1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "jfr/support/jfrThreadId.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "memory/resourceArea.hpp" 31 #include "oops/markOop.hpp" 32 #include "oops/oop.inline.hpp" 33 #include "runtime/atomic.hpp" 34 #include "runtime/handles.inline.hpp" 35 #include "runtime/interfaceSupport.inline.hpp" 36 #include "runtime/mutexLocker.hpp" 37 #include "runtime/objectMonitor.hpp" 38 #include "runtime/objectMonitor.inline.hpp" 39 #include "runtime/orderAccess.hpp" 40 #include "runtime/osThread.hpp" 41 #include "runtime/safepointMechanism.inline.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 #include "runtime/stubRoutines.hpp" 44 #include "runtime/thread.inline.hpp" 45 #include "services/threadService.hpp" 46 #include "utilities/dtrace.hpp" 47 #include "utilities/macros.hpp" 48 #include "utilities/preserveException.hpp" 49 #if INCLUDE_JFR 50 #include "jfr/support/jfrFlush.hpp" 51 #endif 52 53 #ifdef DTRACE_ENABLED 54 55 // Only bother with this argument setup if dtrace is available 56 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 57 58 59 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 60 char* bytes = NULL; \ 61 int len = 0; \ 62 jlong jtid = SharedRuntime::get_java_tid(thread); \ 63 Symbol* klassname = ((oop)obj)->klass()->name(); \ 64 if (klassname != NULL) { \ 65 bytes = (char*)klassname->bytes(); \ 66 len = klassname->utf8_length(); \ 67 } 68 69 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 70 { \ 71 if (DTraceMonitorProbes) { \ 72 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 73 HOTSPOT_MONITOR_WAIT(jtid, \ 74 (monitor), bytes, len, (millis)); \ 75 } \ 76 } 77 78 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 79 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 80 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 81 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 82 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 83 84 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 85 { \ 86 if (DTraceMonitorProbes) { \ 87 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 88 HOTSPOT_MONITOR_##probe(jtid, \ 89 (uintptr_t)(monitor), bytes, len); \ 90 } \ 91 } 92 93 #else // ndef DTRACE_ENABLED 94 95 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 96 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 97 98 #endif // ndef DTRACE_ENABLED 99 100 // Tunables ... 101 // The knob* variables are effectively final. Once set they should 102 // never be modified hence. Consider using __read_mostly with GCC. 103 104 int ObjectMonitor::Knob_ExitRelease = 0; 105 int ObjectMonitor::Knob_InlineNotify = 1; 106 int ObjectMonitor::Knob_Verbose = 0; 107 int ObjectMonitor::Knob_VerifyInUse = 0; 108 int ObjectMonitor::Knob_VerifyMatch = 0; 109 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 110 111 static int Knob_SpinBase = 0; // Floor AKA SpinMin 112 static int Knob_CASPenalty = -1; // Penalty for failed CAS 113 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 114 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 115 static int Knob_SpinEarly = 1; 116 static int Knob_SuccEnabled = 1; // futile wake throttling 117 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 118 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 119 static int Knob_Bonus = 100; // spin success bonus 120 static int Knob_BonusB = 100; // spin success bonus 121 static int Knob_Penalty = 200; // spin failure penalty 122 static int Knob_Poverty = 1000; 123 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 124 static int Knob_FixedSpin = 0; 125 static int Knob_OState = 3; // Spinner checks thread state of _owner 126 static int Knob_UsePause = 1; 127 static int Knob_ExitPolicy = 0; 128 static int Knob_PreSpin = 10; // 20-100 likely better 129 static int Knob_ResetEvent = 0; 130 static int BackOffMask = 0; 131 132 static int Knob_FastHSSEC = 0; 133 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 134 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 135 static volatile int InitDone = 0; 136 137 // ----------------------------------------------------------------------------- 138 // Theory of operations -- Monitors lists, thread residency, etc: 139 // 140 // * A thread acquires ownership of a monitor by successfully 141 // CAS()ing the _owner field from null to non-null. 142 // 143 // * Invariant: A thread appears on at most one monitor list -- 144 // cxq, EntryList or WaitSet -- at any one time. 145 // 146 // * Contending threads "push" themselves onto the cxq with CAS 147 // and then spin/park. 148 // 149 // * After a contending thread eventually acquires the lock it must 150 // dequeue itself from either the EntryList or the cxq. 151 // 152 // * The exiting thread identifies and unparks an "heir presumptive" 153 // tentative successor thread on the EntryList. Critically, the 154 // exiting thread doesn't unlink the successor thread from the EntryList. 155 // After having been unparked, the wakee will recontend for ownership of 156 // the monitor. The successor (wakee) will either acquire the lock or 157 // re-park itself. 158 // 159 // Succession is provided for by a policy of competitive handoff. 160 // The exiting thread does _not_ grant or pass ownership to the 161 // successor thread. (This is also referred to as "handoff" succession"). 162 // Instead the exiting thread releases ownership and possibly wakes 163 // a successor, so the successor can (re)compete for ownership of the lock. 164 // If the EntryList is empty but the cxq is populated the exiting 165 // thread will drain the cxq into the EntryList. It does so by 166 // by detaching the cxq (installing null with CAS) and folding 167 // the threads from the cxq into the EntryList. The EntryList is 168 // doubly linked, while the cxq is singly linked because of the 169 // CAS-based "push" used to enqueue recently arrived threads (RATs). 170 // 171 // * Concurrency invariants: 172 // 173 // -- only the monitor owner may access or mutate the EntryList. 174 // The mutex property of the monitor itself protects the EntryList 175 // from concurrent interference. 176 // -- Only the monitor owner may detach the cxq. 177 // 178 // * The monitor entry list operations avoid locks, but strictly speaking 179 // they're not lock-free. Enter is lock-free, exit is not. 180 // For a description of 'Methods and apparatus providing non-blocking access 181 // to a resource,' see U.S. Pat. No. 7844973. 182 // 183 // * The cxq can have multiple concurrent "pushers" but only one concurrent 184 // detaching thread. This mechanism is immune from the ABA corruption. 185 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 186 // 187 // * Taken together, the cxq and the EntryList constitute or form a 188 // single logical queue of threads stalled trying to acquire the lock. 189 // We use two distinct lists to improve the odds of a constant-time 190 // dequeue operation after acquisition (in the ::enter() epilogue) and 191 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 192 // A key desideratum is to minimize queue & monitor metadata manipulation 193 // that occurs while holding the monitor lock -- that is, we want to 194 // minimize monitor lock holds times. Note that even a small amount of 195 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 196 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 197 // locks and monitor metadata. 198 // 199 // Cxq points to the set of Recently Arrived Threads attempting entry. 200 // Because we push threads onto _cxq with CAS, the RATs must take the form of 201 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 202 // the unlocking thread notices that EntryList is null but _cxq is != null. 203 // 204 // The EntryList is ordered by the prevailing queue discipline and 205 // can be organized in any convenient fashion, such as a doubly-linked list or 206 // a circular doubly-linked list. Critically, we want insert and delete operations 207 // to operate in constant-time. If we need a priority queue then something akin 208 // to Solaris' sleepq would work nicely. Viz., 209 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 210 // Queue discipline is enforced at ::exit() time, when the unlocking thread 211 // drains the cxq into the EntryList, and orders or reorders the threads on the 212 // EntryList accordingly. 213 // 214 // Barring "lock barging", this mechanism provides fair cyclic ordering, 215 // somewhat similar to an elevator-scan. 216 // 217 // * The monitor synchronization subsystem avoids the use of native 218 // synchronization primitives except for the narrow platform-specific 219 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 220 // the semantics of park-unpark. Put another way, this monitor implementation 221 // depends only on atomic operations and park-unpark. The monitor subsystem 222 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 223 // underlying OS manages the READY<->RUN transitions. 224 // 225 // * Waiting threads reside on the WaitSet list -- wait() puts 226 // the caller onto the WaitSet. 227 // 228 // * notify() or notifyAll() simply transfers threads from the WaitSet to 229 // either the EntryList or cxq. Subsequent exit() operations will 230 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 231 // it's likely the notifyee would simply impale itself on the lock held 232 // by the notifier. 233 // 234 // * An interesting alternative is to encode cxq as (List,LockByte) where 235 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 236 // variable, like _recursions, in the scheme. The threads or Events that form 237 // the list would have to be aligned in 256-byte addresses. A thread would 238 // try to acquire the lock or enqueue itself with CAS, but exiting threads 239 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 240 // Note that is is *not* word-tearing, but it does presume that full-word 241 // CAS operations are coherent with intermix with STB operations. That's true 242 // on most common processors. 243 // 244 // * See also http://blogs.sun.com/dave 245 246 247 void* ObjectMonitor::operator new (size_t size) throw() { 248 return AllocateHeap(size, mtInternal); 249 } 250 void* ObjectMonitor::operator new[] (size_t size) throw() { 251 return operator new (size); 252 } 253 void ObjectMonitor::operator delete(void* p) { 254 FreeHeap(p); 255 } 256 void ObjectMonitor::operator delete[] (void *p) { 257 operator delete(p); 258 } 259 260 // ----------------------------------------------------------------------------- 261 // Enter support 262 263 void ObjectMonitor::enter(TRAPS) { 264 // The following code is ordered to check the most common cases first 265 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 266 Thread * const Self = THREAD; 267 268 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 269 if (cur == NULL) { 270 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 271 assert(_recursions == 0, "invariant"); 272 assert(_owner == Self, "invariant"); 273 return; 274 } 275 276 if (cur == Self) { 277 // TODO-FIXME: check for integer overflow! BUGID 6557169. 