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