objectMonitor.cpp revision 6599:d9f77ba99034
1/* 2 * Copyright (c) 1998, 2014, 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 "memory/resourceArea.hpp" 28#include "oops/markOop.hpp" 29#include "oops/oop.inline.hpp" 30#include "runtime/handles.inline.hpp" 31#include "runtime/interfaceSupport.hpp" 32#include "runtime/mutexLocker.hpp" 33#include "runtime/objectMonitor.hpp" 34#include "runtime/objectMonitor.inline.hpp" 35#include "runtime/orderAccess.inline.hpp" 36#include "runtime/osThread.hpp" 37#include "runtime/stubRoutines.hpp" 38#include "runtime/thread.inline.hpp" 39#include "services/threadService.hpp" 40#include "trace/tracing.hpp" 41#include "trace/traceMacros.hpp" 42#include "utilities/dtrace.hpp" 43#include "utilities/macros.hpp" 44#include "utilities/preserveException.hpp" 45#ifdef TARGET_OS_FAMILY_linux 46# include "os_linux.inline.hpp" 47#endif 48#ifdef TARGET_OS_FAMILY_solaris 49# include "os_solaris.inline.hpp" 50#endif 51#ifdef TARGET_OS_FAMILY_windows 52# include "os_windows.inline.hpp" 53#endif 54#ifdef TARGET_OS_FAMILY_bsd 55# include "os_bsd.inline.hpp" 56#endif 57 58#if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 59 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 60 #define ATTR __attribute__((noinline)) 61#else 62 #define ATTR 63#endif 64 65 66#ifdef DTRACE_ENABLED 67 68// Only bother with this argument setup if dtrace is available 69// TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 70 71 72#define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 73 char* bytes = NULL; \ 74 int len = 0; \ 75 jlong jtid = SharedRuntime::get_java_tid(thread); \ 76 Symbol* klassname = ((oop)obj)->klass()->name(); \ 77 if (klassname != NULL) { \ 78 bytes = (char*)klassname->bytes(); \ 79 len = klassname->utf8_length(); \ 80 } 81 82#define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 83 { \ 84 if (DTraceMonitorProbes) { \ 85 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 86 HOTSPOT_MONITOR_WAIT(jtid, \ 87 (monitor), bytes, len, (millis)); \ 88 } \ 89 } 90 91#define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 92#define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 93#define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 94#define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 95#define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 96 97#define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 98 { \ 99 if (DTraceMonitorProbes) { \ 100 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 101 HOTSPOT_MONITOR_##probe(jtid, \ 102 (uintptr_t)(monitor), bytes, len); \ 103 } \ 104 } 105 106#else // ndef DTRACE_ENABLED 107 108#define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 109#define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 110 111#endif // ndef DTRACE_ENABLED 112 113// Tunables ... 114// The knob* variables are effectively final. Once set they should 115// never be modified hence. Consider using __read_mostly with GCC. 116 117int ObjectMonitor::Knob_Verbose = 0; 118int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 119static int Knob_LogSpins = 0; // enable jvmstat tally for spins 120static int Knob_HandOff = 0; 121static int Knob_ReportSettings = 0; 122 123static int Knob_SpinBase = 0; // Floor AKA SpinMin 124static int Knob_SpinBackOff = 0; // spin-loop backoff 125static int Knob_CASPenalty = -1; // Penalty for failed CAS 126static int Knob_OXPenalty = -1; // Penalty for observed _owner change 127static int Knob_SpinSetSucc = 1; // spinners set the _succ field 128static int Knob_SpinEarly = 1; 129static int Knob_SuccEnabled = 1; // futile wake throttling 130static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 131static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 132static int Knob_Bonus = 100; // spin success bonus 133static int Knob_BonusB = 100; // spin success bonus 134static int Knob_Penalty = 200; // spin failure penalty 135static int Knob_Poverty = 1000; 136static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 137static int Knob_FixedSpin = 0; 138static int Knob_OState = 3; // Spinner checks thread state of _owner 139static int Knob_UsePause = 1; 140static int Knob_ExitPolicy = 0; 141static int Knob_PreSpin = 10; // 20-100 likely better 142static int Knob_ResetEvent = 0; 143static int BackOffMask = 0; 144 145static int Knob_FastHSSEC = 0; 146static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 147static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 148static volatile int InitDone = 0; 149 150#define TrySpin TrySpin_VaryDuration 151 152// ----------------------------------------------------------------------------- 153// Theory of operations -- Monitors lists, thread residency, etc: 154// 155// * A thread acquires ownership of a monitor by successfully 156// CAS()ing the _owner field from null to non-null. 157// 158// * Invariant: A thread appears on at most one monitor list -- 159// cxq, EntryList or WaitSet -- at any one time. 160// 161// * Contending threads "push" themselves onto the cxq with CAS 162// and then spin/park. 163// 164// * After a contending thread eventually acquires the lock it must 165// dequeue itself from either the EntryList or the cxq. 166// 167// * The exiting thread identifies and unparks an "heir presumptive" 168// tentative successor thread on the EntryList. Critically, the 169// exiting thread doesn't unlink the successor thread from the EntryList. 170// After having been unparked, the wakee will recontend for ownership of 171// the monitor. The successor (wakee) will either acquire the lock or 172// re-park itself. 173// 174// Succession is provided for by a policy of competitive handoff. 175// The exiting thread does _not_ grant or pass ownership to the 176// successor thread. (This is also referred to as "handoff" succession"). 177// Instead the exiting thread releases ownership and possibly wakes 178// a successor, so the successor can (re)compete for ownership of the lock. 179// If the EntryList is empty but the cxq is populated the exiting 180// thread will drain the cxq into the EntryList. It does so by 181// by detaching the cxq (installing null with CAS) and folding 182// the threads from the cxq into the EntryList. The EntryList is 183// doubly linked, while the cxq is singly linked because of the 184// CAS-based "push" used to enqueue recently arrived threads (RATs). 185// 186// * Concurrency invariants: 187// 188// -- only the monitor owner may access or mutate the EntryList. 189// The mutex property of the monitor itself protects the EntryList 190// from concurrent interference. 191// -- Only the monitor owner may detach the cxq. 192// 193// * The monitor entry list operations avoid locks, but strictly speaking 194// they're not lock-free. Enter is lock-free, exit is not. 195// See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 196// 197// * The cxq can have multiple concurrent "pushers" but only one concurrent 198// detaching thread. This mechanism is immune from the ABA corruption. 199// More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 200// 201// * Taken together, the cxq and the EntryList constitute or form a 202// single logical queue of threads stalled trying to acquire the lock. 203// We use two distinct lists to improve the odds of a constant-time 204// dequeue operation after acquisition (in the ::enter() epilogue) and 205// to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 206// A key desideratum is to minimize queue & monitor metadata manipulation 207// that occurs while holding the monitor lock -- that is, we want to 208// minimize monitor lock holds times. Note that even a small amount of 209// fixed spinning will greatly reduce the # of enqueue-dequeue operations 210// on EntryList|cxq. That is, spinning relieves contention on the "inner" 211// locks and monitor metadata. 212// 213// Cxq points to the the set of Recently Arrived Threads attempting entry. 214// Because we push threads onto _cxq with CAS, the RATs must take the form of 215// a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 216// the unlocking thread notices that EntryList is null but _cxq is != null. 217// 218// The EntryList is ordered by the prevailing queue discipline and 219// can be organized in any convenient fashion, such as a doubly-linked list or 220// a circular doubly-linked list. Critically, we want insert and delete operations 221// to operate in constant-time. If we need a priority queue then something akin 222// to Solaris' sleepq would work nicely. Viz., 223// http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 224// Queue discipline is enforced at ::exit() time, when the unlocking thread 225// drains the cxq into the EntryList, and orders or reorders the threads on the 226// EntryList accordingly. 227// 228// Barring "lock barging", this mechanism provides fair cyclic ordering, 229// somewhat similar to an elevator-scan. 230// 231// * The monitor synchronization subsystem avoids the use of native 232// synchronization primitives except for the narrow platform-specific 233// park-unpark abstraction. See the comments in os_solaris.cpp regarding 234// the semantics of park-unpark. Put another way, this monitor implementation 235// depends only on atomic operations and park-unpark. The monitor subsystem 236// manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 237// underlying OS manages the READY<->RUN transitions. 238// 239// * Waiting threads reside on the WaitSet list -- wait() puts 240// the caller onto the WaitSet. 241// 242// * notify() or notifyAll() simply transfers threads from the WaitSet to 243// either the EntryList or cxq. Subsequent exit() operations will 244// unpark the notifyee. Unparking a notifee in notify() is inefficient - 245// it's likely the notifyee would simply impale itself on the lock held 246// by the notifier. 247// 248// * An interesting alternative is to encode cxq as (List,LockByte) where 249// the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 250// variable, like _recursions, in the scheme. The threads or Events that form 251// the list would have to be aligned in 256-byte addresses. A thread would 252// try to acquire the lock or enqueue itself with CAS, but exiting threads 253// could use a 1-0 protocol and simply STB to set the LockByte to 0. 254// Note that is is *not* word-tearing, but it does presume that full-word 255// CAS operations are coherent with intermix with STB operations. That's true 256// on most common processors. 257// 258// * See also http://blogs.sun.com/dave 259 260 261// ----------------------------------------------------------------------------- 262// Enter support 263 264bool ObjectMonitor::try_enter(Thread* THREAD) { 265 if (THREAD != _owner) { 266 if (THREAD->is_lock_owned ((address)_owner)) { 267 assert(_recursions == 0, "internal state error"); 268 _owner = THREAD; 269 _recursions = 1; 270 OwnerIsThread = 1; 271 return true; 272 } 273 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 274 return false; 275 } 276 return true; 277 } else { 278 _recursions++; 279 return true; 280 } 281} 282 283void ATTR ObjectMonitor::enter(TRAPS) { 284 // The following code is ordered to check the most common cases first 285 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 286 Thread * const Self = THREAD; 287 void * cur; 288 289 cur = Atomic::cmpxchg_ptr(Self, &_owner, NULL); 290 if (cur == NULL) { 291 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 292 assert(_recursions == 0 , "invariant"); 293 assert(_owner == Self, "invariant"); 294 // CONSIDER: set or assert OwnerIsThread == 1 295 return; 296 } 297 298 if (cur == Self) { 299 // TODO-FIXME: check for integer overflow! BUGID 6557169. 300 _recursions++; 301 return; 302 } 303 304 if (Self->is_lock_owned ((address)cur)) { 305 assert(_recursions == 0, "internal state error"); 306 _recursions = 1; 307 // Commute owner from a thread-specific on-stack BasicLockObject address to 308 // a full-fledged "Thread *". 309 _owner = Self; 310 OwnerIsThread = 1; 311 return; 312 } 313 314 // We've encountered genuine contention. 