re-fetch a fresh TState value. 1484 // That is, we fail toward safety. 1485 1486 if (node.TState == ObjectWaiter::TS_WAIT) { 1487 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1488 if (node.TState == ObjectWaiter::TS_WAIT) { 1489 DequeueSpecificWaiter(&node); // unlink from WaitSet 1490 assert(node._notified == 0, "invariant"); 1491 node.TState = ObjectWaiter::TS_RUN; 1492 } 1493 Thread::SpinRelease(&_WaitSetLock); 1494 } 1495 1496 // The thread is now either on off-list (TS_RUN), 1497 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1498 // The Node's TState variable is stable from the perspective of this thread. 1499 // No other threads will asynchronously modify TState. 1500 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1501 OrderAccess::loadload(); 1502 if (_succ == Self) _succ = NULL; 1503 WasNotified = node._notified; 1504 1505 // Reentry phase -- reacquire the monitor. 1506 // re-enter contended monitor after object.wait(). 1507 // retain OBJECT_WAIT state until re-enter successfully completes 1508 // Thread state is thread_in_vm and oop access is again safe, 1509 // although the raw address of the object may have changed. 1510 // (Don't cache naked oops over safepoints, of course). 1511 1512 // post monitor waited event. Note that this is past-tense, we are done waiting. 1513 if (JvmtiExport::should_post_monitor_waited()) { 1514 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1515 1516 if (node._notified != 0 && _succ == Self) { 1517 // In this part of the monitor wait-notify-reenter protocol it 1518 // is possible (and normal) for another thread to do a fastpath 1519 // monitor enter-exit while this thread is still trying to get 1520 // to the reenter portion of the protocol. 1521 // 1522 // The ObjectMonitor was notified and the current thread is 1523 // the successor which also means that an unpark() has already 1524 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1525 // consume the unpark() that was done when the successor was 1526 // set because the same ParkEvent is shared between Java 1527 // monitors and JVM/TI RawMonitors (for now). 1528 // 1529 // We redo the unpark() to ensure forward progress, i.e., we 1530 // don't want all pending threads hanging (parked) with none 1531 // entering the unlocked monitor. 1532 node._event->unpark(); 1533 } 1534 } 1535 1536 if (event.should_commit()) { 1537 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1538 } 1539 1540 OrderAccess::fence(); 1541 1542 assert(Self->_Stalled != 0, "invariant"); 1543 Self->_Stalled = 0; 1544 1545 assert(_owner != Self, "invariant"); 1546 ObjectWaiter::TStates v = node.TState; 1547 if (v == ObjectWaiter::TS_RUN) { 1548 enter(Self); 1549 } else { 1550 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1551 ReenterI(Self, &node); 1552 node.wait_reenter_end(this); 1553 } 1554 1555 // Self has reacquired the lock. 1556 // Lifecycle - the node representing Self must not appear on any queues. 1557 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1558 // want residual elements associated with this thread left on any lists. 1559 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1560 assert(_owner == Self, "invariant"); 1561 assert(_succ != Self, "invariant"); 1562 } // OSThreadWaitState() 1563 1564 jt->set_current_waiting_monitor(NULL); 1565 1566 guarantee(_recursions == 0, "invariant"); 1567 _recursions = save; // restore the old recursion count 1568 _waiters--; // decrement the number of waiters 1569 1570 // Verify a few postconditions 1571 assert(_owner == Self, "invariant"); 1572 assert(_succ != Self, "invariant"); 1573 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1574 1575 // check if the notification happened 1576 if (!WasNotified) { 1577 // no, it could be timeout or Thread.interrupt() or both 1578 // check for interrupt event, otherwise it is timeout 1579 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1580 THROW(vmSymbols::java_lang_InterruptedException()); 1581 } 1582 } 1583 1584 // NOTE: Spurious wake up will be consider as timeout. 1585 // Monitor notify has precedence over thread interrupt. 1586 } 1587 1588 1589 // Consider: 1590 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1591 // then instead of transferring a thread from the WaitSet to the EntryList 1592 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1593 1594 void ObjectMonitor::INotify(Thread * Self) { 1595 const int policy = Knob_MoveNotifyee; 1596 1597 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1598 ObjectWaiter * iterator = DequeueWaiter(); 1599 if (iterator != NULL) { 1600 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1601 guarantee(iterator->_notified == 0, "invariant"); 1602 // Disposition - what might we do with iterator ? 