1 /* 2 * Copyright (c) 1998, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "oops/markOop.hpp" 29 #include "oops/oop.inline.hpp" 30 #include "runtime/atomic.inline.hpp" 31 #include "runtime/handles.inline.hpp" 32 #include "runtime/interfaceSupport.hpp" 33 #include "runtime/mutexLocker.hpp" 34 #include "runtime/objectMonitor.hpp" 35 #include "runtime/objectMonitor.inline.hpp" 36 #include "runtime/orderAccess.inline.hpp" 37 #include "runtime/osThread.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #include "runtime/thread.inline.hpp" 40 #include "services/threadService.hpp" 41 #include "trace/tracing.hpp" 42 #include "trace/traceMacros.hpp" 43 #include "utilities/dtrace.hpp" 44 #include "utilities/macros.hpp" 45 #include "utilities/preserveException.hpp" 46 47 #ifdef DTRACE_ENABLED 48 49 // Only bother with this argument setup if dtrace is available 50 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 51 52 53 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 54 char* bytes = NULL; \ 55 int len = 0; \ 56 jlong jtid = SharedRuntime::get_java_tid(thread); \ 57 Symbol* klassname = ((oop)obj)->klass()->name(); \ 58 if (klassname != NULL) { \ 59 bytes = (char*)klassname->bytes(); \ 60 len = klassname->utf8_length(); \ 61 } 62 63 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 64 { \ 65 if (DTraceMonitorProbes) { \ 66 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 67 HOTSPOT_MONITOR_WAIT(jtid, \ 68 (monitor), bytes, len, (millis)); \ 69 } \ 70 } 71 72 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 73 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 74 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 75 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 76 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 77 78 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 79 { \ 80 if (DTraceMonitorProbes) { \ 81 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 82 HOTSPOT_MONITOR_##probe(jtid, \ 83 (uintptr_t)(monitor), bytes, len); \ 84 } \ 85 } 86 87 #else // ndef DTRACE_ENABLED 88 89 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 90 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 91 92 #endif // ndef DTRACE_ENABLED 93 94 // Tunables ... 95 // The knob* variables are effectively final. Once set they should 96 // never be modified hence. Consider using __read_mostly with GCC. 97 98 int ObjectMonitor::Knob_ExitRelease = 0; 99 int ObjectMonitor::Knob_Verbose = 0; 100 int ObjectMonitor::Knob_VerifyInUse = 0; 101 int ObjectMonitor::Knob_VerifyMatch = 0; 102 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 103 static int Knob_LogSpins = 0; // enable jvmstat tally for spins 104 static int Knob_HandOff = 0; 105 static int Knob_ReportSettings = 0; 106 107 static int Knob_SpinBase = 0; // Floor AKA SpinMin 108 static int Knob_SpinBackOff = 0; // spin-loop backoff 109 static int Knob_CASPenalty = -1; // Penalty for failed CAS 110 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 111 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 112 static int Knob_SpinEarly = 1; 113 static int Knob_SuccEnabled = 1; // futile wake throttling 114 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 115 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 116 static int Knob_Bonus = 100; // spin success bonus 117 static int Knob_BonusB = 100; // spin success bonus 118 static int Knob_Penalty = 200; // spin failure penalty 119 static int Knob_Poverty = 1000; 120 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 121 static int Knob_FixedSpin = 0; 122 static int Knob_OState = 3; // Spinner checks thread state of _owner 123 static int Knob_UsePause = 1; 124 static int Knob_ExitPolicy = 0; 125 static int Knob_PreSpin = 10; // 20-100 likely better 126 static int Knob_ResetEvent = 0; 127 static int BackOffMask = 0; 128 129 static int Knob_FastHSSEC = 0; 130 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 131 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 132 static volatile int InitDone = 0; 133 134 #define TrySpin TrySpin_VaryDuration 135 136 // ----------------------------------------------------------------------------- 137 // Theory of operations -- Monitors lists, thread residency, etc: 138 // 139 // * A thread acquires ownership of a monitor by successfully 140 // CAS()ing the _owner field from null to non-null. 141 // 142 // * Invariant: A thread appears on at most one monitor list -- 143 // cxq, EntryList or WaitSet -- at any one time. 144 // 145 // * Contending threads "push" themselves onto the cxq with CAS 146 // and then spin/park. 147 // 148 // * After a contending thread eventually acquires the lock it must 149 // dequeue itself from either the EntryList or the cxq. 150 // 151 // * The exiting thread identifies and unparks an "heir presumptive" 152 // tentative successor thread on the EntryList. Critically, the 153 // exiting thread doesn't unlink the successor thread from the EntryList. 154 // After having been unparked, the wakee will recontend for ownership of 155 // the monitor. The successor (wakee) will either acquire the lock or 156 // re-park itself. 157 // 158 // Succession is provided for by a policy of competitive handoff. 159 // The exiting thread does _not_ grant or pass ownership to the 160 // successor thread. (This is also referred to as "handoff" succession"). 161 // Instead the exiting thread releases ownership and possibly wakes 162 // a successor, so the successor can (re)compete for ownership of the lock. 163 // If the EntryList is empty but the cxq is populated the exiting 164 // thread will drain the cxq into the EntryList. It does so by 165 // by detaching the cxq (installing null with CAS) and folding 166 // the threads from the cxq into the EntryList. The EntryList is 167 // doubly linked, while the cxq is singly linked because of the 168 // CAS-based "push" used to enqueue recently arrived threads (RATs). 169 // 170 // * Concurrency invariants: 171 // 172 // -- only the monitor owner may access or mutate the EntryList. 173 // The mutex property of the monitor itself protects the EntryList 174 // from concurrent interference. 175 // -- Only the monitor owner may detach the cxq. 176 // 177 // * The monitor entry list operations avoid locks, but strictly speaking 178 // they're not lock-free. Enter is lock-free, exit is not. 179 // For a description of 'Methods and apparatus providing non-blocking access 180 // to a resource,' see U.S. Pat. No. 7844973. 181 // 182 // * The cxq can have multiple concurrent "pushers" but only one concurrent 183 // detaching thread. This mechanism is immune from the ABA corruption. 184 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 185 // 186 // * Taken together, the cxq and the EntryList constitute or form a 187 // single logical queue of threads stalled trying to acquire the lock. 188 // We use two distinct lists to improve the odds of a constant-time 189 // dequeue operation after acquisition (in the ::enter() epilogue) and 190 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 191 // A key desideratum is to minimize queue & monitor metadata manipulation 192 // that occurs while holding the monitor lock -- that is, we want to 193 // minimize monitor lock holds times. Note that even a small amount of 194 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 195 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 196 // locks and monitor metadata. 197 // 198 // Cxq points to the set of Recently Arrived Threads attempting entry. 199 // Because we push threads onto _cxq with CAS, the RATs must take the form of 200 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 201 // the unlocking thread notices that EntryList is null but _cxq is != null. 202 // 203 // The EntryList is ordered by the prevailing queue discipline and 204 // can be organized in any convenient fashion, such as a doubly-linked list or 205 // a circular doubly-linked list. Critically, we want insert and delete operations 206 // to operate in constant-time. If we need a priority queue then something akin 207 // to Solaris' sleepq would work nicely. Viz., 208 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 209 // Queue discipline is enforced at ::exit() time, when the unlocking thread 210 // drains the cxq into the EntryList, and orders or reorders the threads on the 211 // EntryList accordingly. 212 // 213 // Barring "lock barging", this mechanism provides fair cyclic ordering, 214 // somewhat similar to an elevator-scan. 215 // 216 // * The monitor synchronization subsystem avoids the use of native 217 // synchronization primitives except for the narrow platform-specific 218 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 219 // the semantics of park-unpark. Put another way, this monitor implementation 220 // depends only on atomic operations and park-unpark. The monitor subsystem 221 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 222 // underlying OS manages the READY<->RUN transitions. 223 // 224 // * Waiting threads reside on the WaitSet list -- wait() puts 225 // the caller onto the WaitSet. 226 // 227 // * notify() or notifyAll() simply transfers threads from the WaitSet to 228 // either the EntryList or cxq. Subsequent exit() operations will 229 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 230 // it's likely the notifyee would simply impale itself on the lock held 231 // by the notifier. 232 // 233 // * An interesting alternative is to encode cxq as (List,LockByte) where 234 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 235 // variable, like _recursions, in the scheme. The threads or Events that form 236 // the list would have to be aligned in 256-byte addresses. A thread would 237 // try to acquire the lock or enqueue itself with CAS, but exiting threads 238 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 239 // Note that is is *not* word-tearing, but it does presume that full-word 240 // CAS operations are coherent with intermix with STB operations. That's true 241 // on most common processors. 242 // 243 // * See also http://blogs.sun.com/dave 244 245 246 // ----------------------------------------------------------------------------- 247 // Enter support 248 249 void ObjectMonitor::enter(TRAPS) { 250 // The following code is ordered to check the most common cases first 251 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 252 Thread * const Self = THREAD; 253 254 void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL); 255 if (cur == NULL) { 256 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 257 assert(_recursions == 0, "invariant"); 258 assert(_owner == Self, "invariant"); 259 return; 260 } 261 262 if (cur == Self) { 263 // TODO-FIXME: check for integer overflow! BUGID 6557169. 264 _recursions++; 265 return; 266 } 267 268 if (Self->is_lock_owned ((address)cur)) { 269 assert(_recursions == 0, "internal state error"); 270 _recursions = 1; 271 // Commute owner from a thread-specific on-stack BasicLockObject address to 272 // a full-fledged "Thread *". 273 _owner = Self; 274 return; 275 } 276 277 // We've encountered genuine contention. 278 assert(Self->_Stalled == 0, "invariant"); 279 Self->_Stalled = intptr_t(this); 280 281 // Try one round of spinning *before* enqueueing Self 282 // and before going through the awkward and expensive state 283 // transitions. The following spin is strictly optional ... 