315 assert(Self->_Stalled == 0, "invariant"); 316 Self->_Stalled = intptr_t(this); 317 318 // Try one round of spinning *before* enqueueing Self 319 // and before going through the awkward and expensive state 320 // transitions. The following spin is strictly optional ... 321 // Note that if we acquire the monitor from an initial spin 322 // we forgo posting JVMTI events and firing DTRACE probes. 323 if (Knob_SpinEarly && TrySpin (Self) > 0) { 324 assert(_owner == Self , "invariant"); 325 assert(_recursions == 0 , "invariant"); 326 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 327 Self->_Stalled = 0; 328 return; 329 } 330 331 assert(_owner != Self , "invariant"); 332 assert(_succ != Self , "invariant"); 333 assert(Self->is_Java_thread() , "invariant"); 334 JavaThread * jt = (JavaThread *) Self; 335 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 336 assert(jt->thread_state() != _thread_blocked , "invariant"); 337 assert(this->object() != NULL , "invariant"); 338 assert(_count >= 0, "invariant"); 339 340 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 341 // Ensure the object-monitor relationship remains stable while there's contention. 342 Atomic::inc_ptr(&_count); 343 344 EventJavaMonitorEnter event; 345 346 { // Change java thread status to indicate blocked on monitor enter. 347 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 348 349 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 350 if (JvmtiExport::should_post_monitor_contended_enter()) { 351 JvmtiExport::post_monitor_contended_enter(jt, this); 352 353 // The current thread does not yet own the monitor and does not 354 // yet appear on any queues that would get it made the successor. 355 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 356 // handler cannot accidentally consume an unpark() meant for the 357 // ParkEvent associated with this ObjectMonitor. 358 } 359 360 OSThreadContendState osts(Self->osthread()); 361 ThreadBlockInVM tbivm(jt); 362 363 Self->set_current_pending_monitor(this); 364 365 // TODO-FIXME: change the following for(;;) loop to straight-line code. 366 for (;;) { 367 jt->set_suspend_equivalent(); 368 // cleared by handle_special_suspend_equivalent_condition() 369 // or java_suspend_self() 370 371 EnterI(THREAD); 372 373 if (!ExitSuspendEquivalent(jt)) break; 374 375 // 376 // We have acquired the contended monitor, but while we were 377 // waiting another thread suspended us. We don't want to enter 378 // the monitor while suspended because that would surprise the 379 // thread that suspended us. 380 // 381 _recursions = 0; 382 _succ = NULL; 383 exit(false, Self); 384 385 jt->java_suspend_self(); 386 } 387 Self->set_current_pending_monitor(NULL); 388 389 // We cleared the pending monitor info since we've just gotten past 390 // the enter-check-for-suspend dance and we now own the monitor free 391 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 392 // destructor can go to a safepoint at the end of this block. If we 393 // do a thread dump during that safepoint, then this thread will show 394 // as having "-locked" the monitor, but the OS and java.lang.Thread 395 // states will still report that the thread is blocked trying to 396 // acquire it. 397 } 398 399 Atomic::dec_ptr(&_count); 400 assert(_count >= 0, "invariant"); 401 Self->_Stalled = 0; 402 403 // Must either set _recursions = 0 or ASSERT _recursions == 0. 404 assert(_recursions == 0 , "invariant"); 405 assert(_owner == Self , "invariant"); 406 assert(_succ != Self , "invariant"); 407 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 408 409 // The thread -- now the owner -- is back in vm mode. 410 // Report the glorious news via TI,DTrace and jvmstat. 411 // The probe effect is non-trivial. All the reportage occurs 412 // while we hold the monitor, increasing the length of the critical 413 // section. Amdahl's parallel speedup law comes vividly into play. 414 // 415 // Another option might be to aggregate the events (thread local or 416 // per-monitor aggregation) and defer reporting until a more opportune 417 // time -- such as next time some thread encounters contention but has 418 // yet to acquire the lock. While spinning that thread could 419 // spinning we could increment JVMStat counters, etc. 420 421 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 422 if (JvmtiExport::should_post_monitor_contended_entered()) { 423 JvmtiExport::post_monitor_contended_entered(jt, this); 424 425 // The current thread already owns the monitor and is not going to 426 // call park() for the remainder of the monitor enter protocol. So 427 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 428 // event handler consumed an unpark() issued by the thread that 429 // just exited the monitor. 430 } 431 432 if (event.should_commit()) { 433 event.set_klass(((oop)this->object())->klass()); 434 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 435 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 436 event.commit(); 437 } 438 439 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 440 ObjectMonitor::_sync_ContendedLockAttempts->inc(); 441 } 442} 443 444 445// Caveat: TryLock() is not necessarily serializing if it returns failure. 446// Callers must compensate as needed. 447 448int ObjectMonitor::TryLock (Thread * Self) { 449 for (;;) { 450 void * own = _owner; 451 if (own != NULL) return 0; 452 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 453 // Either guarantee _recursions == 0 or set _recursions = 0. 454 assert(_recursions == 0, "invariant"); 455 assert(_owner == Self, "invariant"); 456 // CONSIDER: set or assert that OwnerIsThread == 1 457 return 1; 458 } 459 // The lock had been free momentarily, but we lost the race to the lock. 460 // Interference -- the CAS failed. 461 // We can either return -1 or retry. 462 // Retry doesn't make as much sense because the lock was just acquired. 463 if (true) return -1; 464 } 465} 466 467void ATTR ObjectMonitor::EnterI (TRAPS) { 468 Thread * Self = THREAD; 469 assert(Self->is_Java_thread(), "invariant"); 470 assert(((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant"); 471 472 // Try the lock - TATAS 473 if (TryLock (Self) > 0) { 474 assert(_succ != Self , "invariant"); 475 assert(_owner == Self , "invariant"); 476 assert(_Responsible != Self , "invariant"); 477 return; 478 } 479 480 DeferredInitialize(); 481 482 // We try one round of spinning *before* enqueueing Self. 483 // 484 // If the _owner is ready but OFFPROC we could use a YieldTo() 485 // operation to donate the remainder of this thread's quantum 486 // to the owner. This has subtle but beneficial affinity 487 // effects. 488 489 if (TrySpin (Self) > 0) { 490 assert(_owner == Self , "invariant"); 491 assert(_succ != Self , "invariant"); 492 assert(_Responsible != Self , "invariant"); 493 return; 494 } 495 496 // The Spin failed -- Enqueue and park the thread ... 497 assert(_succ != Self , "invariant"); 498 assert(_owner != Self , "invariant"); 499 assert(_Responsible != Self , "invariant"); 500 501 // Enqueue "Self" on ObjectMonitor's _cxq. 502 // 503 // Node acts as a proxy for Self. 504 // As an aside, if were to ever rewrite the synchronization code mostly 505 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 506 // Java objects. This would avoid awkward lifecycle and liveness issues, 507 // as well as eliminate a subset of ABA issues. 508 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 509 // 510 511 ObjectWaiter node(Self); 512 Self->_ParkEvent->reset(); 513 node._prev = (ObjectWaiter *) 0xBAD; 514 node.TState = ObjectWaiter::TS_CXQ; 515 516 // Push "Self" onto the front of the _cxq. 517 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 518 // Note that spinning tends to reduce the rate at which threads 519 // enqueue and dequeue on EntryList|cxq. 520 ObjectWaiter * nxt; 521 for (;;) { 522 node._next = nxt = _cxq; 523 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break; 524 525 // Interference - the CAS failed because _cxq changed. Just retry. 526 // As an optional optimization we retry the lock. 527 if (TryLock (Self) > 0) { 528 assert(_succ != Self , "invariant"); 529 assert(_owner == Self , "invariant"); 530 assert(_Responsible != Self , "invariant"); 531 return; 532 } 533 } 534 535 // Check for cxq|EntryList edge transition to non-null. This indicates 536 // the onset of contention. While contention persists exiting threads 537 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 538 // operations revert to the faster 1-0 mode. This enter operation may interleave 539 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 540 // arrange for one of the contending thread to use a timed park() operations 541 // to detect and recover from the race. (Stranding is form of progress failure 542 // where the monitor is unlocked but all the contending threads remain parked). 543 // That is, at least one of the contended threads will periodically poll _owner. 544 // One of the contending threads will become the designated "Responsible" thread. 545 // The Responsible thread uses a timed park instead of a normal indefinite park 546 // operation -- it periodically wakes and checks for and recovers from potential 547 // strandings admitted by 1-0 exit operations. We need at most one Responsible 548 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 549 // be responsible for a monitor. 550 // 551 // Currently, one of the contended threads takes on the added role of "Responsible". 552 // A viable alternative would be to use a dedicated "stranding checker" thread 553 // that periodically iterated over all the threads (or active monitors) and unparked 554 // successors where there was risk of stranding. This would help eliminate the 555 // timer scalability issues we see on some platforms as we'd only have one thread 556 // -- the checker -- parked on a timer. 557 558 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 559 // Try to assume the role of responsible thread for the monitor. 560 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 561 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 562 } 563 564 // The lock have been released while this thread was occupied queueing 565 // itself onto _cxq. To close the race and avoid "stranding" and 566 // progress-liveness failure we must resample-retry _owner before parking. 567 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 568 // In this case the ST-MEMBAR is accomplished with CAS(). 569 // 570 // TODO: Defer all thread state transitions until park-time. 571 // Since state transitions are heavy and inefficient we'd like 572 // to defer the state transitions until absolutely necessary, 573 // and in doing so avoid some transitions ... 574 575 TEVENT(Inflated enter - Contention); 576 int nWakeups = 0; 577 int RecheckInterval = 1; 578 579 for (;;) { 580 581 if (TryLock(Self) > 0) break; 582 assert(_owner != Self, "invariant"); 583 584 if ((SyncFlags & 2) && _Responsible == NULL) { 585 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 586 } 587 588 // park self 589 if (_Responsible == Self || (SyncFlags & 1)) { 590 TEVENT(Inflated enter - park TIMED); 591 Self->_ParkEvent->park((jlong) RecheckInterval); 592 // Increase the RecheckInterval, but clamp the value. 593 RecheckInterval *= 8; 594 if (RecheckInterval > 1000) RecheckInterval = 1000; 595 } else { 596 TEVENT(Inflated enter - park UNTIMED); 597 Self->_ParkEvent->park(); 598 } 599 600 if (TryLock(Self) > 0) break; 601 602 // The lock is still contested. 603 // Keep a tally of the # of futile wakeups. 604 // Note that the counter is not protected by a lock or updated by atomics. 605 // That is by design - we trade "lossy" counters which are exposed to 606 // races during updates for a lower probe effect. 607 TEVENT(Inflated enter - Futile wakeup); 608 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 609 ObjectMonitor::_sync_FutileWakeups->inc(); 610 } 611 ++nWakeups; 612 613 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 614 // We can defer clearing _succ until after the spin completes 615 // TrySpin() must tolerate being called with _succ == Self. 616 // Try yet another round of adaptive spinning. 