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/atomic.inline.hpp" 31 #include "runtime/handles.inline.hpp" 32 #include "runtime/interfaceSupport.hpp" 33 #include "runtime/mutexLocker.hpp" 34 #include "runtime/objectMonitor.hpp" 35 #include "runtime/objectMonitor.inline.hpp" 36 #include "runtime/orderAccess.inline.hpp" 37 #include "runtime/osThread.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #include "runtime/thread.inline.hpp" 40 #include "services/threadService.hpp" 41 #include "trace/tracing.hpp" 42 #include "trace/traceMacros.hpp" 43 #include "utilities/dtrace.hpp" 44 #include "utilities/macros.hpp" 45 #include "utilities/preserveException.hpp" 46 47 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 48 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 49 #define NOINLINE __attribute__((noinline)) 50 #else 51 #define NOINLINE 52 #endif 53 54 55 #ifdef DTRACE_ENABLED 56 57 // Only bother with this argument setup if dtrace is available 58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95 #else // ndef DTRACE_ENABLED 96 97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100 #endif // ndef DTRACE_ENABLED 101 102 // Tunables ... 103 // The knob* variables are effectively final. Once set they should 104 // never be modified hence. Consider using __read_mostly with GCC. 105 106 int ObjectMonitor::Knob_Verbose = 0; 107 int ObjectMonitor::Knob_VerifyInUse = 0; 108 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 109 static int Knob_LogSpins = 0; // enable jvmstat tally for spins 110 static int Knob_HandOff = 0; 111 static int Knob_ReportSettings = 0; 112 113 static int Knob_SpinBase = 0; // Floor AKA SpinMin 114 static int Knob_SpinBackOff = 0; // spin-loop backoff 115 static int Knob_CASPenalty = -1; // Penalty for failed CAS 116 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 117 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 118 static int Knob_SpinEarly = 1; 119 static int Knob_SuccEnabled = 1; // futile wake throttling 120 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 121 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 122 static int Knob_Bonus = 100; // spin success bonus 123 static int Knob_BonusB = 100; // spin success bonus 124 static int Knob_Penalty = 200; // spin failure penalty 125 static int Knob_Poverty = 1000; 126 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 127 static int Knob_FixedSpin = 0; 128 static int Knob_OState = 3; // Spinner checks thread state of _owner 129 static int Knob_UsePause = 1; 130 static int Knob_ExitPolicy = 0; 131 static int Knob_PreSpin = 10; // 20-100 likely better 132 static int Knob_ResetEvent = 0; 133 static int BackOffMask = 0; 134 135 static int Knob_FastHSSEC = 0; 136 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 137 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 138 static volatile int InitDone = 0; 139 140 #define TrySpin TrySpin_VaryDuration 141 142 // ----------------------------------------------------------------------------- 143 // Theory of operations -- Monitors lists, thread residency, etc: 144 // 145 // * A thread acquires ownership of a monitor by successfully 146 // CAS()ing the _owner field from null to non-null. 147 // 148 // * Invariant: A thread appears on at most one monitor list -- 149 // cxq, EntryList or WaitSet -- at any one time. 150 // 151 // * Contending threads "push" themselves onto the cxq with CAS 152 // and then spin/park. 153 // 154 // * After a contending thread eventually acquires the lock it must 155 // dequeue itself from either the EntryList or the cxq. 156 // 157 // * The exiting thread identifies and unparks an "heir presumptive" 158 // tentative successor thread on the EntryList. Critically, the 159 // exiting thread doesn't unlink the successor thread from the EntryList. 160 // After having been unparked, the wakee will recontend for ownership of 161 // the monitor. The successor (wakee) will either acquire the lock or 162 // re-park itself. 163 // 164 // Succession is provided for by a policy of competitive handoff. 165 // The exiting thread does _not_ grant or pass ownership to the 166 // successor thread. (This is also referred to as "handoff" succession"). 167 // Instead the exiting thread releases ownership and possibly wakes 168 // a successor, so the successor can (re)compete for ownership of the lock. 169 // If the EntryList is empty but the cxq is populated the exiting 170 // thread will drain the cxq into the EntryList. It does so by 171 // by detaching the cxq (installing null with CAS) and folding 172 // the threads from the cxq into the EntryList. The EntryList is 173 // doubly linked, while the cxq is singly linked because of the 174 // CAS-based "push" used to enqueue recently arrived threads (RATs). 175 // 176 // * Concurrency invariants: 177 // 178 // -- only the monitor owner may access or mutate the EntryList. 179 // The mutex property of the monitor itself protects the EntryList 180 // from concurrent interference. 181 // -- Only the monitor owner may detach the cxq. 182 // 183 // * The monitor entry list operations avoid locks, but strictly speaking 184 // they're not lock-free. Enter is lock-free, exit is not. 185 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 186 // 187 // * The cxq can have multiple concurrent "pushers" but only one concurrent 188 // detaching thread. This mechanism is immune from the ABA corruption. 189 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 190 // 191 // * Taken together, the cxq and the EntryList constitute or form a 192 // single logical queue of threads stalled trying to acquire the lock. 193 // We use two distinct lists to improve the odds of a constant-time 194 // dequeue operation after acquisition (in the ::enter() epilogue) and 195 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 196 // A key desideratum is to minimize queue & monitor metadata manipulation 197 // that occurs while holding the monitor lock -- that is, we want to 198 // minimize monitor lock holds times. Note that even a small amount of 199 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 200 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 201 // locks and monitor metadata. 202 // 203 // Cxq points to the set of Recently Arrived Threads attempting entry. 204 // Because we push threads onto _cxq with CAS, the RATs must take the form of 205 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 206 // the unlocking thread notices that EntryList is null but _cxq is != null. 207 // 208 // The EntryList is ordered by the prevailing queue discipline and 209 // can be organized in any convenient fashion, such as a doubly-linked list or 210 // a circular doubly-linked list. Critically, we want insert and delete operations 211 // to operate in constant-time. If we need a priority queue then something akin 212 // to Solaris' sleepq would work nicely. Viz., 213 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 214 // Queue discipline is enforced at ::exit() time, when the unlocking thread 215 // drains the cxq into the EntryList, and orders or reorders the threads on the 216 // EntryList accordingly. 217 // 218 // Barring "lock barging", this mechanism provides fair cyclic ordering, 219 // somewhat similar to an elevator-scan. 220 // 221 // * The monitor synchronization subsystem avoids the use of native 222 // synchronization primitives except for the narrow platform-specific 223 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 224 // the semantics of park-unpark. Put another way, this monitor implementation 225 // depends only on atomic operations and park-unpark. The monitor subsystem 226 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 227 // underlying OS manages the READY<->RUN transitions. 228 // 229 // * Waiting threads reside on the WaitSet list -- wait() puts 230 // the caller onto the WaitSet. 231 // 232 // * notify() or notifyAll() simply transfers threads from the WaitSet to 233 // either the EntryList or cxq. Subsequent exit() operations will 234 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 235 // it's likely the notifyee would simply impale itself on the lock held 236 // by the notifier. 237 // 238 // * An interesting alternative is to encode cxq as (List,LockByte) where 239 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 240 // variable, like _recursions, in the scheme. The threads or Events that form 241 // the list would have to be aligned in 256-byte addresses. A thread would 242 // try to acquire the lock or enqueue itself with CAS, but exiting threads 243 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 244 // Note that is is *not* word-tearing, but it does presume that full-word 245 // CAS operations are coherent with intermix with STB operations. That's true 246 // on most common processors. 247 // 248 // * See also http://blogs.sun.com/dave 249 250 251 // ----------------------------------------------------------------------------- 252 // Enter support 253 254 bool ObjectMonitor::try_enter(Thread* THREAD) { 255 if (THREAD != _owner) { 256 if (THREAD->is_lock_owned ((address)_owner)) { 257 assert(_recursions == 0, "internal state error"); 258 _owner = THREAD; 259 _recursions = 1; 260 return true; 261 } 262 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 263 return false; 264 } 265 return true; 266 } else { 267 _recursions++; 268 return true; 269 } 270 } 271 272 void NOINLINE ObjectMonitor::enter(TRAPS) { 273 // The following code is ordered to check the most common cases first 274 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 275 Thread * const Self = THREAD; 276 277 void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL); 278 if (cur == NULL) { 279 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 280 assert(_recursions == 0, "invariant"); 281 assert(_owner == Self, "invariant"); 282 return; 283 } 284 285 if (cur == Self) { 286 // TODO-FIXME: check for integer overflow! BUGID 6557169. 287 _recursions++; 288 return; 289 } 290 291 if (Self->is_lock_owned ((address)cur)) { 292 assert(_recursions == 0, "internal state error"); 293 _recursions = 1; 294 // Commute owner from a thread-specific on-stack BasicLockObject address to 295 // a full-fledged "Thread *". 296 _owner = Self; 297 return; 298 } 299 300 // We've encountered genuine contention. 301 assert(Self->_Stalled == 0, "invariant"); 302 Self->_Stalled = intptr_t(this); 303 304 // Try one round of spinning *before* enqueueing Self 305 // and before going through the awkward and expensive state 306 // transitions. The following spin is strictly optional ... 307 // Note that if we acquire the monitor from an initial spin 308 // we forgo posting JVMTI events and firing DTRACE probes. 309 if (Knob_SpinEarly && TrySpin (Self) > 0) { 310 assert(_owner == Self, "invariant"); 311 assert(_recursions == 0, "invariant"); 312 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 313 Self->_Stalled = 0; 314 return; 315 } 316 317 assert(_owner != Self, "invariant"); 318 assert(_succ != Self, "invariant"); 319 assert(Self->is_Java_thread(), "invariant"); 320 JavaThread * jt = (JavaThread *) Self; 321 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 322 assert(jt->thread_state() != _thread_blocked, "invariant"); 323 assert(this->object() != NULL, "invariant"); 324 assert(_count >= 0, "invariant"); 325 326 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 327 // Ensure the object-monitor relationship remains stable while there's contention. 328 Atomic::inc(&_count); 329 330 EventJavaMonitorEnter event; 331 332 { // Change java thread status to indicate blocked on monitor enter. 333 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 334 335 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 336 if (JvmtiExport::should_post_monitor_contended_enter()) { 337 JvmtiExport::post_monitor_contended_enter(jt, this); 338 339 // The current thread does not yet own the monitor and does not 340 // yet appear on any queues that would get it made the successor. 341 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 342 // handler cannot accidentally consume an unpark() meant for the 343 // ParkEvent associated with this ObjectMonitor. 