1603 // a. add it directly to the EntryList - either tail (policy == 1) 1604 // or head (policy == 0). 1605 // b. push it onto the front of the _cxq (policy == 2). 1606 // For now we use (b). 1607 if (policy != 4) { 1608 iterator->TState = ObjectWaiter::TS_ENTER; 1609 } 1610 iterator->_notified = 1; 1611 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1612 1613 ObjectWaiter * list = _EntryList; 1614 if (list != NULL) { 1615 assert(list->_prev == NULL, "invariant"); 1616 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1617 assert(list != iterator, "invariant"); 1618 } 1619 1620 if (policy == 0) { // prepend to EntryList 1621 if (list == NULL) { 1622 iterator->_next = iterator->_prev = NULL; 1623 _EntryList = iterator; 1624 } else { 1625 list->_prev = iterator; 1626 iterator->_next = list; 1627 iterator->_prev = NULL; 1628 _EntryList = iterator; 1629 } 1630 } else if (policy == 1) { // append to EntryList 1631 if (list == NULL) { 1632 iterator->_next = iterator->_prev = NULL; 1633 _EntryList = iterator; 1634 } else { 1635 // CONSIDER: finding the tail currently requires a linear-time walk of 1636 // the EntryList. We can make tail access constant-time by converting to 1637 // a CDLL instead of using our current DLL. 1638 ObjectWaiter * tail; 1639 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1640 assert(tail != NULL && tail->_next == NULL, "invariant"); 1641 tail->_next = iterator; 1642 iterator->_prev = tail; 1643 iterator->_next = NULL; 1644 } 1645 } else if (policy == 2) { // prepend to cxq 1646 if (list == NULL) { 1647 iterator->_next = iterator->_prev = NULL; 1648 _EntryList = iterator; 1649 } else { 1650 iterator->TState = ObjectWaiter::TS_CXQ; 1651 for (;;) { 1652 ObjectWaiter * front = _cxq; 1653 iterator->_next = front; 1654 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1655 break; 1656 } 1657 } 1658 } 1659 } else if (policy == 3) { // append to cxq 1660 iterator->TState = ObjectWaiter::TS_CXQ; 1661 for (;;) { 1662 ObjectWaiter * tail = _cxq; 1663 if (tail == NULL) { 1664 iterator->_next = NULL; 1665 if (Atomic::replace_if_null(iterator, &_cxq)) { 1666 break; 1667 } 1668 } else { 1669 while (tail->_next != NULL) tail = tail->_next; 1670 tail->_next = iterator; 1671 iterator->_prev = tail; 1672 iterator->_next = NULL; 1673 break; 1674 } 1675 } 1676 } else { 1677 ParkEvent * ev = iterator->_event; 1678 iterator->TState = ObjectWaiter::TS_RUN; 1679 OrderAccess::fence(); 1680 ev->unpark(); 1681 } 1682 1683 // _WaitSetLock protects the wait queue, not the EntryList. We could 1684 // move the add-to-EntryList operation, above, outside the critical section 1685 // protected by _WaitSetLock. In practice that's not useful. With the 1686 // exception of wait() timeouts and interrupts the monitor owner 1687 // is the only thread that grabs _WaitSetLock. There's almost no contention 1688 // on _WaitSetLock so it's not profitable to reduce the length of the 1689 // critical section. 1690 1691 if (policy < 4) { 1692 iterator->wait_reenter_begin(this); 1693 } 1694 } 1695 Thread::SpinRelease(&_WaitSetLock); 1696 } 1697 1698 // Consider: a not-uncommon synchronization bug is to use notify() when 1699 // notifyAll() is more appropriate, potentially resulting in stranded 1700 // threads; this is one example of a lost wakeup. A useful diagnostic 1701 // option is to force all notify() operations to behave as notifyAll(). 1702 // 1703 // Note: We can also detect many such problems with a "minimum wait". 1704 // When the "minimum wait" is set to a small non-zero timeout value 1705 // and the program does not hang whereas it did absent "minimum wait", 1706 // that suggests a lost wakeup bug. 1707 1708 void ObjectMonitor::notify(TRAPS) { 1709 CHECK_OWNER(); 1710 if (_WaitSet == NULL) { 1711 return; 1712 } 1713 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1714 INotify(THREAD); 1715 OM_PERFDATA_OP(Notifications, inc(1)); 1716 } 1717 1718 1719 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1720 // from the waitset to the EntryList. This could be done more efficiently with a 1721 // single bulk transfer but in practice it's not time-critical. Beware too, 1722 // that in prepend-mode we invert the order of the waiters. Let's say that the 1723 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1724 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1725 1726 void ObjectMonitor::notifyAll(TRAPS) { 