617 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 618 619 // We can find that we were unpark()ed and redesignated _succ while 620 // we were spinning. That's harmless. If we iterate and call park(), 621 // park() will consume the event and return immediately and we'll 622 // just spin again. This pattern can repeat, leaving _succ to simply 623 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 624 // Alternately, we can sample fired() here, and if set, forgo spinning 625 // in the next iteration. 626 627 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 628 Self->_ParkEvent->reset(); 629 OrderAccess::fence(); 630 } 631 if (_succ == Self) _succ = NULL; 632 633 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 634 OrderAccess::fence(); 635 } 636 637 // Egress : 638 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 639 // Normally we'll find Self on the EntryList . 640 // From the perspective of the lock owner (this thread), the 641 // EntryList is stable and cxq is prepend-only. 642 // The head of cxq is volatile but the interior is stable. 643 // In addition, Self.TState is stable. 644 645 assert(_owner == Self , "invariant"); 646 assert(object() != NULL , "invariant"); 647 // I'd like to write: 648 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 649 // but as we're at a safepoint that's not safe. 650 651 UnlinkAfterAcquire(Self, &node); 652 if (_succ == Self) _succ = NULL; 653 654 assert(_succ != Self, "invariant"); 655 if (_Responsible == Self) { 656 _Responsible = NULL; 657 OrderAccess::fence(); // Dekker pivot-point 658 659 // We may leave threads on cxq|EntryList without a designated 660 // "Responsible" thread. This is benign. When this thread subsequently 661 // exits the monitor it can "see" such preexisting "old" threads -- 662 // threads that arrived on the cxq|EntryList before the fence, above -- 663 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 664 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 665 // non-null and elect a new "Responsible" timer thread. 666 // 667 // This thread executes: 668 // ST Responsible=null; MEMBAR (in enter epilogue - here) 669 // LD cxq|EntryList (in subsequent exit) 670 // 671 // Entering threads in the slow/contended path execute: 672 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 673 // The (ST cxq; MEMBAR) is accomplished with CAS(). 674 // 675 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 676 // exit operation from floating above the ST Responsible=null. 677 } 678 679 // We've acquired ownership with CAS(). 680 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 681 // But since the CAS() this thread may have also stored into _succ, 682 // EntryList, cxq or Responsible. These meta-data updates must be 683 // visible __before this thread subsequently drops the lock. 684 // Consider what could occur if we didn't enforce this constraint -- 685 // STs to monitor meta-data and user-data could reorder with (become 686 // visible after) the ST in exit that drops ownership of the lock. 687 // Some other thread could then acquire the lock, but observe inconsistent 688 // or old monitor meta-data and heap data. That violates the JMM. 689 // To that end, the 1-0 exit() operation must have at least STST|LDST 690 // "release" barrier semantics. Specifically, there must be at least a 691 // STST|LDST barrier in exit() before the ST of null into _owner that drops 692 // the lock. The barrier ensures that changes to monitor meta-data and data 693 // protected by the lock will be visible before we release the lock, and 694 // therefore before some other thread (CPU) has a chance to acquire the lock. 695 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 696 // 697 // Critically, any prior STs to _succ or EntryList must be visible before 698 // the ST of null into _owner in the *subsequent* (following) corresponding 699 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 700 // execute a serializing instruction. 701 702 if (SyncFlags & 8) { 703 OrderAccess::fence(); 704 } 705 return; 706} 707 708// ReenterI() is a specialized inline form of the latter half of the 709// contended slow-path from EnterI(). We use ReenterI() only for 710// monitor reentry in wait(). 711// 712// In the future we should reconcile EnterI() and ReenterI(), adding 713// Knob_Reset and Knob_SpinAfterFutile support and restructuring the 714// loop accordingly. 715 716void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 717 assert(Self != NULL , "invariant"); 718 assert(SelfNode != NULL , "invariant"); 719 assert(SelfNode->_thread == Self , "invariant"); 720 assert(_waiters > 0 , "invariant"); 721 assert(((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant"); 722 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 723 JavaThread * jt = (JavaThread *) Self; 724 725 int nWakeups = 0; 726 for (;;) { 727 ObjectWaiter::TStates v = SelfNode->TState; 728 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 729 assert(_owner != Self, "invariant"); 730 731 if (TryLock(Self) > 0) break; 732 if (TrySpin(Self) > 0) break; 733 734 TEVENT(Wait Reentry - parking); 735 736 // State transition wrappers around park() ... 737 // ReenterI() wisely defers state transitions until 738 // it's clear we must park the thread. 739 { 740 OSThreadContendState osts(Self->osthread()); 741 ThreadBlockInVM tbivm(jt); 742 743 // cleared by handle_special_suspend_equivalent_condition() 744 // or java_suspend_self() 745 jt->set_suspend_equivalent(); 746 if (SyncFlags & 1) { 747 Self->_ParkEvent->park((jlong)1000); 748 } else { 749 Self->_ParkEvent->park(); 750 } 751 752 // were we externally suspended while we were waiting? 753 for (;;) { 754 if (!ExitSuspendEquivalent(jt)) break; 755 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 756 jt->java_suspend_self(); 757 jt->set_suspend_equivalent(); 758 } 759 } 760 761 // Try again, but just so we distinguish between futile wakeups and 762 // successful wakeups. The following test isn't algorithmically 763 // necessary, but it helps us maintain sensible statistics. 764 if (TryLock(Self) > 0) break; 765 766 // The lock is still contested. 767 // Keep a tally of the # of futile wakeups. 768 // Note that the counter is not protected by a lock or updated by atomics. 769 // That is by design - we trade "lossy" counters which are exposed to 770 // races during updates for a lower probe effect. 771 TEVENT(Wait Reentry - futile wakeup); 772 ++nWakeups; 773 774 // Assuming this is not a spurious wakeup we'll normally 775 // find that _succ == Self. 776 if (_succ == Self) _succ = NULL; 777 778 // Invariant: after clearing _succ a contending thread 779 // *must* retry _owner before parking. 780 OrderAccess::fence(); 781 782 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 783 ObjectMonitor::_sync_FutileWakeups->inc(); 784 } 785 } 786 787 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 788 // Normally we'll find Self on the EntryList. 789 // Unlinking from the EntryList is constant-time and atomic-free. 790 // From the perspective of the lock owner (this thread), the 791 // EntryList is stable and cxq is prepend-only. 792 // The head of cxq is volatile but the interior is stable. 793 // In addition, Self.TState is stable. 794 795 assert(_owner == Self, "invariant"); 796 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 797 UnlinkAfterAcquire(Self, SelfNode); 798 if (_succ == Self) _succ = NULL; 799 assert(_succ != Self, "invariant"); 800 SelfNode->TState = ObjectWaiter::TS_RUN; 801 OrderAccess::fence(); // see comments at the end of EnterI() 802} 803 804// after the thread acquires the lock in ::enter(). Equally, we could defer 805// unlinking the thread until ::exit()-time. 806 807void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 808{ 809 assert(_owner == Self, "invariant"); 810 assert(SelfNode->_thread == Self, "invariant"); 811 812 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 813 // Normal case: remove Self from the DLL EntryList . 814 // This is a constant-time operation. 815 ObjectWaiter * nxt = SelfNode->_next; 816 ObjectWaiter * prv = SelfNode->_prev; 817 if (nxt != NULL) nxt->_prev = prv; 818 if (prv != NULL) prv->_next = nxt; 819 if (SelfNode == _EntryList) _EntryList = nxt; 820 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 821 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 822 TEVENT(Unlink from EntryList); 823 } else { 824 guarantee(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 825 // Inopportune interleaving -- Self is still on the cxq. 826 // This usually means the enqueue of self raced an exiting thread. 827 // Normally we'll find Self near the front of the cxq, so 828 // dequeueing is typically fast. If needbe we can accelerate 829 // this with some MCS/CHL-like bidirectional list hints and advisory 830 // back-links so dequeueing from the interior will normally operate 831 // in constant-time. 832 // Dequeue Self from either the head (with CAS) or from the interior 833 // with a linear-time scan and normal non-atomic memory operations. 834 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 835 // and then unlink Self from EntryList. We have to drain eventually, 836 // so it might as well be now. 837 838 ObjectWaiter * v = _cxq; 839 assert(v != NULL, "invariant"); 840 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 841 // The CAS above can fail from interference IFF a "RAT" arrived. 842 // In that case Self must be in the interior and can no longer be 843 // at the head of cxq. 844 if (v == SelfNode) { 845 assert(_cxq != v, "invariant"); 846 v = _cxq; // CAS above failed - start scan at head of list 847 } 848 ObjectWaiter * p; 849 ObjectWaiter * q = NULL; 850 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 851 q = p; 852 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 853 } 854 assert(v != SelfNode, "invariant"); 855 assert(p == SelfNode, "Node not found on cxq"); 856 assert(p != _cxq, "invariant"); 857 assert(q != NULL, "invariant"); 858 assert(q->_next == p, "invariant"); 859 q->_next = p->_next; 860 } 861 TEVENT(Unlink from cxq); 862 } 863 864 // Diagnostic hygiene ... 865 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 866 SelfNode->_next = (ObjectWaiter *) 0xBAD; 867 SelfNode->TState = ObjectWaiter::TS_RUN; 868} 869 870// ----------------------------------------------------------------------------- 871// Exit support 872// 873// exit() 874// ~~~~~~ 875// Note that the collector can't reclaim the objectMonitor or deflate 876// the object out from underneath the thread calling ::exit() as the 877// thread calling ::exit() never transitions to a stable state. 878// This inhibits GC, which in turn inhibits asynchronous (and 879// inopportune) reclamation of "this". 880// 881// We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 882// There's one exception to the claim above, however. EnterI() can call 883// exit() to drop a lock if the acquirer has been externally suspended. 884// In that case exit() is called with _thread_state as _thread_blocked, 885// but the monitor's _count field is > 0, which inhibits reclamation. 886// 887// 1-0 exit 888// ~~~~~~~~ 889// ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 890// the fast-path operators have been optimized so the common ::exit() 891// operation is 1-0. See i486.ad fast_unlock(), for instance. 892// The code emitted by fast_unlock() elides the usual MEMBAR. This 893// greatly improves latency -- MEMBAR and CAS having considerable local 894// latency on modern processors -- but at the cost of "stranding". Absent the 895// MEMBAR, a thread in fast_unlock() can race a thread in the slow 896// ::enter() path, resulting in the entering thread being stranding 897// and a progress-liveness failure. Stranding is extremely rare. 898// We use timers (timed park operations) & periodic polling to detect 899// and recover from stranding. Potentially stranded threads periodically 900// wake up and poll the lock. See the usage of the _Responsible variable. 