344 } 345 346 OSThreadContendState osts(Self->osthread()); 347 ThreadBlockInVM tbivm(jt); 348 349 Self->set_current_pending_monitor(this); 350 351 // TODO-FIXME: change the following for(;;) loop to straight-line code. 352 for (;;) { 353 jt->set_suspend_equivalent(); 354 // cleared by handle_special_suspend_equivalent_condition() 355 // or java_suspend_self() 356 357 EnterI(THREAD); 358 359 if (!ExitSuspendEquivalent(jt)) break; 360 361 // We have acquired the contended monitor, but while we were 362 // waiting another thread suspended us. We don't want to enter 363 // the monitor while suspended because that would surprise the 364 // thread that suspended us. 365 // 366 _recursions = 0; 367 _succ = NULL; 368 exit(false, Self); 369 370 jt->java_suspend_self(); 371 } 372 Self->set_current_pending_monitor(NULL); 373 374 // We cleared the pending monitor info since we've just gotten past 375 // the enter-check-for-suspend dance and we now own the monitor free 376 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 377 // destructor can go to a safepoint at the end of this block. If we 378 // do a thread dump during that safepoint, then this thread will show 379 // as having "-locked" the monitor, but the OS and java.lang.Thread 380 // states will still report that the thread is blocked trying to 381 // acquire it. 382 } 383 384 Atomic::dec(&_count); 385 assert(_count >= 0, "invariant"); 386 Self->_Stalled = 0; 387 388 // Must either set _recursions = 0 or ASSERT _recursions == 0. 389 assert(_recursions == 0, "invariant"); 390 assert(_owner == Self, "invariant"); 391 assert(_succ != Self, "invariant"); 392 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 393 394 // The thread -- now the owner -- is back in vm mode. 395 // Report the glorious news via TI,DTrace and jvmstat. 396 // The probe effect is non-trivial. All the reportage occurs 397 // while we hold the monitor, increasing the length of the critical 398 // section. Amdahl's parallel speedup law comes vividly into play. 399 // 400 // Another option might be to aggregate the events (thread local or 401 // per-monitor aggregation) and defer reporting until a more opportune 402 // time -- such as next time some thread encounters contention but has 403 // yet to acquire the lock. While spinning that thread could 404 // spinning we could increment JVMStat counters, etc. 405 406 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 407 if (JvmtiExport::should_post_monitor_contended_entered()) { 408 JvmtiExport::post_monitor_contended_entered(jt, this); 409 410 // The current thread already owns the monitor and is not going to 411 // call park() for the remainder of the monitor enter protocol. So 412 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 413 // event handler consumed an unpark() issued by the thread that 414 // just exited the monitor. 415 } 416 417 if (event.should_commit()) { 418 event.set_klass(((oop)this->object())->klass()); 419 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 420 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 421 event.commit(); 422 } 423 424 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 425 ObjectMonitor::_sync_ContendedLockAttempts->inc(); 426 } 427 } 428 429 430 // Caveat: TryLock() is not necessarily serializing if it returns failure. 431 // Callers must compensate as needed. 432 433 int ObjectMonitor::TryLock(Thread * Self) { 434 void * own = _owner; 435 if (own != NULL) return 0; 436 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 437 // Either guarantee _recursions == 0 or set _recursions = 0. 438 assert(_recursions == 0, "invariant"); 439 assert(_owner == Self, "invariant"); 440 return 1; 441 } 442 // The lock had been free momentarily, but we lost the race to the lock. 443 // Interference -- the CAS failed. 444 // We can either return -1 or retry. 445 // Retry doesn't make as much sense because the lock was just acquired. 446 return -1; 447 } 448 449 void NOINLINE ObjectMonitor::EnterI(TRAPS) { 450 Thread * const Self = THREAD; 451 assert(Self->is_Java_thread(), "invariant"); 452 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 453 454 // Try the lock - TATAS 455 if (TryLock (Self) > 0) { 456 assert(_succ != Self, "invariant"); 457 assert(_owner == Self, "invariant"); 458 assert(_Responsible != Self, "invariant"); 459 return; 460 } 461 462 DeferredInitialize(); 463 464 // We try one round of spinning *before* enqueueing Self. 465 // 466 // If the _owner is ready but OFFPROC we could use a YieldTo() 467 // operation to donate the remainder of this thread's quantum 468 // to the owner. This has subtle but beneficial affinity 469 // effects. 470 471 if (TrySpin (Self) > 0) { 472 assert(_owner == Self, "invariant"); 473 assert(_succ != Self, "invariant"); 474 assert(_Responsible != Self, "invariant"); 475 return; 476 } 477 478 // The Spin failed -- Enqueue and park the thread ... 479 assert(_succ != Self, "invariant"); 480 assert(_owner != Self, "invariant"); 481 assert(_Responsible != Self, "invariant"); 482 483 // Enqueue "Self" on ObjectMonitor's _cxq. 484 // 485 // Node acts as a proxy for Self. 486 // As an aside, if were to ever rewrite the synchronization code mostly 487 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 488 // Java objects. This would avoid awkward lifecycle and liveness issues, 489 // as well as eliminate a subset of ABA issues. 490 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 491 492 ObjectWaiter node(Self); 493 Self->_ParkEvent->reset(); 494 node._prev = (ObjectWaiter *) 0xBAD; 495 node.TState = ObjectWaiter::TS_CXQ; 496 497 // Push "Self" onto the front of the _cxq. 498 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 499 // Note that spinning tends to reduce the rate at which threads 500 // enqueue and dequeue on EntryList|cxq. 501 ObjectWaiter * nxt; 502 for (;;) { 503 node._next = nxt = _cxq; 504 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break; 505 506 // Interference - the CAS failed because _cxq changed. Just retry. 507 // As an optional optimization we retry the lock. 508 if (TryLock (Self) > 0) { 509 assert(_succ != Self, "invariant"); 510 assert(_owner == Self, "invariant"); 511 assert(_Responsible != Self, "invariant"); 512 return; 513 } 514 } 515 516 // Check for cxq|EntryList edge transition to non-null. This indicates 517 // the onset of contention. While contention persists exiting threads 518 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 519 // operations revert to the faster 1-0 mode. This enter operation may interleave 520 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 521 // arrange for one of the contending thread to use a timed park() operations 522 // to detect and recover from the race. (Stranding is form of progress failure 523 // where the monitor is unlocked but all the contending threads remain parked). 524 // That is, at least one of the contended threads will periodically poll _owner. 525 // One of the contending threads will become the designated "Responsible" thread. 526 // The Responsible thread uses a timed park instead of a normal indefinite park 527 // operation -- it periodically wakes and checks for and recovers from potential 528 // strandings admitted by 1-0 exit operations. We need at most one Responsible 529 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 530 // be responsible for a monitor. 531 // 532 // Currently, one of the contended threads takes on the added role of "Responsible". 533 // A viable alternative would be to use a dedicated "stranding checker" thread 534 // that periodically iterated over all the threads (or active monitors) and unparked 535 // successors where there was risk of stranding. This would help eliminate the 536 // timer scalability issues we see on some platforms as we'd only have one thread 537 // -- the checker -- parked on a timer. 538 539 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 540 // Try to assume the role of responsible thread for the monitor. 541 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 542 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 543 } 544 545 // The lock might have been released while this thread was occupied queueing 546 // itself onto _cxq. To close the race and avoid "stranding" and 547 // progress-liveness failure we must resample-retry _owner before parking. 548 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 549 // In this case the ST-MEMBAR is accomplished with CAS(). 550 // 551 // TODO: Defer all thread state transitions until park-time. 552 // Since state transitions are heavy and inefficient we'd like 553 // to defer the state transitions until absolutely necessary, 554 // and in doing so avoid some transitions ... 555 556 TEVENT(Inflated enter - Contention); 557 int nWakeups = 0; 558 int RecheckInterval = 1; 559 560 for (;;) { 561 562 if (TryLock(Self) > 0) break; 563 assert(_owner != Self, "invariant"); 564 565 if ((SyncFlags & 2) && _Responsible == NULL) { 566 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 567 } 568 569 // park self 570 if (_Responsible == Self || (SyncFlags & 1)) { 571 TEVENT(Inflated enter - park TIMED); 572 Self->_ParkEvent->park((jlong) RecheckInterval); 573 // Increase the RecheckInterval, but clamp the value. 574 RecheckInterval *= 8; 575 if (RecheckInterval > 1000) RecheckInterval = 1000; 576 } else { 577 TEVENT(Inflated enter - park UNTIMED); 578 Self->_ParkEvent->park(); 579 } 580 581 if (TryLock(Self) > 0) break; 582 583 // The lock is still contested. 584 // Keep a tally of the # of futile wakeups. 585 // Note that the counter is not protected by a lock or updated by atomics. 586 // That is by design - we trade "lossy" counters which are exposed to 587 // races during updates for a lower probe effect. 588 TEVENT(Inflated enter - Futile wakeup); 589 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 590 ObjectMonitor::_sync_FutileWakeups->inc(); 591 } 592 ++nWakeups; 593 594 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 595 // We can defer clearing _succ until after the spin completes 596 // TrySpin() must tolerate being called with _succ == Self. 597 // Try yet another round of adaptive spinning. 598 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 599 600 // We can find that we were unpark()ed and redesignated _succ while 601 // we were spinning. That's harmless. If we iterate and call park(), 602 // park() will consume the event and return immediately and we'll 603 // just spin again. This pattern can repeat, leaving _succ to simply 604 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 605 // Alternately, we can sample fired() here, and if set, forgo spinning 606 // in the next iteration. 607 608 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 609 Self->_ParkEvent->reset(); 610 OrderAccess::fence(); 611 } 612 if (_succ == Self) _succ = NULL; 613 614 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 615 OrderAccess::fence(); 616 } 617 618 // Egress : 619 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 620 // Normally we'll find Self on the EntryList . 621 // From the perspective of the lock owner (this thread), the 622 // EntryList is stable and cxq is prepend-only. 623 // The head of cxq is volatile but the interior is stable. 624 // In addition, Self.TState is stable. 625 626 assert(_owner == Self, "invariant"); 627 assert(object() != NULL, "invariant"); 628 // I'd like to write: 629 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 630 // but as we're at a safepoint that's not safe. 631 632 UnlinkAfterAcquire(Self, &node); 633 if (_succ == Self) _succ = NULL; 634 635 assert(_succ != Self, "invariant"); 636 if (_Responsible == Self) { 637 _Responsible = NULL; 638 OrderAccess::fence(); // Dekker pivot-point 639 640 // We may leave threads on cxq|EntryList without a designated 641 // "Responsible" thread. This is benign. When this thread subsequently 642 // exits the monitor it can "see" such preexisting "old" threads -- 643 // threads that arrived on the cxq|EntryList before the fence, above -- 644 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 645 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 646 // non-null and elect a new "Responsible" timer thread. 