278 _recursions++; 279 return; 280 } 281 282 if (Self->is_lock_owned ((address)cur)) { 283 assert(_recursions == 0, "internal state error"); 284 _recursions = 1; 285 // Commute owner from a thread-specific on-stack BasicLockObject address to 286 // a full-fledged "Thread *". 287 _owner = Self; 288 return; 289 } 290 291 // We've encountered genuine contention. 292 assert(Self->_Stalled == 0, "invariant"); 293 Self->_Stalled = intptr_t(this); 294 295 // Try one round of spinning *before* enqueueing Self 296 // and before going through the awkward and expensive state 297 // transitions. The following spin is strictly optional ... 298 // Note that if we acquire the monitor from an initial spin 299 // we forgo posting JVMTI events and firing DTRACE probes. 300 if (Knob_SpinEarly && TrySpin (Self) > 0) { 301 assert(_owner == Self, "invariant"); 302 assert(_recursions == 0, "invariant"); 303 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 304 Self->_Stalled = 0; 305 return; 306 } 307 308 assert(_owner != Self, "invariant"); 309 assert(_succ != Self, "invariant"); 310 assert(Self->is_Java_thread(), "invariant"); 311 JavaThread * jt = (JavaThread *) Self; 312 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 313 assert(jt->thread_state() != _thread_blocked, "invariant"); 314 assert(this->object() != NULL, "invariant"); 315 assert(_count >= 0, "invariant"); 316 317 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 318 // Ensure the object-monitor relationship remains stable while there's contention. 319 Atomic::inc(&_count); 320 321 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 322 EventJavaMonitorEnter event; 323 if (event.should_commit()) { 324 event.set_monitorClass(((oop)this->object())->klass()); 325 event.set_address((uintptr_t)(this->object_addr())); 326 } 327 328 { // Change java thread status to indicate blocked on monitor enter. 329 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 330 331 Self->set_current_pending_monitor(this); 332 333 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 334 if (JvmtiExport::should_post_monitor_contended_enter()) { 335 JvmtiExport::post_monitor_contended_enter(jt, this); 336 337 // The current thread does not yet own the monitor and does not 338 // yet appear on any queues that would get it made the successor. 339 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 340 // handler cannot accidentally consume an unpark() meant for the 341 // ParkEvent associated with this ObjectMonitor. 342 } 343 344 OSThreadContendState osts(Self->osthread()); 345 ThreadBlockInVM tbivm(jt); 346 347 // TODO-FIXME: change the following for(;;) loop to straight-line code. 348 for (;;) { 349 jt->set_suspend_equivalent(); 350 // cleared by handle_special_suspend_equivalent_condition() 351 // or java_suspend_self() 352 353 EnterI(THREAD); 354 355 if (!ExitSuspendEquivalent(jt)) break; 356 357 // We have acquired the contended monitor, but while we were 358 // waiting another thread suspended us. We don't want to enter 359 // the monitor while suspended because that would surprise the 360 // thread that suspended us. 361 // 362 _recursions = 0; 363 _succ = NULL; 364 exit(false, Self); 365 366 jt->java_suspend_self(); 367 } 368 Self->set_current_pending_monitor(NULL); 369 370 // We cleared the pending monitor info since we've just gotten past 371 // the enter-check-for-suspend dance and we now own the monitor free 372 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 373 // destructor can go to a safepoint at the end of this block. If we 374 // do a thread dump during that safepoint, then this thread will show 375 // as having "-locked" the monitor, but the OS and java.lang.Thread 376 // states will still report that the thread is blocked trying to 377 // acquire it. 378 } 379 380 Atomic::dec(&_count); 381 assert(_count >= 0, "invariant"); 382 Self->_Stalled = 0; 383 384 // Must either set _recursions = 0 or ASSERT _recursions == 0. 385 assert(_recursions == 0, "invariant"); 386 assert(_owner == Self, "invariant"); 387 assert(_succ != Self, "invariant"); 388 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 389 390 // The thread -- now the owner -- is back in vm mode. 391 // Report the glorious news via TI,DTrace and jvmstat. 392 // The probe effect is non-trivial. All the reportage occurs 393 // while we hold the monitor, increasing the length of the critical 394 // section. Amdahl's parallel speedup law comes vividly into play. 395 // 396 // Another option might be to aggregate the events (thread local or 397 // per-monitor aggregation) and defer reporting until a more opportune 398 // time -- such as next time some thread encounters contention but has 399 // yet to acquire the lock. While spinning that thread could 400 // spinning we could increment JVMStat counters, etc. 401 402 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 403 if (JvmtiExport::should_post_monitor_contended_entered()) { 404 JvmtiExport::post_monitor_contended_entered(jt, this); 405 406 // The current thread already owns the monitor and is not going to 407 // call park() for the remainder of the monitor enter protocol. So 408 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 409 // event handler consumed an unpark() issued by the thread that 410 // just exited the monitor. 411 } 412 if (event.should_commit()) { 413 event.set_previousOwner((uintptr_t)_previous_owner_tid); 414 event.commit(); 415 } 416 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 417 } 418 419 // Caveat: TryLock() is not necessarily serializing if it returns failure. 420 // Callers must compensate as needed. 421 422 int ObjectMonitor::TryLock(Thread * Self) { 423 void * own = _owner; 424 if (own != NULL) return 0; 425 if (Atomic::replace_if_null(Self, &_owner)) { 426 // Either guarantee _recursions == 0 or set _recursions = 0. 427 assert(_recursions == 0, "invariant"); 428 assert(_owner == Self, "invariant"); 429 return 1; 430 } 431 // The lock had been free momentarily, but we lost the race to the lock. 432 // Interference -- the CAS failed. 433 // We can either return -1 or retry. 434 // Retry doesn't make as much sense because the lock was just acquired. 435 return -1; 436 } 437 438 #define MAX_RECHECK_INTERVAL 1000 439 440 void ObjectMonitor::EnterI(TRAPS) { 441 Thread * const Self = THREAD; 442 assert(Self->is_Java_thread(), "invariant"); 443 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 444 445 // Try the lock - TATAS 446 if (TryLock (Self) > 0) { 447 assert(_succ != Self, "invariant"); 448 assert(_owner == Self, "invariant"); 449 assert(_Responsible != Self, "invariant"); 450 return; 451 } 452 453 DeferredInitialize(); 454 455 // We try one round of spinning *before* enqueueing Self. 456 // 457 // If the _owner is ready but OFFPROC we could use a YieldTo() 458 // operation to donate the remainder of this thread's quantum 459 // to the owner. This has subtle but beneficial affinity 460 // effects. 461 462 if (TrySpin (Self) > 0) { 463 assert(_owner == Self, "invariant"); 464 assert(_succ != Self, "invariant"); 465 assert(_Responsible != Self, "invariant"); 466 return; 467 } 468 469 // The Spin failed -- Enqueue and park the thread ... 470 assert(_succ != Self, "invariant"); 471 assert(_owner != Self, "invariant"); 472 assert(_Responsible != Self, "invariant"); 473 474 // Enqueue "Self" on ObjectMonitor's _cxq. 475 // 476 // Node acts as a proxy for Self. 477 // As an aside, if were to ever rewrite the synchronization code mostly 478 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 479 // Java objects. This would avoid awkward lifecycle and liveness issues, 480 // as well as eliminate a subset of ABA issues. 481 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 482 483 ObjectWaiter node(Self); 484 Self->_ParkEvent->reset(); 485 node._prev = (ObjectWaiter *) 0xBAD; 486 node.TState = ObjectWaiter::TS_CXQ; 487 488 // Push "Self" onto the front of the _cxq. 489 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 490 // Note that spinning tends to reduce the rate at which threads 491 // enqueue and dequeue on EntryList|cxq. 492 ObjectWaiter * nxt; 493 for (;;) { 494 node._next = nxt = _cxq; 495 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 496 497 // Interference - the CAS failed because _cxq changed. Just retry. 498 // As an optional optimization we retry the lock. 499 if (TryLock (Self) > 0) { 500 assert(_succ != Self, "invariant"); 501 assert(_owner == Self, "invariant"); 502 assert(_Responsible != Self, "invariant"); 503 return; 504 } 505 } 506 507 // Check for cxq|EntryList edge transition to non-null. This indicates 508 // the onset of contention. While contention persists exiting threads 509 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 510 // operations revert to the faster 1-0 mode. This enter operation may interleave 511 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 512 // arrange for one of the contending thread to use a timed park() operations 513 // to detect and recover from the race. (Stranding is form of progress failure 514 // where the monitor is unlocked but all the contending threads remain parked). 515 // That is, at least one of the contended threads will periodically poll _owner. 516 // One of the contending threads will become the designated "Responsible" thread. 517 // The Responsible thread uses a timed park instead of a normal indefinite park 518 // operation -- it periodically wakes and checks for and recovers from potential 519 // strandings admitted by 1-0 exit operations. We need at most one Responsible 520 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 521 // be responsible for a monitor. 522 // 523 // Currently, one of the contended threads takes on the added role of "Responsible". 524 // A viable alternative would be to use a dedicated "stranding checker" thread 525 // that periodically iterated over all the threads (or active monitors) and unparked 526 // successors where there was risk of stranding. This would help eliminate the 527 // timer scalability issues we see on some platforms as we'd only have one thread 528 // -- the checker -- parked on a timer. 529 530 if (nxt == NULL && _EntryList == NULL) { 531 // Try to assume the role of responsible thread for the monitor. 532 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 533 Atomic::replace_if_null(Self, &_Responsible); 534 } 535 536 // The lock might have been released while this thread was occupied queueing 537 // itself onto _cxq. To close the race and avoid "stranding" and 538 // progress-liveness failure we must resample-retry _owner before parking. 539 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 540 // In this case the ST-MEMBAR is accomplished with CAS(). 541 // 542 // TODO: Defer all thread state transitions until park-time. 543 // Since state transitions are heavy and inefficient we'd like 544 // to defer the state transitions until absolutely necessary, 545 // and in doing so avoid some transitions ... 