901// 902// The CAS() in enter provides for safety and exclusion, while the CAS or 903// MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 904// eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 905// We detect and recover from stranding with timers. 906// 907// If a thread transiently strands it'll park until (a) another 908// thread acquires the lock and then drops the lock, at which time the 909// exiting thread will notice and unpark the stranded thread, or, (b) 910// the timer expires. If the lock is high traffic then the stranding latency 911// will be low due to (a). If the lock is low traffic then the odds of 912// stranding are lower, although the worst-case stranding latency 913// is longer. Critically, we don't want to put excessive load in the 914// platform's timer subsystem. We want to minimize both the timer injection 915// rate (timers created/sec) as well as the number of timers active at 916// any one time. (more precisely, we want to minimize timer-seconds, which is 917// the integral of the # of active timers at any instant over time). 918// Both impinge on OS scalability. Given that, at most one thread parked on 919// a monitor will use a timer. 920 921void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 922 Thread * Self = THREAD; 923 if (THREAD != _owner) { 924 if (THREAD->is_lock_owned((address) _owner)) { 925 // Transmute _owner from a BasicLock pointer to a Thread address. 926 // We don't need to hold _mutex for this transition. 927 // Non-null to Non-null is safe as long as all readers can 928 // tolerate either flavor. 929 assert(_recursions == 0, "invariant"); 930 _owner = THREAD; 931 _recursions = 0; 932 OwnerIsThread = 1; 933 } else { 934 // NOTE: we need to handle unbalanced monitor enter/exit 935 // in native code by throwing an exception. 936 // TODO: Throw an IllegalMonitorStateException ? 937 TEVENT(Exit - Throw IMSX); 938 assert(false, "Non-balanced monitor enter/exit!"); 939 if (false) { 940 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 941 } 942 return; 943 } 944 } 945 946 if (_recursions != 0) { 947 _recursions--; // this is simple recursive enter 948 TEVENT(Inflated exit - recursive); 949 return; 950 } 951 952 // Invariant: after setting Responsible=null an thread must execute 953 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 954 if ((SyncFlags & 4) == 0) { 955 _Responsible = NULL; 956 } 957 958#if INCLUDE_TRACE 959 // get the owner's thread id for the MonitorEnter event 960 // if it is enabled and the thread isn't suspended 961 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 962 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 963 } 964#endif 965 966 for (;;) { 967 assert(THREAD == _owner, "invariant"); 968 969 970 if (Knob_ExitPolicy == 0) { 971 // release semantics: prior loads and stores from within the critical section 972 // must not float (reorder) past the following store that drops the lock. 973 // On SPARC that requires MEMBAR #loadstore|#storestore. 974 // But of course in TSO #loadstore|#storestore is not required. 975 // I'd like to write one of the following: 976 // A. OrderAccess::release() ; _owner = NULL 977 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 978 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 979 // store into a _dummy variable. That store is not needed, but can result 980 // in massive wasteful coherency traffic on classic SMP systems. 981 // Instead, I use release_store(), which is implemented as just a simple 982 // ST on x64, x86 and SPARC. 983 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 984 OrderAccess::storeload(); // See if we need to wake a successor 985 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 986 TEVENT(Inflated exit - simple egress); 987 return; 988 } 989 TEVENT(Inflated exit - complex egress); 990 991 // Normally the exiting thread is responsible for ensuring succession, 992 // but if other successors are ready or other entering threads are spinning 993 // then this thread can simply store NULL into _owner and exit without 994 // waking a successor. The existence of spinners or ready successors 995 // guarantees proper succession (liveness). Responsibility passes to the 996 // ready or running successors. The exiting thread delegates the duty. 997 // More precisely, if a successor already exists this thread is absolved 998 // of the responsibility of waking (unparking) one. 999 // 1000 // The _succ variable is critical to reducing futile wakeup frequency. 1001 // _succ identifies the "heir presumptive" thread that has been made 1002 // ready (unparked) but that has not yet run. We need only one such 1003 // successor thread to guarantee progress. 1004 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1005 // section 3.3 "Futile Wakeup Throttling" for details. 1006 // 1007 // Note that spinners in Enter() also set _succ non-null. 1008 // In the current implementation spinners opportunistically set 1009 // _succ so that exiting threads might avoid waking a successor. 1010 // Another less appealing alternative would be for the exiting thread 1011 // to drop the lock and then spin briefly to see if a spinner managed 1012 // to acquire the lock. If so, the exiting thread could exit 1013 // immediately without waking a successor, otherwise the exiting 1014 // thread would need to dequeue and wake a successor. 1015 // (Note that we'd need to make the post-drop spin short, but no 1016 // shorter than the worst-case round-trip cache-line migration time. 1017 // The dropped lock needs to become visible to the spinner, and then 1018 // the acquisition of the lock by the spinner must become visible to 1019 // the exiting thread). 1020 // 1021 1022 // It appears that an heir-presumptive (successor) must be made ready. 1023 // Only the current lock owner can manipulate the EntryList or 1024 // drain _cxq, so we need to reacquire the lock. If we fail 1025 // to reacquire the lock the responsibility for ensuring succession 1026 // falls to the new owner. 1027 // 1028 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1029 return; 1030 } 1031 TEVENT(Exit - Reacquired); 1032 } else { 1033 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1034 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1035 OrderAccess::storeload(); 1036 // Ratify the previously observed values. 1037 if (_cxq == NULL || _succ != NULL) { 1038 TEVENT(Inflated exit - simple egress); 1039 return; 1040 } 1041 1042 // inopportune interleaving -- the exiting thread (this thread) 1043 // in the fast-exit path raced an entering thread in the slow-enter 1044 // path. 1045 // We have two choices: 1046 // A. Try to reacquire the lock. 1047 // If the CAS() fails return immediately, otherwise 1048 // we either restart/rerun the exit operation, or simply 1049 // fall-through into the code below which wakes a successor. 1050 // B. If the elements forming the EntryList|cxq are TSM 1051 // we could simply unpark() the lead thread and return 1052 // without having set _succ. 1053 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1054 TEVENT(Inflated exit - reacquired succeeded); 1055 return; 1056 } 1057 TEVENT(Inflated exit - reacquired failed); 1058 } else { 1059 TEVENT(Inflated exit - complex egress); 1060 } 1061 } 1062 1063 guarantee(_owner == THREAD, "invariant"); 1064 1065 ObjectWaiter * w = NULL; 1066 int QMode = Knob_QMode; 1067 1068 if (QMode == 2 && _cxq != NULL) { 1069 // QMode == 2 : cxq has precedence over EntryList. 1070 // Try to directly wake a successor from the cxq. 1071 // If successful, the successor will need to unlink itself from cxq. 1072 w = _cxq; 1073 assert(w != NULL, "invariant"); 1074 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1075 ExitEpilog(Self, w); 1076 return; 1077 } 1078 1079 if (QMode == 3 && _cxq != NULL) { 1080 // Aggressively drain cxq into EntryList at the first opportunity. 1081 // This policy ensure that recently-run threads live at the head of EntryList. 1082 // Drain _cxq into EntryList - bulk transfer. 1083 // First, detach _cxq. 1084 // The following loop is tantamount to: w = swap (&cxq, NULL) 1085 w = _cxq; 1086 for (;;) { 1087 assert(w != NULL, "Invariant"); 1088 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1089 if (u == w) break; 1090 w = u; 1091 } 1092 assert(w != NULL , "invariant"); 1093 1094 ObjectWaiter * q = NULL; 1095 ObjectWaiter * p; 1096 for (p = w; p != NULL; p = p->_next) { 1097 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1098 p->TState = ObjectWaiter::TS_ENTER; 1099 p->_prev = q; 1100 q = p; 1101 } 1102 1103 // Append the RATs to the EntryList 1104 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1105 ObjectWaiter * Tail; 1106 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; Tail = Tail->_next); 1107 if (Tail == NULL) { 1108 _EntryList = w; 1109 } else { 1110 Tail->_next = w; 1111 w->_prev = Tail; 1112 } 1113 1114 // Fall thru into code that tries to wake a successor from EntryList 1115 } 1116 1117 if (QMode == 4 && _cxq != NULL) { 1118 // Aggressively drain cxq into EntryList at the first opportunity. 1119 // This policy ensure that recently-run threads live at the head of EntryList. 1120 1121 // Drain _cxq into EntryList - bulk transfer. 1122 // First, detach _cxq. 1123 // The following loop is tantamount to: w = swap (&cxq, NULL) 1124 w = _cxq; 1125 for (;;) { 1126 assert(w != NULL, "Invariant"); 1127 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1128 if (u == w) break; 1129 w = u; 1130 } 1131 assert(w != NULL , "invariant"); 1132 1133 ObjectWaiter * q = NULL; 1134 ObjectWaiter * p; 1135 for (p = w; p != NULL; p = p->_next) { 1136 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1137 p->TState = ObjectWaiter::TS_ENTER; 1138 p->_prev = q; 1139 q = p; 1140 } 1141 1142 // Prepend the RATs to the EntryList 1143 if (_EntryList != NULL) { 1144 q->_next = _EntryList; 1145 _EntryList->_prev = q; 1146 } 1147 _EntryList = w; 1148 1149 // Fall thru into code that tries to wake a successor from EntryList 1150 } 1151 1152 w = _EntryList; 1153 if (w != NULL) { 1154 // I'd like to write: guarantee (w->_thread != Self). 1155 // But in practice an exiting thread may find itself on the EntryList. 1156 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1157 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1158 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1159 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1160 // release the lock "O". T2 resumes immediately after the ST of null into 1161 // _owner, above. T2 notices that the EntryList is populated, so it 1162 // reacquires the lock and then finds itself on the EntryList. 1163 // Given all that, we have to tolerate the circumstance where "w" is 1164 // associated with Self. 1165 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1166 ExitEpilog(Self, w); 1167 return; 1168 } 1169 1170 // If we find that both _cxq and EntryList are null then just 1171 // re-run the exit protocol from the top. 1172 w = _cxq; 1173 if (w == NULL) continue; 1174 1175 // Drain _cxq into EntryList - bulk transfer. 1176 // First, detach _cxq. 1177 // The following loop is tantamount to: w = swap (&cxq, NULL) 1178 for (;;) { 1179 assert(w != NULL, "Invariant"); 1180 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1181 if (u == w) break; 1182 w = u; 1183 } 1184 TEVENT(Inflated exit - drain cxq into EntryList); 1185 1186 assert(w != NULL , "invariant"); 1187 assert(_EntryList == NULL , "invariant"); 1188 1189 // Convert the LIFO SLL anchored by _cxq into a DLL. 1190 // The list reorganization step operates in O(LENGTH(w)) time. 1191 // It's critical that this step operate quickly as 1192 // "Self" still holds the outer-lock, restricting parallelism 1193 // and effectively lengthening the critical section. 