284 // Note that if we acquire the monitor from an initial spin 285 // we forgo posting JVMTI events and firing DTRACE probes. 286 if (Knob_SpinEarly && TrySpin (Self) > 0) { 287 assert(_owner == Self, "invariant"); 288 assert(_recursions == 0, "invariant"); 289 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 290 Self->_Stalled = 0; 291 return; 292 } 293 294 assert(_owner != Self, "invariant"); 295 assert(_succ != Self, "invariant"); 296 assert(Self->is_Java_thread(), "invariant"); 297 JavaThread * jt = (JavaThread *) Self; 298 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 299 assert(jt->thread_state() != _thread_blocked, "invariant"); 300 assert(this->object() != NULL, "invariant"); 301 assert(_count >= 0, "invariant"); 302 303 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 304 // Ensure the object-monitor relationship remains stable while there's contention. 305 Atomic::inc(&_count); 306 307 EventJavaMonitorEnter event; 308 309 { // Change java thread status to indicate blocked on monitor enter. 310 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 311 312 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 313 if (JvmtiExport::should_post_monitor_contended_enter()) { 314 JvmtiExport::post_monitor_contended_enter(jt, this); 315 316 // The current thread does not yet own the monitor and does not 317 // yet appear on any queues that would get it made the successor. 318 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 319 // handler cannot accidentally consume an unpark() meant for the 320 // ParkEvent associated with this ObjectMonitor. 321 } 322 323 OSThreadContendState osts(Self->osthread()); 324 ThreadBlockInVM tbivm(jt); 325 326 Self->set_current_pending_monitor(this); 327 328 // TODO-FIXME: change the following for(;;) loop to straight-line code. 329 for (;;) { 330 jt->set_suspend_equivalent(); 331 // cleared by handle_special_suspend_equivalent_condition() 332 // or java_suspend_self() 333 334 EnterI(THREAD); 335 336 if (!ExitSuspendEquivalent(jt)) break; 337 338 // We have acquired the contended monitor, but while we were 339 // waiting another thread suspended us. We don't want to enter 340 // the monitor while suspended because that would surprise the 341 // thread that suspended us. 342 // 343 _recursions = 0; 344 _succ = NULL; 345 exit(false, Self); 346 347 jt->java_suspend_self(); 348 } 349 Self->set_current_pending_monitor(NULL); 350 351 // We cleared the pending monitor info since we've just gotten past 352 // the enter-check-for-suspend dance and we now own the monitor free 353 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 354 // destructor can go to a safepoint at the end of this block. If we 355 // do a thread dump during that safepoint, then this thread will show 356 // as having "-locked" the monitor, but the OS and java.lang.Thread 357 // states will still report that the thread is blocked trying to 358 // acquire it. 359 } 360 361 Atomic::dec(&_count); 362 assert(_count >= 0, "invariant"); 363 Self->_Stalled = 0; 364 365 // Must either set _recursions = 0 or ASSERT _recursions == 0. 366 assert(_recursions == 0, "invariant"); 367 assert(_owner == Self, "invariant"); 368 assert(_succ != Self, "invariant"); 369 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 370 371 // The thread -- now the owner -- is back in vm mode. 372 // Report the glorious news via TI,DTrace and jvmstat. 373 // The probe effect is non-trivial. All the reportage occurs 374 // while we hold the monitor, increasing the length of the critical 375 // section. Amdahl's parallel speedup law comes vividly into play. 376 // 377 // Another option might be to aggregate the events (thread local or 378 // per-monitor aggregation) and defer reporting until a more opportune 379 // time -- such as next time some thread encounters contention but has 380 // yet to acquire the lock. While spinning that thread could 381 // spinning we could increment JVMStat counters, etc. 382 383 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 384 if (JvmtiExport::should_post_monitor_contended_entered()) { 385 JvmtiExport::post_monitor_contended_entered(jt, this); 386 387 // The current thread already owns the monitor and is not going to 388 // call park() for the remainder of the monitor enter protocol. So 389 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 390 // event handler consumed an unpark() issued by the thread that 391 // just exited the monitor. 392 } 393 394 if (event.should_commit()) { 395 event.set_klass(((oop)this->object())->klass()); 396 event.set_previousOwner((TYPE_THREAD)_previous_owner_tid); 397 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 398 event.commit(); 399 } 400 401 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 402 } 403 404 405 // Caveat: TryLock() is not necessarily serializing if it returns failure. 406 // Callers must compensate as needed. 407 408 int ObjectMonitor::TryLock(Thread * Self) { 409 void * own = _owner; 410 if (own != NULL) return 0; 411 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 412 // Either guarantee _recursions == 0 or set _recursions = 0. 413 assert(_recursions == 0, "invariant"); 414 assert(_owner == Self, "invariant"); 415 return 1; 416 } 417 // The lock had been free momentarily, but we lost the race to the lock. 418 // Interference -- the CAS failed. 419 // We can either return -1 or retry. 420 // Retry doesn't make as much sense because the lock was just acquired. 421 return -1; 422 } 423 424 #define MAX_RECHECK_INTERVAL 1000 425 426 void ObjectMonitor::EnterI(TRAPS) { 427 Thread * const Self = THREAD; 428 assert(Self->is_Java_thread(), "invariant"); 429 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 430 431 // Try the lock - TATAS 432 if (TryLock (Self) > 0) { 433 assert(_succ != Self, "invariant"); 434 assert(_owner == Self, "invariant"); 435 assert(_Responsible != Self, "invariant"); 436 return; 437 } 438 439 DeferredInitialize(); 440 441 // We try one round of spinning *before* enqueueing Self. 442 // 443 // If the _owner is ready but OFFPROC we could use a YieldTo() 444 // operation to donate the remainder of this thread's quantum 445 // to the owner. This has subtle but beneficial affinity 446 // effects. 447 448 if (TrySpin (Self) > 0) { 449 assert(_owner == Self, "invariant"); 450 assert(_succ != Self, "invariant"); 451 assert(_Responsible != Self, "invariant"); 452 return; 453 } 454 455 // The Spin failed -- Enqueue and park the thread ... 456 assert(_succ != Self, "invariant"); 457 assert(_owner != Self, "invariant"); 458 assert(_Responsible != Self, "invariant"); 459 460 // Enqueue "Self" on ObjectMonitor's _cxq. 461 // 462 // Node acts as a proxy for Self. 463 // As an aside, if were to ever rewrite the synchronization code mostly 464 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 465 // Java objects. This would avoid awkward lifecycle and liveness issues, 466 // as well as eliminate a subset of ABA issues. 467 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 468 469 ObjectWaiter node(Self); 470 Self->_ParkEvent->reset(); 471 node._prev = (ObjectWaiter *) 0xBAD; 472 node.TState = ObjectWaiter::TS_CXQ; 473 474 // Push "Self" onto the front of the _cxq. 475 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 476 // Note that spinning tends to reduce the rate at which threads 477 // enqueue and dequeue on EntryList|cxq. 478 ObjectWaiter * nxt; 479 for (;;) { 480 node._next = nxt = _cxq; 481 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break; 482 483 // Interference - the CAS failed because _cxq changed. Just retry. 484 // As an optional optimization we retry the lock. 485 if (TryLock (Self) > 0) { 486 assert(_succ != Self, "invariant"); 487 assert(_owner == Self, "invariant"); 488 assert(_Responsible != Self, "invariant"); 489 return; 490 } 491 } 492 493 // Check for cxq|EntryList edge transition to non-null. This indicates 494 // the onset of contention. While contention persists exiting threads 495 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 496 // operations revert to the faster 1-0 mode. This enter operation may interleave 497 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 498 // arrange for one of the contending thread to use a timed park() operations 499 // to detect and recover from the race. (Stranding is form of progress failure 500 // where the monitor is unlocked but all the contending threads remain parked). 501 // That is, at least one of the contended threads will periodically poll _owner. 502 // One of the contending threads will become the designated "Responsible" thread. 503 // The Responsible thread uses a timed park instead of a normal indefinite park 504 // operation -- it periodically wakes and checks for and recovers from potential 505 // strandings admitted by 1-0 exit operations. We need at most one Responsible 506 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 507 // be responsible for a monitor. 508 // 509 // Currently, one of the contended threads takes on the added role of "Responsible". 510 // A viable alternative would be to use a dedicated "stranding checker" thread 511 // that periodically iterated over all the threads (or active monitors) and unparked 512 // successors where there was risk of stranding. This would help eliminate the 513 // timer scalability issues we see on some platforms as we'd only have one thread 514 // -- the checker -- parked on a timer. 515 516 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 517 // Try to assume the role of responsible thread for the monitor. 518 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 519 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 520 } 521 522 // The lock might have been released while this thread was occupied queueing 523 // itself onto _cxq. To close the race and avoid "stranding" and 524 // progress-liveness failure we must resample-retry _owner before parking. 525 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 526 // In this case the ST-MEMBAR is accomplished with CAS(). 527 // 528 // TODO: Defer all thread state transitions until park-time. 529 // Since state transitions are heavy and inefficient we'd like 530 // to defer the state transitions until absolutely necessary, 531 // and in doing so avoid some transitions ... 532 533 TEVENT(Inflated enter - Contention); 534 int nWakeups = 0; 535 int recheckInterval = 1; 536 537 for (;;) { 538 539 if (TryLock(Self) > 0) break; 540 assert(_owner != Self, "invariant"); 541 542 if ((SyncFlags & 2) && _Responsible == NULL) { 543 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 544 } 545 546 // park self 547 if (_Responsible == Self || (SyncFlags & 1)) { 548 TEVENT(Inflated enter - park TIMED); 549 Self->_ParkEvent->park((jlong) recheckInterval); 550 // Increase the recheckInterval, but clamp the value. 551 recheckInterval *= 8; 552 if (recheckInterval > MAX_RECHECK_INTERVAL) { 553 recheckInterval = MAX_RECHECK_INTERVAL; 554 } 555 } else { 556 TEVENT(Inflated enter - park UNTIMED); 557 Self->_ParkEvent->park(); 558 } 559 560 if (TryLock(Self) > 0) break; 561 562 // The lock is still contested. 563 // Keep a tally of the # of futile wakeups. 