647 // 648 // This thread executes: 649 // ST Responsible=null; MEMBAR (in enter epilogue - here) 650 // LD cxq|EntryList (in subsequent exit) 651 // 652 // Entering threads in the slow/contended path execute: 653 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 654 // The (ST cxq; MEMBAR) is accomplished with CAS(). 655 // 656 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 657 // exit operation from floating above the ST Responsible=null. 658 } 659 660 // We've acquired ownership with CAS(). 661 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 662 // But since the CAS() this thread may have also stored into _succ, 663 // EntryList, cxq or Responsible. These meta-data updates must be 664 // visible __before this thread subsequently drops the lock. 665 // Consider what could occur if we didn't enforce this constraint -- 666 // STs to monitor meta-data and user-data could reorder with (become 667 // visible after) the ST in exit that drops ownership of the lock. 668 // Some other thread could then acquire the lock, but observe inconsistent 669 // or old monitor meta-data and heap data. That violates the JMM. 670 // To that end, the 1-0 exit() operation must have at least STST|LDST 671 // "release" barrier semantics. Specifically, there must be at least a 672 // STST|LDST barrier in exit() before the ST of null into _owner that drops 673 // the lock. The barrier ensures that changes to monitor meta-data and data 674 // protected by the lock will be visible before we release the lock, and 675 // therefore before some other thread (CPU) has a chance to acquire the lock. 676 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 677 // 678 // Critically, any prior STs to _succ or EntryList must be visible before 679 // the ST of null into _owner in the *subsequent* (following) corresponding 680 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 681 // execute a serializing instruction. 682 683 if (SyncFlags & 8) { 684 OrderAccess::fence(); 685 } 686 return; 687 } 688 689 // ReenterI() is a specialized inline form of the latter half of the 690 // contended slow-path from EnterI(). We use ReenterI() only for 691 // monitor reentry in wait(). 692 // 693 // In the future we should reconcile EnterI() and ReenterI(), adding 694 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 695 // loop accordingly. 696 697 void NOINLINE ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 698 assert(Self != NULL, "invariant"); 699 assert(SelfNode != NULL, "invariant"); 700 assert(SelfNode->_thread == Self, "invariant"); 701 assert(_waiters > 0, "invariant"); 702 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 703 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 704 JavaThread * jt = (JavaThread *) Self; 705 706 int nWakeups = 0; 707 for (;;) { 708 ObjectWaiter::TStates v = SelfNode->TState; 709 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 710 assert(_owner != Self, "invariant"); 711 712 if (TryLock(Self) > 0) break; 713 if (TrySpin(Self) > 0) break; 714 715 TEVENT(Wait Reentry - parking); 716 717 // State transition wrappers around park() ... 718 // ReenterI() wisely defers state transitions until 719 // it's clear we must park the thread. 720 { 721 OSThreadContendState osts(Self->osthread()); 722 ThreadBlockInVM tbivm(jt); 723 724 // cleared by handle_special_suspend_equivalent_condition() 725 // or java_suspend_self() 726 jt->set_suspend_equivalent(); 727 if (SyncFlags & 1) { 728 Self->_ParkEvent->park((jlong)1000); 729 } else { 730 Self->_ParkEvent->park(); 731 } 732 733 // were we externally suspended while we were waiting? 734 for (;;) { 735 if (!ExitSuspendEquivalent(jt)) break; 736 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 737 jt->java_suspend_self(); 738 jt->set_suspend_equivalent(); 739 } 740 } 741 742 // Try again, but just so we distinguish between futile wakeups and 743 // successful wakeups. The following test isn't algorithmically 744 // necessary, but it helps us maintain sensible statistics. 745 if (TryLock(Self) > 0) break; 746 747 // The lock is still contested. 748 // Keep a tally of the # of futile wakeups. 749 // Note that the counter is not protected by a lock or updated by atomics. 750 // That is by design - we trade "lossy" counters which are exposed to 751 // races during updates for a lower probe effect. 752 TEVENT(Wait Reentry - futile wakeup); 753 ++nWakeups; 754 755 // Assuming this is not a spurious wakeup we'll normally 756 // find that _succ == Self. 757 if (_succ == Self) _succ = NULL; 758 759 // Invariant: after clearing _succ a contending thread 760 // *must* retry _owner before parking. 761 OrderAccess::fence(); 762 763 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 764 ObjectMonitor::_sync_FutileWakeups->inc(); 765 } 766 } 767 768 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 769 // Normally we'll find Self on the EntryList. 770 // Unlinking from the EntryList is constant-time and atomic-free. 771 // From the perspective of the lock owner (this thread), the 772 // EntryList is stable and cxq is prepend-only. 773 // The head of cxq is volatile but the interior is stable. 774 // In addition, Self.TState is stable. 775 776 assert(_owner == Self, "invariant"); 777 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 778 UnlinkAfterAcquire(Self, SelfNode); 779 if (_succ == Self) _succ = NULL; 780 assert(_succ != Self, "invariant"); 781 SelfNode->TState = ObjectWaiter::TS_RUN; 782 OrderAccess::fence(); // see comments at the end of EnterI() 783 } 784 785 // By convention we unlink a contending thread from EntryList|cxq immediately 786 // after the thread acquires the lock in ::enter(). Equally, we could defer 787 // unlinking the thread until ::exit()-time. 788 789 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 790 assert(_owner == Self, "invariant"); 791 assert(SelfNode->_thread == Self, "invariant"); 792 793 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 794 // Normal case: remove Self from the DLL EntryList . 795 // This is a constant-time operation. 796 ObjectWaiter * nxt = SelfNode->_next; 797 ObjectWaiter * prv = SelfNode->_prev; 798 if (nxt != NULL) nxt->_prev = prv; 799 if (prv != NULL) prv->_next = nxt; 800 if (SelfNode == _EntryList) _EntryList = nxt; 801 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 802 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 803 TEVENT(Unlink from EntryList); 804 } else { 805 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 806 // Inopportune interleaving -- Self is still on the cxq. 807 // This usually means the enqueue of self raced an exiting thread. 808 // Normally we'll find Self near the front of the cxq, so 809 // dequeueing is typically fast. If needbe we can accelerate 810 // this with some MCS/CHL-like bidirectional list hints and advisory 811 // back-links so dequeueing from the interior will normally operate 812 // in constant-time. 813 // Dequeue Self from either the head (with CAS) or from the interior 814 // with a linear-time scan and normal non-atomic memory operations. 815 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 816 // and then unlink Self from EntryList. We have to drain eventually, 817 // so it might as well be now. 818 819 ObjectWaiter * v = _cxq; 820 assert(v != NULL, "invariant"); 821 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 822 // The CAS above can fail from interference IFF a "RAT" arrived. 823 // In that case Self must be in the interior and can no longer be 824 // at the head of cxq. 825 if (v == SelfNode) { 826 assert(_cxq != v, "invariant"); 827 v = _cxq; // CAS above failed - start scan at head of list 828 } 829 ObjectWaiter * p; 830 ObjectWaiter * q = NULL; 831 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 832 q = p; 833 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 834 } 835 assert(v != SelfNode, "invariant"); 836 assert(p == SelfNode, "Node not found on cxq"); 837 assert(p != _cxq, "invariant"); 838 assert(q != NULL, "invariant"); 839 assert(q->_next == p, "invariant"); 840 q->_next = p->_next; 841 } 842 TEVENT(Unlink from cxq); 843 } 844 845 #ifdef ASSERT 846 // Diagnostic hygiene ... 847 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 848 SelfNode->_next = (ObjectWaiter *) 0xBAD; 849 SelfNode->TState = ObjectWaiter::TS_RUN; 850 #endif 851 } 852 853 // ----------------------------------------------------------------------------- 854 // Exit support 855 // 856 // exit() 857 // ~~~~~~ 858 // Note that the collector can't reclaim the objectMonitor or deflate 859 // the object out from underneath the thread calling ::exit() as the 860 // thread calling ::exit() never transitions to a stable state. 861 // This inhibits GC, which in turn inhibits asynchronous (and 862 // inopportune) reclamation of "this". 863 // 864 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 865 // There's one exception to the claim above, however. EnterI() can call 866 // exit() to drop a lock if the acquirer has been externally suspended. 867 // In that case exit() is called with _thread_state as _thread_blocked, 868 // but the monitor's _count field is > 0, which inhibits reclamation. 869 // 870 // 1-0 exit 871 // ~~~~~~~~ 872 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 873 // the fast-path operators have been optimized so the common ::exit() 874 // operation is 1-0. See i486.ad fast_unlock(), for instance. 875 // The code emitted by fast_unlock() elides the usual MEMBAR. This 876 // greatly improves latency -- MEMBAR and CAS having considerable local 877 // latency on modern processors -- but at the cost of "stranding". Absent the 878 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 879 // ::enter() path, resulting in the entering thread being stranding 880 // and a progress-liveness failure. Stranding is extremely rare. 881 // We use timers (timed park operations) & periodic polling to detect 882 // and recover from stranding. Potentially stranded threads periodically 883 // wake up and poll the lock. See the usage of the _Responsible variable. 884 // 885 // The CAS() in enter provides for safety and exclusion, while the CAS or 886 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 887 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 888 // We detect and recover from stranding with timers. 889 // 890 // If a thread transiently strands it'll park until (a) another 891 // thread acquires the lock and then drops the lock, at which time the 892 // exiting thread will notice and unpark the stranded thread, or, (b) 893 // the timer expires. If the lock is high traffic then the stranding latency 894 // will be low due to (a). If the lock is low traffic then the odds of 895 // stranding are lower, although the worst-case stranding latency 896 // is longer. Critically, we don't want to put excessive load in the 897 // platform's timer subsystem. We want to minimize both the timer injection 898 // rate (timers created/sec) as well as the number of timers active at 899 // any one time. (more precisely, we want to minimize timer-seconds, which is 900 // the integral of the # of active timers at any instant over time). 901 // Both impinge on OS scalability. Given that, at most one thread parked on 902 // a monitor will use a timer. 903 // 904 // There is also the risk of a futile wake-up. If we drop the lock 905 // another thread can reacquire the lock immediately, and we can 906 // then wake a thread unnecessarily. This is benign, and we've 907 // structured the code so the windows are short and the frequency 908 // of such futile wakups is low. 909 910 void NOINLINE ObjectMonitor::exit(bool not_suspended, TRAPS) { 911 Thread * const Self = THREAD; 912 if (THREAD != _owner) { 913 if (THREAD->is_lock_owned((address) _owner)) { 914 // Transmute _owner from a BasicLock pointer to a Thread address. 915 // We don't need to hold _mutex for this transition. 916 // Non-null to Non-null is safe as long as all readers can 917 // tolerate either flavor. 918 assert(_recursions == 0, "invariant"); 919 _owner = THREAD; 920 _recursions = 0; 921 } else { 922 // Apparent unbalanced locking ... 923 // Naively we'd like to throw IllegalMonitorStateException. 