1194 // Invariant: s chases t chases u. 1195 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1196 // we have faster access to the tail. 1197 1198 if (QMode == 1) { 1199 // QMode == 1 : drain cxq to EntryList, reversing order 1200 // We also reverse the order of the list. 1201 ObjectWaiter * s = NULL; 1202 ObjectWaiter * t = w; 1203 ObjectWaiter * u = NULL; 1204 while (t != NULL) { 1205 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1206 t->TState = ObjectWaiter::TS_ENTER; 1207 u = t->_next; 1208 t->_prev = u; 1209 t->_next = s; 1210 s = t; 1211 t = u; 1212 } 1213 _EntryList = s; 1214 assert(s != NULL, "invariant"); 1215 } else { 1216 // QMode == 0 or QMode == 2 1217 _EntryList = w; 1218 ObjectWaiter * q = NULL; 1219 ObjectWaiter * p; 1220 for (p = w; p != NULL; p = p->_next) { 1221 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1222 p->TState = ObjectWaiter::TS_ENTER; 1223 p->_prev = q; 1224 q = p; 1225 } 1226 } 1227 1228 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1229 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1230 1231 // See if we can abdicate to a spinner instead of waking a thread. 1232 // A primary goal of the implementation is to reduce the 1233 // context-switch rate. 1234 if (_succ != NULL) continue; 1235 1236 w = _EntryList; 1237 if (w != NULL) { 1238 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1239 ExitEpilog(Self, w); 1240 return; 1241 } 1242 } 1243} 1244 1245// ExitSuspendEquivalent: 1246// A faster alternate to handle_special_suspend_equivalent_condition() 1247// 1248// handle_special_suspend_equivalent_condition() unconditionally 1249// acquires the SR_lock. On some platforms uncontended MutexLocker() 1250// operations have high latency. Note that in ::enter() we call HSSEC 1251// while holding the monitor, so we effectively lengthen the critical sections. 1252// 1253// There are a number of possible solutions: 1254// 1255// A. To ameliorate the problem we might also defer state transitions 1256// to as late as possible -- just prior to parking. 1257// Given that, we'd call HSSEC after having returned from park(), 1258// but before attempting to acquire the monitor. This is only a 1259// partial solution. It avoids calling HSSEC while holding the 1260// monitor (good), but it still increases successor reacquisition latency -- 1261// the interval between unparking a successor and the time the successor 1262// resumes and retries the lock. See ReenterI(), which defers state transitions. 1263// If we use this technique we can also avoid EnterI()-exit() loop 1264// in ::enter() where we iteratively drop the lock and then attempt 1265// to reacquire it after suspending. 1266// 1267// B. In the future we might fold all the suspend bits into a 1268// composite per-thread suspend flag and then update it with CAS(). 1269// Alternately, a Dekker-like mechanism with multiple variables 1270// would suffice: 1271// ST Self->_suspend_equivalent = false 1272// MEMBAR 1273// LD Self_>_suspend_flags 1274// 1275 1276 1277bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1278 int Mode = Knob_FastHSSEC; 1279 if (Mode && !jSelf->is_external_suspend()) { 1280 assert(jSelf->is_suspend_equivalent(), "invariant"); 1281 jSelf->clear_suspend_equivalent(); 1282 if (2 == Mode) OrderAccess::storeload(); 1283 if (!jSelf->is_external_suspend()) return false; 1284 // We raced a suspension -- fall thru into the slow path 1285 TEVENT(ExitSuspendEquivalent - raced); 1286 jSelf->set_suspend_equivalent(); 1287 } 1288 return jSelf->handle_special_suspend_equivalent_condition(); 1289} 1290 1291 1292void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1293 assert(_owner == Self, "invariant"); 1294 1295 // Exit protocol: 1296 // 1. ST _succ = wakee 1297 // 2. membar #loadstore|#storestore; 1298 // 2. ST _owner = NULL 1299 // 3. unpark(wakee) 1300 1301 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1302 ParkEvent * Trigger = Wakee->_event; 1303 1304 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1305 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1306 // out-of-scope (non-extant). 1307 Wakee = NULL; 1308 1309 // Drop the lock 1310 OrderAccess::release_store_ptr(&_owner, NULL); 1311 OrderAccess::fence(); // ST _owner vs LD in unpark() 1312 1313 if (SafepointSynchronize::do_call_back()) { 1314 TEVENT(unpark before SAFEPOINT); 1315 } 1316 1317 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1318 Trigger->unpark(); 1319 1320 // Maintain stats and report events to JVMTI 1321 if (ObjectMonitor::_sync_Parks != NULL) { 1322 ObjectMonitor::_sync_Parks->inc(); 1323 } 1324} 1325 1326 1327// ----------------------------------------------------------------------------- 1328// Class Loader deadlock handling. 1329// 1330// complete_exit exits a lock returning recursion count 1331// complete_exit/reenter operate as a wait without waiting 1332// complete_exit requires an inflated monitor 1333// The _owner field is not always the Thread addr even with an 1334// inflated monitor, e.g. the monitor can be inflated by a non-owning 1335// thread due to contention. 1336intptr_t ObjectMonitor::complete_exit(TRAPS) { 1337 Thread * const Self = THREAD; 1338 assert(Self->is_Java_thread(), "Must be Java thread!"); 1339 JavaThread *jt = (JavaThread *)THREAD; 1340 1341 DeferredInitialize(); 1342 1343 if (THREAD != _owner) { 1344 if (THREAD->is_lock_owned ((address)_owner)) { 1345 assert(_recursions == 0, "internal state error"); 1346 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ 1347 _recursions = 0; 1348 OwnerIsThread = 1; 1349 } 1350 } 1351 1352 guarantee(Self == _owner, "complete_exit not owner"); 1353 intptr_t save = _recursions; // record the old recursion count 1354 _recursions = 0; // set the recursion level to be 0 1355 exit(true, Self); // exit the monitor 1356 guarantee(_owner != Self, "invariant"); 1357 return save; 1358} 1359 1360// reenter() enters a lock and sets recursion count 1361// complete_exit/reenter operate as a wait without waiting 1362void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1363 Thread * const Self = THREAD; 1364 assert(Self->is_Java_thread(), "Must be Java thread!"); 1365 JavaThread *jt = (JavaThread *)THREAD; 1366 1367 guarantee(_owner != Self, "reenter already owner"); 1368 enter(THREAD); // enter the monitor 1369 guarantee(_recursions == 0, "reenter recursion"); 1370 _recursions = recursions; 1371 return; 1372} 1373 1374 1375// ----------------------------------------------------------------------------- 1376// A macro is used below because there may already be a pending 1377// exception which should not abort the execution of the routines 1378// which use this (which is why we don't put this into check_slow and 1379// call it with a CHECK argument). 1380 1381#define CHECK_OWNER() \ 1382 do { \ 1383 if (THREAD != _owner) { \ 1384 if (THREAD->is_lock_owned((address) _owner)) { \ 1385 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1386 _recursions = 0; \ 1387 OwnerIsThread = 1; \ 1388 } else { \ 1389 TEVENT(Throw IMSX); \ 1390 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1391 } \ 1392 } \ 1393 } while (false) 1394 1395// check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1396// TODO-FIXME: remove check_slow() -- it's likely dead. 1397 1398void ObjectMonitor::check_slow(TRAPS) { 1399 TEVENT(check_slow - throw IMSX); 1400 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1401 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1402} 1403 1404static int Adjust (volatile int * adr, int dx) { 1405 int v; 1406 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr); 1407 return v; 1408} 1409 1410// helper method for posting a monitor wait event 1411void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1412 jlong notifier_tid, 1413 jlong timeout, 1414 bool timedout) { 1415 event->set_klass(((oop)this->object())->klass()); 1416 event->set_timeout((TYPE_ULONG)timeout); 1417 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1418 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1419 event->set_timedOut((TYPE_BOOLEAN)timedout); 1420 event->commit(); 1421} 1422 1423// ----------------------------------------------------------------------------- 1424// Wait/Notify/NotifyAll 1425// 1426// Note: a subset of changes to ObjectMonitor::wait() 1427// will need to be replicated in complete_exit above 1428void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1429 Thread * const Self = THREAD; 1430 assert(Self->is_Java_thread(), "Must be Java thread!"); 1431 JavaThread *jt = (JavaThread *)THREAD; 1432 1433 DeferredInitialize(); 1434 1435 // Throw IMSX or IEX. 1436 CHECK_OWNER(); 1437 1438 EventJavaMonitorWait event; 1439 1440 // check for a pending interrupt 1441 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1442 // post monitor waited event. Note that this is past-tense, we are done waiting. 1443 if (JvmtiExport::should_post_monitor_waited()) { 1444 // Note: 'false' parameter is passed here because the 1445 // wait was not timed out due to thread interrupt. 1446 JvmtiExport::post_monitor_waited(jt, this, false); 1447 1448 // In this short circuit of the monitor wait protocol, the 1449 // current thread never drops ownership of the monitor and 1450 // never gets added to the wait queue so the current thread 1451 // cannot be made the successor. This means that the 1452 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1453 // consume an unpark() meant for the ParkEvent associated with 1454 // this ObjectMonitor. 1455 } 1456 if (event.should_commit()) { 1457 post_monitor_wait_event(&event, 0, millis, false); 1458 } 1459 TEVENT(Wait - Throw IEX); 1460 THROW(vmSymbols::java_lang_InterruptedException()); 1461 return; 1462 } 1463 1464 TEVENT(Wait); 1465 1466 assert(Self->_Stalled == 0, "invariant"); 1467 Self->_Stalled = intptr_t(this); 1468 jt->set_current_waiting_monitor(this); 1469 1470 // create a node to be put into the queue 1471 // Critically, after we reset() the event but prior to park(), we must check 1472 // for a pending interrupt. 1473 ObjectWaiter node(Self); 1474 node.TState = ObjectWaiter::TS_WAIT; 1475 Self->_ParkEvent->reset(); 1476 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1477 1478 // Enter the waiting queue, which is a circular doubly linked list in this case 1479 // but it could be a priority queue or any data structure. 1480 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1481 // by the the owner of the monitor *except* in the case where park() 1482 // returns because of a timeout of interrupt. Contention is exceptionally rare 1483 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1484 1485 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1486 AddWaiter(&node); 1487 Thread::SpinRelease(&_WaitSetLock); 1488 1489 if ((SyncFlags & 4) == 0) { 1490 _Responsible = NULL; 1491 } 1492 intptr_t save = _recursions; // record the old recursion count 1493 _waiters++; // increment the number of waiters 1494 _recursions = 0; // set the recursion level to be 1 1495 exit(true, Self); // exit the monitor 1496 guarantee(_owner != Self, "invariant"); 1497 1498 // The thread is on the WaitSet list - now park() it. 1499 // On MP systems it's conceivable that a brief spin before we park 1500 // could be profitable. 1501 // 1502 // TODO-FIXME: change the following logic to a loop of the form 1503 // while (!timeout && !interrupted && _notified == 0) park() 1504 1505 int ret = OS_OK; 1506 int WasNotified = 0; 1507 { // State transition wrappers 1508 OSThread* osthread = Self->osthread(); 1509 OSThreadWaitState osts(osthread, true); 1510 { 1511 ThreadBlockInVM tbivm(jt); 1512 // Thread is in thread_blocked state and oop access is unsafe. 1513 jt->set_suspend_equivalent(); 1514 1515 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1516 // Intentionally empty 1517 } else 1518 if (node._notified == 0) { 1519 if (millis <= 0) { 1520 Self->_ParkEvent->park(); 1521 } else { 1522 ret = Self->_ParkEvent->park(millis); 1523 } 1524 } 1525 1526 // were we externally suspended while we were waiting? 1527 if (ExitSuspendEquivalent (jt)) { 1528 // TODO-FIXME: add -- if succ == Self then succ = null. 1529 jt->java_suspend_self(); 1530 } 1531 1532 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1533 1534 1535 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1536 // from the WaitSet to the EntryList. 1537 // See if we need to remove Node from the WaitSet. 1538 // We use double-checked locking to avoid grabbing _WaitSetLock 1539 // if the thread is not on the wait queue. 