1727 CHECK_OWNER(); 1728 if (_WaitSet == NULL) { 1729 return; 1730 } 1731 1732 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1733 int tally = 0; 1734 while (_WaitSet != NULL) { 1735 tally++; 1736 INotify(THREAD); 1737 } 1738 1739 OM_PERFDATA_OP(Notifications, inc(tally)); 1740 } 1741 1742 // ----------------------------------------------------------------------------- 1743 // Adaptive Spinning Support 1744 // 1745 // Adaptive spin-then-block - rational spinning 1746 // 1747 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1748 // algorithm. On high order SMP systems it would be better to start with 1749 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1750 // a contending thread could enqueue itself on the cxq and then spin locally 1751 // on a thread-specific variable such as its ParkEvent._Event flag. 1752 // That's left as an exercise for the reader. Note that global spinning is 1753 // not problematic on Niagara, as the L2 cache serves the interconnect and 1754 // has both low latency and massive bandwidth. 1755 // 1756 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1757 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1758 // (duration) or we can fix the count at approximately the duration of 1759 // a context switch and vary the frequency. Of course we could also 1760 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1761 // For a description of 'Adaptive spin-then-block mutual exclusion in 1762 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1763 // 1764 // This implementation varies the duration "D", where D varies with 1765 // the success rate of recent spin attempts. (D is capped at approximately 1766 // length of a round-trip context switch). The success rate for recent 1767 // spin attempts is a good predictor of the success rate of future spin 1768 // attempts. The mechanism adapts automatically to varying critical 1769 // section length (lock modality), system load and degree of parallelism. 1770 // D is maintained per-monitor in _SpinDuration and is initialized 1771 // optimistically. Spin frequency is fixed at 100%. 1772 // 1773 // Note that _SpinDuration is volatile, but we update it without locks 1774 // or atomics. The code is designed so that _SpinDuration stays within 1775 // a reasonable range even in the presence of races. The arithmetic 1776 // operations on _SpinDuration are closed over the domain of legal values, 1777 // so at worst a race will install and older but still legal value. 1778 // At the very worst this introduces some apparent non-determinism. 1779 // We might spin when we shouldn't or vice-versa, but since the spin 1780 // count are relatively short, even in the worst case, the effect is harmless. 1781 // 1782 // Care must be taken that a low "D" value does not become an 1783 // an absorbing state. Transient spinning failures -- when spinning 1784 // is overall profitable -- should not cause the system to converge 1785 // on low "D" values. We want spinning to be stable and predictable 1786 // and fairly responsive to change and at the same time we don't want 1787 // it to oscillate, become metastable, be "too" non-deterministic, 1788 // or converge on or enter undesirable stable absorbing states. 1789 // 1790 // We implement a feedback-based control system -- using past behavior 1791 // to predict future behavior. We face two issues: (a) if the 1792 // input signal is random then the spin predictor won't provide optimal 1793 // results, and (b) if the signal frequency is too high then the control 1794 // system, which has some natural response lag, will "chase" the signal. 1795 // (b) can arise from multimodal lock hold times. Transient preemption 1796 // can also result in apparent bimodal lock hold times. 1797 // Although sub-optimal, neither condition is particularly harmful, as 1798 // in the worst-case we'll spin when we shouldn't or vice-versa. 1799 // The maximum spin duration is rather short so the failure modes aren't bad. 1800 // To be conservative, I've tuned the gain in system to bias toward 1801 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1802 // "rings" or oscillates between spinning and not spinning. This happens 1803 // when spinning is just on the cusp of profitability, however, so the 1804 // situation is not dire. The state is benign -- there's no need to add 1805 // hysteresis control to damp the transition rate between spinning and 1806 // not spinning. 