546 547 int nWakeups = 0; 548 int recheckInterval = 1; 549 550 for (;;) { 551 552 if (TryLock(Self) > 0) break; 553 assert(_owner != Self, "invariant"); 554 555 // park self 556 if (_Responsible == Self) { 557 Self->_ParkEvent->park((jlong) recheckInterval); 558 // Increase the recheckInterval, but clamp the value. 559 recheckInterval *= 8; 560 if (recheckInterval > MAX_RECHECK_INTERVAL) { 561 recheckInterval = MAX_RECHECK_INTERVAL; 562 } 563 } else { 564 Self->_ParkEvent->park(); 565 } 566 567 if (TryLock(Self) > 0) break; 568 569 // The lock is still contested. 570 // Keep a tally of the # of futile wakeups. 571 // Note that the counter is not protected by a lock or updated by atomics. 572 // That is by design - we trade "lossy" counters which are exposed to 573 // races during updates for a lower probe effect. 574 575 // This PerfData object can be used in parallel with a safepoint. 576 // See the work around in PerfDataManager::destroy(). 577 OM_PERFDATA_OP(FutileWakeups, inc()); 578 ++nWakeups; 579 580 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 581 // We can defer clearing _succ until after the spin completes 582 // TrySpin() must tolerate being called with _succ == Self. 583 // Try yet another round of adaptive spinning. 584 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 585 586 // We can find that we were unpark()ed and redesignated _succ while 587 // we were spinning. That's harmless. If we iterate and call park(), 588 // park() will consume the event and return immediately and we'll 589 // just spin again. This pattern can repeat, leaving _succ to simply 590 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 591 // Alternately, we can sample fired() here, and if set, forgo spinning 592 // in the next iteration. 593 594 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 595 Self->_ParkEvent->reset(); 596 OrderAccess::fence(); 597 } 598 if (_succ == Self) _succ = NULL; 599 600 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 601 OrderAccess::fence(); 602 } 603 604 // Egress : 605 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 606 // Normally we'll find Self on the EntryList . 607 // From the perspective of the lock owner (this thread), the 608 // EntryList is stable and cxq is prepend-only. 609 // The head of cxq is volatile but the interior is stable. 610 // In addition, Self.TState is stable. 611 612 assert(_owner == Self, "invariant"); 613 assert(object() != NULL, "invariant"); 614 // I'd like to write: 615 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 616 // but as we're at a safepoint that's not safe. 617 618 UnlinkAfterAcquire(Self, &node); 619 if (_succ == Self) _succ = NULL; 620 621 assert(_succ != Self, "invariant"); 622 if (_Responsible == Self) { 623 _Responsible = NULL; 624 OrderAccess::fence(); // Dekker pivot-point 625 626 // We may leave threads on cxq|EntryList without a designated 627 // "Responsible" thread. This is benign. When this thread subsequently 628 // exits the monitor it can "see" such preexisting "old" threads -- 629 // threads that arrived on the cxq|EntryList before the fence, above -- 630 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 631 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 632 // non-null and elect a new "Responsible" timer thread. 633 // 634 // This thread executes: 635 // ST Responsible=null; MEMBAR (in enter epilogue - here) 636 // LD cxq|EntryList (in subsequent exit) 637 // 638 // Entering threads in the slow/contended path execute: 639 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 640 // The (ST cxq; MEMBAR) is accomplished with CAS(). 641 // 642 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 643 // exit operation from floating above the ST Responsible=null. 644 } 645 646 // We've acquired ownership with CAS(). 647 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 648 // But since the CAS() this thread may have also stored into _succ, 649 // EntryList, cxq or Responsible. These meta-data updates must be 650 // visible __before this thread subsequently drops the lock. 651 // Consider what could occur if we didn't enforce this constraint -- 652 // STs to monitor meta-data and user-data could reorder with (become 653 // visible after) the ST in exit that drops ownership of the lock. 654 // Some other thread could then acquire the lock, but observe inconsistent 655 // or old monitor meta-data and heap data. That violates the JMM. 656 // To that end, the 1-0 exit() operation must have at least STST|LDST 657 // "release" barrier semantics. Specifically, there must be at least a 658 // STST|LDST barrier in exit() before the ST of null into _owner that drops 659 // the lock. The barrier ensures that changes to monitor meta-data and data 660 // protected by the lock will be visible before we release the lock, and 661 // therefore before some other thread (CPU) has a chance to acquire the lock. 662 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 663 // 664 // Critically, any prior STs to _succ or EntryList must be visible before 665 // the ST of null into _owner in the *subsequent* (following) corresponding 666 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 667 // execute a serializing instruction. 668 669 return; 670 } 671 672 // ReenterI() is a specialized inline form of the latter half of the 673 // contended slow-path from EnterI(). We use ReenterI() only for 674 // monitor reentry in wait(). 675 // 676 // In the future we should reconcile EnterI() and ReenterI(), adding 677 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 678 // loop accordingly. 679 680 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 681 assert(Self != NULL, "invariant"); 682 assert(SelfNode != NULL, "invariant"); 683 assert(SelfNode->_thread == Self, "invariant"); 684 assert(_waiters > 0, "invariant"); 685 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 686 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 687 JavaThread * jt = (JavaThread *) Self; 688 689 int nWakeups = 0; 690 for (;;) { 691 ObjectWaiter::TStates v = SelfNode->TState; 692 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 693 assert(_owner != Self, "invariant"); 694 695 if (TryLock(Self) > 0) break; 696 if (TrySpin(Self) > 0) break; 697 698 // State transition wrappers around park() ... 699 // ReenterI() wisely defers state transitions until 700 // it's clear we must park the thread. 701 { 702 OSThreadContendState osts(Self->osthread()); 703 ThreadBlockInVM tbivm(jt); 704 705 // cleared by handle_special_suspend_equivalent_condition() 706 // or java_suspend_self() 707 jt->set_suspend_equivalent(); 708 Self->_ParkEvent->park(); 709 710 // were we externally suspended while we were waiting? 711 for (;;) { 712 if (!ExitSuspendEquivalent(jt)) break; 713 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 714 jt->java_suspend_self(); 715 jt->set_suspend_equivalent(); 716 } 717 } 718 719 // Try again, but just so we distinguish between futile wakeups and 720 // successful wakeups. The following test isn't algorithmically 721 // necessary, but it helps us maintain sensible statistics. 722 if (TryLock(Self) > 0) break; 723 724 // The lock is still contested. 725 // Keep a tally of the # of futile wakeups. 726 // Note that the counter is not protected by a lock or updated by atomics. 727 // That is by design - we trade "lossy" counters which are exposed to 728 // races during updates for a lower probe effect. 729 ++nWakeups; 730 731 // Assuming this is not a spurious wakeup we'll normally 732 // find that _succ == Self. 733 if (_succ == Self) _succ = NULL; 734 735 // Invariant: after clearing _succ a contending thread 736 // *must* retry _owner before parking. 737 OrderAccess::fence(); 738 739 // This PerfData object can be used in parallel with a safepoint. 740 // See the work around in PerfDataManager::destroy(). 741 OM_PERFDATA_OP(FutileWakeups, inc()); 742 } 743 744 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 745 // Normally we'll find Self on the EntryList. 746 // Unlinking from the EntryList is constant-time and atomic-free. 747 // From the perspective of the lock owner (this thread), the 748 // EntryList is stable and cxq is prepend-only. 749 // The head of cxq is volatile but the interior is stable. 750 // In addition, Self.TState is stable. 751 752 assert(_owner == Self, "invariant"); 753 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 754 UnlinkAfterAcquire(Self, SelfNode); 755 if (_succ == Self) _succ = NULL; 756 assert(_succ != Self, "invariant"); 757 SelfNode->TState = ObjectWaiter::TS_RUN; 758 OrderAccess::fence(); // see comments at the end of EnterI() 759 } 760 761 // By convention we unlink a contending thread from EntryList|cxq immediately 762 // after the thread acquires the lock in ::enter(). Equally, we could defer 763 // unlinking the thread until ::exit()-time. 764 765 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 766 assert(_owner == Self, "invariant"); 767 assert(SelfNode->_thread == Self, "invariant"); 768 769 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 770 // Normal case: remove Self from the DLL EntryList . 771 // This is a constant-time operation. 772 ObjectWaiter * nxt = SelfNode->_next; 773 ObjectWaiter * prv = SelfNode->_prev; 774 if (nxt != NULL) nxt->_prev = prv; 775 if (prv != NULL) prv->_next = nxt; 776 if (SelfNode == _EntryList) _EntryList = nxt; 777 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 778 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 779 } else { 780 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 781 // Inopportune interleaving -- Self is still on the cxq. 782 // This usually means the enqueue of self raced an exiting thread. 783 // Normally we'll find Self near the front of the cxq, so 784 // dequeueing is typically fast. If needbe we can accelerate 785 // this with some MCS/CHL-like bidirectional list hints and advisory 786 // back-links so dequeueing from the interior will normally operate 787 // in constant-time. 788 // Dequeue Self from either the head (with CAS) or from the interior 789 // with a linear-time scan and normal non-atomic memory operations. 790 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 791 // and then unlink Self from EntryList. We have to drain eventually, 792 // so it might as well be now. 793 794 ObjectWaiter * v = _cxq; 795 assert(v != NULL, "invariant"); 796 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 797 // The CAS above can fail from interference IFF a "RAT" arrived. 798 // In that case Self must be in the interior and can no longer be 799 // at the head of cxq. 800 if (v == SelfNode) { 801 assert(_cxq != v, "invariant"); 802 v = _cxq; // CAS above failed - start scan at head of list 803 } 804 ObjectWaiter * p; 805 ObjectWaiter * q = NULL; 806 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 807 q = p; 808 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 809 } 810 assert(v != SelfNode, "invariant"); 811 assert(p == SelfNode, "Node not found on cxq"); 812 assert(p != _cxq, "invariant"); 813 assert(q != NULL, "invariant"); 814 assert(q->_next == p, "invariant"); 815 q->_next = p->_next; 816 } 817 } 818 819 #ifdef ASSERT 820 // Diagnostic hygiene ... 