924 // As a practical matter we can neither allocate nor throw an 925 // exception as ::exit() can be called from leaf routines. 926 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 927 // Upon deeper reflection, however, in a properly run JVM the only 928 // way we should encounter this situation is in the presence of 929 // unbalanced JNI locking. TODO: CheckJNICalls. 930 // See also: CR4414101 931 TEVENT(Exit - Throw IMSX); 932 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 933 return; 934 } 935 } 936 937 if (_recursions != 0) { 938 _recursions--; // this is simple recursive enter 939 TEVENT(Inflated exit - recursive); 940 return; 941 } 942 943 // Invariant: after setting Responsible=null an thread must execute 944 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 945 if ((SyncFlags & 4) == 0) { 946 _Responsible = NULL; 947 } 948 949 #if INCLUDE_TRACE 950 // get the owner's thread id for the MonitorEnter event 951 // if it is enabled and the thread isn't suspended 952 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 953 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 954 } 955 #endif 956 957 for (;;) { 958 assert(THREAD == _owner, "invariant"); 959 960 961 if (Knob_ExitPolicy == 0) { 962 // release semantics: prior loads and stores from within the critical section 963 // must not float (reorder) past the following store that drops the lock. 964 // On SPARC that requires MEMBAR #loadstore|#storestore. 965 // But of course in TSO #loadstore|#storestore is not required. 966 // I'd like to write one of the following: 967 // A. OrderAccess::release() ; _owner = NULL 968 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 969 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 970 // store into a _dummy variable. That store is not needed, but can result 971 // in massive wasteful coherency traffic on classic SMP systems. 972 // Instead, I use release_store(), which is implemented as just a simple 973 // ST on x64, x86 and SPARC. 974 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 975 OrderAccess::storeload(); // See if we need to wake a successor 976 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 977 TEVENT(Inflated exit - simple egress); 978 return; 979 } 980 TEVENT(Inflated exit - complex egress); 981 // Other threads are blocked trying to acquire the lock. 982 983 // Normally the exiting thread is responsible for ensuring succession, 984 // but if other successors are ready or other entering threads are spinning 985 // then this thread can simply store NULL into _owner and exit without 986 // waking a successor. The existence of spinners or ready successors 987 // guarantees proper succession (liveness). Responsibility passes to the 988 // ready or running successors. The exiting thread delegates the duty. 989 // More precisely, if a successor already exists this thread is absolved 990 // of the responsibility of waking (unparking) one. 991 // 992 // The _succ variable is critical to reducing futile wakeup frequency. 993 // _succ identifies the "heir presumptive" thread that has been made 994 // ready (unparked) but that has not yet run. We need only one such 995 // successor thread to guarantee progress. 996 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 997 // section 3.3 "Futile Wakeup Throttling" for details. 998 // 999 // Note that spinners in Enter() also set _succ non-null. 1000 // In the current implementation spinners opportunistically set 1001 // _succ so that exiting threads might avoid waking a successor. 1002 // Another less appealing alternative would be for the exiting thread 1003 // to drop the lock and then spin briefly to see if a spinner managed 1004 // to acquire the lock. If so, the exiting thread could exit 1005 // immediately without waking a successor, otherwise the exiting 1006 // thread would need to dequeue and wake a successor. 1007 // (Note that we'd need to make the post-drop spin short, but no 1008 // shorter than the worst-case round-trip cache-line migration time. 1009 // The dropped lock needs to become visible to the spinner, and then 1010 // the acquisition of the lock by the spinner must become visible to 1011 // the exiting thread). 1012 1013 // It appears that an heir-presumptive (successor) must be made ready. 1014 // Only the current lock owner can manipulate the EntryList or 1015 // drain _cxq, so we need to reacquire the lock. If we fail 1016 // to reacquire the lock the responsibility for ensuring succession 1017 // falls to the new owner. 1018 // 1019 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1020 return; 1021 } 1022 TEVENT(Exit - Reacquired); 1023 } else { 1024 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1025 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1026 OrderAccess::storeload(); 1027 // Ratify the previously observed values. 1028 if (_cxq == NULL || _succ != NULL) { 1029 TEVENT(Inflated exit - simple egress); 1030 return; 1031 } 1032 1033 // inopportune interleaving -- the exiting thread (this thread) 1034 // in the fast-exit path raced an entering thread in the slow-enter 1035 // path. 1036 // We have two choices: 1037 // A. Try to reacquire the lock. 1038 // If the CAS() fails return immediately, otherwise 1039 // we either restart/rerun the exit operation, or simply 1040 // fall-through into the code below which wakes a successor. 1041 // B. If the elements forming the EntryList|cxq are TSM 1042 // we could simply unpark() the lead thread and return 1043 // without having set _succ. 1044 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1045 TEVENT(Inflated exit - reacquired succeeded); 1046 return; 1047 } 1048 TEVENT(Inflated exit - reacquired failed); 1049 } else { 1050 TEVENT(Inflated exit - complex egress); 1051 } 1052 } 1053 1054 guarantee(_owner == THREAD, "invariant"); 1055 1056 ObjectWaiter * w = NULL; 1057 int QMode = Knob_QMode; 1058 1059 if (QMode == 2 && _cxq != NULL) { 1060 // QMode == 2 : cxq has precedence over EntryList. 1061 // Try to directly wake a successor from the cxq. 1062 // If successful, the successor will need to unlink itself from cxq. 1063 w = _cxq; 1064 assert(w != NULL, "invariant"); 1065 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1066 ExitEpilog(Self, w); 1067 return; 1068 } 1069 1070 if (QMode == 3 && _cxq != NULL) { 1071 // Aggressively drain cxq into EntryList at the first opportunity. 1072 // This policy ensure that recently-run threads live at the head of EntryList. 1073 // Drain _cxq into EntryList - bulk transfer. 1074 // First, detach _cxq. 1075 // The following loop is tantamount to: w = swap(&cxq, NULL) 1076 w = _cxq; 1077 for (;;) { 1078 assert(w != NULL, "Invariant"); 1079 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1080 if (u == w) break; 1081 w = u; 1082 } 1083 assert(w != NULL, "invariant"); 1084 1085 ObjectWaiter * q = NULL; 1086 ObjectWaiter * p; 1087 for (p = w; p != NULL; p = p->_next) { 1088 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1089 p->TState = ObjectWaiter::TS_ENTER; 1090 p->_prev = q; 1091 q = p; 1092 } 1093 1094 // Append the RATs to the EntryList 1095 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1096 ObjectWaiter * Tail; 1097 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1098 Tail = Tail->_next) 1099 /* empty */; 1100 if (Tail == NULL) { 1101 _EntryList = w; 1102 } else { 1103 Tail->_next = w; 1104 w->_prev = Tail; 1105 } 1106 1107 // Fall thru into code that tries to wake a successor from EntryList 1108 } 1109 1110 if (QMode == 4 && _cxq != NULL) { 1111 // Aggressively drain cxq into EntryList at the first opportunity. 1112 // This policy ensure that recently-run threads live at the head of EntryList. 1113 1114 // Drain _cxq into EntryList - bulk transfer. 1115 // First, detach _cxq. 1116 // The following loop is tantamount to: w = swap(&cxq, NULL) 1117 w = _cxq; 1118 for (;;) { 1119 assert(w != NULL, "Invariant"); 1120 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1121 if (u == w) break; 1122 w = u; 1123 } 1124 assert(w != NULL, "invariant"); 1125 1126 ObjectWaiter * q = NULL; 1127 ObjectWaiter * p; 1128 for (p = w; p != NULL; p = p->_next) { 1129 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1130 p->TState = ObjectWaiter::TS_ENTER; 1131 p->_prev = q; 1132 q = p; 1133 } 1134 1135 // Prepend the RATs to the EntryList 1136 if (_EntryList != NULL) { 1137 q->_next = _EntryList; 1138 _EntryList->_prev = q; 1139 } 1140 _EntryList = w; 1141 1142 // Fall thru into code that tries to wake a successor from EntryList 1143 } 1144 1145 w = _EntryList; 1146 if (w != NULL) { 1147 // I'd like to write: guarantee (w->_thread != Self). 1148 // But in practice an exiting thread may find itself on the EntryList. 1149 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1150 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1151 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1152 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1153 // release the lock "O". T2 resumes immediately after the ST of null into 1154 // _owner, above. T2 notices that the EntryList is populated, so it 1155 // reacquires the lock and then finds itself on the EntryList. 1156 // Given all that, we have to tolerate the circumstance where "w" is 1157 // associated with Self. 1158 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1159 ExitEpilog(Self, w); 1160 return; 1161 } 1162 1163 // If we find that both _cxq and EntryList are null then just 1164 // re-run the exit protocol from the top. 1165 w = _cxq; 1166 if (w == NULL) continue; 1167 1168 // Drain _cxq into EntryList - bulk transfer. 1169 // First, detach _cxq. 1170 // The following loop is tantamount to: w = swap(&cxq, NULL) 1171 for (;;) { 1172 assert(w != NULL, "Invariant"); 1173 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1174 if (u == w) break; 1175 w = u; 1176 } 1177 TEVENT(Inflated exit - drain cxq into EntryList); 1178 1179 assert(w != NULL, "invariant"); 1180 assert(_EntryList == NULL, "invariant"); 1181 1182 // Convert the LIFO SLL anchored by _cxq into a DLL. 1183 // The list reorganization step operates in O(LENGTH(w)) time. 1184 // It's critical that this step operate quickly as 1185 // "Self" still holds the outer-lock, restricting parallelism 1186 // and effectively lengthening the critical section. 1187 // Invariant: s chases t chases u. 1188 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1189 // we have faster access to the tail. 1190 1191 if (QMode == 1) { 1192 // QMode == 1 : drain cxq to EntryList, reversing order 1193 // We also reverse the order of the list. 1194 ObjectWaiter * s = NULL; 1195 ObjectWaiter * t = w; 1196 ObjectWaiter * u = NULL; 1197 while (t != NULL) { 1198 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1199 t->TState = ObjectWaiter::TS_ENTER; 1200 u = t->_next; 1201 t->_prev = u; 1202 t->_next = s; 1203 s = t; 1204 t = u; 1205 } 1206 _EntryList = s; 1207 assert(s != NULL, "invariant"); 1208 } else { 1209 // QMode == 0 or QMode == 2 1210 _EntryList = w; 1211 ObjectWaiter * q = NULL; 1212 ObjectWaiter * p; 1213 for (p = w; p != NULL; p = p->_next) { 1214 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1215 p->TState = ObjectWaiter::TS_ENTER; 1216 p->_prev = q; 1217 q = p; 1218 } 1219 } 1220 1221 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1222 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1223 1224 // See if we can abdicate to a spinner instead of waking a thread. 1225 // A primary goal of the implementation is to reduce the 1226 // context-switch rate. 1227 if (_succ != NULL) continue; 1228 1229 w = _EntryList; 1230 if (w != NULL) { 1231 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1232 ExitEpilog(Self, w); 1233 return; 1234 } 1235 } 1236 } 1237 1238 // ExitSuspendEquivalent: 1239 // A faster alternate to handle_special_suspend_equivalent_condition() 1240 // 1241 // handle_special_suspend_equivalent_condition() unconditionally 1242 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1243 // operations have high latency. Note that in ::enter() we call HSSEC 1244 // while holding the monitor, so we effectively lengthen the critical sections. 