1807 1808 // Spinning: Fixed frequency (100%), vary duration 1809 int ObjectMonitor::TrySpin(Thread * Self) { 1810 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1811 int ctr = Knob_FixedSpin; 1812 if (ctr != 0) { 1813 while (--ctr >= 0) { 1814 if (TryLock(Self) > 0) return 1; 1815 SpinPause(); 1816 } 1817 return 0; 1818 } 1819 1820 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1821 if (TryLock(Self) > 0) { 1822 // Increase _SpinDuration ... 1823 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1824 // Raising _SpurDuration to the poverty line is key. 1825 int x = _SpinDuration; 1826 if (x < Knob_SpinLimit) { 1827 if (x < Knob_Poverty) x = Knob_Poverty; 1828 _SpinDuration = x + Knob_BonusB; 1829 } 1830 return 1; 1831 } 1832 SpinPause(); 1833 } 1834 1835 // Admission control - verify preconditions for spinning 1836 // 1837 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1838 // becoming an absorbing state. Put another way, we spin briefly to 1839 // sample, just in case the system load, parallelism, contention, or lock 1840 // modality changed. 1841 // 1842 // Consider the following alternative: 1843 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1844 // spin attempt. "Periodically" might mean after a tally of 1845 // the # of failed spin attempts (or iterations) reaches some threshold. 1846 // This takes us into the realm of 1-out-of-N spinning, where we 1847 // hold the duration constant but vary the frequency. 1848 1849 ctr = _SpinDuration; 1850 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1851 if (ctr <= 0) return 0; 1852 1853 if (Knob_SuccRestrict && _succ != NULL) return 0; 1854 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1855 return 0; 1856 } 1857 1858 int MaxSpin = Knob_MaxSpinners; 1859 if (MaxSpin >= 0) { 1860 if (_Spinner > MaxSpin) { 1861 return 0; 1862 } 1863 // Slightly racy, but benign ... 1864 Adjust(&_Spinner, 1); 1865 } 1866 1867 // We're good to spin ... spin ingress. 1868 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1869 // when preparing to LD...CAS _owner, etc and the CAS is likely 1870 // to succeed. 1871 int caspty = Knob_CASPenalty; 1872 int oxpty = Knob_OXPenalty; 1873 int sss = Knob_SpinSetSucc; 1874 if (sss && _succ == NULL) _succ = Self; 1875 Thread * prv = NULL; 1876 1877 // There are three ways to exit the following loop: 1878 // 1. A successful spin where this thread has acquired the lock. 1879 // 2. Spin failure with prejudice 1880 // 3. Spin failure without prejudice 1881 1882 while (--ctr >= 0) { 1883 1884 // Periodic polling -- Check for pending GC 1885 // Threads may spin while they're unsafe. 1886 // We don't want spinning threads to delay the JVM from reaching 1887 // a stop-the-world safepoint or to steal cycles from GC. 1888 // If we detect a pending safepoint we abort in order that 1889 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1890 // this thread, if safe, doesn't steal cycles from GC. 1891 // This is in keeping with the "no loitering in runtime" rule. 1892 // We periodically check to see if there's a safepoint pending. 1893 if ((ctr & 0xFF) == 0) { 1894 if (SafepointMechanism::poll(Self)) { 1895 goto Abort; // abrupt spin egress 1896 } 1897 if (Knob_UsePause & 1) SpinPause(); 1898 } 1899 1900 if (Knob_UsePause & 2) SpinPause(); 1901 1902 // Probe _owner with TATAS 1903 // If this thread observes the monitor transition or flicker 1904 // from locked to unlocked to locked, then the odds that this 1905 // thread will acquire the lock in this spin attempt go down 1906 // considerably. The same argument applies if the CAS fails 1907 // or if we observe _owner change from one non-null value to 1908 // another non-null value. In such cases we might abort 1909 // the spin without prejudice or apply a "penalty" to the 1910 // spin count-down variable "ctr", reducing it by 100, say. 1911 1912 Thread * ox = (Thread *) _owner; 1913 if (ox == NULL) { 1914 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1915 if (ox == NULL) { 1916 // The CAS succeeded -- this thread acquired ownership 1917 // Take care of some bookkeeping to exit spin state. 1918 if (sss && _succ == Self) { 1919 _succ = NULL; 1920 } 1921 if (MaxSpin > 0) Adjust(&_Spinner, -1); 1922 1923 // Increase _SpinDuration : 1924 // The spin was successful (profitable) so we tend toward 1925 // longer spin attempts in the future. 1926 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1927 // If we acquired the lock early in the spin cycle it 1928 // makes sense to increase _SpinDuration proportionally. 