1540 // 1541 // Note that we don't need a fence before the fetch of TState. 1542 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1543 // written by the is thread. (perhaps the fetch might even be satisfied 1544 // by a look-aside into the processor's own store buffer, although given 1545 // the length of the code path between the prior ST and this load that's 1546 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1547 // then we'll acquire the lock and then re-fetch a fresh TState value. 1548 // That is, we fail toward safety. 1549 1550 if (node.TState == ObjectWaiter::TS_WAIT) { 1551 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1552 if (node.TState == ObjectWaiter::TS_WAIT) { 1553 DequeueSpecificWaiter(&node); // unlink from WaitSet 1554 assert(node._notified == 0, "invariant"); 1555 node.TState = ObjectWaiter::TS_RUN; 1556 } 1557 Thread::SpinRelease(&_WaitSetLock); 1558 } 1559 1560 // The thread is now either on off-list (TS_RUN), 1561 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1562 // The Node's TState variable is stable from the perspective of this thread. 1563 // No other threads will asynchronously modify TState. 1564 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1565 OrderAccess::loadload(); 1566 if (_succ == Self) _succ = NULL; 1567 WasNotified = node._notified; 1568 1569 // Reentry phase -- reacquire the monitor. 1570 // re-enter contended monitor after object.wait(). 1571 // retain OBJECT_WAIT state until re-enter successfully completes 1572 // Thread state is thread_in_vm and oop access is again safe, 1573 // although the raw address of the object may have changed. 1574 // (Don't cache naked oops over safepoints, of course). 1575 1576 // post monitor waited event. Note that this is past-tense, we are done waiting. 1577 if (JvmtiExport::should_post_monitor_waited()) { 1578 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1579 1580 if (node._notified != 0 && _succ == Self) { 1581 // In this part of the monitor wait-notify-reenter protocol it 1582 // is possible (and normal) for another thread to do a fastpath 1583 // monitor enter-exit while this thread is still trying to get 1584 // to the reenter portion of the protocol. 1585 // 1586 // The ObjectMonitor was notified and the current thread is 1587 // the successor which also means that an unpark() has already 1588 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1589 // consume the unpark() that was done when the successor was 1590 // set because the same ParkEvent is shared between Java 1591 // monitors and JVM/TI RawMonitors (for now). 1592 // 1593 // We redo the unpark() to ensure forward progress, i.e., we 1594 // don't want all pending threads hanging (parked) with none 1595 // entering the unlocked monitor. 1596 node._event->unpark(); 1597 } 1598 } 1599 1600 if (event.should_commit()) { 1601 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1602 } 1603 1604 OrderAccess::fence(); 1605 1606 assert(Self->_Stalled != 0, "invariant"); 1607 Self->_Stalled = 0; 1608 1609 assert(_owner != Self, "invariant"); 1610 ObjectWaiter::TStates v = node.TState; 1611 if (v == ObjectWaiter::TS_RUN) { 1612 enter(Self); 1613 } else { 1614 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1615 ReenterI(Self, &node); 1616 node.wait_reenter_end(this); 1617 } 1618 1619 // Self has reacquired the lock. 1620 // Lifecycle - the node representing Self must not appear on any queues. 1621 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1622 // want residual elements associated with this thread left on any lists. 1623 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1624 assert(_owner == Self, "invariant"); 1625 assert(_succ != Self , "invariant"); 1626 } // OSThreadWaitState() 1627 1628 jt->set_current_waiting_monitor(NULL); 1629 1630 guarantee(_recursions == 0, "invariant"); 1631 _recursions = save; // restore the old recursion count 1632 _waiters--; // decrement the number of waiters 1633 1634 // Verify a few postconditions 1635 assert(_owner == Self , "invariant"); 1636 assert(_succ != Self , "invariant"); 1637 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1638 1639 if (SyncFlags & 32) { 1640 OrderAccess::fence(); 1641 } 1642 1643 // check if the notification happened 1644 if (!WasNotified) { 1645 // no, it could be timeout or Thread.interrupt() or both 1646 // check for interrupt event, otherwise it is timeout 1647 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1648 TEVENT(Wait - throw IEX from epilog); 1649 THROW(vmSymbols::java_lang_InterruptedException()); 1650 } 1651 } 1652 1653 // NOTE: Spurious wake up will be consider as timeout. 1654 // Monitor notify has precedence over thread interrupt. 1655} 1656 1657 1658// Consider: 1659// If the lock is cool (cxq == null && succ == null) and we're on an MP system 1660// then instead of transferring a thread from the WaitSet to the EntryList 1661// we might just dequeue a thread from the WaitSet and directly unpark() it. 1662 1663void ObjectMonitor::notify(TRAPS) { 1664 CHECK_OWNER(); 1665 if (_WaitSet == NULL) { 1666 TEVENT(Empty-Notify); 1667 return; 1668 } 1669 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1670 1671 int Policy = Knob_MoveNotifyee; 1672 1673 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1674 ObjectWaiter * iterator = DequeueWaiter(); 1675 if (iterator != NULL) { 1676 TEVENT(Notify1 - Transfer); 1677 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1678 guarantee(iterator->_notified == 0, "invariant"); 1679 if (Policy != 4) { 1680 iterator->TState = ObjectWaiter::TS_ENTER; 1681 } 1682 iterator->_notified = 1; 1683 Thread * Self = THREAD; 1684 iterator->_notifier_tid = Self->osthread()->thread_id(); 1685 1686 ObjectWaiter * List = _EntryList; 1687 if (List != NULL) { 1688 assert(List->_prev == NULL, "invariant"); 1689 assert(List->TState == ObjectWaiter::TS_ENTER, "invariant"); 1690 assert(List != iterator, "invariant"); 1691 } 1692 1693 if (Policy == 0) { // prepend to EntryList 1694 if (List == NULL) { 1695 iterator->_next = iterator->_prev = NULL; 1696 _EntryList = iterator; 1697 } else { 1698 List->_prev = iterator; 1699 iterator->_next = List; 1700 iterator->_prev = NULL; 1701 _EntryList = iterator; 1702 } 1703 } else 1704 if (Policy == 1) { // append to EntryList 1705 if (List == NULL) { 1706 iterator->_next = iterator->_prev = NULL; 1707 _EntryList = iterator; 1708 } else { 1709 // CONSIDER: finding the tail currently requires a linear-time walk of 1710 // the EntryList. We can make tail access constant-time by converting to 1711 // a CDLL instead of using our current DLL. 1712 ObjectWaiter * Tail; 1713 for (Tail = List; Tail->_next != NULL; Tail = Tail->_next); 1714 assert(Tail != NULL && Tail->_next == NULL, "invariant"); 1715 Tail->_next = iterator; 1716 iterator->_prev = Tail; 1717 iterator->_next = NULL; 1718 } 1719 } else 1720 if (Policy == 2) { // prepend to cxq 1721 // prepend to cxq 1722 if (List == NULL) { 1723 iterator->_next = iterator->_prev = NULL; 1724 _EntryList = iterator; 1725 } else { 1726 iterator->TState = ObjectWaiter::TS_CXQ; 1727 for (;;) { 1728 ObjectWaiter * Front = _cxq; 1729 iterator->_next = Front; 1730 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1731 break; 1732 } 1733 } 1734 } 1735 } else 1736 if (Policy == 3) { // append to cxq 1737 iterator->TState = ObjectWaiter::TS_CXQ; 1738 for (;;) { 1739 ObjectWaiter * Tail; 1740 Tail = _cxq; 1741 if (Tail == NULL) { 1742 iterator->_next = NULL; 1743 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1744 break; 1745 } 1746 } else { 1747 while (Tail->_next != NULL) Tail = Tail->_next; 1748 Tail->_next = iterator; 1749 iterator->_prev = Tail; 1750 iterator->_next = NULL; 1751 break; 1752 } 1753 } 1754 } else { 1755 ParkEvent * ev = iterator->_event; 1756 iterator->TState = ObjectWaiter::TS_RUN; 1757 OrderAccess::fence(); 1758 ev->unpark(); 1759 } 1760 1761 if (Policy < 4) { 1762 iterator->wait_reenter_begin(this); 1763 } 1764 1765 // _WaitSetLock protects the wait queue, not the EntryList. We could 1766 // move the add-to-EntryList operation, above, outside the critical section 1767 // protected by _WaitSetLock. In practice that's not useful. With the 1768 // exception of wait() timeouts and interrupts the monitor owner 1769 // is the only thread that grabs _WaitSetLock. There's almost no contention 1770 // on _WaitSetLock so it's not profitable to reduce the length of the 1771 // critical section. 1772 } 1773 1774 Thread::SpinRelease(&_WaitSetLock); 1775 1776 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1777 ObjectMonitor::_sync_Notifications->inc(); 1778 } 1779} 1780 1781 1782void ObjectMonitor::notifyAll(TRAPS) { 1783 CHECK_OWNER(); 1784 ObjectWaiter* iterator; 1785 if (_WaitSet == NULL) { 1786 TEVENT(Empty-NotifyAll); 1787 return; 1788 } 1789 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1790 1791 int Policy = Knob_MoveNotifyee; 1792 int Tally = 0; 1793 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notifyall"); 1794 1795 for (;;) { 1796 iterator = DequeueWaiter(); 1797 if (iterator == NULL) break; 1798 TEVENT(NotifyAll - Transfer1); 1799 ++Tally; 1800 1801 // Disposition - what might we do with iterator ? 1802 // a. add it directly to the EntryList - either tail or head. 1803 // b. push it onto the front of the _cxq. 1804 // For now we use (a). 1805 1806 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1807 guarantee(iterator->_notified == 0, "invariant"); 1808 iterator->_notified = 1; 1809 Thread * Self = THREAD; 1810 iterator->_notifier_tid = Self->osthread()->thread_id(); 1811 if (Policy != 4) { 1812 iterator->TState = ObjectWaiter::TS_ENTER; 1813 } 1814 1815 ObjectWaiter * List = _EntryList; 1816 if (List != NULL) { 1817 assert(List->_prev == NULL, "invariant"); 1818 assert(List->TState == ObjectWaiter::TS_ENTER, "invariant"); 1819 assert(List != iterator, "invariant"); 1820 } 1821 1822 if (Policy == 0) { // prepend to EntryList 1823 if (List == NULL) { 1824 iterator->_next = iterator->_prev = NULL; 1825 _EntryList = iterator; 1826 } else { 1827 List->_prev = iterator; 1828 iterator->_next = List; 1829 iterator->_prev = NULL; 1830 _EntryList = iterator; 1831 } 1832 } else 1833 if (Policy == 1) { // append to EntryList 1834 if (List == NULL) { 1835 iterator->_next = iterator->_prev = NULL; 1836 _EntryList = iterator; 1837 } else { 1838 // CONSIDER: finding the tail currently requires a linear-time walk of 1839 // the EntryList. We can make tail access constant-time by converting to 1840 // a CDLL instead of using our current DLL. 1841 ObjectWaiter * Tail; 1842 for (Tail = List; Tail->_next != NULL; Tail = Tail->_next); 1843 assert(Tail != NULL && Tail->_next == NULL, "invariant"); 1844 Tail->_next = iterator; 1845 iterator->_prev = Tail; 1846 iterator->_next = NULL; 1847 } 1848 } else 1849 if (Policy == 2) { // prepend to cxq 1850 // prepend to cxq 1851 iterator->TState = ObjectWaiter::TS_CXQ; 1852 for (;;) { 1853 ObjectWaiter * Front = _cxq; 1854 iterator->_next = Front; 1855 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1856 break; 1857 } 1858 } 1859 } else 1860 if (Policy == 3) { // append to cxq 1861 iterator->TState = ObjectWaiter::TS_CXQ; 1862 for (;;) { 1863 ObjectWaiter * Tail; 1864 Tail = _cxq; 1865 if (Tail == NULL) { 1866 iterator->_next = NULL; 1867 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1868 break; 1869 } 1870 } else { 1871 while (Tail->_next != NULL) Tail = Tail->_next; 1872 Tail->_next = iterator; 1873 iterator->_prev = Tail; 1874 iterator->_next = NULL; 1875 break; 1876 } 1877 } 1878 } else { 1879 ParkEvent * ev = iterator->_event; 1880 iterator->TState = ObjectWaiter::TS_RUN; 1881 OrderAccess::fence(); 1882 ev->unpark(); 1883 } 1884 1885 if (Policy < 4) { 1886 iterator->wait_reenter_begin(this); 1887 } 1888 1889 // _WaitSetLock protects the wait queue, not the EntryList. We could 1890 // move the add-to-EntryList operation, above, outside the critical section 1891 // protected by _WaitSetLock. In practice that's not useful. With the 1892 // exception of wait() timeouts and interrupts the monitor owner 1893 // is the only thread that grabs _WaitSetLock. There's almost no contention 1894 // on _WaitSetLock so it's not profitable to reduce the length of the 1895 // critical section. 