564 // Note that the counter is not protected by a lock or updated by atomics. 565 // That is by design - we trade "lossy" counters which are exposed to 566 // races during updates for a lower probe effect. 567 TEVENT(Inflated enter - Futile wakeup); 568 // This PerfData object can be used in parallel with a safepoint. 569 // See the work around in PerfDataManager::destroy(). 570 OM_PERFDATA_OP(FutileWakeups, inc()); 571 ++nWakeups; 572 573 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 574 // We can defer clearing _succ until after the spin completes 575 // TrySpin() must tolerate being called with _succ == Self. 576 // Try yet another round of adaptive spinning. 577 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 578 579 // We can find that we were unpark()ed and redesignated _succ while 580 // we were spinning. That's harmless. If we iterate and call park(), 581 // park() will consume the event and return immediately and we'll 582 // just spin again. This pattern can repeat, leaving _succ to simply 583 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 584 // Alternately, we can sample fired() here, and if set, forgo spinning 585 // in the next iteration. 586 587 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 588 Self->_ParkEvent->reset(); 589 OrderAccess::fence(); 590 } 591 if (_succ == Self) _succ = NULL; 592 593 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 594 OrderAccess::fence(); 595 } 596 597 // Egress : 598 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 599 // Normally we'll find Self on the EntryList . 600 // From the perspective of the lock owner (this thread), the 601 // EntryList is stable and cxq is prepend-only. 602 // The head of cxq is volatile but the interior is stable. 603 // In addition, Self.TState is stable. 604 605 assert(_owner == Self, "invariant"); 606 assert(object() != NULL, "invariant"); 607 // I'd like to write: 608 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 609 // but as we're at a safepoint that's not safe. 610 611 UnlinkAfterAcquire(Self, &node); 612 if (_succ == Self) _succ = NULL; 613 614 assert(_succ != Self, "invariant"); 615 if (_Responsible == Self) { 616 _Responsible = NULL; 617 OrderAccess::fence(); // Dekker pivot-point 618 619 // We may leave threads on cxq|EntryList without a designated 620 // "Responsible" thread. This is benign. When this thread subsequently 621 // exits the monitor it can "see" such preexisting "old" threads -- 622 // threads that arrived on the cxq|EntryList before the fence, above -- 623 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 624 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 625 // non-null and elect a new "Responsible" timer thread. 626 // 627 // This thread executes: 628 // ST Responsible=null; MEMBAR (in enter epilogue - here) 629 // LD cxq|EntryList (in subsequent exit) 630 // 631 // Entering threads in the slow/contended path execute: 632 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 633 // The (ST cxq; MEMBAR) is accomplished with CAS(). 634 // 635 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 636 // exit operation from floating above the ST Responsible=null. 637 } 638 639 // We've acquired ownership with CAS(). 640 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 641 // But since the CAS() this thread may have also stored into _succ, 642 // EntryList, cxq or Responsible. These meta-data updates must be 643 // visible __before this thread subsequently drops the lock. 644 // Consider what could occur if we didn't enforce this constraint -- 645 // STs to monitor meta-data and user-data could reorder with (become 646 // visible after) the ST in exit that drops ownership of the lock. 647 // Some other thread could then acquire the lock, but observe inconsistent 648 // or old monitor meta-data and heap data. That violates the JMM. 649 // To that end, the 1-0 exit() operation must have at least STST|LDST 650 // "release" barrier semantics. Specifically, there must be at least a 651 // STST|LDST barrier in exit() before the ST of null into _owner that drops 652 // the lock. The barrier ensures that changes to monitor meta-data and data 653 // protected by the lock will be visible before we release the lock, and 654 // therefore before some other thread (CPU) has a chance to acquire the lock. 655 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 656 // 657 // Critically, any prior STs to _succ or EntryList must be visible before 658 // the ST of null into _owner in the *subsequent* (following) corresponding 659 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 660 // execute a serializing instruction. 661 662 if (SyncFlags & 8) { 663 OrderAccess::fence(); 664 } 665 return; 666 } 667 668 // ReenterI() is a specialized inline form of the latter half of the 669 // contended slow-path from EnterI(). We use ReenterI() only for 670 // monitor reentry in wait(). 671 // 672 // In the future we should reconcile EnterI() and ReenterI(), adding 673 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 674 // loop accordingly. 675 676 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 677 assert(Self != NULL, "invariant"); 678 assert(SelfNode != NULL, "invariant"); 679 assert(SelfNode->_thread == Self, "invariant"); 680 assert(_waiters > 0, "invariant"); 681 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 682 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 683 JavaThread * jt = (JavaThread *) Self; 684 685 int nWakeups = 0; 686 for (;;) { 687 ObjectWaiter::TStates v = SelfNode->TState; 688 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 689 assert(_owner != Self, "invariant"); 690 691 if (TryLock(Self) > 0) break; 692 if (TrySpin(Self) > 0) break; 693 694 TEVENT(Wait Reentry - parking); 695 696 // State transition wrappers around park() ... 697 // ReenterI() wisely defers state transitions until 698 // it's clear we must park the thread. 699 { 700 OSThreadContendState osts(Self->osthread()); 701 ThreadBlockInVM tbivm(jt); 702 703 // cleared by handle_special_suspend_equivalent_condition() 704 // or java_suspend_self() 705 jt->set_suspend_equivalent(); 706 if (SyncFlags & 1) { 707 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL); 708 } else { 709 Self->_ParkEvent->park(); 710 } 711 712 // were we externally suspended while we were waiting? 713 for (;;) { 714 if (!ExitSuspendEquivalent(jt)) break; 715 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 716 jt->java_suspend_self(); 717 jt->set_suspend_equivalent(); 718 } 719 } 720 721 // Try again, but just so we distinguish between futile wakeups and 722 // successful wakeups. The following test isn't algorithmically 723 // necessary, but it helps us maintain sensible statistics. 724 if (TryLock(Self) > 0) break; 725 726 // The lock is still contested. 727 // Keep a tally of the # of futile wakeups. 728 // Note that the counter is not protected by a lock or updated by atomics. 729 // That is by design - we trade "lossy" counters which are exposed to 730 // races during updates for a lower probe effect. 731 TEVENT(Wait Reentry - futile wakeup); 732 ++nWakeups; 733 734 // Assuming this is not a spurious wakeup we'll normally 735 // find that _succ == Self. 736 if (_succ == Self) _succ = NULL; 737 738 // Invariant: after clearing _succ a contending thread 739 // *must* retry _owner before parking. 740 OrderAccess::fence(); 741 742 // This PerfData object can be used in parallel with a safepoint. 743 // See the work around in PerfDataManager::destroy(). 744 OM_PERFDATA_OP(FutileWakeups, inc()); 745 } 746 747 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 748 // Normally we'll find Self on the EntryList. 749 // Unlinking from the EntryList is constant-time and atomic-free. 750 // From the perspective of the lock owner (this thread), the 751 // EntryList is stable and cxq is prepend-only. 752 // The head of cxq is volatile but the interior is stable. 753 // In addition, Self.TState is stable. 754 755 assert(_owner == Self, "invariant"); 756 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 757 UnlinkAfterAcquire(Self, SelfNode); 758 if (_succ == Self) _succ = NULL; 759 assert(_succ != Self, "invariant"); 760 SelfNode->TState = ObjectWaiter::TS_RUN; 761 OrderAccess::fence(); // see comments at the end of EnterI() 762 } 763 764 // By convention we unlink a contending thread from EntryList|cxq immediately 765 // after the thread acquires the lock in ::enter(). Equally, we could defer 766 // unlinking the thread until ::exit()-time. 767 768 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 769 assert(_owner == Self, "invariant"); 770 assert(SelfNode->_thread == Self, "invariant"); 771 772 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 773 // Normal case: remove Self from the DLL EntryList . 774 // This is a constant-time operation. 775 ObjectWaiter * nxt = SelfNode->_next; 776 ObjectWaiter * prv = SelfNode->_prev; 777 if (nxt != NULL) nxt->_prev = prv; 778 if (prv != NULL) prv->_next = nxt; 779 if (SelfNode == _EntryList) _EntryList = nxt; 780 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 781 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 782 TEVENT(Unlink from EntryList); 783 } else { 784 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 785 // Inopportune interleaving -- Self is still on the cxq. 786 // This usually means the enqueue of self raced an exiting thread. 787 // Normally we'll find Self near the front of the cxq, so 788 // dequeueing is typically fast. If needbe we can accelerate 789 // this with some MCS/CHL-like bidirectional list hints and advisory 790 // back-links so dequeueing from the interior will normally operate 791 // in constant-time. 792 // Dequeue Self from either the head (with CAS) or from the interior 793 // with a linear-time scan and normal non-atomic memory operations. 794 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 795 // and then unlink Self from EntryList. We have to drain eventually, 796 // so it might as well be now. 797 798 ObjectWaiter * v = _cxq; 799 assert(v != NULL, "invariant"); 800 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 801 // The CAS above can fail from interference IFF a "RAT" arrived. 802 // In that case Self must be in the interior and can no longer be 803 // at the head of cxq. 804 if (v == SelfNode) { 805 assert(_cxq != v, "invariant"); 806 v = _cxq; // CAS above failed - start scan at head of list 807 } 808 ObjectWaiter * p; 809 ObjectWaiter * q = NULL; 810 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 811 q = p; 812 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 813 } 814 assert(v != SelfNode, "invariant"); 815 assert(p == SelfNode, "Node not found on cxq"); 816 assert(p != _cxq, "invariant"); 817 assert(q != NULL, "invariant"); 818 assert(q->_next == p, "invariant"); 819 q->_next = p->_next; 820 } 821 TEVENT(Unlink from cxq); 822 } 823 824 #ifdef ASSERT 825 // Diagnostic hygiene ... 826 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 827 SelfNode->_next = (ObjectWaiter *) 0xBAD; 828 SelfNode->TState = ObjectWaiter::TS_RUN; 829 #endif 830 } 831 832 // ----------------------------------------------------------------------------- 833 // Exit support 834 // 835 // exit() 836 // ~~~~~~ 837 // Note that the collector can't reclaim the objectMonitor or deflate 838 // the object out from underneath the thread calling ::exit() as the 839 // thread calling ::exit() never transitions to a stable state. 