821 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 822 SelfNode->_next = (ObjectWaiter *) 0xBAD; 823 SelfNode->TState = ObjectWaiter::TS_RUN; 824 #endif 825 } 826 827 // ----------------------------------------------------------------------------- 828 // Exit support 829 // 830 // exit() 831 // ~~~~~~ 832 // Note that the collector can't reclaim the objectMonitor or deflate 833 // the object out from underneath the thread calling ::exit() as the 834 // thread calling ::exit() never transitions to a stable state. 835 // This inhibits GC, which in turn inhibits asynchronous (and 836 // inopportune) reclamation of "this". 837 // 838 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 839 // There's one exception to the claim above, however. EnterI() can call 840 // exit() to drop a lock if the acquirer has been externally suspended. 841 // In that case exit() is called with _thread_state as _thread_blocked, 842 // but the monitor's _count field is > 0, which inhibits reclamation. 843 // 844 // 1-0 exit 845 // ~~~~~~~~ 846 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 847 // the fast-path operators have been optimized so the common ::exit() 848 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 849 // The code emitted by fast_unlock() elides the usual MEMBAR. This 850 // greatly improves latency -- MEMBAR and CAS having considerable local 851 // latency on modern processors -- but at the cost of "stranding". Absent the 852 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 853 // ::enter() path, resulting in the entering thread being stranding 854 // and a progress-liveness failure. Stranding is extremely rare. 855 // We use timers (timed park operations) & periodic polling to detect 856 // and recover from stranding. Potentially stranded threads periodically 857 // wake up and poll the lock. See the usage of the _Responsible variable. 858 // 859 // The CAS() in enter provides for safety and exclusion, while the CAS or 860 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 861 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 862 // We detect and recover from stranding with timers. 863 // 864 // If a thread transiently strands it'll park until (a) another 865 // thread acquires the lock and then drops the lock, at which time the 866 // exiting thread will notice and unpark the stranded thread, or, (b) 867 // the timer expires. If the lock is high traffic then the stranding latency 868 // will be low due to (a). If the lock is low traffic then the odds of 869 // stranding are lower, although the worst-case stranding latency 870 // is longer. Critically, we don't want to put excessive load in the 871 // platform's timer subsystem. We want to minimize both the timer injection 872 // rate (timers created/sec) as well as the number of timers active at 873 // any one time. (more precisely, we want to minimize timer-seconds, which is 874 // the integral of the # of active timers at any instant over time). 875 // Both impinge on OS scalability. Given that, at most one thread parked on 876 // a monitor will use a timer. 877 // 878 // There is also the risk of a futile wake-up. If we drop the lock 879 // another thread can reacquire the lock immediately, and we can 880 // then wake a thread unnecessarily. This is benign, and we've 881 // structured the code so the windows are short and the frequency 882 // of such futile wakups is low. 883 884 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 885 Thread * const Self = THREAD; 886 if (THREAD != _owner) { 887 if (THREAD->is_lock_owned((address) _owner)) { 888 // Transmute _owner from a BasicLock pointer to a Thread address. 889 // We don't need to hold _mutex for this transition. 890 // Non-null to Non-null is safe as long as all readers can 891 // tolerate either flavor. 892 assert(_recursions == 0, "invariant"); 893 _owner = THREAD; 894 _recursions = 0; 895 } else { 896 // Apparent unbalanced locking ... 897 // Naively we'd like to throw IllegalMonitorStateException. 898 // As a practical matter we can neither allocate nor throw an 899 // exception as ::exit() can be called from leaf routines. 900 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 901 // Upon deeper reflection, however, in a properly run JVM the only 902 // way we should encounter this situation is in the presence of 903 // unbalanced JNI locking. TODO: CheckJNICalls. 904 // See also: CR4414101 905 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 906 return; 907 } 908 } 909 910 if (_recursions != 0) { 911 _recursions--; // this is simple recursive enter 912 return; 913 } 914 915 // Invariant: after setting Responsible=null an thread must execute 916 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 917 _Responsible = NULL; 918 919 #if INCLUDE_JFR 920 // get the owner's thread id for the MonitorEnter event 921 // if it is enabled and the thread isn't suspended 922 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 923 _previous_owner_tid = JFR_THREAD_ID(Self); 924 } 925 #endif 926 927 for (;;) { 928 assert(THREAD == _owner, "invariant"); 929 930 if (Knob_ExitPolicy == 0) { 931 // release semantics: prior loads and stores from within the critical section 932 // must not float (reorder) past the following store that drops the lock. 933 // On SPARC that requires MEMBAR #loadstore|#storestore. 934 // But of course in TSO #loadstore|#storestore is not required. 935 // I'd like to write one of the following: 936 // A. OrderAccess::release() ; _owner = NULL 937 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 938 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 939 // store into a _dummy variable. That store is not needed, but can result 940 // in massive wasteful coherency traffic on classic SMP systems. 941 // Instead, I use release_store(), which is implemented as just a simple 942 // ST on x64, x86 and SPARC. 943 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 944 OrderAccess::storeload(); // See if we need to wake a successor 945 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 946 return; 947 } 948 // Other threads are blocked trying to acquire the lock. 949 950 // Normally the exiting thread is responsible for ensuring succession, 951 // but if other successors are ready or other entering threads are spinning 952 // then this thread can simply store NULL into _owner and exit without 953 // waking a successor. The existence of spinners or ready successors 954 // guarantees proper succession (liveness). Responsibility passes to the 955 // ready or running successors. The exiting thread delegates the duty. 956 // More precisely, if a successor already exists this thread is absolved 957 // of the responsibility of waking (unparking) one. 958 // 959 // The _succ variable is critical to reducing futile wakeup frequency. 960 // _succ identifies the "heir presumptive" thread that has been made 961 // ready (unparked) but that has not yet run. We need only one such 962 // successor thread to guarantee progress. 963 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 964 // section 3.3 "Futile Wakeup Throttling" for details. 965 // 966 // Note that spinners in Enter() also set _succ non-null. 967 // In the current implementation spinners opportunistically set 968 // _succ so that exiting threads might avoid waking a successor. 969 // Another less appealing alternative would be for the exiting thread 970 // to drop the lock and then spin briefly to see if a spinner managed 971 // to acquire the lock. If so, the exiting thread could exit 972 // immediately without waking a successor, otherwise the exiting 973 // thread would need to dequeue and wake a successor. 974 // (Note that we'd need to make the post-drop spin short, but no 975 // shorter than the worst-case round-trip cache-line migration time. 976 // The dropped lock needs to become visible to the spinner, and then 977 // the acquisition of the lock by the spinner must become visible to 978 // the exiting thread). 979 980 // It appears that an heir-presumptive (successor) must be made ready. 981 // Only the current lock owner can manipulate the EntryList or 982 // drain _cxq, so we need to reacquire the lock. If we fail 983 // to reacquire the lock the responsibility for ensuring succession 984 // falls to the new owner. 985 // 986 if (!Atomic::replace_if_null(THREAD, &_owner)) { 987 return; 988 } 989 } else { 990 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 991 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 992 OrderAccess::storeload(); 993 // Ratify the previously observed values. 994 if (_cxq == NULL || _succ != NULL) { 995 return; 996 } 997 998 // inopportune interleaving -- the exiting thread (this thread) 999 // in the fast-exit path raced an entering thread in the slow-enter 1000 // path. 1001 // We have two choices: 1002 // A. Try to reacquire the lock. 1003 // If the CAS() fails return immediately, otherwise 1004 // we either restart/rerun the exit operation, or simply 1005 // fall-through into the code below which wakes a successor. 1006 // B. If the elements forming the EntryList|cxq are TSM 1007 // we could simply unpark() the lead thread and return 1008 // without having set _succ. 1009 if (!Atomic::replace_if_null(THREAD, &_owner)) { 1010 return; 1011 } 1012 } 1013 } 1014 1015 guarantee(_owner == THREAD, "invariant"); 1016 1017 ObjectWaiter * w = NULL; 1018 int QMode = Knob_QMode; 1019 1020 if (QMode == 2 && _cxq != NULL) { 1021 // QMode == 2 : cxq has precedence over EntryList. 1022 // Try to directly wake a successor from the cxq. 1023 // If successful, the successor will need to unlink itself from cxq. 1024 w = _cxq; 1025 assert(w != NULL, "invariant"); 1026 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1027 ExitEpilog(Self, w); 1028 return; 1029 } 1030 1031 if (QMode == 3 && _cxq != NULL) { 1032 // Aggressively drain cxq into EntryList at the first opportunity. 1033 // This policy ensure that recently-run threads live at the head of EntryList. 1034 // Drain _cxq into EntryList - bulk transfer. 1035 // First, detach _cxq. 1036 // The following loop is tantamount to: w = swap(&cxq, NULL) 1037 w = _cxq; 1038 for (;;) { 1039 assert(w != NULL, "Invariant"); 1040 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1041 if (u == w) break; 1042 w = u; 1043 } 1044 assert(w != NULL, "invariant"); 1045 1046 ObjectWaiter * q = NULL; 1047 ObjectWaiter * p; 1048 for (p = w; p != NULL; p = p->_next) { 1049 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1050 p->TState = ObjectWaiter::TS_ENTER; 1051 p->_prev = q; 1052 q = p; 1053 } 1054 1055 // Append the RATs to the EntryList 1056 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1057 ObjectWaiter * Tail; 1058 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1059 Tail = Tail->_next) 1060 /* empty */; 1061 if (Tail == NULL) { 1062 _EntryList = w; 1063 } else { 1064 Tail->_next = w; 1065 w->_prev = Tail; 1066 } 1067 1068 // Fall thru into code that tries to wake a successor from EntryList 1069 } 1070 1071 if (QMode == 4 && _cxq != NULL) { 1072 // Aggressively drain cxq into EntryList at the first opportunity. 