1245 // 1246 // There are a number of possible solutions: 1247 // 1248 // A. To ameliorate the problem we might also defer state transitions 1249 // to as late as possible -- just prior to parking. 1250 // Given that, we'd call HSSEC after having returned from park(), 1251 // but before attempting to acquire the monitor. This is only a 1252 // partial solution. It avoids calling HSSEC while holding the 1253 // monitor (good), but it still increases successor reacquisition latency -- 1254 // the interval between unparking a successor and the time the successor 1255 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1256 // If we use this technique we can also avoid EnterI()-exit() loop 1257 // in ::enter() where we iteratively drop the lock and then attempt 1258 // to reacquire it after suspending. 1259 // 1260 // B. In the future we might fold all the suspend bits into a 1261 // composite per-thread suspend flag and then update it with CAS(). 1262 // Alternately, a Dekker-like mechanism with multiple variables 1263 // would suffice: 1264 // ST Self->_suspend_equivalent = false 1265 // MEMBAR 1266 // LD Self_>_suspend_flags 1267 // 1268 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1269 // with a more efficient implementation, the need to use "FastHSSEC" has 1270 // decreased. - Dave 1271 1272 1273 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1274 const int Mode = Knob_FastHSSEC; 1275 if (Mode && !jSelf->is_external_suspend()) { 1276 assert(jSelf->is_suspend_equivalent(), "invariant"); 1277 jSelf->clear_suspend_equivalent(); 1278 if (2 == Mode) OrderAccess::storeload(); 1279 if (!jSelf->is_external_suspend()) return false; 1280 // We raced a suspension -- fall thru into the slow path 1281 TEVENT(ExitSuspendEquivalent - raced); 1282 jSelf->set_suspend_equivalent(); 1283 } 1284 return jSelf->handle_special_suspend_equivalent_condition(); 1285 } 1286 1287 1288 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1289 assert(_owner == Self, "invariant"); 1290 1291 // Exit protocol: 1292 // 1. ST _succ = wakee 1293 // 2. membar #loadstore|#storestore; 1294 // 2. ST _owner = NULL 1295 // 3. unpark(wakee) 1296 1297 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1298 ParkEvent * Trigger = Wakee->_event; 1299 1300 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1301 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1302 // out-of-scope (non-extant). 1303 Wakee = NULL; 1304 1305 // Drop the lock 1306 OrderAccess::release_store_ptr(&_owner, NULL); 1307 OrderAccess::fence(); // ST _owner vs LD in unpark() 1308 1309 if (SafepointSynchronize::do_call_back()) { 1310 TEVENT(unpark before SAFEPOINT); 1311 } 1312 1313 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1314 Trigger->unpark(); 1315 1316 // Maintain stats and report events to JVMTI 1317 if (ObjectMonitor::_sync_Parks != NULL) { 1318 ObjectMonitor::_sync_Parks->inc(); 1319 } 1320 } 1321 1322 1323 // ----------------------------------------------------------------------------- 1324 // Class Loader deadlock handling. 1325 // 1326 // complete_exit exits a lock returning recursion count 1327 // complete_exit/reenter operate as a wait without waiting 1328 // complete_exit requires an inflated monitor 1329 // The _owner field is not always the Thread addr even with an 1330 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1331 // thread due to contention. 1332 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1333 Thread * const Self = THREAD; 1334 assert(Self->is_Java_thread(), "Must be Java thread!"); 1335 JavaThread *jt = (JavaThread *)THREAD; 1336 1337 DeferredInitialize(); 1338 1339 if (THREAD != _owner) { 1340 if (THREAD->is_lock_owned ((address)_owner)) { 1341 assert(_recursions == 0, "internal state error"); 1342 _owner = THREAD; // Convert from basiclock addr to Thread addr 1343 _recursions = 0; 1344 } 1345 } 1346 1347 guarantee(Self == _owner, "complete_exit not owner"); 1348 intptr_t save = _recursions; // record the old recursion count 1349 _recursions = 0; // set the recursion level to be 0 1350 exit(true, Self); // exit the monitor 1351 guarantee(_owner != Self, "invariant"); 1352 return save; 1353 } 1354 1355 // reenter() enters a lock and sets recursion count 1356 // complete_exit/reenter operate as a wait without waiting 1357 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1358 Thread * const Self = THREAD; 1359 assert(Self->is_Java_thread(), "Must be Java thread!"); 1360 JavaThread *jt = (JavaThread *)THREAD; 1361 1362 guarantee(_owner != Self, "reenter already owner"); 1363 enter(THREAD); // enter the monitor 1364 guarantee(_recursions == 0, "reenter recursion"); 1365 _recursions = recursions; 1366 return; 1367 } 1368 1369 1370 // ----------------------------------------------------------------------------- 1371 // A macro is used below because there may already be a pending 1372 // exception which should not abort the execution of the routines 1373 // which use this (which is why we don't put this into check_slow and 1374 // call it with a CHECK argument). 1375 1376 #define CHECK_OWNER() \ 1377 do { \ 1378 if (THREAD != _owner) { \ 1379 if (THREAD->is_lock_owned((address) _owner)) { \ 1380 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1381 _recursions = 0; \ 1382 } else { \ 1383 TEVENT(Throw IMSX); \ 1384 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1385 } \ 1386 } \ 1387 } while (false) 1388 1389 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1390 // TODO-FIXME: remove check_slow() -- it's likely dead. 1391 1392 void ObjectMonitor::check_slow(TRAPS) { 1393 TEVENT(check_slow - throw IMSX); 1394 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1395 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1396 } 1397 1398 static int Adjust(volatile int * adr, int dx) { 1399 int v; 1400 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1401 return v; 1402 } 1403 1404 // helper method for posting a monitor wait event 1405 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1406 jlong notifier_tid, 1407 jlong timeout, 1408 bool timedout) { 1409 event->set_klass(((oop)this->object())->klass()); 1410 event->set_timeout((TYPE_ULONG)timeout); 1411 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1412 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1413 event->set_timedOut((TYPE_BOOLEAN)timedout); 1414 event->commit(); 1415 } 1416 1417 // ----------------------------------------------------------------------------- 1418 // Wait/Notify/NotifyAll 1419 // 1420 // Note: a subset of changes to ObjectMonitor::wait() 1421 // will need to be replicated in complete_exit 1422 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1423 Thread * const Self = THREAD; 1424 assert(Self->is_Java_thread(), "Must be Java thread!"); 1425 JavaThread *jt = (JavaThread *)THREAD; 1426 1427 DeferredInitialize(); 1428 1429 // Throw IMSX or IEX. 1430 CHECK_OWNER(); 1431 1432 EventJavaMonitorWait event; 1433 1434 // check for a pending interrupt 1435 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1436 // post monitor waited event. Note that this is past-tense, we are done waiting. 1437 if (JvmtiExport::should_post_monitor_waited()) { 1438 // Note: 'false' parameter is passed here because the 1439 // wait was not timed out due to thread interrupt. 1440 JvmtiExport::post_monitor_waited(jt, this, false); 1441 1442 // In this short circuit of the monitor wait protocol, the 1443 // current thread never drops ownership of the monitor and 1444 // never gets added to the wait queue so the current thread 1445 // cannot be made the successor. This means that the 1446 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1447 // consume an unpark() meant for the ParkEvent associated with 1448 // this ObjectMonitor. 1449 } 1450 if (event.should_commit()) { 1451 post_monitor_wait_event(&event, 0, millis, false); 1452 } 1453 TEVENT(Wait - Throw IEX); 1454 THROW(vmSymbols::java_lang_InterruptedException()); 1455 return; 1456 } 1457 1458 TEVENT(Wait); 1459 1460 assert(Self->_Stalled == 0, "invariant"); 1461 Self->_Stalled = intptr_t(this); 1462 jt->set_current_waiting_monitor(this); 1463 1464 // create a node to be put into the queue 1465 // Critically, after we reset() the event but prior to park(), we must check 1466 // for a pending interrupt. 1467 ObjectWaiter node(Self); 1468 node.TState = ObjectWaiter::TS_WAIT; 1469 Self->_ParkEvent->reset(); 1470 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1471 1472 // Enter the waiting queue, which is a circular doubly linked list in this case 1473 // but it could be a priority queue or any data structure. 1474 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1475 // by the the owner of the monitor *except* in the case where park() 1476 // returns because of a timeout of interrupt. Contention is exceptionally rare 1477 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1478 1479 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1480 AddWaiter(&node); 1481 Thread::SpinRelease(&_WaitSetLock); 1482 1483 if ((SyncFlags & 4) == 0) { 1484 _Responsible = NULL; 1485 } 1486 intptr_t save = _recursions; // record the old recursion count 1487 _waiters++; // increment the number of waiters 1488 _recursions = 0; // set the recursion level to be 1 1489 exit(true, Self); // exit the monitor 1490 guarantee(_owner != Self, "invariant"); 1491 1492 // The thread is on the WaitSet list - now park() it. 1493 // On MP systems it's conceivable that a brief spin before we park 1494 // could be profitable. 1495 // 1496 // TODO-FIXME: change the following logic to a loop of the form 1497 // while (!timeout && !interrupted && _notified == 0) park() 1498 1499 int ret = OS_OK; 1500 int WasNotified = 0; 1501 { // State transition wrappers 1502 OSThread* osthread = Self->osthread(); 1503 OSThreadWaitState osts(osthread, true); 1504 { 1505 ThreadBlockInVM tbivm(jt); 1506 // Thread is in thread_blocked state and oop access is unsafe. 1507 jt->set_suspend_equivalent(); 1508 1509 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1510 // Intentionally empty 1511 } else if (node._notified == 0) { 1512 if (millis <= 0) { 1513 Self->_ParkEvent->park(); 1514 } else { 1515 ret = Self->_ParkEvent->park(millis); 1516 } 1517 } 1518 1519 // were we externally suspended while we were waiting? 1520 if (ExitSuspendEquivalent (jt)) { 1521 // TODO-FIXME: add -- if succ == Self then succ = null. 1522 jt->java_suspend_self(); 1523 } 1524 1525 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1526 1527 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1528 // from the WaitSet to the EntryList. 1529 // See if we need to remove Node from the WaitSet. 1530 // We use double-checked locking to avoid grabbing _WaitSetLock 1531 // if the thread is not on the wait queue. 1532 // 1533 // Note that we don't need a fence before the fetch of TState. 1534 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1535 // written by the is thread. (perhaps the fetch might even be satisfied 1536 // by a look-aside into the processor's own store buffer, although given 1537 // the length of the code path between the prior ST and this load that's 1538 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1539 // then we'll acquire the lock and then re-fetch a fresh TState value. 1540 // That is, we fail toward safety. 1541 1542 if (node.TState == ObjectWaiter::TS_WAIT) { 1543 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1544 if (node.TState == ObjectWaiter::TS_WAIT) { 1545 DequeueSpecificWaiter(&node); // unlink from WaitSet 1546 assert(node._notified == 0, "invariant"); 1547 node.TState = ObjectWaiter::TS_RUN; 1548 } 1549 Thread::SpinRelease(&_WaitSetLock); 1550 } 1551 1552 // The thread is now either on off-list (TS_RUN), 1553 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1554 // The Node's TState variable is stable from the perspective of this thread. 1555 // No other threads will asynchronously modify TState. 1556 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1557 OrderAccess::loadload(); 1558 if (_succ == Self) _succ = NULL; 1559 WasNotified = node._notified; 1560 1561 // Reentry phase -- reacquire the monitor. 1562 // re-enter contended monitor after object.wait(). 