840 // This inhibits GC, which in turn inhibits asynchronous (and 841 // inopportune) reclamation of "this". 842 // 843 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 844 // There's one exception to the claim above, however. EnterI() can call 845 // exit() to drop a lock if the acquirer has been externally suspended. 846 // In that case exit() is called with _thread_state as _thread_blocked, 847 // but the monitor's _count field is > 0, which inhibits reclamation. 848 // 849 // 1-0 exit 850 // ~~~~~~~~ 851 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 852 // the fast-path operators have been optimized so the common ::exit() 853 // operation is 1-0. See i486.ad fast_unlock(), for instance. 854 // The code emitted by fast_unlock() elides the usual MEMBAR. This 855 // greatly improves latency -- MEMBAR and CAS having considerable local 856 // latency on modern processors -- but at the cost of "stranding". Absent the 857 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 858 // ::enter() path, resulting in the entering thread being stranding 859 // and a progress-liveness failure. Stranding is extremely rare. 860 // We use timers (timed park operations) & periodic polling to detect 861 // and recover from stranding. Potentially stranded threads periodically 862 // wake up and poll the lock. See the usage of the _Responsible variable. 863 // 864 // The CAS() in enter provides for safety and exclusion, while the CAS or 865 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 866 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 867 // We detect and recover from stranding with timers. 868 // 869 // If a thread transiently strands it'll park until (a) another 870 // thread acquires the lock and then drops the lock, at which time the 871 // exiting thread will notice and unpark the stranded thread, or, (b) 872 // the timer expires. If the lock is high traffic then the stranding latency 873 // will be low due to (a). If the lock is low traffic then the odds of 874 // stranding are lower, although the worst-case stranding latency 875 // is longer. Critically, we don't want to put excessive load in the 876 // platform's timer subsystem. We want to minimize both the timer injection 877 // rate (timers created/sec) as well as the number of timers active at 878 // any one time. (more precisely, we want to minimize timer-seconds, which is 879 // the integral of the # of active timers at any instant over time). 880 // Both impinge on OS scalability. Given that, at most one thread parked on 881 // a monitor will use a timer. 882 // 883 // There is also the risk of a futile wake-up. If we drop the lock 884 // another thread can reacquire the lock immediately, and we can 885 // then wake a thread unnecessarily. This is benign, and we've 886 // structured the code so the windows are short and the frequency 887 // of such futile wakups is low. 888 889 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 890 Thread * const Self = THREAD; 891 if (THREAD != _owner) { 892 if (THREAD->is_lock_owned((address) _owner)) { 893 // Transmute _owner from a BasicLock pointer to a Thread address. 894 // We don't need to hold _mutex for this transition. 895 // Non-null to Non-null is safe as long as all readers can 896 // tolerate either flavor. 897 assert(_recursions == 0, "invariant"); 898 _owner = THREAD; 899 _recursions = 0; 900 } else { 901 // Apparent unbalanced locking ... 902 // Naively we'd like to throw IllegalMonitorStateException. 903 // As a practical matter we can neither allocate nor throw an 904 // exception as ::exit() can be called from leaf routines. 905 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 906 // Upon deeper reflection, however, in a properly run JVM the only 907 // way we should encounter this situation is in the presence of 908 // unbalanced JNI locking. TODO: CheckJNICalls. 909 // See also: CR4414101 910 TEVENT(Exit - Throw IMSX); 911 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 912 return; 913 } 914 } 915 916 if (_recursions != 0) { 917 _recursions--; // this is simple recursive enter 918 TEVENT(Inflated exit - recursive); 919 return; 920 } 921 922 // Invariant: after setting Responsible=null an thread must execute 923 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 924 if ((SyncFlags & 4) == 0) { 925 _Responsible = NULL; 926 } 927 928 #if INCLUDE_TRACE 929 // get the owner's thread id for the MonitorEnter event 930 // if it is enabled and the thread isn't suspended 931 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 932 _previous_owner_tid = THREAD_TRACE_ID(Self); 933 } 934 #endif 935 936 for (;;) { 937 assert(THREAD == _owner, "invariant"); 938 939 940 if (Knob_ExitPolicy == 0) { 941 // release semantics: prior loads and stores from within the critical section 942 // must not float (reorder) past the following store that drops the lock. 943 // On SPARC that requires MEMBAR #loadstore|#storestore. 944 // But of course in TSO #loadstore|#storestore is not required. 945 // I'd like to write one of the following: 946 // A. OrderAccess::release() ; _owner = NULL 947 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 948 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 949 // store into a _dummy variable. That store is not needed, but can result 950 // in massive wasteful coherency traffic on classic SMP systems. 951 // Instead, I use release_store(), which is implemented as just a simple 952 // ST on x64, x86 and SPARC. 953 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 954 OrderAccess::storeload(); // See if we need to wake a successor 955 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 956 TEVENT(Inflated exit - simple egress); 957 return; 958 } 959 TEVENT(Inflated exit - complex egress); 960 // Other threads are blocked trying to acquire the lock. 961 962 // Normally the exiting thread is responsible for ensuring succession, 963 // but if other successors are ready or other entering threads are spinning 964 // then this thread can simply store NULL into _owner and exit without 965 // waking a successor. The existence of spinners or ready successors 966 // guarantees proper succession (liveness). Responsibility passes to the 967 // ready or running successors. The exiting thread delegates the duty. 968 // More precisely, if a successor already exists this thread is absolved 969 // of the responsibility of waking (unparking) one. 970 // 971 // The _succ variable is critical to reducing futile wakeup frequency. 972 // _succ identifies the "heir presumptive" thread that has been made 973 // ready (unparked) but that has not yet run. We need only one such 974 // successor thread to guarantee progress. 975 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 976 // section 3.3 "Futile Wakeup Throttling" for details. 977 // 978 // Note that spinners in Enter() also set _succ non-null. 979 // In the current implementation spinners opportunistically set 980 // _succ so that exiting threads might avoid waking a successor. 981 // Another less appealing alternative would be for the exiting thread 982 // to drop the lock and then spin briefly to see if a spinner managed 983 // to acquire the lock. If so, the exiting thread could exit 984 // immediately without waking a successor, otherwise the exiting 985 // thread would need to dequeue and wake a successor. 986 // (Note that we'd need to make the post-drop spin short, but no 987 // shorter than the worst-case round-trip cache-line migration time. 988 // The dropped lock needs to become visible to the spinner, and then 989 // the acquisition of the lock by the spinner must become visible to 990 // the exiting thread). 991 992 // It appears that an heir-presumptive (successor) must be made ready. 993 // Only the current lock owner can manipulate the EntryList or 994 // drain _cxq, so we need to reacquire the lock. If we fail 995 // to reacquire the lock the responsibility for ensuring succession 996 // falls to the new owner. 997 // 998 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 999 return; 1000 } 1001 TEVENT(Exit - Reacquired); 1002 } else { 1003 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1004 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1005 OrderAccess::storeload(); 1006 // Ratify the previously observed values. 1007 if (_cxq == NULL || _succ != NULL) { 1008 TEVENT(Inflated exit - simple egress); 1009 return; 1010 } 1011 1012 // inopportune interleaving -- the exiting thread (this thread) 1013 // in the fast-exit path raced an entering thread in the slow-enter 1014 // path. 1015 // We have two choices: 1016 // A. Try to reacquire the lock. 1017 // If the CAS() fails return immediately, otherwise 1018 // we either restart/rerun the exit operation, or simply 1019 // fall-through into the code below which wakes a successor. 1020 // B. If the elements forming the EntryList|cxq are TSM 1021 // we could simply unpark() the lead thread and return 1022 // without having set _succ. 1023 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1024 TEVENT(Inflated exit - reacquired succeeded); 1025 return; 1026 } 1027 TEVENT(Inflated exit - reacquired failed); 1028 } else { 1029 TEVENT(Inflated exit - complex egress); 1030 } 1031 } 1032 1033 guarantee(_owner == THREAD, "invariant"); 1034 1035 ObjectWaiter * w = NULL; 1036 int QMode = Knob_QMode; 1037 1038 if (QMode == 2 && _cxq != NULL) { 1039 // QMode == 2 : cxq has precedence over EntryList. 1040 // Try to directly wake a successor from the cxq. 1041 // If successful, the successor will need to unlink itself from cxq. 1042 w = _cxq; 1043 assert(w != NULL, "invariant"); 1044 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1045 ExitEpilog(Self, w); 1046 return; 1047 } 1048 1049 if (QMode == 3 && _cxq != NULL) { 1050 // Aggressively drain cxq into EntryList at the first opportunity. 1051 // This policy ensure that recently-run threads live at the head of EntryList. 1052 // Drain _cxq into EntryList - bulk transfer. 1053 // First, detach _cxq. 1054 // The following loop is tantamount to: w = swap(&cxq, NULL) 1055 w = _cxq; 1056 for (;;) { 1057 assert(w != NULL, "Invariant"); 1058 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1059 if (u == w) break; 1060 w = u; 1061 } 1062 assert(w != NULL, "invariant"); 1063 1064 ObjectWaiter * q = NULL; 1065 ObjectWaiter * p; 1066 for (p = w; p != NULL; p = p->_next) { 1067 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1068 p->TState = ObjectWaiter::TS_ENTER; 1069 p->_prev = q; 1070 q = p; 1071 } 1072 1073 // Append the RATs to the EntryList 1074 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1075 ObjectWaiter * Tail; 1076 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1077 Tail = Tail->_next) 1078 /* empty */; 1079 if (Tail == NULL) { 1080 _EntryList = w; 1081 } else { 1082 Tail->_next = w; 1083 w->_prev = Tail; 1084 } 1085 1086 // Fall thru into code that tries to wake a successor from EntryList 1087 } 1088 1089 if (QMode == 4 && _cxq != NULL) { 1090 // Aggressively drain cxq into EntryList at the first opportunity. 