1073 // This policy ensure that recently-run threads live at the head of EntryList. 1074 1075 // Drain _cxq into EntryList - bulk transfer. 1076 // First, detach _cxq. 1077 // The following loop is tantamount to: w = swap(&cxq, NULL) 1078 w = _cxq; 1079 for (;;) { 1080 assert(w != NULL, "Invariant"); 1081 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1082 if (u == w) break; 1083 w = u; 1084 } 1085 assert(w != NULL, "invariant"); 1086 1087 ObjectWaiter * q = NULL; 1088 ObjectWaiter * p; 1089 for (p = w; p != NULL; p = p->_next) { 1090 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1091 p->TState = ObjectWaiter::TS_ENTER; 1092 p->_prev = q; 1093 q = p; 1094 } 1095 1096 // Prepend the RATs to the EntryList 1097 if (_EntryList != NULL) { 1098 q->_next = _EntryList; 1099 _EntryList->_prev = q; 1100 } 1101 _EntryList = w; 1102 1103 // Fall thru into code that tries to wake a successor from EntryList 1104 } 1105 1106 w = _EntryList; 1107 if (w != NULL) { 1108 // I'd like to write: guarantee (w->_thread != Self). 1109 // But in practice an exiting thread may find itself on the EntryList. 1110 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1111 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1112 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1113 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1114 // release the lock "O". T2 resumes immediately after the ST of null into 1115 // _owner, above. T2 notices that the EntryList is populated, so it 1116 // reacquires the lock and then finds itself on the EntryList. 1117 // Given all that, we have to tolerate the circumstance where "w" is 1118 // associated with Self. 1119 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1120 ExitEpilog(Self, w); 1121 return; 1122 } 1123 1124 // If we find that both _cxq and EntryList are null then just 1125 // re-run the exit protocol from the top. 1126 w = _cxq; 1127 if (w == NULL) continue; 1128 1129 // Drain _cxq into EntryList - bulk transfer. 1130 // First, detach _cxq. 1131 // The following loop is tantamount to: w = swap(&cxq, NULL) 1132 for (;;) { 1133 assert(w != NULL, "Invariant"); 1134 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1135 if (u == w) break; 1136 w = u; 1137 } 1138 1139 assert(w != NULL, "invariant"); 1140 assert(_EntryList == NULL, "invariant"); 1141 1142 // Convert the LIFO SLL anchored by _cxq into a DLL. 1143 // The list reorganization step operates in O(LENGTH(w)) time. 1144 // It's critical that this step operate quickly as 1145 // "Self" still holds the outer-lock, restricting parallelism 1146 // and effectively lengthening the critical section. 1147 // Invariant: s chases t chases u. 1148 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1149 // we have faster access to the tail. 1150 1151 if (QMode == 1) { 1152 // QMode == 1 : drain cxq to EntryList, reversing order 1153 // We also reverse the order of the list. 1154 ObjectWaiter * s = NULL; 1155 ObjectWaiter * t = w; 1156 ObjectWaiter * u = NULL; 1157 while (t != NULL) { 1158 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1159 t->TState = ObjectWaiter::TS_ENTER; 1160 u = t->_next; 1161 t->_prev = u; 1162 t->_next = s; 1163 s = t; 1164 t = u; 1165 } 1166 _EntryList = s; 1167 assert(s != NULL, "invariant"); 1168 } else { 1169 // QMode == 0 or QMode == 2 1170 _EntryList = w; 1171 ObjectWaiter * q = NULL; 1172 ObjectWaiter * p; 1173 for (p = w; p != NULL; p = p->_next) { 1174 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1175 p->TState = ObjectWaiter::TS_ENTER; 1176 p->_prev = q; 1177 q = p; 1178 } 1179 } 1180 1181 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1182 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1183 1184 // See if we can abdicate to a spinner instead of waking a thread. 1185 // A primary goal of the implementation is to reduce the 1186 // context-switch rate. 1187 if (_succ != NULL) continue; 1188 1189 w = _EntryList; 1190 if (w != NULL) { 1191 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1192 ExitEpilog(Self, w); 1193 return; 1194 } 1195 } 1196 } 1197 1198 // ExitSuspendEquivalent: 1199 // A faster alternate to handle_special_suspend_equivalent_condition() 1200 // 1201 // handle_special_suspend_equivalent_condition() unconditionally 1202 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1203 // operations have high latency. Note that in ::enter() we call HSSEC 1204 // while holding the monitor, so we effectively lengthen the critical sections. 1205 // 1206 // There are a number of possible solutions: 1207 // 1208 // A. To ameliorate the problem we might also defer state transitions 1209 // to as late as possible -- just prior to parking. 1210 // Given that, we'd call HSSEC after having returned from park(), 1211 // but before attempting to acquire the monitor. This is only a 1212 // partial solution. It avoids calling HSSEC while holding the 1213 // monitor (good), but it still increases successor reacquisition latency -- 1214 // the interval between unparking a successor and the time the successor 1215 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1216 // If we use this technique we can also avoid EnterI()-exit() loop 1217 // in ::enter() where we iteratively drop the lock and then attempt 1218 // to reacquire it after suspending. 1219 // 1220 // B. In the future we might fold all the suspend bits into a 1221 // composite per-thread suspend flag and then update it with CAS(). 1222 // Alternately, a Dekker-like mechanism with multiple variables 1223 // would suffice: 1224 // ST Self->_suspend_equivalent = false 1225 // MEMBAR 1226 // LD Self_>_suspend_flags 1227 // 1228 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1229 // with a more efficient implementation, the need to use "FastHSSEC" has 1230 // decreased. - Dave 1231 1232 1233 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1234 const int Mode = Knob_FastHSSEC; 1235 if (Mode && !jSelf->is_external_suspend()) { 1236 assert(jSelf->is_suspend_equivalent(), "invariant"); 1237 jSelf->clear_suspend_equivalent(); 1238 if (2 == Mode) OrderAccess::storeload(); 1239 if (!jSelf->is_external_suspend()) return false; 1240 // We raced a suspension -- fall thru into the slow path 1241 jSelf->set_suspend_equivalent(); 1242 } 1243 return jSelf->handle_special_suspend_equivalent_condition(); 1244 } 1245 1246 1247 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1248 assert(_owner == Self, "invariant"); 1249 1250 // Exit protocol: 1251 // 1. ST _succ = wakee 1252 // 2. membar #loadstore|#storestore; 1253 // 2. ST _owner = NULL 1254 // 3. unpark(wakee) 1255 1256 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1257 ParkEvent * Trigger = Wakee->_event; 1258 1259 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1260 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1261 // out-of-scope (non-extant). 1262 Wakee = NULL; 1263 1264 // Drop the lock 1265 OrderAccess::release_store(&_owner, (void*)NULL); 1266 OrderAccess::fence(); // ST _owner vs LD in unpark() 1267 1268 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1269 Trigger->unpark(); 1270 1271 // Maintain stats and report events to JVMTI 1272 OM_PERFDATA_OP(Parks, inc()); 1273 } 1274 1275 1276 // ----------------------------------------------------------------------------- 1277 // Class Loader deadlock handling. 1278 // 1279 // complete_exit exits a lock returning recursion count 1280 // complete_exit/reenter operate as a wait without waiting 1281 // complete_exit requires an inflated monitor 1282 // The _owner field is not always the Thread addr even with an 1283 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1284 // thread due to contention. 1285 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1286 Thread * const Self = THREAD; 1287 assert(Self->is_Java_thread(), "Must be Java thread!"); 1288 JavaThread *jt = (JavaThread *)THREAD; 1289 1290 DeferredInitialize(); 1291 1292 if (THREAD != _owner) { 1293 if (THREAD->is_lock_owned ((address)_owner)) { 1294 assert(_recursions == 0, "internal state error"); 1295 _owner = THREAD; // Convert from basiclock addr to Thread addr 1296 _recursions = 0; 1297 } 1298 } 1299 1300 guarantee(Self == _owner, "complete_exit not owner"); 1301 intptr_t save = _recursions; // record the old recursion count 1302 _recursions = 0; // set the recursion level to be 0 1303 exit(true, Self); // exit the monitor 1304 guarantee(_owner != Self, "invariant"); 1305 return save; 1306 } 1307 1308 // reenter() enters a lock and sets recursion count 1309 // complete_exit/reenter operate as a wait without waiting 1310 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1311 Thread * const Self = THREAD; 1312 assert(Self->is_Java_thread(), "Must be Java thread!"); 1313 JavaThread *jt = (JavaThread *)THREAD; 1314 1315 guarantee(_owner != Self, "reenter already owner"); 1316 enter(THREAD); // enter the monitor 1317 guarantee(_recursions == 0, "reenter recursion"); 1318 _recursions = recursions; 1319 return; 1320 } 1321 1322 1323 // ----------------------------------------------------------------------------- 1324 // A macro is used below because there may already be a pending 1325 // exception which should not abort the execution of the routines 1326 // which use this (which is why we don't put this into check_slow and 1327 // call it with a CHECK argument). 1328 1329 #define CHECK_OWNER() \ 1330 do { \ 1331 if (THREAD != _owner) { \ 1332 if (THREAD->is_lock_owned((address) _owner)) { \ 1333 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1334 _recursions = 0; \ 1335 } else { \ 1336 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1337 } \ 1338 } \ 1339 } while (false) 1340 1341 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1342 // TODO-FIXME: remove check_slow() -- it's likely dead. 1343 1344 void ObjectMonitor::check_slow(TRAPS) { 1345 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1346 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1347 } 1348 1349 static int Adjust(volatile int * adr, int dx) { 1350 int v; 1351 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1352 return v; 1353 } 1354 1355 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1356 ObjectMonitor* monitor, 1357 jlong notifier_tid, 1358 jlong timeout, 1359 bool timedout) { 1360 assert(event != NULL, "invariant"); 1361 assert(monitor != NULL, "invariant"); 1362 event->set_monitorClass(((oop)monitor->object())->klass()); 1363 event->set_timeout(timeout); 1364 event->set_address((uintptr_t)monitor->object_addr()); 1365 event->set_notifier(notifier_tid); 1366 event->set_timedOut(timedout); 1367 event->commit(); 1368 } 1369 1370 // ----------------------------------------------------------------------------- 1371 // Wait/Notify/NotifyAll 1372 // 1373 // Note: a subset of changes to ObjectMonitor::wait() 1374 // will need to be replicated in complete_exit 1375 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1376 Thread * const Self = THREAD; 1377 assert(Self->is_Java_thread(), "Must be Java thread!"); 1378 JavaThread *jt = (JavaThread *)THREAD; 1379 1380 DeferredInitialize(); 1381 1382 // Throw IMSX or IEX. 1383 CHECK_OWNER(); 1384 1385 EventJavaMonitorWait event; 1386 1387 // check for a pending interrupt 1388 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1389 // post monitor waited event. Note that this is past-tense, we are done waiting. 