1896 } 1897 1898 Thread::SpinRelease(&_WaitSetLock); 1899 1900 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1901 ObjectMonitor::_sync_Notifications->inc(Tally); 1902 } 1903} 1904 1905// ----------------------------------------------------------------------------- 1906// Adaptive Spinning Support 1907// 1908// Adaptive spin-then-block - rational spinning 1909// 1910// Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1911// algorithm. On high order SMP systems it would be better to start with 1912// a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1913// a contending thread could enqueue itself on the cxq and then spin locally 1914// on a thread-specific variable such as its ParkEvent._Event flag. 1915// That's left as an exercise for the reader. Note that global spinning is 1916// not problematic on Niagara, as the L2$ serves the interconnect and has both 1917// low latency and massive bandwidth. 1918// 1919// Broadly, we can fix the spin frequency -- that is, the % of contended lock 1920// acquisition attempts where we opt to spin -- at 100% and vary the spin count 1921// (duration) or we can fix the count at approximately the duration of 1922// a context switch and vary the frequency. Of course we could also 1923// vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1924// See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1925// 1926// This implementation varies the duration "D", where D varies with 1927// the success rate of recent spin attempts. (D is capped at approximately 1928// length of a round-trip context switch). The success rate for recent 1929// spin attempts is a good predictor of the success rate of future spin 1930// attempts. The mechanism adapts automatically to varying critical 1931// section length (lock modality), system load and degree of parallelism. 1932// D is maintained per-monitor in _SpinDuration and is initialized 1933// optimistically. Spin frequency is fixed at 100%. 1934// 1935// Note that _SpinDuration is volatile, but we update it without locks 1936// or atomics. The code is designed so that _SpinDuration stays within 1937// a reasonable range even in the presence of races. The arithmetic 1938// operations on _SpinDuration are closed over the domain of legal values, 1939// so at worst a race will install and older but still legal value. 1940// At the very worst this introduces some apparent non-determinism. 1941// We might spin when we shouldn't or vice-versa, but since the spin 1942// count are relatively short, even in the worst case, the effect is harmless. 1943// 1944// Care must be taken that a low "D" value does not become an 1945// an absorbing state. Transient spinning failures -- when spinning 1946// is overall profitable -- should not cause the system to converge 1947// on low "D" values. We want spinning to be stable and predictable 1948// and fairly responsive to change and at the same time we don't want 1949// it to oscillate, become metastable, be "too" non-deterministic, 1950// or converge on or enter undesirable stable absorbing states. 1951// 1952// We implement a feedback-based control system -- using past behavior 1953// to predict future behavior. We face two issues: (a) if the 1954// input signal is random then the spin predictor won't provide optimal 1955// results, and (b) if the signal frequency is too high then the control 1956// system, which has some natural response lag, will "chase" the signal. 1957// (b) can arise from multimodal lock hold times. Transient preemption 1958// can also result in apparent bimodal lock hold times. 1959// Although sub-optimal, neither condition is particularly harmful, as 1960// in the worst-case we'll spin when we shouldn't or vice-versa. 1961// The maximum spin duration is rather short so the failure modes aren't bad. 1962// To be conservative, I've tuned the gain in system to bias toward 1963// _not spinning. Relatedly, the system can sometimes enter a mode where it 1964// "rings" or oscillates between spinning and not spinning. This happens 1965// when spinning is just on the cusp of profitability, however, so the 1966// situation is not dire. The state is benign -- there's no need to add 1967// hysteresis control to damp the transition rate between spinning and 1968// not spinning. 1969// 1970 1971intptr_t ObjectMonitor::SpinCallbackArgument = 0; 1972int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL; 1973 1974// Spinning: Fixed frequency (100%), vary duration 1975 1976 1977int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1978 1979 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1980 int ctr = Knob_FixedSpin; 1981 if (ctr != 0) { 1982 while (--ctr >= 0) { 1983 if (TryLock(Self) > 0) return 1; 1984 SpinPause(); 1985 } 1986 return 0; 1987 } 1988 1989 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1990 if (TryLock(Self) > 0) { 1991 // Increase _SpinDuration ... 1992 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1993 // Raising _SpurDuration to the poverty line is key. 1994 int x = _SpinDuration; 1995 if (x < Knob_SpinLimit) { 1996 if (x < Knob_Poverty) x = Knob_Poverty; 1997 _SpinDuration = x + Knob_BonusB; 1998 } 1999 return 1; 2000 } 2001 SpinPause(); 2002 } 2003 2004 // Admission control - verify preconditions for spinning 2005 // 2006 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2007 // becoming an absorbing state. Put another way, we spin briefly to 2008 // sample, just in case the system load, parallelism, contention, or lock 2009 // modality changed. 2010 // 2011 // Consider the following alternative: 2012 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2013 // spin attempt. "Periodically" might mean after a tally of 2014 // the # of failed spin attempts (or iterations) reaches some threshold. 2015 // This takes us into the realm of 1-out-of-N spinning, where we 2016 // hold the duration constant but vary the frequency. 2017 2018 ctr = _SpinDuration; 2019 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 2020 if (ctr <= 0) return 0; 2021 2022 if (Knob_SuccRestrict && _succ != NULL) return 0; 2023 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2024 TEVENT(Spin abort - notrunnable [TOP]); 2025 return 0; 2026 } 2027 2028 int MaxSpin = Knob_MaxSpinners; 2029 if (MaxSpin >= 0) { 2030 if (_Spinner > MaxSpin) { 2031 TEVENT(Spin abort -- too many spinners); 2032 return 0; 2033 } 2034 // Slightly racy, but benign ... 2035 Adjust(&_Spinner, 1); 2036 } 2037 2038 // We're good to spin ... spin ingress. 2039 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2040 // when preparing to LD...CAS _owner, etc and the CAS is likely 2041 // to succeed. 2042 int hits = 0; 2043 int msk = 0; 2044 int caspty = Knob_CASPenalty; 2045 int oxpty = Knob_OXPenalty; 2046 int sss = Knob_SpinSetSucc; 2047 if (sss && _succ == NULL) _succ = Self; 2048 Thread * prv = NULL; 2049 2050 // There are three ways to exit the following loop: 2051 // 1. A successful spin where this thread has acquired the lock. 2052 // 2. Spin failure with prejudice 2053 // 3. Spin failure without prejudice 2054 2055 while (--ctr >= 0) { 2056 2057 // Periodic polling -- Check for pending GC 2058 // Threads may spin while they're unsafe. 2059 // We don't want spinning threads to delay the JVM from reaching 2060 // a stop-the-world safepoint or to steal cycles from GC. 2061 // If we detect a pending safepoint we abort in order that 2062 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2063 // this thread, if safe, doesn't steal cycles from GC. 2064 // This is in keeping with the "no loitering in runtime" rule. 2065 // We periodically check to see if there's a safepoint pending. 2066 if ((ctr & 0xFF) == 0) { 2067 if (SafepointSynchronize::do_call_back()) { 2068 TEVENT(Spin: safepoint); 2069 goto Abort; // abrupt spin egress 2070 } 2071 if (Knob_UsePause & 1) SpinPause(); 2072 2073 int (*scb)(intptr_t,int) = SpinCallbackFunction; 2074 if (hits > 50 && scb != NULL) { 2075 int abend = (*scb)(SpinCallbackArgument, 0); 2076 } 2077 } 2078 2079 if (Knob_UsePause & 2) SpinPause(); 2080 2081 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2082 // This is useful on classic SMP systems, but is of less utility on 2083 // N1-style CMT platforms. 2084 // 2085 // Trade-off: lock acquisition latency vs coherency bandwidth. 2086 // Lock hold times are typically short. A histogram 2087 // of successful spin attempts shows that we usually acquire 2088 // the lock early in the spin. That suggests we want to 2089 // sample _owner frequently in the early phase of the spin, 2090 // but then back-off and sample less frequently as the spin 2091 // progresses. The back-off makes a good citizen on SMP big 2092 // SMP systems. Oversampling _owner can consume excessive 2093 // coherency bandwidth. Relatedly, if we _oversample _owner we 2094 // can inadvertently interfere with the the ST m->owner=null. 2095 // executed by the lock owner. 2096 if (ctr & msk) continue; 2097 ++hits; 2098 if ((hits & 0xF) == 0) { 2099 // The 0xF, above, corresponds to the exponent. 2100 // Consider: (msk+1)|msk 2101 msk = ((msk << 2)|3) & BackOffMask; 2102 } 2103 2104 // Probe _owner with TATAS 2105 // If this thread observes the monitor transition or flicker 2106 // from locked to unlocked to locked, then the odds that this 2107 // thread will acquire the lock in this spin attempt go down 2108 // considerably. The same argument applies if the CAS fails 2109 // or if we observe _owner change from one non-null value to 2110 // another non-null value. In such cases we might abort 2111 // the spin without prejudice or apply a "penalty" to the 2112 // spin count-down variable "ctr", reducing it by 100, say. 2113 2114 Thread * ox = (Thread *) _owner; 2115 if (ox == NULL) { 2116 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 2117 if (ox == NULL) { 2118 // The CAS succeeded -- this thread acquired ownership 2119 // Take care of some bookkeeping to exit spin state. 2120 if (sss && _succ == Self) { 2121 _succ = NULL; 2122 } 2123 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2124 2125 // Increase _SpinDuration : 2126 // The spin was successful (profitable) so we tend toward 2127 // longer spin attempts in the future. 2128 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2129 // If we acquired the lock early in the spin cycle it 2130 // makes sense to increase _SpinDuration proportionally. 2131 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2132 int x = _SpinDuration; 2133 if (x < Knob_SpinLimit) { 2134 if (x < Knob_Poverty) x = Knob_Poverty; 2135 _SpinDuration = x + Knob_Bonus; 2136 } 2137 return 1; 2138 } 2139 2140 // The CAS failed ... we can take any of the following actions: 2141 // * penalize: ctr -= Knob_CASPenalty 2142 // * exit spin with prejudice -- goto Abort; 2143 // * exit spin without prejudice. 2144 // * Since CAS is high-latency, retry again immediately. 2145 prv = ox; 2146 TEVENT(Spin: cas failed); 2147 if (caspty == -2) break; 2148 if (caspty == -1) goto Abort; 2149 ctr -= caspty; 2150 continue; 2151 } 2152 2153 // Did lock ownership change hands ? 2154 if (ox != prv && prv != NULL) { 2155 TEVENT(spin: Owner changed) 2156 if (oxpty == -2) break; 2157 if (oxpty == -1) goto Abort; 2158 ctr -= oxpty; 2159 } 2160 prv = ox; 2161 2162 // Abort the spin if the owner is not executing. 2163 // The owner must be executing in order to drop the lock. 2164 // Spinning while the owner is OFFPROC is idiocy. 2165 // Consider: ctr -= RunnablePenalty ; 2166 if (Knob_OState && NotRunnable (Self, ox)) { 2167 TEVENT(Spin abort - notrunnable); 2168 goto Abort; 2169 } 2170 if (sss && _succ == NULL) _succ = Self; 2171 } 2172 2173 // Spin failed with prejudice -- reduce _SpinDuration. 2174 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2175 // AIMD is globally stable. 2176 TEVENT(Spin failure); 2177 { 2178 int x = _SpinDuration; 2179 if (x > 0) { 2180 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2181 // This is globally sample and tends to damp the response. 