1091 // This policy ensure that recently-run threads live at the head of EntryList. 1092 1093 // Drain _cxq into EntryList - bulk transfer. 1094 // First, detach _cxq. 1095 // The following loop is tantamount to: w = swap(&cxq, NULL) 1096 w = _cxq; 1097 for (;;) { 1098 assert(w != NULL, "Invariant"); 1099 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1100 if (u == w) break; 1101 w = u; 1102 } 1103 assert(w != NULL, "invariant"); 1104 1105 ObjectWaiter * q = NULL; 1106 ObjectWaiter * p; 1107 for (p = w; p != NULL; p = p->_next) { 1108 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1109 p->TState = ObjectWaiter::TS_ENTER; 1110 p->_prev = q; 1111 q = p; 1112 } 1113 1114 // Prepend the RATs to the EntryList 1115 if (_EntryList != NULL) { 1116 q->_next = _EntryList; 1117 _EntryList->_prev = q; 1118 } 1119 _EntryList = w; 1120 1121 // Fall thru into code that tries to wake a successor from EntryList 1122 } 1123 1124 w = _EntryList; 1125 if (w != NULL) { 1126 // I'd like to write: guarantee (w->_thread != Self). 1127 // But in practice an exiting thread may find itself on the EntryList. 1128 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1129 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1130 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1131 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1132 // release the lock "O". T2 resumes immediately after the ST of null into 1133 // _owner, above. T2 notices that the EntryList is populated, so it 1134 // reacquires the lock and then finds itself on the EntryList. 1135 // Given all that, we have to tolerate the circumstance where "w" is 1136 // associated with Self. 1137 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1138 ExitEpilog(Self, w); 1139 return; 1140 } 1141 1142 // If we find that both _cxq and EntryList are null then just 1143 // re-run the exit protocol from the top. 1144 w = _cxq; 1145 if (w == NULL) continue; 1146 1147 // Drain _cxq into EntryList - bulk transfer. 1148 // First, detach _cxq. 1149 // The following loop is tantamount to: w = swap(&cxq, NULL) 1150 for (;;) { 1151 assert(w != NULL, "Invariant"); 1152 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1153 if (u == w) break; 1154 w = u; 1155 } 1156 TEVENT(Inflated exit - drain cxq into EntryList); 1157 1158 assert(w != NULL, "invariant"); 1159 assert(_EntryList == NULL, "invariant"); 1160 1161 // Convert the LIFO SLL anchored by _cxq into a DLL. 1162 // The list reorganization step operates in O(LENGTH(w)) time. 1163 // It's critical that this step operate quickly as 1164 // "Self" still holds the outer-lock, restricting parallelism 1165 // and effectively lengthening the critical section. 1166 // Invariant: s chases t chases u. 1167 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1168 // we have faster access to the tail. 1169 1170 if (QMode == 1) { 1171 // QMode == 1 : drain cxq to EntryList, reversing order 1172 // We also reverse the order of the list. 1173 ObjectWaiter * s = NULL; 1174 ObjectWaiter * t = w; 1175 ObjectWaiter * u = NULL; 1176 while (t != NULL) { 1177 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1178 t->TState = ObjectWaiter::TS_ENTER; 1179 u = t->_next; 1180 t->_prev = u; 1181 t->_next = s; 1182 s = t; 1183 t = u; 1184 } 1185 _EntryList = s; 1186 assert(s != NULL, "invariant"); 1187 } else { 1188 // QMode == 0 or QMode == 2 1189 _EntryList = w; 1190 ObjectWaiter * q = NULL; 1191 ObjectWaiter * p; 1192 for (p = w; p != NULL; p = p->_next) { 1193 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1194 p->TState = ObjectWaiter::TS_ENTER; 1195 p->_prev = q; 1196 q = p; 1197 } 1198 } 1199 1200 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1201 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1202 1203 // See if we can abdicate to a spinner instead of waking a thread. 1204 // A primary goal of the implementation is to reduce the 1205 // context-switch rate. 1206 if (_succ != NULL) continue; 1207 1208 w = _EntryList; 1209 if (w != NULL) { 1210 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1211 ExitEpilog(Self, w); 1212 return; 1213 } 1214 } 1215 } 1216 1217 // ExitSuspendEquivalent: 1218 // A faster alternate to handle_special_suspend_equivalent_condition() 1219 // 1220 // handle_special_suspend_equivalent_condition() unconditionally 1221 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1222 // operations have high latency. Note that in ::enter() we call HSSEC 1223 // while holding the monitor, so we effectively lengthen the critical sections. 1224 // 1225 // There are a number of possible solutions: 1226 // 1227 // A. To ameliorate the problem we might also defer state transitions 1228 // to as late as possible -- just prior to parking. 1229 // Given that, we'd call HSSEC after having returned from park(), 1230 // but before attempting to acquire the monitor. This is only a 1231 // partial solution. It avoids calling HSSEC while holding the 1232 // monitor (good), but it still increases successor reacquisition latency -- 1233 // the interval between unparking a successor and the time the successor 1234 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1235 // If we use this technique we can also avoid EnterI()-exit() loop 1236 // in ::enter() where we iteratively drop the lock and then attempt 1237 // to reacquire it after suspending. 1238 // 1239 // B. In the future we might fold all the suspend bits into a 1240 // composite per-thread suspend flag and then update it with CAS(). 1241 // Alternately, a Dekker-like mechanism with multiple variables 1242 // would suffice: 1243 // ST Self->_suspend_equivalent = false 1244 // MEMBAR 1245 // LD Self_>_suspend_flags 1246 // 1247 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1248 // with a more efficient implementation, the need to use "FastHSSEC" has 1249 // decreased. - Dave 1250 1251 1252 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1253 const int Mode = Knob_FastHSSEC; 1254 if (Mode && !jSelf->is_external_suspend()) { 1255 assert(jSelf->is_suspend_equivalent(), "invariant"); 1256 jSelf->clear_suspend_equivalent(); 1257 if (2 == Mode) OrderAccess::storeload(); 1258 if (!jSelf->is_external_suspend()) return false; 1259 // We raced a suspension -- fall thru into the slow path 1260 TEVENT(ExitSuspendEquivalent - raced); 1261 jSelf->set_suspend_equivalent(); 1262 } 1263 return jSelf->handle_special_suspend_equivalent_condition(); 1264 } 1265 1266 1267 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1268 assert(_owner == Self, "invariant"); 1269 1270 // Exit protocol: 1271 // 1. ST _succ = wakee 1272 // 2. membar #loadstore|#storestore; 1273 // 2. ST _owner = NULL 1274 // 3. unpark(wakee) 1275 1276 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1277 ParkEvent * Trigger = Wakee->_event; 1278 1279 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1280 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1281 // out-of-scope (non-extant). 1282 Wakee = NULL; 1283 1284 // Drop the lock 1285 OrderAccess::release_store_ptr(&_owner, NULL); 1286 OrderAccess::fence(); // ST _owner vs LD in unpark() 1287 1288 if (SafepointSynchronize::do_call_back()) { 1289 TEVENT(unpark before SAFEPOINT); 1290 } 1291 1292 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1293 Trigger->unpark(); 1294 1295 // Maintain stats and report events to JVMTI 1296 OM_PERFDATA_OP(Parks, inc()); 1297 } 1298 1299 1300 // ----------------------------------------------------------------------------- 1301 // Class Loader deadlock handling. 1302 // 1303 // complete_exit exits a lock returning recursion count 1304 // complete_exit/reenter operate as a wait without waiting 1305 // complete_exit requires an inflated monitor 1306 // The _owner field is not always the Thread addr even with an 1307 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1308 // thread due to contention. 1309 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1310 Thread * const Self = THREAD; 1311 assert(Self->is_Java_thread(), "Must be Java thread!"); 1312 JavaThread *jt = (JavaThread *)THREAD; 1313 1314 DeferredInitialize(); 1315 1316 if (THREAD != _owner) { 1317 if (THREAD->is_lock_owned ((address)_owner)) { 1318 assert(_recursions == 0, "internal state error"); 1319 _owner = THREAD; // Convert from basiclock addr to Thread addr 1320 _recursions = 0; 1321 } 1322 } 1323 1324 guarantee(Self == _owner, "complete_exit not owner"); 1325 intptr_t save = _recursions; // record the old recursion count 1326 _recursions = 0; // set the recursion level to be 0 1327 exit(true, Self); // exit the monitor 1328 guarantee(_owner != Self, "invariant"); 1329 return save; 1330 } 1331 1332 // reenter() enters a lock and sets recursion count 1333 // complete_exit/reenter operate as a wait without waiting 1334 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1335 Thread * const Self = THREAD; 1336 assert(Self->is_Java_thread(), "Must be Java thread!"); 1337 JavaThread *jt = (JavaThread *)THREAD; 1338 1339 guarantee(_owner != Self, "reenter already owner"); 1340 enter(THREAD); // enter the monitor 1341 guarantee(_recursions == 0, "reenter recursion"); 1342 _recursions = recursions; 1343 return; 1344 } 1345 1346 1347 // ----------------------------------------------------------------------------- 1348 // A macro is used below because there may already be a pending 1349 // exception which should not abort the execution of the routines 1350 // which use this (which is why we don't put this into check_slow and 1351 // call it with a CHECK argument). 1352 1353 #define CHECK_OWNER() \ 1354 do { \ 1355 if (THREAD != _owner) { \ 1356 if (THREAD->is_lock_owned((address) _owner)) { \ 1357 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1358 _recursions = 0; \ 1359 } else { \ 1360 TEVENT(Throw IMSX); \ 1361 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1362 } \ 1363 } \ 1364 } while (false) 1365 1366 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1367 // TODO-FIXME: remove check_slow() -- it's likely dead. 1368 1369 void ObjectMonitor::check_slow(TRAPS) { 1370 TEVENT(check_slow - throw IMSX); 1371 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1372 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1373 } 1374 1375 static int Adjust(volatile int * adr, int dx) { 1376 int v; 1377 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1378 return v; 1379 } 1380 1381 // helper method for posting a monitor wait event 1382 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1383 jlong notifier_tid, 1384 jlong timeout, 1385 bool timedout) { 1386 assert(event != NULL, "invariant"); 1387 event->set_klass(((oop)this->object())->klass()); 1388 event->set_timeout(timeout); 1389 event->set_address((TYPE_ADDRESS)this->object_addr()); 1390 event->set_notifier(notifier_tid); 1391 event->set_timedOut(timedout); 1392 event->commit(); 1393 } 1394 1395 // ----------------------------------------------------------------------------- 1396 // Wait/Notify/NotifyAll 1397 // 1398 // Note: a subset of changes to ObjectMonitor::wait() 1399 // will need to be replicated in complete_exit 1400 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1401 Thread * const Self = THREAD; 1402 assert(Self->is_Java_thread(), "Must be Java thread!"); 1403 JavaThread *jt = (JavaThread *)THREAD; 1404 1405 DeferredInitialize(); 1406 1407 // Throw IMSX or IEX. 1408 CHECK_OWNER(); 1409 1410 EventJavaMonitorWait event; 1411 1412 // check for a pending interrupt 1413 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1414 // post monitor waited event. Note that this is past-tense, we are done waiting. 