1390 if (JvmtiExport::should_post_monitor_waited()) { 1391 // Note: 'false' parameter is passed here because the 1392 // wait was not timed out due to thread interrupt. 1393 JvmtiExport::post_monitor_waited(jt, this, false); 1394 1395 // In this short circuit of the monitor wait protocol, the 1396 // current thread never drops ownership of the monitor and 1397 // never gets added to the wait queue so the current thread 1398 // cannot be made the successor. This means that the 1399 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1400 // consume an unpark() meant for the ParkEvent associated with 1401 // this ObjectMonitor. 1402 } 1403 if (event.should_commit()) { 1404 post_monitor_wait_event(&event, this, 0, millis, false); 1405 } 1406 THROW(vmSymbols::java_lang_InterruptedException()); 1407 return; 1408 } 1409 1410 assert(Self->_Stalled == 0, "invariant"); 1411 Self->_Stalled = intptr_t(this); 1412 jt->set_current_waiting_monitor(this); 1413 1414 // create a node to be put into the queue 1415 // Critically, after we reset() the event but prior to park(), we must check 1416 // for a pending interrupt. 1417 ObjectWaiter node(Self); 1418 node.TState = ObjectWaiter::TS_WAIT; 1419 Self->_ParkEvent->reset(); 1420 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1421 1422 // Enter the waiting queue, which is a circular doubly linked list in this case 1423 // but it could be a priority queue or any data structure. 1424 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1425 // by the the owner of the monitor *except* in the case where park() 1426 // returns because of a timeout of interrupt. Contention is exceptionally rare 1427 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1428 1429 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1430 AddWaiter(&node); 1431 Thread::SpinRelease(&_WaitSetLock); 1432 1433 _Responsible = NULL; 1434 1435 intptr_t save = _recursions; // record the old recursion count 1436 _waiters++; // increment the number of waiters 1437 _recursions = 0; // set the recursion level to be 1 1438 exit(true, Self); // exit the monitor 1439 guarantee(_owner != Self, "invariant"); 1440 1441 // The thread is on the WaitSet list - now park() it. 1442 // On MP systems it's conceivable that a brief spin before we park 1443 // could be profitable. 1444 // 1445 // TODO-FIXME: change the following logic to a loop of the form 1446 // while (!timeout && !interrupted && _notified == 0) park() 1447 1448 int ret = OS_OK; 1449 int WasNotified = 0; 1450 { // State transition wrappers 1451 OSThread* osthread = Self->osthread(); 1452 OSThreadWaitState osts(osthread, true); 1453 { 1454 ThreadBlockInVM tbivm(jt); 1455 // Thread is in thread_blocked state and oop access is unsafe. 1456 jt->set_suspend_equivalent(); 1457 1458 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1459 // Intentionally empty 1460 } else if (node._notified == 0) { 1461 if (millis <= 0) { 1462 Self->_ParkEvent->park(); 1463 } else { 1464 ret = Self->_ParkEvent->park(millis); 1465 } 1466 } 1467 1468 // were we externally suspended while we were waiting? 1469 if (ExitSuspendEquivalent (jt)) { 1470 // TODO-FIXME: add -- if succ == Self then succ = null. 1471 jt->java_suspend_self(); 1472 } 1473 1474 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1475 1476 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1477 // from the WaitSet to the EntryList. 1478 // See if we need to remove Node from the WaitSet. 1479 // We use double-checked locking to avoid grabbing _WaitSetLock 1480 // if the thread is not on the wait queue. 1481 // 1482 // Note that we don't need a fence before the fetch of TState. 1483 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1484 // written by the is thread. (perhaps the fetch might even be satisfied 1485 // by a look-aside into the processor's own store buffer, although given 1486 // the length of the code path between the prior ST and this load that's 1487 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1488 // then we'll acquire the lock and then re-fetch a fresh TState value. 1489 // That is, we fail toward safety. 1490 1491 if (node.TState == ObjectWaiter::TS_WAIT) { 1492 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1493 if (node.TState == ObjectWaiter::TS_WAIT) { 1494 DequeueSpecificWaiter(&node); // unlink from WaitSet 1495 assert(node._notified == 0, "invariant"); 1496 node.TState = ObjectWaiter::TS_RUN; 1497 } 1498 Thread::SpinRelease(&_WaitSetLock); 1499 } 1500 1501 // The thread is now either on off-list (TS_RUN), 1502 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1503 // The Node's TState variable is stable from the perspective of this thread. 1504 // No other threads will asynchronously modify TState. 1505 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1506 OrderAccess::loadload(); 1507 if (_succ == Self) _succ = NULL; 1508 WasNotified = node._notified; 1509 1510 // Reentry phase -- reacquire the monitor. 1511 // re-enter contended monitor after object.wait(). 1512 // retain OBJECT_WAIT state until re-enter successfully completes 1513 // Thread state is thread_in_vm and oop access is again safe, 1514 // although the raw address of the object may have changed. 1515 // (Don't cache naked oops over safepoints, of course). 1516 1517 // post monitor waited event. Note that this is past-tense, we are done waiting. 1518 if (JvmtiExport::should_post_monitor_waited()) { 1519 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1520 1521 if (node._notified != 0 && _succ == Self) { 1522 // In this part of the monitor wait-notify-reenter protocol it 1523 // is possible (and normal) for another thread to do a fastpath 1524 // monitor enter-exit while this thread is still trying to get 1525 // to the reenter portion of the protocol. 1526 // 1527 // The ObjectMonitor was notified and the current thread is 1528 // the successor which also means that an unpark() has already 1529 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1530 // consume the unpark() that was done when the successor was 1531 // set because the same ParkEvent is shared between Java 1532 // monitors and JVM/TI RawMonitors (for now). 1533 // 1534 // We redo the unpark() to ensure forward progress, i.e., we 1535 // don't want all pending threads hanging (parked) with none 1536 // entering the unlocked monitor. 1537 node._event->unpark(); 1538 } 1539 } 1540 1541 if (event.should_commit()) { 1542 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1543 } 1544 1545 OrderAccess::fence(); 1546 1547 assert(Self->_Stalled != 0, "invariant"); 1548 Self->_Stalled = 0; 1549 1550 assert(_owner != Self, "invariant"); 1551 ObjectWaiter::TStates v = node.TState; 1552 if (v == ObjectWaiter::TS_RUN) { 1553 enter(Self); 1554 } else { 1555 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1556 ReenterI(Self, &node); 1557 node.wait_reenter_end(this); 1558 } 1559 1560 // Self has reacquired the lock. 1561 // Lifecycle - the node representing Self must not appear on any queues. 1562 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1563 // want residual elements associated with this thread left on any lists. 1564 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1565 assert(_owner == Self, "invariant"); 1566 assert(_succ != Self, "invariant"); 1567 } // OSThreadWaitState() 1568 1569 jt->set_current_waiting_monitor(NULL); 1570 1571 guarantee(_recursions == 0, "invariant"); 1572 _recursions = save; // restore the old recursion count 1573 _waiters--; // decrement the number of waiters 1574 1575 // Verify a few postconditions 1576 assert(_owner == Self, "invariant"); 1577 assert(_succ != Self, "invariant"); 1578 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1579 1580 // check if the notification happened 1581 if (!WasNotified) { 1582 // no, it could be timeout or Thread.interrupt() or both 1583 // check for interrupt event, otherwise it is timeout 1584 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1585 THROW(vmSymbols::java_lang_InterruptedException()); 1586 } 1587 } 1588 1589 // NOTE: Spurious wake up will be consider as timeout. 1590 // Monitor notify has precedence over thread interrupt. 1591 } 1592 1593 1594 // Consider: 1595 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1596 // then instead of transferring a thread from the WaitSet to the EntryList 1597 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1598 1599 void ObjectMonitor::INotify(Thread * Self) { 1600 const int policy = Knob_MoveNotifyee; 1601 1602 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1603 ObjectWaiter * iterator = DequeueWaiter(); 1604 if (iterator != NULL) { 1605 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1606 guarantee(iterator->_notified == 0, "invariant"); 1607 // Disposition - what might we do with iterator ? 1608 // a. add it directly to the EntryList - either tail (policy == 1) 1609 // or head (policy == 0). 1610 // b. push it onto the front of the _cxq (policy == 2). 1611 // For now we use (b). 1612 if (policy != 4) { 1613 iterator->TState = ObjectWaiter::TS_ENTER; 1614 } 1615 iterator->_notified = 1; 1616 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1617 1618 ObjectWaiter * list = _EntryList; 1619 if (list != NULL) { 1620 assert(list->_prev == NULL, "invariant"); 1621 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1622 assert(list != iterator, "invariant"); 1623 } 1624 1625 if (policy == 0) { // prepend to EntryList 1626 if (list == NULL) { 1627 iterator->_next = iterator->_prev = NULL; 1628 _EntryList = iterator; 1629 } else { 1630 list->_prev = iterator; 1631 iterator->_next = list; 1632 iterator->_prev = NULL; 1633 _EntryList = iterator; 1634 } 1635 } else if (policy == 1) { // append to EntryList 1636 if (list == NULL) { 1637 iterator->_next = iterator->_prev = NULL; 1638 _EntryList = iterator; 1639 } else { 1640 // CONSIDER: finding the tail currently requires a linear-time walk of 1641 // the EntryList. We can make tail access constant-time by converting to 1642 // a CDLL instead of using our current DLL. 1643 ObjectWaiter * tail; 1644 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1645 assert(tail != NULL && tail->_next == NULL, "invariant"); 1646 tail->_next = iterator; 1647 iterator->_prev = tail; 1648 iterator->_next = NULL; 1649 } 1650 } else if (policy == 2) { // prepend to cxq 1651 if (list == NULL) { 1652 iterator->_next = iterator->_prev = NULL; 1653 _EntryList = iterator; 1654 } else { 1655 iterator->TState = ObjectWaiter::TS_CXQ; 1656 for (;;) { 1657 ObjectWaiter * front = _cxq; 1658 iterator->_next = front; 1659 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1660 break; 1661 } 1662 } 1663 } 1664 } else if (policy == 3) { // append to cxq 1665 iterator->TState = ObjectWaiter::TS_CXQ; 1666 for (;;) { 1667 ObjectWaiter * tail = _cxq; 1668 if (tail == NULL) { 1669 iterator->_next = NULL; 1670 if (Atomic::replace_if_null(iterator, &_cxq)) { 1671 break; 1672 } 1673 } else { 1674 while (tail->_next != NULL) tail = tail->_next; 1675 tail->_next = iterator; 1676 iterator->_prev = tail; 1677 iterator->_next = NULL; 1678 break; 1679 } 1680 } 1681 } else { 1682 ParkEvent * ev = iterator->_event; 1683 iterator->TState = ObjectWaiter::TS_RUN; 1684 OrderAccess::fence(); 1685 ev->unpark(); 1686 } 1687 1688 // _WaitSetLock protects the wait queue, not the EntryList. We could 1689 // move the add-to-EntryList operation, above, outside the critical section 1690 // protected by _WaitSetLock. In practice that's not useful. With the 1691 // exception of wait() timeouts and interrupts the monitor owner 1692 // is the only thread that grabs _WaitSetLock. There's almost no contention 1693 // on _WaitSetLock so it's not profitable to reduce the length of the 1694 // critical section. 