2182 x -= Knob_Penalty; 2183 if (x < 0) x = 0; 2184 _SpinDuration = x; 2185 } 2186 } 2187 2188 Abort: 2189 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2190 if (sss && _succ == Self) { 2191 _succ = NULL; 2192 // Invariant: after setting succ=null a contending thread 2193 // must recheck-retry _owner before parking. This usually happens 2194 // in the normal usage of TrySpin(), but it's safest 2195 // to make TrySpin() as foolproof as possible. 2196 OrderAccess::fence(); 2197 if (TryLock(Self) > 0) return 1; 2198 } 2199 return 0; 2200} 2201 2202// NotRunnable() -- informed spinning 2203// 2204// Don't bother spinning if the owner is not eligible to drop the lock. 2205// Peek at the owner's schedctl.sc_state and Thread._thread_values and 2206// spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2207// The thread must be runnable in order to drop the lock in timely fashion. 2208// If the _owner is not runnable then spinning will not likely be 2209// successful (profitable). 2210// 2211// Beware -- the thread referenced by _owner could have died 2212// so a simply fetch from _owner->_thread_state might trap. 2213// Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2214// Because of the lifecycle issues the schedctl and _thread_state values 2215// observed by NotRunnable() might be garbage. NotRunnable must 2216// tolerate this and consider the observed _thread_state value 2217// as advisory. 2218// 2219// Beware too, that _owner is sometimes a BasicLock address and sometimes 2220// a thread pointer. We differentiate the two cases with OwnerIsThread. 2221// Alternately, we might tag the type (thread pointer vs basiclock pointer) 2222// with the LSB of _owner. Another option would be to probablistically probe 2223// the putative _owner->TypeTag value. 2224// 2225// Checking _thread_state isn't perfect. Even if the thread is 2226// in_java it might be blocked on a page-fault or have been preempted 2227// and sitting on a ready/dispatch queue. _thread state in conjunction 2228// with schedctl.sc_state gives us a good picture of what the 2229// thread is doing, however. 2230// 2231// TODO: check schedctl.sc_state. 2232// We'll need to use SafeFetch32() to read from the schedctl block. 2233// See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2234// 2235// The return value from NotRunnable() is *advisory* -- the 2236// result is based on sampling and is not necessarily coherent. 2237// The caller must tolerate false-negative and false-positive errors. 2238// Spinning, in general, is probabilistic anyway. 2239 2240 2241int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2242 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2243 if (!OwnerIsThread) return 0; 2244 2245 if (ox == NULL) return 0; 2246 2247 // Avoid transitive spinning ... 2248 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2249 // Immediately after T1 acquires L it's possible that T2, also 2250 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2251 // This occurs transiently after T1 acquired L but before 2252 // T1 managed to clear T1.Stalled. T2 does not need to abort 2253 // its spin in this circumstance. 2254 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2255 2256 if (BlockedOn == 1) return 1; 2257 if (BlockedOn != 0) { 2258 return BlockedOn != intptr_t(this) && _owner == ox; 2259 } 2260 2261 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2262 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2263 // consider also: jst != _thread_in_Java -- but that's overspecific. 2264 return jst == _thread_blocked || jst == _thread_in_native; 2265} 2266 2267 2268// ----------------------------------------------------------------------------- 2269// WaitSet management ... 2270 2271ObjectWaiter::ObjectWaiter(Thread* thread) { 2272 _next = NULL; 2273 _prev = NULL; 2274 _notified = 0; 2275 TState = TS_RUN; 2276 _thread = thread; 2277 _event = thread->_ParkEvent; 2278 _active = false; 2279 assert(_event != NULL, "invariant"); 2280} 2281 2282void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2283 JavaThread *jt = (JavaThread *)this->_thread; 2284 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2285} 2286 2287void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2288 JavaThread *jt = (JavaThread *)this->_thread; 2289 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2290} 2291 2292inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2293 assert(node != NULL, "should not dequeue NULL node"); 2294 assert(node->_prev == NULL, "node already in list"); 2295 assert(node->_next == NULL, "node already in list"); 2296 // put node at end of queue (circular doubly linked list) 2297 if (_WaitSet == NULL) { 2298 _WaitSet = node; 2299 node->_prev = node; 2300 node->_next = node; 2301 } else { 2302 ObjectWaiter* head = _WaitSet; 2303 ObjectWaiter* tail = head->_prev; 2304 assert(tail->_next == head, "invariant check"); 2305 tail->_next = node; 2306 head->_prev = node; 2307 node->_next = head; 2308 node->_prev = tail; 2309 } 2310} 2311 2312inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2313 // dequeue the very first waiter 2314 ObjectWaiter* waiter = _WaitSet; 2315 if (waiter) { 2316 DequeueSpecificWaiter(waiter); 2317 } 2318 return waiter; 2319} 2320 2321inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2322 assert(node != NULL, "should not dequeue NULL node"); 2323 assert(node->_prev != NULL, "node already removed from list"); 2324 assert(node->_next != NULL, "node already removed from list"); 2325 // when the waiter has woken up because of interrupt, 2326 // timeout or other spurious wake-up, dequeue the 2327 // waiter from waiting list 2328 ObjectWaiter* next = node->_next; 2329 if (next == node) { 2330 assert(node->_prev == node, "invariant check"); 2331 _WaitSet = NULL; 2332 } else { 2333 ObjectWaiter* prev = node->_prev; 2334 assert(prev->_next == node, "invariant check"); 2335 assert(next->_prev == node, "invariant check"); 2336 next->_prev = prev; 2337 prev->_next = next; 2338 if (_WaitSet == node) { 2339 _WaitSet = next; 2340 } 2341 } 2342 node->_next = NULL; 2343 node->_prev = NULL; 2344} 2345 2346// ----------------------------------------------------------------------------- 2347// PerfData support 2348PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2349PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2350PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2351PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2352PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2353PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2354PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2355PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2356PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2357PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2358PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2359PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2360PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2361PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2362PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2363PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2364PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2365PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2366 2367// One-shot global initialization for the sync subsystem. 2368// We could also defer initialization and initialize on-demand 2369// the first time we call inflate(). Initialization would 2370// be protected - like so many things - by the MonitorCache_lock. 2371 2372void ObjectMonitor::Initialize() { 2373 static int InitializationCompleted = 0; 2374 assert(InitializationCompleted == 0, "invariant"); 2375 InitializationCompleted = 1; 2376 if (UsePerfData) { 2377 EXCEPTION_MARK; 2378 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2379 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2380 NEWPERFCOUNTER(_sync_Inflations); 2381 NEWPERFCOUNTER(_sync_Deflations); 2382 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2383 NEWPERFCOUNTER(_sync_FutileWakeups); 2384 NEWPERFCOUNTER(_sync_Parks); 2385 NEWPERFCOUNTER(_sync_EmptyNotifications); 2386 NEWPERFCOUNTER(_sync_Notifications); 2387 NEWPERFCOUNTER(_sync_SlowEnter); 2388 NEWPERFCOUNTER(_sync_SlowExit); 2389 NEWPERFCOUNTER(_sync_SlowNotify); 2390 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2391 NEWPERFCOUNTER(_sync_FailedSpins); 2392 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2393 NEWPERFCOUNTER(_sync_PrivateA); 2394 NEWPERFCOUNTER(_sync_PrivateB); 2395 NEWPERFCOUNTER(_sync_MonInCirculation); 2396 NEWPERFCOUNTER(_sync_MonScavenged); 2397 NEWPERFVARIABLE(_sync_MonExtant); 2398 #undef NEWPERFCOUNTER 2399 } 2400} 2401 2402 2403// Compile-time asserts 2404// When possible, it's better to catch errors deterministically at 2405// compile-time than at runtime. The down-side to using compile-time 2406// asserts is that error message -- often something about negative array 2407// indices -- is opaque. 2408 2409#define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2410 2411void ObjectMonitor::ctAsserts() { 2412 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2413} 2414 2415 2416static char * kvGet (char * kvList, const char * Key) { 2417 if (kvList == NULL) return NULL; 2418 size_t n = strlen(Key); 2419 char * Search; 2420 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2421 if (strncmp (Search, Key, n) == 0) { 2422 if (Search[n] == '=') return Search + n + 1; 2423 if (Search[n] == 0) return(char *) "1"; 2424 } 2425 } 2426 return NULL; 2427} 2428 2429static int kvGetInt (char * kvList, const char * Key, int Default) { 2430 char * v = kvGet(kvList, Key); 2431 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2432 if (Knob_ReportSettings && v != NULL) { 2433 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2434 ::fflush(stdout); 2435 } 2436 return rslt; 2437} 2438 2439void ObjectMonitor::DeferredInitialize() { 2440 if (InitDone > 0) return; 2441 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2442 while (InitDone != 1); 2443 return; 2444 } 2445 2446 // One-shot global initialization ... 2447 // The initialization is idempotent, so we don't need locks. 2448 // In the future consider doing this via os::init_2(). 2449 // SyncKnobs consist of <Key>=<Value> pairs in the style 2450 // of environment variables. Start by converting ':' to NUL. 2451 2452 if (SyncKnobs == NULL) SyncKnobs = ""; 2453 2454 size_t sz = strlen(SyncKnobs); 2455 char * knobs = (char *) malloc(sz + 2); 2456 if (knobs == NULL) { 2457 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2458 guarantee(0, "invariant"); 2459 } 2460 strcpy(knobs, SyncKnobs); 2461 knobs[sz+1] = 0; 2462 for (char * p = knobs; *p; p++) { 2463 if (*p == ':') *p = 0; 2464 } 2465 2466 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2467 SETKNOB(ReportSettings); 2468 SETKNOB(Verbose); 2469 SETKNOB(FixedSpin); 2470 SETKNOB(SpinLimit); 2471 SETKNOB(SpinBase); 2472 SETKNOB(SpinBackOff); 2473 SETKNOB(CASPenalty); 2474 SETKNOB(OXPenalty); 2475 SETKNOB(LogSpins); 2476 SETKNOB(SpinSetSucc); 2477 SETKNOB(SuccEnabled); 2478 SETKNOB(SuccRestrict); 2479 SETKNOB(Penalty); 2480 SETKNOB(Bonus); 2481 SETKNOB(BonusB); 2482 SETKNOB(Poverty); 2483 SETKNOB(SpinAfterFutile); 2484 SETKNOB(UsePause); 2485 SETKNOB(SpinEarly); 2486 SETKNOB(OState); 2487 SETKNOB(MaxSpinners); 2488 SETKNOB(PreSpin); 2489 SETKNOB(ExitPolicy); 2490 SETKNOB(QMode); 2491 SETKNOB(ResetEvent); 2492 SETKNOB(MoveNotifyee); 2493 SETKNOB(FastHSSEC); 2494 #undef SETKNOB 2495 2496 if (os::is_MP()) { 2497 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2498 if (Knob_ReportSettings) ::printf("BackOffMask=%X\n", BackOffMask); 2499 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2500 } else { 2501 Knob_SpinLimit = 0; 2502 Knob_SpinBase = 0; 2503 Knob_PreSpin = 0; 2504 Knob_FixedSpin = -1; 2505 } 2506 2507 if (Knob_LogSpins == 0) { 2508 ObjectMonitor::_sync_FailedSpins = NULL; 2509 } 2510 2511 free(knobs); 2512 OrderAccess::fence(); 2513 InitDone = 1; 2514} 2515 2516#ifndef PRODUCT 2517void ObjectMonitor::verify() { 2518} 2519 2520void ObjectMonitor::print() { 2521} 2522#endif 2523