1415 if (JvmtiExport::should_post_monitor_waited()) { 1416 // Note: 'false' parameter is passed here because the 1417 // wait was not timed out due to thread interrupt. 1418 JvmtiExport::post_monitor_waited(jt, this, false); 1419 1420 // In this short circuit of the monitor wait protocol, the 1421 // current thread never drops ownership of the monitor and 1422 // never gets added to the wait queue so the current thread 1423 // cannot be made the successor. This means that the 1424 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1425 // consume an unpark() meant for the ParkEvent associated with 1426 // this ObjectMonitor. 1427 } 1428 if (event.should_commit()) { 1429 post_monitor_wait_event(&event, 0, millis, false); 1430 } 1431 TEVENT(Wait - Throw IEX); 1432 THROW(vmSymbols::java_lang_InterruptedException()); 1433 return; 1434 } 1435 1436 TEVENT(Wait); 1437 1438 assert(Self->_Stalled == 0, "invariant"); 1439 Self->_Stalled = intptr_t(this); 1440 jt->set_current_waiting_monitor(this); 1441 1442 // create a node to be put into the queue 1443 // Critically, after we reset() the event but prior to park(), we must check 1444 // for a pending interrupt. 1445 ObjectWaiter node(Self); 1446 node.TState = ObjectWaiter::TS_WAIT; 1447 Self->_ParkEvent->reset(); 1448 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1449 1450 // Enter the waiting queue, which is a circular doubly linked list in this case 1451 // but it could be a priority queue or any data structure. 1452 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1453 // by the the owner of the monitor *except* in the case where park() 1454 // returns because of a timeout of interrupt. Contention is exceptionally rare 1455 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1456 1457 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1458 AddWaiter(&node); 1459 Thread::SpinRelease(&_WaitSetLock); 1460 1461 if ((SyncFlags & 4) == 0) { 1462 _Responsible = NULL; 1463 } 1464 intptr_t save = _recursions; // record the old recursion count 1465 _waiters++; // increment the number of waiters 1466 _recursions = 0; // set the recursion level to be 1 1467 exit(true, Self); // exit the monitor 1468 guarantee(_owner != Self, "invariant"); 1469 1470 // The thread is on the WaitSet list - now park() it. 1471 // On MP systems it's conceivable that a brief spin before we park 1472 // could be profitable. 1473 // 1474 // TODO-FIXME: change the following logic to a loop of the form 1475 // while (!timeout && !interrupted && _notified == 0) park() 1476 1477 int ret = OS_OK; 1478 int WasNotified = 0; 1479 { // State transition wrappers 1480 OSThread* osthread = Self->osthread(); 1481 OSThreadWaitState osts(osthread, true); 1482 { 1483 ThreadBlockInVM tbivm(jt); 1484 // Thread is in thread_blocked state and oop access is unsafe. 1485 jt->set_suspend_equivalent(); 1486 1487 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1488 // Intentionally empty 1489 } else if (node._notified == 0) { 1490 if (millis <= 0) { 1491 Self->_ParkEvent->park(); 1492 } else { 1493 ret = Self->_ParkEvent->park(millis); 1494 } 1495 } 1496 1497 // were we externally suspended while we were waiting? 1498 if (ExitSuspendEquivalent (jt)) { 1499 // TODO-FIXME: add -- if succ == Self then succ = null. 1500 jt->java_suspend_self(); 1501 } 1502 1503 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1504 1505 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1506 // from the WaitSet to the EntryList. 1507 // See if we need to remove Node from the WaitSet. 1508 // We use double-checked locking to avoid grabbing _WaitSetLock 1509 // if the thread is not on the wait queue. 1510 // 1511 // Note that we don't need a fence before the fetch of TState. 1512 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1513 // written by the is thread. (perhaps the fetch might even be satisfied 1514 // by a look-aside into the processor's own store buffer, although given 1515 // the length of the code path between the prior ST and this load that's 1516 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1517 // then we'll acquire the lock and then re-fetch a fresh TState value. 1518 // That is, we fail toward safety. 1519 1520 if (node.TState == ObjectWaiter::TS_WAIT) { 1521 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1522 if (node.TState == ObjectWaiter::TS_WAIT) { 1523 DequeueSpecificWaiter(&node); // unlink from WaitSet 1524 assert(node._notified == 0, "invariant"); 1525 node.TState = ObjectWaiter::TS_RUN; 1526 } 1527 Thread::SpinRelease(&_WaitSetLock); 1528 } 1529 1530 // The thread is now either on off-list (TS_RUN), 1531 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1532 // The Node's TState variable is stable from the perspective of this thread. 1533 // No other threads will asynchronously modify TState. 1534 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1535 OrderAccess::loadload(); 1536 if (_succ == Self) _succ = NULL; 1537 WasNotified = node._notified; 1538 1539 // Reentry phase -- reacquire the monitor. 1540 // re-enter contended monitor after object.wait(). 1541 // retain OBJECT_WAIT state until re-enter successfully completes 1542 // Thread state is thread_in_vm and oop access is again safe, 1543 // although the raw address of the object may have changed. 1544 // (Don't cache naked oops over safepoints, of course). 1545 1546 // post monitor waited event. Note that this is past-tense, we are done waiting. 1547 if (JvmtiExport::should_post_monitor_waited()) { 1548 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1549 1550 if (node._notified != 0 && _succ == Self) { 1551 // In this part of the monitor wait-notify-reenter protocol it 1552 // is possible (and normal) for another thread to do a fastpath 1553 // monitor enter-exit while this thread is still trying to get 1554 // to the reenter portion of the protocol. 1555 // 1556 // The ObjectMonitor was notified and the current thread is 1557 // the successor which also means that an unpark() has already 1558 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1559 // consume the unpark() that was done when the successor was 1560 // set because the same ParkEvent is shared between Java 1561 // monitors and JVM/TI RawMonitors (for now). 1562 // 1563 // We redo the unpark() to ensure forward progress, i.e., we 1564 // don't want all pending threads hanging (parked) with none 1565 // entering the unlocked monitor. 1566 node._event->unpark(); 1567 } 1568 } 1569 1570 if (event.should_commit()) { 1571 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1572 } 1573 1574 OrderAccess::fence(); 1575 1576 assert(Self->_Stalled != 0, "invariant"); 1577 Self->_Stalled = 0; 1578 1579 assert(_owner != Self, "invariant"); 1580 ObjectWaiter::TStates v = node.TState; 1581 if (v == ObjectWaiter::TS_RUN) { 1582 enter(Self); 1583 } else { 1584 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1585 ReenterI(Self, &node); 1586 node.wait_reenter_end(this); 1587 } 1588 1589 // Self has reacquired the lock. 1590 // Lifecycle - the node representing Self must not appear on any queues. 1591 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1592 // want residual elements associated with this thread left on any lists. 1593 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1594 assert(_owner == Self, "invariant"); 1595 assert(_succ != Self, "invariant"); 1596 } // OSThreadWaitState() 1597 1598 jt->set_current_waiting_monitor(NULL); 1599 1600 guarantee(_recursions == 0, "invariant"); 1601 _recursions = save; // restore the old recursion count 1602 _waiters--; // decrement the number of waiters 1603 1604 // Verify a few postconditions 1605 assert(_owner == Self, "invariant"); 1606 assert(_succ != Self, "invariant"); 1607 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1608 1609 if (SyncFlags & 32) { 1610 OrderAccess::fence(); 1611 } 1612 1613 // check if the notification happened 1614 if (!WasNotified) { 1615 // no, it could be timeout or Thread.interrupt() or both 1616 // check for interrupt event, otherwise it is timeout 1617 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1618 TEVENT(Wait - throw IEX from epilog); 1619 THROW(vmSymbols::java_lang_InterruptedException()); 1620 } 1621 } 1622 1623 // NOTE: Spurious wake up will be consider as timeout. 1624 // Monitor notify has precedence over thread interrupt. 1625 } 1626 1627 1628 // Consider: 1629 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1630 // then instead of transferring a thread from the WaitSet to the EntryList 1631 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1632 1633 void ObjectMonitor::INotify(Thread * Self) { 1634 const int policy = Knob_MoveNotifyee; 1635 1636 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1637 ObjectWaiter * iterator = DequeueWaiter(); 1638 if (iterator != NULL) { 1639 TEVENT(Notify1 - Transfer); 1640 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1641 guarantee(iterator->_notified == 0, "invariant"); 1642 // Disposition - what might we do with iterator ? 1643 // a. add it directly to the EntryList - either tail (policy == 1) 1644 // or head (policy == 0). 1645 // b. push it onto the front of the _cxq (policy == 2). 1646 // For now we use (b). 1647 if (policy != 4) { 1648 iterator->TState = ObjectWaiter::TS_ENTER; 1649 } 1650 iterator->_notified = 1; 1651 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1652 1653 ObjectWaiter * list = _EntryList; 1654 if (list != NULL) { 1655 assert(list->_prev == NULL, "invariant"); 1656 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1657 assert(list != iterator, "invariant"); 1658 } 1659 1660 if (policy == 0) { // prepend to EntryList 1661 if (list == NULL) { 1662 iterator->_next = iterator->_prev = NULL; 1663 _EntryList = iterator; 1664 } else { 1665 list->_prev = iterator; 1666 iterator->_next = list; 1667 iterator->_prev = NULL; 1668 _EntryList = iterator; 1669 } 1670 } else if (policy == 1) { // append to EntryList 1671 if (list == NULL) { 1672 iterator->_next = iterator->_prev = NULL; 1673 _EntryList = iterator; 1674 } else { 1675 // CONSIDER: finding the tail currently requires a linear-time walk of 1676 // the EntryList. We can make tail access constant-time by converting to 1677 // a CDLL instead of using our current DLL. 1678 ObjectWaiter * tail; 1679 for (tail = list; tail->_next != NULL; tail = tail->_next) /* empty */; 1680 assert(tail != NULL && tail->_next == NULL, "invariant"); 1681 tail->_next = iterator; 1682 iterator->_prev = tail; 1683 iterator->_next = NULL; 1684 } 1685 } else if (policy == 2) { // prepend to cxq 1686 if (list == NULL) { 1687 iterator->_next = iterator->_prev = NULL; 1688 _EntryList = iterator; 1689 } else { 1690 iterator->TState = ObjectWaiter::TS_CXQ; 1691 for (;;) { 1692 ObjectWaiter * front = _cxq; 1693 iterator->_next = front; 1694 if (Atomic::cmpxchg_ptr(iterator, &_cxq, front) == front) { 1695 break; 1696 } 1697 } 1698 } 1699 } else if (policy == 3) { // append to cxq 1700 iterator->TState = ObjectWaiter::TS_CXQ; 1701 for (;;) { 1702 ObjectWaiter * tail = _cxq; 1703 if (tail == NULL) { 1704 iterator->_next = NULL; 1705 if (Atomic::cmpxchg_ptr(iterator, &_cxq, NULL) == NULL) { 1706 break; 1707 } 1708 } else { 1709 while (tail->_next != NULL) tail = tail->_next; 1710 tail->_next = iterator; 1711 iterator->_prev = tail; 1712 iterator->_next = NULL; 1713 break; 1714 } 1715 } 1716 } else { 1717 ParkEvent * ev = iterator->_event; 1718 iterator->TState = ObjectWaiter::TS_RUN; 1719 OrderAccess::fence(); 1720 ev->unpark(); 1721 } 1722 1723 // _WaitSetLock protects the wait queue, not the EntryList. We could 1724 // move the add-to-EntryList operation, above, outside the critical section 1725 // protected by _WaitSetLock. In practice that's not useful. With the 1726 // exception of wait() timeouts and interrupts the monitor owner 1727 // is the only thread that grabs _WaitSetLock. There's almost no contention 1728 // on _WaitSetLock so it's not profitable to reduce the length of the 1729 // critical section. 