1695 1696 if (policy < 4) { 1697 iterator->wait_reenter_begin(this); 1698 } 1699 } 1700 Thread::SpinRelease(&_WaitSetLock); 1701 } 1702 1703 // Consider: a not-uncommon synchronization bug is to use notify() when 1704 // notifyAll() is more appropriate, potentially resulting in stranded 1705 // threads; this is one example of a lost wakeup. A useful diagnostic 1706 // option is to force all notify() operations to behave as notifyAll(). 1707 // 1708 // Note: We can also detect many such problems with a "minimum wait". 1709 // When the "minimum wait" is set to a small non-zero timeout value 1710 // and the program does not hang whereas it did absent "minimum wait", 1711 // that suggests a lost wakeup bug. 1712 1713 void ObjectMonitor::notify(TRAPS) { 1714 CHECK_OWNER(); 1715 if (_WaitSet == NULL) { 1716 return; 1717 } 1718 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1719 INotify(THREAD); 1720 OM_PERFDATA_OP(Notifications, inc(1)); 1721 } 1722 1723 1724 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1725 // from the waitset to the EntryList. This could be done more efficiently with a 1726 // single bulk transfer but in practice it's not time-critical. Beware too, 1727 // that in prepend-mode we invert the order of the waiters. Let's say that the 1728 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1729 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1730 1731 void ObjectMonitor::notifyAll(TRAPS) { 1732 CHECK_OWNER(); 1733 if (_WaitSet == NULL) { 1734 return; 1735 } 1736 1737 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1738 int tally = 0; 1739 while (_WaitSet != NULL) { 1740 tally++; 1741 INotify(THREAD); 1742 } 1743 1744 OM_PERFDATA_OP(Notifications, inc(tally)); 1745 } 1746 1747 // ----------------------------------------------------------------------------- 1748 // Adaptive Spinning Support 1749 // 1750 // Adaptive spin-then-block - rational spinning 1751 // 1752 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1753 // algorithm. On high order SMP systems it would be better to start with 1754 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1755 // a contending thread could enqueue itself on the cxq and then spin locally 1756 // on a thread-specific variable such as its ParkEvent._Event flag. 1757 // That's left as an exercise for the reader. Note that global spinning is 1758 // not problematic on Niagara, as the L2 cache serves the interconnect and 1759 // has both low latency and massive bandwidth. 1760 // 1761 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1762 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1763 // (duration) or we can fix the count at approximately the duration of 1764 // a context switch and vary the frequency. Of course we could also 1765 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1766 // For a description of 'Adaptive spin-then-block mutual exclusion in 1767 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1768 // 1769 // This implementation varies the duration "D", where D varies with 1770 // the success rate of recent spin attempts. (D is capped at approximately 1771 // length of a round-trip context switch). The success rate for recent 1772 // spin attempts is a good predictor of the success rate of future spin 1773 // attempts. The mechanism adapts automatically to varying critical 1774 // section length (lock modality), system load and degree of parallelism. 1775 // D is maintained per-monitor in _SpinDuration and is initialized 1776 // optimistically. Spin frequency is fixed at 100%. 1777 // 1778 // Note that _SpinDuration is volatile, but we update it without locks 1779 // or atomics. The code is designed so that _SpinDuration stays within 1780 // a reasonable range even in the presence of races. The arithmetic 1781 // operations on _SpinDuration are closed over the domain of legal values, 1782 // so at worst a race will install and older but still legal value. 1783 // At the very worst this introduces some apparent non-determinism. 1784 // We might spin when we shouldn't or vice-versa, but since the spin 1785 // count are relatively short, even in the worst case, the effect is harmless. 1786 // 1787 // Care must be taken that a low "D" value does not become an 1788 // an absorbing state. Transient spinning failures -- when spinning 1789 // is overall profitable -- should not cause the system to converge 1790 // on low "D" values. We want spinning to be stable and predictable 1791 // and fairly responsive to change and at the same time we don't want 1792 // it to oscillate, become metastable, be "too" non-deterministic, 1793 // or converge on or enter undesirable stable absorbing states. 1794 // 1795 // We implement a feedback-based control system -- using past behavior 1796 // to predict future behavior. We face two issues: (a) if the 1797 // input signal is random then the spin predictor won't provide optimal 1798 // results, and (b) if the signal frequency is too high then the control 1799 // system, which has some natural response lag, will "chase" the signal. 1800 // (b) can arise from multimodal lock hold times. Transient preemption 1801 // can also result in apparent bimodal lock hold times. 1802 // Although sub-optimal, neither condition is particularly harmful, as 1803 // in the worst-case we'll spin when we shouldn't or vice-versa. 1804 // The maximum spin duration is rather short so the failure modes aren't bad. 1805 // To be conservative, I've tuned the gain in system to bias toward 1806 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1807 // "rings" or oscillates between spinning and not spinning. This happens 1808 // when spinning is just on the cusp of profitability, however, so the 1809 // situation is not dire. The state is benign -- there's no need to add 1810 // hysteresis control to damp the transition rate between spinning and 1811 // not spinning. 1812 1813 // Spinning: Fixed frequency (100%), vary duration 1814 int ObjectMonitor::TrySpin(Thread * Self) { 1815 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1816 int ctr = Knob_FixedSpin; 1817 if (ctr != 0) { 1818 while (--ctr >= 0) { 1819 if (TryLock(Self) > 0) return 1; 1820 SpinPause(); 1821 } 1822 return 0; 1823 } 1824 1825 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1826 if (TryLock(Self) > 0) { 1827 // Increase _SpinDuration ... 1828 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1829 // Raising _SpurDuration to the poverty line is key. 1830 int x = _SpinDuration; 1831 if (x < Knob_SpinLimit) { 1832 if (x < Knob_Poverty) x = Knob_Poverty; 1833 _SpinDuration = x + Knob_BonusB; 1834 } 1835 return 1; 1836 } 1837 SpinPause(); 1838 } 1839 1840 // Admission control - verify preconditions for spinning 1841 // 1842 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1843 // becoming an absorbing state. Put another way, we spin briefly to 1844 // sample, just in case the system load, parallelism, contention, or lock 1845 // modality changed. 1846 // 1847 // Consider the following alternative: 1848 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1849 // spin attempt. "Periodically" might mean after a tally of 1850 // the # of failed spin attempts (or iterations) reaches some threshold. 1851 // This takes us into the realm of 1-out-of-N spinning, where we 1852 // hold the duration constant but vary the frequency. 1853 1854 ctr = _SpinDuration; 1855 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1856 if (ctr <= 0) return 0; 1857 1858 if (Knob_SuccRestrict && _succ != NULL) return 0; 1859 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1860 return 0; 1861 } 1862 1863 int MaxSpin = Knob_MaxSpinners; 1864 if (MaxSpin >= 0) { 1865 if (_Spinner > MaxSpin) { 1866 return 0; 1867 } 1868 // Slightly racy, but benign ... 1869 Adjust(&_Spinner, 1); 1870 } 1871 1872 // We're good to spin ... spin ingress. 1873 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1874 // when preparing to LD...CAS _owner, etc and the CAS is likely 1875 // to succeed. 1876 int hits = 0; 1877 int msk = 0; 1878 int caspty = Knob_CASPenalty; 1879 int oxpty = Knob_OXPenalty; 1880 int sss = Knob_SpinSetSucc; 1881 if (sss && _succ == NULL) _succ = Self; 1882 Thread * prv = NULL; 1883 1884 // There are three ways to exit the following loop: 1885 // 1. A successful spin where this thread has acquired the lock. 1886 // 2. Spin failure with prejudice 1887 // 3. Spin failure without prejudice 1888 1889 while (--ctr >= 0) { 1890 1891 // Periodic polling -- Check for pending GC 1892 // Threads may spin while they're unsafe. 1893 // We don't want spinning threads to delay the JVM from reaching 1894 // a stop-the-world safepoint or to steal cycles from GC. 1895 // If we detect a pending safepoint we abort in order that 1896 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1897 // this thread, if safe, doesn't steal cycles from GC. 1898 // This is in keeping with the "no loitering in runtime" rule. 1899 // We periodically check to see if there's a safepoint pending. 1900 if ((ctr & 0xFF) == 0) { 1901 if (SafepointMechanism::poll(Self)) { 1902 goto Abort; // abrupt spin egress 1903 } 1904 if (Knob_UsePause & 1) SpinPause(); 1905 } 1906 1907 if (Knob_UsePause & 2) SpinPause(); 1908 1909 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1910 // This is useful on classic SMP systems, but is of less utility on 1911 // N1-style CMT platforms. 1912 // 1913 // Trade-off: lock acquisition latency vs coherency bandwidth. 1914 // Lock hold times are typically short. A histogram 1915 // of successful spin attempts shows that we usually acquire 1916 // the lock early in the spin. That suggests we want to 1917 // sample _owner frequently in the early phase of the spin, 1918 // but then back-off and sample less frequently as the spin 1919 // progresses. The back-off makes a good citizen on SMP big 1920 // SMP systems. Oversampling _owner can consume excessive 1921 // coherency bandwidth. Relatedly, if we _oversample _owner we 1922 // can inadvertently interfere with the the ST m->owner=null. 1923 // executed by the lock owner. 