1929 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1930 int x = _SpinDuration; 1931 if (x < Knob_SpinLimit) { 1932 if (x < Knob_Poverty) x = Knob_Poverty; 1933 _SpinDuration = x + Knob_Bonus; 1934 } 1935 return 1; 1936 } 1937 1938 // The CAS failed ... we can take any of the following actions: 1939 // * penalize: ctr -= Knob_CASPenalty 1940 // * exit spin with prejudice -- goto Abort; 1941 // * exit spin without prejudice. 1942 // * Since CAS is high-latency, retry again immediately. 1943 prv = ox; 1944 if (caspty == -2) break; 1945 if (caspty == -1) goto Abort; 1946 ctr -= caspty; 1947 continue; 1948 } 1949 1950 // Did lock ownership change hands ? 1951 if (ox != prv && prv != NULL) { 1952 if (oxpty == -2) break; 1953 if (oxpty == -1) goto Abort; 1954 ctr -= oxpty; 1955 } 1956 prv = ox; 1957 1958 // Abort the spin if the owner is not executing. 1959 // The owner must be executing in order to drop the lock. 1960 // Spinning while the owner is OFFPROC is idiocy. 1961 // Consider: ctr -= RunnablePenalty ; 1962 if (Knob_OState && NotRunnable (Self, ox)) { 1963 goto Abort; 1964 } 1965 if (sss && _succ == NULL) _succ = Self; 1966 } 1967 1968 // Spin failed with prejudice -- reduce _SpinDuration. 1969 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1970 // AIMD is globally stable. 1971 { 1972 int x = _SpinDuration; 1973 if (x > 0) { 1974 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1975 // This is globally sample and tends to damp the response. 1976 x -= Knob_Penalty; 1977 if (x < 0) x = 0; 1978 _SpinDuration = x; 1979 } 1980 } 1981 1982 Abort: 1983 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 1984 if (sss && _succ == Self) { 1985 _succ = NULL; 1986 // Invariant: after setting succ=null a contending thread 1987 // must recheck-retry _owner before parking. This usually happens 1988 // in the normal usage of TrySpin(), but it's safest 1989 // to make TrySpin() as foolproof as possible. 1990 OrderAccess::fence(); 1991 if (TryLock(Self) > 0) return 1; 1992 } 1993 return 0; 1994 } 1995 1996 // NotRunnable() -- informed spinning 1997 // 1998 // Don't bother spinning if the owner is not eligible to drop the lock. 1999 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2000 // The thread must be runnable in order to drop the lock in timely fashion. 2001 // If the _owner is not runnable then spinning will not likely be 2002 // successful (profitable). 2003 // 2004 // Beware -- the thread referenced by _owner could have died 2005 // so a simply fetch from _owner->_thread_state might trap. 2006 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2007 // Because of the lifecycle issues, the _thread_state values 2008 // observed by NotRunnable() might be garbage. NotRunnable must 2009 // tolerate this and consider the observed _thread_state value 2010 // as advisory. 2011 // 2012 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2013 // a thread pointer. 2014 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2015 // with the LSB of _owner. Another option would be to probabilistically probe 2016 // the putative _owner->TypeTag value. 2017 // 2018 // Checking _thread_state isn't perfect. Even if the thread is 2019 // in_java it might be blocked on a page-fault or have been preempted 2020 // and sitting on a ready/dispatch queue. 2021 // 2022 // The return value from NotRunnable() is *advisory* -- the 2023 // result is based on sampling and is not necessarily coherent. 2024 // The caller must tolerate false-negative and false-positive errors. 2025 // Spinning, in general, is probabilistic anyway. 2026 2027 2028 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2029 // Check ox->TypeTag == 2BAD. 2030 if (ox == NULL) return 0; 2031 2032 // Avoid transitive spinning ... 2033 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2034 // Immediately after T1 acquires L it's possible that T2, also 2035 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2036 // This occurs transiently after T1 acquired L but before 2037 // T1 managed to clear T1.Stalled. T2 does not need to abort 2038 // its spin in this circumstance. 2039 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2040 2041 if (BlockedOn == 1) return 1; 2042 if (BlockedOn != 0) { 2043 return BlockedOn != intptr_t(this) && _owner == ox; 2044 } 2045 2046 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2047 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2048 // consider also: jst != _thread_in_Java -- but that's overspecific. 2049 return jst == _thread_blocked || jst == _thread_in_native; 2050 } 2051 2052 2053 // ----------------------------------------------------------------------------- 2054 // WaitSet management ... 