1730 1731 if (policy < 4) { 1732 iterator->wait_reenter_begin(this); 1733 } 1734 } 1735 Thread::SpinRelease(&_WaitSetLock); 1736 } 1737 1738 // Consider: a not-uncommon synchronization bug is to use notify() when 1739 // notifyAll() is more appropriate, potentially resulting in stranded 1740 // threads; this is one example of a lost wakeup. A useful diagnostic 1741 // option is to force all notify() operations to behave as notifyAll(). 1742 // 1743 // Note: We can also detect many such problems with a "minimum wait". 1744 // When the "minimum wait" is set to a small non-zero timeout value 1745 // and the program does not hang whereas it did absent "minimum wait", 1746 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses 1747 // a "minimum wait" for all park() operations; see the recheckInterval 1748 // variable and MAX_RECHECK_INTERVAL. 1749 1750 void ObjectMonitor::notify(TRAPS) { 1751 CHECK_OWNER(); 1752 if (_WaitSet == NULL) { 1753 TEVENT(Empty-Notify); 1754 return; 1755 } 1756 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1757 INotify(THREAD); 1758 OM_PERFDATA_OP(Notifications, inc(1)); 1759 } 1760 1761 1762 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1763 // from the waitset to the EntryList. This could be done more efficiently with a 1764 // single bulk transfer but in practice it's not time-critical. Beware too, 1765 // that in prepend-mode we invert the order of the waiters. Let's say that the 1766 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1767 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1768 1769 void ObjectMonitor::notifyAll(TRAPS) { 1770 CHECK_OWNER(); 1771 if (_WaitSet == NULL) { 1772 TEVENT(Empty-NotifyAll); 1773 return; 1774 } 1775 1776 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1777 int tally = 0; 1778 while (_WaitSet != NULL) { 1779 tally++; 1780 INotify(THREAD); 1781 } 1782 1783 OM_PERFDATA_OP(Notifications, inc(tally)); 1784 } 1785 1786 // ----------------------------------------------------------------------------- 1787 // Adaptive Spinning Support 1788 // 1789 // Adaptive spin-then-block - rational spinning 1790 // 1791 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1792 // algorithm. On high order SMP systems it would be better to start with 1793 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1794 // a contending thread could enqueue itself on the cxq and then spin locally 1795 // on a thread-specific variable such as its ParkEvent._Event flag. 1796 // That's left as an exercise for the reader. Note that global spinning is 1797 // not problematic on Niagara, as the L2 cache serves the interconnect and 1798 // has both low latency and massive bandwidth. 1799 // 1800 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1801 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1802 // (duration) or we can fix the count at approximately the duration of 1803 // a context switch and vary the frequency. Of course we could also 1804 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1805 // For a description of 'Adaptive spin-then-block mutual exclusion in 1806 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1807 // 1808 // This implementation varies the duration "D", where D varies with 1809 // the success rate of recent spin attempts. (D is capped at approximately 1810 // length of a round-trip context switch). The success rate for recent 1811 // spin attempts is a good predictor of the success rate of future spin 1812 // attempts. The mechanism adapts automatically to varying critical 1813 // section length (lock modality), system load and degree of parallelism. 1814 // D is maintained per-monitor in _SpinDuration and is initialized 1815 // optimistically. Spin frequency is fixed at 100%. 1816 // 1817 // Note that _SpinDuration is volatile, but we update it without locks 1818 // or atomics. The code is designed so that _SpinDuration stays within 1819 // a reasonable range even in the presence of races. The arithmetic 1820 // operations on _SpinDuration are closed over the domain of legal values, 1821 // so at worst a race will install and older but still legal value. 1822 // At the very worst this introduces some apparent non-determinism. 1823 // We might spin when we shouldn't or vice-versa, but since the spin 1824 // count are relatively short, even in the worst case, the effect is harmless. 1825 // 1826 // Care must be taken that a low "D" value does not become an 1827 // an absorbing state. Transient spinning failures -- when spinning 1828 // is overall profitable -- should not cause the system to converge 1829 // on low "D" values. We want spinning to be stable and predictable 1830 // and fairly responsive to change and at the same time we don't want 1831 // it to oscillate, become metastable, be "too" non-deterministic, 1832 // or converge on or enter undesirable stable absorbing states. 1833 // 1834 // We implement a feedback-based control system -- using past behavior 1835 // to predict future behavior. We face two issues: (a) if the 1836 // input signal is random then the spin predictor won't provide optimal 1837 // results, and (b) if the signal frequency is too high then the control 1838 // system, which has some natural response lag, will "chase" the signal. 1839 // (b) can arise from multimodal lock hold times. Transient preemption 1840 // can also result in apparent bimodal lock hold times. 1841 // Although sub-optimal, neither condition is particularly harmful, as 1842 // in the worst-case we'll spin when we shouldn't or vice-versa. 1843 // The maximum spin duration is rather short so the failure modes aren't bad. 1844 // To be conservative, I've tuned the gain in system to bias toward 1845 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1846 // "rings" or oscillates between spinning and not spinning. This happens 1847 // when spinning is just on the cusp of profitability, however, so the 1848 // situation is not dire. The state is benign -- there's no need to add 1849 // hysteresis control to damp the transition rate between spinning and 1850 // not spinning. 1851 1852 intptr_t ObjectMonitor::SpinCallbackArgument = 0; 1853 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL; 1854 1855 // Spinning: Fixed frequency (100%), vary duration 1856 1857 1858 int ObjectMonitor::TrySpin_VaryDuration(Thread * Self) { 1859 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1860 int ctr = Knob_FixedSpin; 1861 if (ctr != 0) { 1862 while (--ctr >= 0) { 1863 if (TryLock(Self) > 0) return 1; 1864 SpinPause(); 1865 } 1866 return 0; 1867 } 1868 1869 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1870 if (TryLock(Self) > 0) { 1871 // Increase _SpinDuration ... 1872 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1873 // Raising _SpurDuration to the poverty line is key. 1874 int x = _SpinDuration; 1875 if (x < Knob_SpinLimit) { 1876 if (x < Knob_Poverty) x = Knob_Poverty; 1877 _SpinDuration = x + Knob_BonusB; 1878 } 1879 return 1; 1880 } 1881 SpinPause(); 1882 } 1883 1884 // Admission control - verify preconditions for spinning 1885 // 1886 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1887 // becoming an absorbing state. Put another way, we spin briefly to 1888 // sample, just in case the system load, parallelism, contention, or lock 1889 // modality changed. 1890 // 1891 // Consider the following alternative: 1892 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1893 // spin attempt. "Periodically" might mean after a tally of 1894 // the # of failed spin attempts (or iterations) reaches some threshold. 1895 // This takes us into the realm of 1-out-of-N spinning, where we 1896 // hold the duration constant but vary the frequency. 1897 1898 ctr = _SpinDuration; 1899 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1900 if (ctr <= 0) return 0; 1901 1902 if (Knob_SuccRestrict && _succ != NULL) return 0; 1903 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1904 TEVENT(Spin abort - notrunnable [TOP]); 1905 return 0; 1906 } 1907 1908 int MaxSpin = Knob_MaxSpinners; 1909 if (MaxSpin >= 0) { 1910 if (_Spinner > MaxSpin) { 1911 TEVENT(Spin abort -- too many spinners); 1912 return 0; 1913 } 1914 // Slightly racy, but benign ... 1915 Adjust(&_Spinner, 1); 1916 } 1917 1918 // We're good to spin ... spin ingress. 1919 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1920 // when preparing to LD...CAS _owner, etc and the CAS is likely 1921 // to succeed. 1922 int hits = 0; 1923 int msk = 0; 1924 int caspty = Knob_CASPenalty; 1925 int oxpty = Knob_OXPenalty; 1926 int sss = Knob_SpinSetSucc; 1927 if (sss && _succ == NULL) _succ = Self; 1928 Thread * prv = NULL; 1929 1930 // There are three ways to exit the following loop: 1931 // 1. A successful spin where this thread has acquired the lock. 1932 // 2. Spin failure with prejudice 1933 // 3. Spin failure without prejudice 1934 1935 while (--ctr >= 0) { 1936 1937 // Periodic polling -- Check for pending GC 1938 // Threads may spin while they're unsafe. 1939 // We don't want spinning threads to delay the JVM from reaching 1940 // a stop-the-world safepoint or to steal cycles from GC. 1941 // If we detect a pending safepoint we abort in order that 1942 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1943 // this thread, if safe, doesn't steal cycles from GC. 1944 // This is in keeping with the "no loitering in runtime" rule. 1945 // We periodically check to see if there's a safepoint pending. 1946 if ((ctr & 0xFF) == 0) { 1947 if (SafepointSynchronize::do_call_back()) { 1948 TEVENT(Spin: safepoint); 1949 goto Abort; // abrupt spin egress 1950 } 1951 if (Knob_UsePause & 1) SpinPause(); 1952 1953 int (*scb)(intptr_t,int) = SpinCallbackFunction; 1954 if (hits > 50 && scb != NULL) { 1955 int abend = (*scb)(SpinCallbackArgument, 0); 1956 } 1957 } 1958 1959 if (Knob_UsePause & 2) SpinPause(); 1960 1961 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1962 // This is useful on classic SMP systems, but is of less utility on 1963 // N1-style CMT platforms. 1964 // 1965 // Trade-off: lock acquisition latency vs coherency bandwidth. 