1563 // retain OBJECT_WAIT state until re-enter successfully completes 1564 // Thread state is thread_in_vm and oop access is again safe, 1565 // although the raw address of the object may have changed. 1566 // (Don't cache naked oops over safepoints, of course). 1567 1568 // post monitor waited event. Note that this is past-tense, we are done waiting. 1569 if (JvmtiExport::should_post_monitor_waited()) { 1570 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1571 1572 if (node._notified != 0 && _succ == Self) { 1573 // In this part of the monitor wait-notify-reenter protocol it 1574 // is possible (and normal) for another thread to do a fastpath 1575 // monitor enter-exit while this thread is still trying to get 1576 // to the reenter portion of the protocol. 1577 // 1578 // The ObjectMonitor was notified and the current thread is 1579 // the successor which also means that an unpark() has already 1580 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1581 // consume the unpark() that was done when the successor was 1582 // set because the same ParkEvent is shared between Java 1583 // monitors and JVM/TI RawMonitors (for now). 1584 // 1585 // We redo the unpark() to ensure forward progress, i.e., we 1586 // don't want all pending threads hanging (parked) with none 1587 // entering the unlocked monitor. 1588 node._event->unpark(); 1589 } 1590 } 1591 1592 if (event.should_commit()) { 1593 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1594 } 1595 1596 OrderAccess::fence(); 1597 1598 assert(Self->_Stalled != 0, "invariant"); 1599 Self->_Stalled = 0; 1600 1601 assert(_owner != Self, "invariant"); 1602 ObjectWaiter::TStates v = node.TState; 1603 if (v == ObjectWaiter::TS_RUN) { 1604 enter(Self); 1605 } else { 1606 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1607 ReenterI(Self, &node); 1608 node.wait_reenter_end(this); 1609 } 1610 1611 // Self has reacquired the lock. 1612 // Lifecycle - the node representing Self must not appear on any queues. 1613 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1614 // want residual elements associated with this thread left on any lists. 1615 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1616 assert(_owner == Self, "invariant"); 1617 assert(_succ != Self, "invariant"); 1618 } // OSThreadWaitState() 1619 1620 jt->set_current_waiting_monitor(NULL); 1621 1622 guarantee(_recursions == 0, "invariant"); 1623 _recursions = save; // restore the old recursion count 1624 _waiters--; // decrement the number of waiters 1625 1626 // Verify a few postconditions 1627 assert(_owner == Self, "invariant"); 1628 assert(_succ != Self, "invariant"); 1629 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1630 1631 if (SyncFlags & 32) { 1632 OrderAccess::fence(); 1633 } 1634 1635 // check if the notification happened 1636 if (!WasNotified) { 1637 // no, it could be timeout or Thread.interrupt() or both 1638 // check for interrupt event, otherwise it is timeout 1639 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1640 TEVENT(Wait - throw IEX from epilog); 1641 THROW(vmSymbols::java_lang_InterruptedException()); 1642 } 1643 } 1644 1645 // NOTE: Spurious wake up will be consider as timeout. 1646 // Monitor notify has precedence over thread interrupt. 1647 } 1648 1649 1650 // Consider: 1651 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1652 // then instead of transferring a thread from the WaitSet to the EntryList 1653 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1654 1655 void ObjectMonitor::notify(TRAPS) { 1656 CHECK_OWNER(); 1657 if (_WaitSet == NULL) { 1658 TEVENT(Empty-Notify); 1659 return; 1660 } 1661 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1662 1663 int Policy = Knob_MoveNotifyee; 1664 1665 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1666 ObjectWaiter * iterator = DequeueWaiter(); 1667 if (iterator != NULL) { 1668 TEVENT(Notify1 - Transfer); 1669 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1670 guarantee(iterator->_notified == 0, "invariant"); 1671 if (Policy != 4) { 1672 iterator->TState = ObjectWaiter::TS_ENTER; 1673 } 1674 iterator->_notified = 1; 1675 Thread * Self = THREAD; 1676 iterator->_notifier_tid = Self->osthread()->thread_id(); 1677 1678 ObjectWaiter * List = _EntryList; 1679 if (List != NULL) { 1680 assert(List->_prev == NULL, "invariant"); 1681 assert(List->TState == ObjectWaiter::TS_ENTER, "invariant"); 1682 assert(List != iterator, "invariant"); 1683 } 1684 1685 if (Policy == 0) { // prepend to EntryList 1686 if (List == NULL) { 1687 iterator->_next = iterator->_prev = NULL; 1688 _EntryList = iterator; 1689 } else { 1690 List->_prev = iterator; 1691 iterator->_next = List; 1692 iterator->_prev = NULL; 1693 _EntryList = iterator; 1694 } 1695 } else if (Policy == 1) { // append to EntryList 1696 if (List == NULL) { 1697 iterator->_next = iterator->_prev = NULL; 1698 _EntryList = iterator; 1699 } else { 1700 // CONSIDER: finding the tail currently requires a linear-time walk of 1701 // the EntryList. We can make tail access constant-time by converting to 1702 // a CDLL instead of using our current DLL. 1703 ObjectWaiter * Tail; 1704 for (Tail = List; Tail->_next != NULL; Tail = Tail->_next) /* empty */; 1705 assert(Tail != NULL && Tail->_next == NULL, "invariant"); 1706 Tail->_next = iterator; 1707 iterator->_prev = Tail; 1708 iterator->_next = NULL; 1709 } 1710 } else if (Policy == 2) { // prepend to cxq 1711 // prepend to cxq 1712 if (List == NULL) { 1713 iterator->_next = iterator->_prev = NULL; 1714 _EntryList = iterator; 1715 } else { 1716 iterator->TState = ObjectWaiter::TS_CXQ; 1717 for (;;) { 1718 ObjectWaiter * Front = _cxq; 1719 iterator->_next = Front; 1720 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1721 break; 1722 } 1723 } 1724 } 1725 } else if (Policy == 3) { // append to cxq 1726 iterator->TState = ObjectWaiter::TS_CXQ; 1727 for (;;) { 1728 ObjectWaiter * Tail; 1729 Tail = _cxq; 1730 if (Tail == NULL) { 1731 iterator->_next = NULL; 1732 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1733 break; 1734 } 1735 } else { 1736 while (Tail->_next != NULL) Tail = Tail->_next; 1737 Tail->_next = iterator; 1738 iterator->_prev = Tail; 1739 iterator->_next = NULL; 1740 break; 1741 } 1742 } 1743 } else { 1744 ParkEvent * ev = iterator->_event; 1745 iterator->TState = ObjectWaiter::TS_RUN; 1746 OrderAccess::fence(); 1747 ev->unpark(); 1748 } 1749 1750 if (Policy < 4) { 1751 iterator->wait_reenter_begin(this); 1752 } 1753 1754 // _WaitSetLock protects the wait queue, not the EntryList. We could 1755 // move the add-to-EntryList operation, above, outside the critical section 1756 // protected by _WaitSetLock. In practice that's not useful. With the 1757 // exception of wait() timeouts and interrupts the monitor owner 1758 // is the only thread that grabs _WaitSetLock. There's almost no contention 1759 // on _WaitSetLock so it's not profitable to reduce the length of the 1760 // critical section. 1761 } 1762 1763 Thread::SpinRelease(&_WaitSetLock); 1764 1765 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1766 ObjectMonitor::_sync_Notifications->inc(); 1767 } 1768 } 1769 1770 1771 void ObjectMonitor::notifyAll(TRAPS) { 1772 CHECK_OWNER(); 1773 ObjectWaiter* iterator; 1774 if (_WaitSet == NULL) { 1775 TEVENT(Empty-NotifyAll); 1776 return; 1777 } 1778 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1779 1780 int Policy = Knob_MoveNotifyee; 1781 int Tally = 0; 1782 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notifyall"); 1783 1784 for (;;) { 1785 iterator = DequeueWaiter(); 1786 if (iterator == NULL) break; 1787 TEVENT(NotifyAll - Transfer1); 1788 ++Tally; 1789 1790 // Disposition - what might we do with iterator ? 1791 // a. add it directly to the EntryList - either tail or head. 1792 // b. push it onto the front of the _cxq. 1793 // For now we use (a). 1794 1795 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1796 guarantee(iterator->_notified == 0, "invariant"); 1797 iterator->_notified = 1; 1798 Thread * Self = THREAD; 1799 iterator->_notifier_tid = Self->osthread()->thread_id(); 1800 if (Policy != 4) { 1801 iterator->TState = ObjectWaiter::TS_ENTER; 1802 } 1803 1804 ObjectWaiter * List = _EntryList; 1805 if (List != NULL) { 1806 assert(List->_prev == NULL, "invariant"); 1807 assert(List->TState == ObjectWaiter::TS_ENTER, "invariant"); 1808 assert(List != iterator, "invariant"); 1809 } 1810 1811 if (Policy == 0) { // prepend to EntryList 1812 if (List == NULL) { 1813 iterator->_next = iterator->_prev = NULL; 1814 _EntryList = iterator; 1815 } else { 1816 List->_prev = iterator; 1817 iterator->_next = List; 1818 iterator->_prev = NULL; 1819 _EntryList = iterator; 1820 } 1821 } else if (Policy == 1) { // append to EntryList 1822 if (List == NULL) { 1823 iterator->_next = iterator->_prev = NULL; 1824 _EntryList = iterator; 1825 } else { 1826 // CONSIDER: finding the tail currently requires a linear-time walk of 1827 // the EntryList. We can make tail access constant-time by converting to 1828 // a CDLL instead of using our current DLL. 1829 ObjectWaiter * Tail; 1830 for (Tail = List; Tail->_next != NULL; Tail = Tail->_next) /* empty */; 1831 assert(Tail != NULL && Tail->_next == NULL, "invariant"); 1832 Tail->_next = iterator; 1833 iterator->_prev = Tail; 1834 iterator->_next = NULL; 1835 } 1836 } else if (Policy == 2) { // prepend to cxq 1837 // prepend to cxq 1838 iterator->TState = ObjectWaiter::TS_CXQ; 1839 for (;;) { 1840 ObjectWaiter * Front = _cxq; 1841 iterator->_next = Front; 1842 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1843 break; 1844 } 1845 } 1846 } else if (Policy == 3) { // append to cxq 1847 iterator->TState = ObjectWaiter::TS_CXQ; 1848 for (;;) { 1849 ObjectWaiter * Tail; 1850 Tail = _cxq; 1851 if (Tail == NULL) { 1852 iterator->_next = NULL; 1853 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1854 break; 1855 } 1856 } else { 1857 while (Tail->_next != NULL) Tail = Tail->_next; 1858 Tail->_next = iterator; 1859 iterator->_prev = Tail; 1860 iterator->_next = NULL; 1861 break; 1862 } 1863 } 1864 } else { 1865 ParkEvent * ev = iterator->_event; 1866 iterator->TState = ObjectWaiter::TS_RUN; 1867 OrderAccess::fence(); 1868 ev->unpark(); 1869 } 1870 1871 if (Policy < 4) { 1872 iterator->wait_reenter_begin(this); 1873 } 1874 1875 // _WaitSetLock protects the wait queue, not the EntryList. We could 1876 // move the add-to-EntryList operation, above, outside the critical section 1877 // protected by _WaitSetLock. In practice that's not useful. With the 1878 // exception of wait() timeouts and interrupts the monitor owner 1879 // is the only thread that grabs _WaitSetLock. There's almost no contention 1880 // on _WaitSetLock so it's not profitable to reduce the length of the 1881 // critical section. 1882 } 1883 1884 Thread::SpinRelease(&_WaitSetLock); 1885 1886 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1887 ObjectMonitor::_sync_Notifications->inc(Tally); 1888 } 1889 } 1890 1891 // ----------------------------------------------------------------------------- 1892 // Adaptive Spinning Support 1893 // 1894 // Adaptive spin-then-block - rational spinning 1895 // 1896 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1897 // algorithm. On high order SMP systems it would be better to start with 1898 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1899 // a contending thread could enqueue itself on the cxq and then spin locally 1900 // on a thread-specific variable such as its ParkEvent._Event flag. 1901 // That's left as an exercise for the reader. Note that global spinning is 1902 // not problematic on Niagara, as the L2 cache serves the interconnect and 1903 // has both low latency and massive bandwidth. 1904 // 1905 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1906 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1907 // (duration) or we can fix the count at approximately the duration of 1908 // a context switch and vary the frequency. Of course we could also 1909 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1910 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1911 // 1912 // This implementation varies the duration "D", where D varies with 1913 // the success rate of recent spin attempts. (D is capped at approximately 1914 // length of a round-trip context switch). The success rate for recent 1915 // spin attempts is a good predictor of the success rate of future spin 1916 // attempts. The mechanism adapts automatically to varying critical 1917 // section length (lock modality), system load and degree of parallelism. 