2055 2056 ObjectWaiter::ObjectWaiter(Thread* thread) { 2057 _next = NULL; 2058 _prev = NULL; 2059 _notified = 0; 2060 _notifier_tid = 0; 2061 TState = TS_RUN; 2062 _thread = thread; 2063 _event = thread->_ParkEvent; 2064 _active = false; 2065 assert(_event != NULL, "invariant"); 2066 } 2067 2068 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2069 JavaThread *jt = (JavaThread *)this->_thread; 2070 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2071 } 2072 2073 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2074 JavaThread *jt = (JavaThread *)this->_thread; 2075 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2076 } 2077 2078 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2079 assert(node != NULL, "should not add NULL node"); 2080 assert(node->_prev == NULL, "node already in list"); 2081 assert(node->_next == NULL, "node already in list"); 2082 // put node at end of queue (circular doubly linked list) 2083 if (_WaitSet == NULL) { 2084 _WaitSet = node; 2085 node->_prev = node; 2086 node->_next = node; 2087 } else { 2088 ObjectWaiter* head = _WaitSet; 2089 ObjectWaiter* tail = head->_prev; 2090 assert(tail->_next == head, "invariant check"); 2091 tail->_next = node; 2092 head->_prev = node; 2093 node->_next = head; 2094 node->_prev = tail; 2095 } 2096 } 2097 2098 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2099 // dequeue the very first waiter 2100 ObjectWaiter* waiter = _WaitSet; 2101 if (waiter) { 2102 DequeueSpecificWaiter(waiter); 2103 } 2104 return waiter; 2105 } 2106 2107 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2108 assert(node != NULL, "should not dequeue NULL node"); 2109 assert(node->_prev != NULL, "node already removed from list"); 2110 assert(node->_next != NULL, "node already removed from list"); 2111 // when the waiter has woken up because of interrupt, 2112 // timeout or other spurious wake-up, dequeue the 2113 // waiter from waiting list 2114 ObjectWaiter* next = node->_next; 2115 if (next == node) { 2116 assert(node->_prev == node, "invariant check"); 2117 _WaitSet = NULL; 2118 } else { 2119 ObjectWaiter* prev = node->_prev; 2120 assert(prev->_next == node, "invariant check"); 2121 assert(next->_prev == node, "invariant check"); 2122 next->_prev = prev; 2123 prev->_next = next; 2124 if (_WaitSet == node) { 2125 _WaitSet = next; 2126 } 2127 } 2128 node->_next = NULL; 2129 node->_prev = NULL; 2130 } 2131 2132 // ----------------------------------------------------------------------------- 2133 // PerfData support 2134 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2135 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2136 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2137 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2138 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2139 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2140 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2141 2142 // One-shot global initialization for the sync subsystem. 2143 // We could also defer initialization and initialize on-demand 2144 // the first time we call inflate(). Initialization would 2145 // be protected - like so many things - by the MonitorCache_lock. 2146 2147 void ObjectMonitor::Initialize() { 2148 static int InitializationCompleted = 0; 2149 assert(InitializationCompleted == 0, "invariant"); 2150 InitializationCompleted = 1; 2151 if (UsePerfData) { 2152 EXCEPTION_MARK; 2153 #define NEWPERFCOUNTER(n) \ 2154 { \ 2155 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2156 CHECK); \ 2157 } 2158 #define NEWPERFVARIABLE(n) \ 2159 { \ 2160 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2161 CHECK); \ 2162 } 2163 NEWPERFCOUNTER(_sync_Inflations); 2164 NEWPERFCOUNTER(_sync_Deflations); 2165 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2166 NEWPERFCOUNTER(_sync_FutileWakeups); 2167 NEWPERFCOUNTER(_sync_Parks); 2168 NEWPERFCOUNTER(_sync_Notifications); 2169 NEWPERFVARIABLE(_sync_MonExtant); 2170 #undef NEWPERFCOUNTER 2171 #undef NEWPERFVARIABLE 2172 } 2173 } 2174 2175 void ObjectMonitor::DeferredInitialize() { 2176 if (InitDone > 0) return; 2177 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2178 while (InitDone != 1) /* empty */; 2179 return; 2180 } 2181 2182 // One-shot global initialization ... 2183 // The initialization is idempotent, so we don't need locks. 2184 // In the future consider doing this via os::init_2(). 2185 2186 if (!os::is_MP()) { 2187 Knob_SpinLimit = 0; 2188 Knob_SpinBase = 0; 2189 Knob_PreSpin = 0; 2190 Knob_FixedSpin = -1; 2191 } 2192 2193 OrderAccess::fence(); 2194 InitDone = 1; 2195 } 2196