1924 if (ctr & msk) continue; 1925 ++hits; 1926 if ((hits & 0xF) == 0) { 1927 // The 0xF, above, corresponds to the exponent. 1928 // Consider: (msk+1)|msk 1929 msk = ((msk << 2)|3) & BackOffMask; 1930 } 1931 1932 // Probe _owner with TATAS 1933 // If this thread observes the monitor transition or flicker 1934 // from locked to unlocked to locked, then the odds that this 1935 // thread will acquire the lock in this spin attempt go down 1936 // considerably. The same argument applies if the CAS fails 1937 // or if we observe _owner change from one non-null value to 1938 // another non-null value. In such cases we might abort 1939 // the spin without prejudice or apply a "penalty" to the 1940 // spin count-down variable "ctr", reducing it by 100, say. 1941 1942 Thread * ox = (Thread *) _owner; 1943 if (ox == NULL) { 1944 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1945 if (ox == NULL) { 1946 // The CAS succeeded -- this thread acquired ownership 1947 // Take care of some bookkeeping to exit spin state. 1948 if (sss && _succ == Self) { 1949 _succ = NULL; 1950 } 1951 if (MaxSpin > 0) Adjust(&_Spinner, -1); 1952 1953 // Increase _SpinDuration : 1954 // The spin was successful (profitable) so we tend toward 1955 // longer spin attempts in the future. 1956 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1957 // If we acquired the lock early in the spin cycle it 1958 // makes sense to increase _SpinDuration proportionally. 1959 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1960 int x = _SpinDuration; 1961 if (x < Knob_SpinLimit) { 1962 if (x < Knob_Poverty) x = Knob_Poverty; 1963 _SpinDuration = x + Knob_Bonus; 1964 } 1965 return 1; 1966 } 1967 1968 // The CAS failed ... we can take any of the following actions: 1969 // * penalize: ctr -= Knob_CASPenalty 1970 // * exit spin with prejudice -- goto Abort; 1971 // * exit spin without prejudice. 1972 // * Since CAS is high-latency, retry again immediately. 1973 prv = ox; 1974 if (caspty == -2) break; 1975 if (caspty == -1) goto Abort; 1976 ctr -= caspty; 1977 continue; 1978 } 1979 1980 // Did lock ownership change hands ? 1981 if (ox != prv && prv != NULL) { 1982 if (oxpty == -2) break; 1983 if (oxpty == -1) goto Abort; 1984 ctr -= oxpty; 1985 } 1986 prv = ox; 1987 1988 // Abort the spin if the owner is not executing. 1989 // The owner must be executing in order to drop the lock. 1990 // Spinning while the owner is OFFPROC is idiocy. 1991 // Consider: ctr -= RunnablePenalty ; 1992 if (Knob_OState && NotRunnable (Self, ox)) { 1993 goto Abort; 1994 } 1995 if (sss && _succ == NULL) _succ = Self; 1996 } 1997 1998 // Spin failed with prejudice -- reduce _SpinDuration. 1999 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2000 // AIMD is globally stable. 2001 { 2002 int x = _SpinDuration; 2003 if (x > 0) { 2004 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2005 // This is globally sample and tends to damp the response. 2006 x -= Knob_Penalty; 2007 if (x < 0) x = 0; 2008 _SpinDuration = x; 2009 } 2010 } 2011 2012 Abort: 2013 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2014 if (sss && _succ == Self) { 2015 _succ = NULL; 2016 // Invariant: after setting succ=null a contending thread 2017 // must recheck-retry _owner before parking. This usually happens 2018 // in the normal usage of TrySpin(), but it's safest 2019 // to make TrySpin() as foolproof as possible. 2020 OrderAccess::fence(); 2021 if (TryLock(Self) > 0) return 1; 2022 } 2023 return 0; 2024 } 2025 2026 // NotRunnable() -- informed spinning 2027 // 2028 // Don't bother spinning if the owner is not eligible to drop the lock. 2029 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2030 // The thread must be runnable in order to drop the lock in timely fashion. 2031 // If the _owner is not runnable then spinning will not likely be 2032 // successful (profitable). 2033 // 2034 // Beware -- the thread referenced by _owner could have died 2035 // so a simply fetch from _owner->_thread_state might trap. 2036 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2037 // Because of the lifecycle issues, the _thread_state values 2038 // observed by NotRunnable() might be garbage. NotRunnable must 2039 // tolerate this and consider the observed _thread_state value 2040 // as advisory. 2041 // 2042 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2043 // a thread pointer. 2044 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2045 // with the LSB of _owner. Another option would be to probabilistically probe 2046 // the putative _owner->TypeTag value. 2047 // 2048 // Checking _thread_state isn't perfect. Even if the thread is 2049 // in_java it might be blocked on a page-fault or have been preempted 2050 // and sitting on a ready/dispatch queue. 2051 // 2052 // The return value from NotRunnable() is *advisory* -- the 2053 // result is based on sampling and is not necessarily coherent. 2054 // The caller must tolerate false-negative and false-positive errors. 2055 // Spinning, in general, is probabilistic anyway. 2056 2057 2058 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2059 // Check ox->TypeTag == 2BAD. 2060 if (ox == NULL) return 0; 2061 2062 // Avoid transitive spinning ... 2063 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2064 // Immediately after T1 acquires L it's possible that T2, also 2065 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2066 // This occurs transiently after T1 acquired L but before 2067 // T1 managed to clear T1.Stalled. T2 does not need to abort 2068 // its spin in this circumstance. 2069 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2070 2071 if (BlockedOn == 1) return 1; 2072 if (BlockedOn != 0) { 2073 return BlockedOn != intptr_t(this) && _owner == ox; 2074 } 2075 2076 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2077 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2078 // consider also: jst != _thread_in_Java -- but that's overspecific. 2079 return jst == _thread_blocked || jst == _thread_in_native; 2080 } 2081 2082 2083 // ----------------------------------------------------------------------------- 2084 // WaitSet management ... 2085 2086 ObjectWaiter::ObjectWaiter(Thread* thread) { 2087 _next = NULL; 2088 _prev = NULL; 2089 _notified = 0; 2090 _notifier_tid = 0; 2091 TState = TS_RUN; 2092 _thread = thread; 2093 _event = thread->_ParkEvent; 2094 _active = false; 2095 assert(_event != NULL, "invariant"); 2096 } 2097 2098 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2099 JavaThread *jt = (JavaThread *)this->_thread; 2100 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2101 } 2102 2103 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2104 JavaThread *jt = (JavaThread *)this->_thread; 2105 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2106 } 2107 2108 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2109 assert(node != NULL, "should not add NULL node"); 2110 assert(node->_prev == NULL, "node already in list"); 2111 assert(node->_next == NULL, "node already in list"); 2112 // put node at end of queue (circular doubly linked list) 2113 if (_WaitSet == NULL) { 2114 _WaitSet = node; 2115 node->_prev = node; 2116 node->_next = node; 2117 } else { 2118 ObjectWaiter* head = _WaitSet; 2119 ObjectWaiter* tail = head->_prev; 2120 assert(tail->_next == head, "invariant check"); 2121 tail->_next = node; 2122 head->_prev = node; 2123 node->_next = head; 2124 node->_prev = tail; 2125 } 2126 } 2127 2128 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2129 // dequeue the very first waiter 2130 ObjectWaiter* waiter = _WaitSet; 2131 if (waiter) { 2132 DequeueSpecificWaiter(waiter); 2133 } 2134 return waiter; 2135 } 2136 2137 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2138 assert(node != NULL, "should not dequeue NULL node"); 2139 assert(node->_prev != NULL, "node already removed from list"); 2140 assert(node->_next != NULL, "node already removed from list"); 2141 // when the waiter has woken up because of interrupt, 2142 // timeout or other spurious wake-up, dequeue the 2143 // waiter from waiting list 2144 ObjectWaiter* next = node->_next; 2145 if (next == node) { 2146 assert(node->_prev == node, "invariant check"); 2147 _WaitSet = NULL; 2148 } else { 2149 ObjectWaiter* prev = node->_prev; 2150 assert(prev->_next == node, "invariant check"); 2151 assert(next->_prev == node, "invariant check"); 2152 next->_prev = prev; 2153 prev->_next = next; 2154 if (_WaitSet == node) { 2155 _WaitSet = next; 2156 } 2157 } 2158 node->_next = NULL; 2159 node->_prev = NULL; 2160 } 2161 2162 // ----------------------------------------------------------------------------- 2163 // PerfData support 2164 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2165 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2166 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2167 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2168 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2169 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2170 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2171 2172 // One-shot global initialization for the sync subsystem. 2173 // We could also defer initialization and initialize on-demand 2174 // the first time we call inflate(). Initialization would 2175 // be protected - like so many things - by the MonitorCache_lock. 2176 2177 void ObjectMonitor::Initialize() { 2178 static int InitializationCompleted = 0; 2179 assert(InitializationCompleted == 0, "invariant"); 2180 InitializationCompleted = 1; 2181 if (UsePerfData) { 2182 EXCEPTION_MARK; 2183 #define NEWPERFCOUNTER(n) \ 2184 { \ 2185 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2186 CHECK); \ 2187 } 2188 #define NEWPERFVARIABLE(n) \ 2189 { \ 2190 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2191 CHECK); \ 2192 } 2193 NEWPERFCOUNTER(_sync_Inflations); 2194 NEWPERFCOUNTER(_sync_Deflations); 2195 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2196 NEWPERFCOUNTER(_sync_FutileWakeups); 2197 NEWPERFCOUNTER(_sync_Parks); 2198 NEWPERFCOUNTER(_sync_Notifications); 2199 NEWPERFVARIABLE(_sync_MonExtant); 2200 #undef NEWPERFCOUNTER 2201 #undef NEWPERFVARIABLE 2202 } 2203 } 2204 2205 void ObjectMonitor::DeferredInitialize() { 2206 if (InitDone > 0) return; 2207 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2208 while (InitDone != 1) /* empty */; 2209 return; 2210 } 2211 2212 // One-shot global initialization ... 2213 // The initialization is idempotent, so we don't need locks. 2214 // In the future consider doing this via os::init_2(). 2215 2216 if (os::is_MP()) { 2217 BackOffMask = 0; 2218 } else { 2219 Knob_SpinLimit = 0; 2220 Knob_SpinBase = 0; 2221 Knob_PreSpin = 0; 2222 Knob_FixedSpin = -1; 2223 } 2224 2225 OrderAccess::fence(); 2226 InitDone = 1; 2227 } 2228