1918 // D is maintained per-monitor in _SpinDuration and is initialized 1919 // optimistically. Spin frequency is fixed at 100%. 1920 // 1921 // Note that _SpinDuration is volatile, but we update it without locks 1922 // or atomics. The code is designed so that _SpinDuration stays within 1923 // a reasonable range even in the presence of races. The arithmetic 1924 // operations on _SpinDuration are closed over the domain of legal values, 1925 // so at worst a race will install and older but still legal value. 1926 // At the very worst this introduces some apparent non-determinism. 1927 // We might spin when we shouldn't or vice-versa, but since the spin 1928 // count are relatively short, even in the worst case, the effect is harmless. 1929 // 1930 // Care must be taken that a low "D" value does not become an 1931 // an absorbing state. Transient spinning failures -- when spinning 1932 // is overall profitable -- should not cause the system to converge 1933 // on low "D" values. We want spinning to be stable and predictable 1934 // and fairly responsive to change and at the same time we don't want 1935 // it to oscillate, become metastable, be "too" non-deterministic, 1936 // or converge on or enter undesirable stable absorbing states. 1937 // 1938 // We implement a feedback-based control system -- using past behavior 1939 // to predict future behavior. We face two issues: (a) if the 1940 // input signal is random then the spin predictor won't provide optimal 1941 // results, and (b) if the signal frequency is too high then the control 1942 // system, which has some natural response lag, will "chase" the signal. 1943 // (b) can arise from multimodal lock hold times. Transient preemption 1944 // can also result in apparent bimodal lock hold times. 1945 // Although sub-optimal, neither condition is particularly harmful, as 1946 // in the worst-case we'll spin when we shouldn't or vice-versa. 1947 // The maximum spin duration is rather short so the failure modes aren't bad. 1948 // To be conservative, I've tuned the gain in system to bias toward 1949 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1950 // "rings" or oscillates between spinning and not spinning. This happens 1951 // when spinning is just on the cusp of profitability, however, so the 1952 // situation is not dire. The state is benign -- there's no need to add 1953 // hysteresis control to damp the transition rate between spinning and 1954 // not spinning. 1955 1956 intptr_t ObjectMonitor::SpinCallbackArgument = 0; 1957 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL; 1958 1959 // Spinning: Fixed frequency (100%), vary duration 1960 1961 1962 int ObjectMonitor::TrySpin_VaryDuration(Thread * Self) { 1963 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1964 int ctr = Knob_FixedSpin; 1965 if (ctr != 0) { 1966 while (--ctr >= 0) { 1967 if (TryLock(Self) > 0) return 1; 1968 SpinPause(); 1969 } 1970 return 0; 1971 } 1972 1973 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1974 if (TryLock(Self) > 0) { 1975 // Increase _SpinDuration ... 1976 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1977 // Raising _SpurDuration to the poverty line is key. 1978 int x = _SpinDuration; 1979 if (x < Knob_SpinLimit) { 1980 if (x < Knob_Poverty) x = Knob_Poverty; 1981 _SpinDuration = x + Knob_BonusB; 1982 } 1983 return 1; 1984 } 1985 SpinPause(); 1986 } 1987 1988 // Admission control - verify preconditions for spinning 1989 // 1990 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1991 // becoming an absorbing state. Put another way, we spin briefly to 1992 // sample, just in case the system load, parallelism, contention, or lock 1993 // modality changed. 1994 // 1995 // Consider the following alternative: 1996 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1997 // spin attempt. "Periodically" might mean after a tally of 1998 // the # of failed spin attempts (or iterations) reaches some threshold. 1999 // This takes us into the realm of 1-out-of-N spinning, where we 2000 // hold the duration constant but vary the frequency. 2001 2002 ctr = _SpinDuration; 2003 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 2004 if (ctr <= 0) return 0; 2005 2006 if (Knob_SuccRestrict && _succ != NULL) return 0; 2007 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2008 TEVENT(Spin abort - notrunnable [TOP]); 2009 return 0; 2010 } 2011 2012 int MaxSpin = Knob_MaxSpinners; 2013 if (MaxSpin >= 0) { 2014 if (_Spinner > MaxSpin) { 2015 TEVENT(Spin abort -- too many spinners); 2016 return 0; 2017 } 2018 // Slightly racy, but benign ... 2019 Adjust(&_Spinner, 1); 2020 } 2021 2022 // We're good to spin ... spin ingress. 2023 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2024 // when preparing to LD...CAS _owner, etc and the CAS is likely 2025 // to succeed. 2026 int hits = 0; 2027 int msk = 0; 2028 int caspty = Knob_CASPenalty; 2029 int oxpty = Knob_OXPenalty; 2030 int sss = Knob_SpinSetSucc; 2031 if (sss && _succ == NULL) _succ = Self; 2032 Thread * prv = NULL; 2033 2034 // There are three ways to exit the following loop: 2035 // 1. A successful spin where this thread has acquired the lock. 2036 // 2. Spin failure with prejudice 2037 // 3. Spin failure without prejudice 2038 2039 while (--ctr >= 0) { 2040 2041 // Periodic polling -- Check for pending GC 2042 // Threads may spin while they're unsafe. 2043 // We don't want spinning threads to delay the JVM from reaching 2044 // a stop-the-world safepoint or to steal cycles from GC. 2045 // If we detect a pending safepoint we abort in order that 2046 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2047 // this thread, if safe, doesn't steal cycles from GC. 2048 // This is in keeping with the "no loitering in runtime" rule. 2049 // We periodically check to see if there's a safepoint pending. 2050 if ((ctr & 0xFF) == 0) { 2051 if (SafepointSynchronize::do_call_back()) { 2052 TEVENT(Spin: safepoint); 2053 goto Abort; // abrupt spin egress 2054 } 2055 if (Knob_UsePause & 1) SpinPause(); 2056 2057 int (*scb)(intptr_t,int) = SpinCallbackFunction; 2058 if (hits > 50 && scb != NULL) { 2059 int abend = (*scb)(SpinCallbackArgument, 0); 2060 } 2061 } 2062 2063 if (Knob_UsePause & 2) SpinPause(); 2064 2065 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2066 // This is useful on classic SMP systems, but is of less utility on 2067 // N1-style CMT platforms. 2068 // 2069 // Trade-off: lock acquisition latency vs coherency bandwidth. 2070 // Lock hold times are typically short. A histogram 2071 // of successful spin attempts shows that we usually acquire 2072 // the lock early in the spin. That suggests we want to 2073 // sample _owner frequently in the early phase of the spin, 2074 // but then back-off and sample less frequently as the spin 2075 // progresses. The back-off makes a good citizen on SMP big 2076 // SMP systems. Oversampling _owner can consume excessive 2077 // coherency bandwidth. Relatedly, if we _oversample _owner we 2078 // can inadvertently interfere with the the ST m->owner=null. 2079 // executed by the lock owner. 2080 if (ctr & msk) continue; 2081 ++hits; 2082 if ((hits & 0xF) == 0) { 2083 // The 0xF, above, corresponds to the exponent. 2084 // Consider: (msk+1)|msk 2085 msk = ((msk << 2)|3) & BackOffMask; 2086 } 2087 2088 // Probe _owner with TATAS 2089 // If this thread observes the monitor transition or flicker 2090 // from locked to unlocked to locked, then the odds that this 2091 // thread will acquire the lock in this spin attempt go down 2092 // considerably. The same argument applies if the CAS fails 2093 // or if we observe _owner change from one non-null value to 2094 // another non-null value. In such cases we might abort 2095 // the spin without prejudice or apply a "penalty" to the 2096 // spin count-down variable "ctr", reducing it by 100, say. 2097 2098 Thread * ox = (Thread *) _owner; 2099 if (ox == NULL) { 2100 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 2101 if (ox == NULL) { 2102 // The CAS succeeded -- this thread acquired ownership 2103 // Take care of some bookkeeping to exit spin state. 2104 if (sss && _succ == Self) { 2105 _succ = NULL; 2106 } 2107 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2108 2109 // Increase _SpinDuration : 2110 // The spin was successful (profitable) so we tend toward 2111 // longer spin attempts in the future. 2112 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2113 // If we acquired the lock early in the spin cycle it 2114 // makes sense to increase _SpinDuration proportionally. 2115 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2116 int x = _SpinDuration; 2117 if (x < Knob_SpinLimit) { 2118 if (x < Knob_Poverty) x = Knob_Poverty; 2119 _SpinDuration = x + Knob_Bonus; 2120 } 2121 return 1; 2122 } 2123 2124 // The CAS failed ... we can take any of the following actions: 2125 // * penalize: ctr -= Knob_CASPenalty 2126 // * exit spin with prejudice -- goto Abort; 2127 // * exit spin without prejudice. 2128 // * Since CAS is high-latency, retry again immediately. 2129 prv = ox; 2130 TEVENT(Spin: cas failed); 2131 if (caspty == -2) break; 2132 if (caspty == -1) goto Abort; 2133 ctr -= caspty; 2134 continue; 2135 } 2136 2137 // Did lock ownership change hands ? 2138 if (ox != prv && prv != NULL) { 2139 TEVENT(spin: Owner changed) 2140 if (oxpty == -2) break; 2141 if (oxpty == -1) goto Abort; 2142 ctr -= oxpty; 2143 } 2144 prv = ox; 2145 2146 // Abort the spin if the owner is not executing. 2147 // The owner must be executing in order to drop the lock. 2148 // Spinning while the owner is OFFPROC is idiocy. 2149 // Consider: ctr -= RunnablePenalty ; 2150 if (Knob_OState && NotRunnable (Self, ox)) { 2151 TEVENT(Spin abort - notrunnable); 2152 goto Abort; 2153 } 2154 if (sss && _succ == NULL) _succ = Self; 2155 } 2156 2157 // Spin failed with prejudice -- reduce _SpinDuration. 2158 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2159 // AIMD is globally stable. 2160 TEVENT(Spin failure); 2161 { 2162 int x = _SpinDuration; 2163 if (x > 0) { 2164 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2165 // This is globally sample and tends to damp the response. 2166 x -= Knob_Penalty; 2167 if (x < 0) x = 0; 2168 _SpinDuration = x; 2169 } 2170 } 2171 2172 Abort: 2173 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2174 if (sss && _succ == Self) { 2175 _succ = NULL; 2176 // Invariant: after setting succ=null a contending thread 2177 // must recheck-retry _owner before parking. This usually happens 2178 // in the normal usage of TrySpin(), but it's safest 2179 // to make TrySpin() as foolproof as possible. 2180 OrderAccess::fence(); 2181 if (TryLock(Self) > 0) return 1; 2182 } 2183 return 0; 2184 } 2185 2186 // NotRunnable() -- informed spinning 2187 // 2188 // Don't bother spinning if the owner is not eligible to drop the lock. 2189 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2190 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2191 // The thread must be runnable in order to drop the lock in timely fashion. 2192 // If the _owner is not runnable then spinning will not likely be 2193 // successful (profitable). 2194 // 2195 // Beware -- the thread referenced by _owner could have died 2196 // so a simply fetch from _owner->_thread_state might trap. 2197 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2198 // Because of the lifecycle issues the schedctl and _thread_state values 2199 // observed by NotRunnable() might be garbage. NotRunnable must 2200 // tolerate this and consider the observed _thread_state value 2201 // as advisory. 2202 // 2203 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2204 // a thread pointer. 2205 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2206 // with the LSB of _owner. Another option would be to probablistically probe 2207 // the putative _owner->TypeTag value. 2208 // 2209 // Checking _thread_state isn't perfect. Even if the thread is 2210 // in_java it might be blocked on a page-fault or have been preempted 2211 // and sitting on a ready/dispatch queue. _thread state in conjunction 2212 // with schedctl.sc_state gives us a good picture of what the 2213 // thread is doing, however. 2214 // 2215 // TODO: check schedctl.sc_state. 2216 // We'll need to use SafeFetch32() to read from the schedctl block. 2217 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2218 // 2219 // The return value from NotRunnable() is *advisory* -- the 2220 // result is based on sampling and is not necessarily coherent. 2221 // The caller must tolerate false-negative and false-positive errors. 2222 // Spinning, in general, is probabilistic anyway. 2223 2224 2225 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2226 // Check ox->TypeTag == 2BAD. 2227 if (ox == NULL) return 0; 2228 2229 // Avoid transitive spinning ... 2230 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2231 // Immediately after T1 acquires L it's possible that T2, also 2232 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2233 // This occurs transiently after T1 acquired L but before 2234 // T1 managed to clear T1.Stalled. T2 does not need to abort 2235 // its spin in this circumstance. 2236 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2237 2238 if (BlockedOn == 1) return 1; 2239 if (BlockedOn != 0) { 2240 return BlockedOn != intptr_t(this) && _owner == ox; 2241 } 2242 2243 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2244 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2245 // consider also: jst != _thread_in_Java -- but that's overspecific. 2246 return jst == _thread_blocked || jst == _thread_in_native; 2247 } 2248 2249 2250 // ----------------------------------------------------------------------------- 2251 // WaitSet management ... 2252 2253 ObjectWaiter::ObjectWaiter(Thread* thread) { 2254 _next = NULL; 2255 _prev = NULL; 2256 _notified = 0; 2257 TState = TS_RUN; 2258 _thread = thread; 2259 _event = thread->_ParkEvent; 2260 _active = false; 2261 assert(_event != NULL, "invariant"); 2262 } 2263 2264 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2265 JavaThread *jt = (JavaThread *)this->_thread; 2266 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2267 } 2268 2269 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2270 JavaThread *jt = (JavaThread *)this->_thread; 2271 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2272 } 2273 2274 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2275 assert(node != NULL, "should not dequeue NULL node"); 2276 assert(node->_prev == NULL, "node already in list"); 2277 assert(node->_next == NULL, "node already in list"); 2278 // put node at end of queue (circular doubly linked list) 2279 if (_WaitSet == NULL) { 2280 _WaitSet = node; 2281 node->_prev = node; 2282 node->_next = node; 2283 } else { 2284 ObjectWaiter* head = _WaitSet; 2285 ObjectWaiter* tail = head->_prev; 2286 assert(tail->_next == head, "invariant check"); 2287 tail->_next = node; 2288 head->_prev = node; 2289 node->_next = head; 2290 node->_prev = tail; 2291 } 2292 } 2293 2294 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2295 // dequeue the very first waiter 2296 ObjectWaiter* waiter = _WaitSet; 2297 if (waiter) { 2298 DequeueSpecificWaiter(waiter); 2299 } 2300 return waiter; 2301 } 2302 2303 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2304 assert(node != NULL, "should not dequeue NULL node"); 2305 assert(node->_prev != NULL, "node already removed from list"); 2306 assert(node->_next != NULL, "node already removed from list"); 2307 // when the waiter has woken up because of interrupt, 2308 // timeout or other spurious wake-up, dequeue the 2309 // waiter from waiting list 2310 ObjectWaiter* next = node->_next; 2311 if (next == node) { 2312 assert(node->_prev == node, "invariant check"); 2313 _WaitSet = NULL; 2314 } else { 2315 ObjectWaiter* prev = node->_prev; 2316 assert(prev->_next == node, "invariant check"); 2317 assert(next->_prev == node, "invariant check"); 2318 next->_prev = prev; 2319 prev->_next = next; 2320 if (_WaitSet == node) { 2321 _WaitSet = next; 2322 } 2323 } 2324 node->_next = NULL; 2325 node->_prev = NULL; 2326 } 2327 2328 // ----------------------------------------------------------------------------- 2329 // PerfData support 2330 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2331 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2332 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2333 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2334 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2335 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2336 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2337 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2338 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2339 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2340 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2341 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2342 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2343 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2344 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2345 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2346 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2347 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2348 2349 // One-shot global initialization for the sync subsystem. 2350 // We could also defer initialization and initialize on-demand 2351 // the first time we call inflate(). Initialization would 2352 // be protected - like so many things - by the MonitorCache_lock. 2353 2354 void ObjectMonitor::Initialize() { 2355 static int InitializationCompleted = 0; 2356 assert(InitializationCompleted == 0, "invariant"); 2357 InitializationCompleted = 1; 2358 if (UsePerfData) { 2359 EXCEPTION_MARK; 2360 #define NEWPERFCOUNTER(n) \ 2361 { \ 2362 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2363 CHECK); \ 2364 } 2365 #define NEWPERFVARIABLE(n) \ 2366 { \ 2367 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2368 CHECK); \ 2369 } 2370 NEWPERFCOUNTER(_sync_Inflations); 2371 NEWPERFCOUNTER(_sync_Deflations); 2372 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2373 NEWPERFCOUNTER(_sync_FutileWakeups); 2374 NEWPERFCOUNTER(_sync_Parks); 2375 NEWPERFCOUNTER(_sync_EmptyNotifications); 2376 NEWPERFCOUNTER(_sync_Notifications); 2377 NEWPERFCOUNTER(_sync_SlowEnter); 2378 NEWPERFCOUNTER(_sync_SlowExit); 2379 NEWPERFCOUNTER(_sync_SlowNotify); 2380 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2381 NEWPERFCOUNTER(_sync_FailedSpins); 2382 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2383 NEWPERFCOUNTER(_sync_PrivateA); 2384 NEWPERFCOUNTER(_sync_PrivateB); 2385 NEWPERFCOUNTER(_sync_MonInCirculation); 2386 NEWPERFCOUNTER(_sync_MonScavenged); 2387 NEWPERFVARIABLE(_sync_MonExtant); 2388 #undef NEWPERFCOUNTER 2389 #undef NEWPERFVARIABLE 2390 } 2391 } 2392 2393 static char * kvGet(char * kvList, const char * Key) { 2394 if (kvList == NULL) return NULL; 2395 size_t n = strlen(Key); 2396 char * Search; 2397 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2398 if (strncmp (Search, Key, n) == 0) { 2399 if (Search[n] == '=') return Search + n + 1; 2400 if (Search[n] == 0) return(char *) "1"; 2401 } 2402 } 2403 return NULL; 2404 } 2405 2406 static int kvGetInt(char * kvList, const char * Key, int Default) { 2407 char * v = kvGet(kvList, Key); 2408 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2409 if (Knob_ReportSettings && v != NULL) { 2410 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2411 ::fflush(stdout); 2412 } 2413 return rslt; 2414 } 2415 2416 void ObjectMonitor::DeferredInitialize() { 2417 if (InitDone > 0) return; 2418 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2419 while (InitDone != 1) /* empty */; 2420 return; 2421 } 2422 2423 // One-shot global initialization ... 2424 // The initialization is idempotent, so we don't need locks. 2425 // In the future consider doing this via os::init_2(). 2426 // SyncKnobs consist of <Key>=<Value> pairs in the style 2427 // of environment variables. Start by converting ':' to NUL. 2428 2429 if (SyncKnobs == NULL) SyncKnobs = ""; 2430 2431 size_t sz = strlen(SyncKnobs); 2432 char * knobs = (char *) malloc(sz + 2); 2433 if (knobs == NULL) { 2434 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2435 guarantee(0, "invariant"); 2436 } 2437 strcpy(knobs, SyncKnobs); 2438 knobs[sz+1] = 0; 2439 for (char * p = knobs; *p; p++) { 2440 if (*p == ':') *p = 0; 2441 } 2442 2443 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2444 SETKNOB(ReportSettings); 2445 SETKNOB(Verbose); 2446 SETKNOB(VerifyInUse); 2447 SETKNOB(FixedSpin); 2448 SETKNOB(SpinLimit); 2449 SETKNOB(SpinBase); 2450 SETKNOB(SpinBackOff); 2451 SETKNOB(CASPenalty); 2452 SETKNOB(OXPenalty); 2453 SETKNOB(LogSpins); 2454 SETKNOB(SpinSetSucc); 2455 SETKNOB(SuccEnabled); 2456 SETKNOB(SuccRestrict); 2457 SETKNOB(Penalty); 2458 SETKNOB(Bonus); 2459 SETKNOB(BonusB); 2460 SETKNOB(Poverty); 2461 SETKNOB(SpinAfterFutile); 2462 SETKNOB(UsePause); 2463 SETKNOB(SpinEarly); 2464 SETKNOB(OState); 2465 SETKNOB(MaxSpinners); 2466 SETKNOB(PreSpin); 2467 SETKNOB(ExitPolicy); 2468 SETKNOB(QMode); 2469 SETKNOB(ResetEvent); 2470 SETKNOB(MoveNotifyee); 2471 SETKNOB(FastHSSEC); 2472 #undef SETKNOB 2473 2474 if (Knob_Verbose) { 2475 sanity_checks(); 2476 } 2477 2478 if (os::is_MP()) { 2479 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2480 if (Knob_ReportSettings) ::printf("BackOffMask=%X\n", BackOffMask); 2481 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2482 } else { 2483 Knob_SpinLimit = 0; 2484 Knob_SpinBase = 0; 2485 Knob_PreSpin = 0; 2486 Knob_FixedSpin = -1; 2487 } 2488 2489 if (Knob_LogSpins == 0) { 2490 ObjectMonitor::_sync_FailedSpins = NULL; 2491 } 2492 2493 free(knobs); 2494 OrderAccess::fence(); 2495 InitDone = 1; 2496 } 2497 2498 void ObjectMonitor::sanity_checks() { 2499 int error_cnt = 0; 2500 int warning_cnt = 0; 2501 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2502 2503 if (verbose) { 2504 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2505 sizeof(ObjectMonitor)); 2506 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT, 2507 sizeof(PaddedEnd<ObjectMonitor>)); 2508 } 2509 2510 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2511 if (verbose) { 2512 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2513 } 2514 2515 ObjectMonitor dummy; 2516 u_char *addr_begin = (u_char*)&dummy; 2517 u_char *addr_header = (u_char*)&dummy._header; 2518 u_char *addr_owner = (u_char*)&dummy._owner; 2519 2520 uint offset_header = (uint)(addr_header - addr_begin); 2521 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2522 2523 uint offset_owner = (uint)(addr_owner - addr_begin); 2524 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2525 2526 if ((uint)(addr_header - addr_begin) != 0) { 2527 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2528 error_cnt++; 2529 } 2530 2531 if (cache_line_size != 0) { 2532 // We were able to determine the L1 data cache line size so 2533 // do some cache line specific sanity checks 2534 2535 if ((offset_owner - offset_header) < cache_line_size) { 2536 tty->print_cr("WARNING: the _header and _owner fields are closer " 2537 "than a cache line which permits false sharing."); 2538 warning_cnt++; 2539 } 2540 2541 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) { 2542 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a " 2543 "multiple of a cache line which permits false sharing."); 2544 warning_cnt++; 2545 } 2546 } 2547 2548 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2549 &warning_cnt); 2550 2551 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2552 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2553 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2554 } 2555 2556 guarantee(error_cnt == 0, 2557 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2558 } 2559 2560 #ifndef PRODUCT 2561 void ObjectMonitor::verify() { 2562 } 2563 2564 void ObjectMonitor::print() { 2565 } 2566 #endif