1966 // Lock hold times are typically short. A histogram 1967 // of successful spin attempts shows that we usually acquire 1968 // the lock early in the spin. That suggests we want to 1969 // sample _owner frequently in the early phase of the spin, 1970 // but then back-off and sample less frequently as the spin 1971 // progresses. The back-off makes a good citizen on SMP big 1972 // SMP systems. Oversampling _owner can consume excessive 1973 // coherency bandwidth. Relatedly, if we _oversample _owner we 1974 // can inadvertently interfere with the the ST m->owner=null. 1975 // executed by the lock owner. 1976 if (ctr & msk) continue; 1977 ++hits; 1978 if ((hits & 0xF) == 0) { 1979 // The 0xF, above, corresponds to the exponent. 1980 // Consider: (msk+1)|msk 1981 msk = ((msk << 2)|3) & BackOffMask; 1982 } 1983 1984 // Probe _owner with TATAS 1985 // If this thread observes the monitor transition or flicker 1986 // from locked to unlocked to locked, then the odds that this 1987 // thread will acquire the lock in this spin attempt go down 1988 // considerably. The same argument applies if the CAS fails 1989 // or if we observe _owner change from one non-null value to 1990 // another non-null value. In such cases we might abort 1991 // the spin without prejudice or apply a "penalty" to the 1992 // spin count-down variable "ctr", reducing it by 100, say. 1993 1994 Thread * ox = (Thread *) _owner; 1995 if (ox == NULL) { 1996 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 1997 if (ox == NULL) { 1998 // The CAS succeeded -- this thread acquired ownership 1999 // Take care of some bookkeeping to exit spin state. 2000 if (sss && _succ == Self) { 2001 _succ = NULL; 2002 } 2003 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2004 2005 // Increase _SpinDuration : 2006 // The spin was successful (profitable) so we tend toward 2007 // longer spin attempts in the future. 2008 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2009 // If we acquired the lock early in the spin cycle it 2010 // makes sense to increase _SpinDuration proportionally. 2011 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2012 int x = _SpinDuration; 2013 if (x < Knob_SpinLimit) { 2014 if (x < Knob_Poverty) x = Knob_Poverty; 2015 _SpinDuration = x + Knob_Bonus; 2016 } 2017 return 1; 2018 } 2019 2020 // The CAS failed ... we can take any of the following actions: 2021 // * penalize: ctr -= Knob_CASPenalty 2022 // * exit spin with prejudice -- goto Abort; 2023 // * exit spin without prejudice. 2024 // * Since CAS is high-latency, retry again immediately. 2025 prv = ox; 2026 TEVENT(Spin: cas failed); 2027 if (caspty == -2) break; 2028 if (caspty == -1) goto Abort; 2029 ctr -= caspty; 2030 continue; 2031 } 2032 2033 // Did lock ownership change hands ? 2034 if (ox != prv && prv != NULL) { 2035 TEVENT(spin: Owner changed) 2036 if (oxpty == -2) break; 2037 if (oxpty == -1) goto Abort; 2038 ctr -= oxpty; 2039 } 2040 prv = ox; 2041 2042 // Abort the spin if the owner is not executing. 2043 // The owner must be executing in order to drop the lock. 2044 // Spinning while the owner is OFFPROC is idiocy. 2045 // Consider: ctr -= RunnablePenalty ; 2046 if (Knob_OState && NotRunnable (Self, ox)) { 2047 TEVENT(Spin abort - notrunnable); 2048 goto Abort; 2049 } 2050 if (sss && _succ == NULL) _succ = Self; 2051 } 2052 2053 // Spin failed with prejudice -- reduce _SpinDuration. 2054 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2055 // AIMD is globally stable. 2056 TEVENT(Spin failure); 2057 { 2058 int x = _SpinDuration; 2059 if (x > 0) { 2060 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2061 // This is globally sample and tends to damp the response. 2062 x -= Knob_Penalty; 2063 if (x < 0) x = 0; 2064 _SpinDuration = x; 2065 } 2066 } 2067 2068 Abort: 2069 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2070 if (sss && _succ == Self) { 2071 _succ = NULL; 2072 // Invariant: after setting succ=null a contending thread 2073 // must recheck-retry _owner before parking. This usually happens 2074 // in the normal usage of TrySpin(), but it's safest 2075 // to make TrySpin() as foolproof as possible. 2076 OrderAccess::fence(); 2077 if (TryLock(Self) > 0) return 1; 2078 } 2079 return 0; 2080 } 2081 2082 // NotRunnable() -- informed spinning 2083 // 2084 // Don't bother spinning if the owner is not eligible to drop the lock. 2085 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2086 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2087 // The thread must be runnable in order to drop the lock in timely fashion. 2088 // If the _owner is not runnable then spinning will not likely be 2089 // successful (profitable). 2090 // 2091 // Beware -- the thread referenced by _owner could have died 2092 // so a simply fetch from _owner->_thread_state might trap. 2093 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2094 // Because of the lifecycle issues the schedctl and _thread_state values 2095 // observed by NotRunnable() might be garbage. NotRunnable must 2096 // tolerate this and consider the observed _thread_state value 2097 // as advisory. 2098 // 2099 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2100 // a thread pointer. 2101 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2102 // with the LSB of _owner. Another option would be to probablistically probe 2103 // the putative _owner->TypeTag value. 2104 // 2105 // Checking _thread_state isn't perfect. Even if the thread is 2106 // in_java it might be blocked on a page-fault or have been preempted 2107 // and sitting on a ready/dispatch queue. _thread state in conjunction 2108 // with schedctl.sc_state gives us a good picture of what the 2109 // thread is doing, however. 2110 // 2111 // TODO: check schedctl.sc_state. 2112 // We'll need to use SafeFetch32() to read from the schedctl block. 2113 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2114 // 2115 // The return value from NotRunnable() is *advisory* -- the 2116 // result is based on sampling and is not necessarily coherent. 2117 // The caller must tolerate false-negative and false-positive errors. 2118 // Spinning, in general, is probabilistic anyway. 2119 2120 2121 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2122 // Check ox->TypeTag == 2BAD. 2123 if (ox == NULL) return 0; 2124 2125 // Avoid transitive spinning ... 2126 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2127 // Immediately after T1 acquires L it's possible that T2, also 2128 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2129 // This occurs transiently after T1 acquired L but before 2130 // T1 managed to clear T1.Stalled. T2 does not need to abort 2131 // its spin in this circumstance. 2132 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2133 2134 if (BlockedOn == 1) return 1; 2135 if (BlockedOn != 0) { 2136 return BlockedOn != intptr_t(this) && _owner == ox; 2137 } 2138 2139 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2140 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2141 // consider also: jst != _thread_in_Java -- but that's overspecific. 2142 return jst == _thread_blocked || jst == _thread_in_native; 2143 } 2144 2145 2146 // ----------------------------------------------------------------------------- 2147 // WaitSet management ... 2148 2149 ObjectWaiter::ObjectWaiter(Thread* thread) { 2150 _next = NULL; 2151 _prev = NULL; 2152 _notified = 0; 2153 TState = TS_RUN; 2154 _thread = thread; 2155 _event = thread->_ParkEvent; 2156 _active = false; 2157 assert(_event != NULL, "invariant"); 2158 } 2159 2160 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2161 JavaThread *jt = (JavaThread *)this->_thread; 2162 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2163 } 2164 2165 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2166 JavaThread *jt = (JavaThread *)this->_thread; 2167 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2168 } 2169 2170 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2171 assert(node != NULL, "should not add NULL node"); 2172 assert(node->_prev == NULL, "node already in list"); 2173 assert(node->_next == NULL, "node already in list"); 2174 // put node at end of queue (circular doubly linked list) 2175 if (_WaitSet == NULL) { 2176 _WaitSet = node; 2177 node->_prev = node; 2178 node->_next = node; 2179 } else { 2180 ObjectWaiter* head = _WaitSet; 2181 ObjectWaiter* tail = head->_prev; 2182 assert(tail->_next == head, "invariant check"); 2183 tail->_next = node; 2184 head->_prev = node; 2185 node->_next = head; 2186 node->_prev = tail; 2187 } 2188 } 2189 2190 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2191 // dequeue the very first waiter 2192 ObjectWaiter* waiter = _WaitSet; 2193 if (waiter) { 2194 DequeueSpecificWaiter(waiter); 2195 } 2196 return waiter; 2197 } 2198 2199 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2200 assert(node != NULL, "should not dequeue NULL node"); 2201 assert(node->_prev != NULL, "node already removed from list"); 2202 assert(node->_next != NULL, "node already removed from list"); 2203 // when the waiter has woken up because of interrupt, 2204 // timeout or other spurious wake-up, dequeue the 2205 // waiter from waiting list 2206 ObjectWaiter* next = node->_next; 2207 if (next == node) { 2208 assert(node->_prev == node, "invariant check"); 2209 _WaitSet = NULL; 2210 } else { 2211 ObjectWaiter* prev = node->_prev; 2212 assert(prev->_next == node, "invariant check"); 2213 assert(next->_prev == node, "invariant check"); 2214 next->_prev = prev; 2215 prev->_next = next; 2216 if (_WaitSet == node) { 2217 _WaitSet = next; 2218 } 2219 } 2220 node->_next = NULL; 2221 node->_prev = NULL; 2222 } 2223 2224 // ----------------------------------------------------------------------------- 2225 // PerfData support 2226 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2227 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2228 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2229 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2230 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2231 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2232 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2233 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2234 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2235 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2236 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2237 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2238 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2239 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2240 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2241 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2242 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2243 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2244 2245 // One-shot global initialization for the sync subsystem. 2246 // We could also defer initialization and initialize on-demand 2247 // the first time we call inflate(). Initialization would 2248 // be protected - like so many things - by the MonitorCache_lock. 2249 2250 void ObjectMonitor::Initialize() { 2251 static int InitializationCompleted = 0; 2252 assert(InitializationCompleted == 0, "invariant"); 2253 InitializationCompleted = 1; 2254 if (UsePerfData) { 2255 EXCEPTION_MARK; 2256 #define NEWPERFCOUNTER(n) \ 2257 { \ 2258 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2259 CHECK); \ 2260 } 2261 #define NEWPERFVARIABLE(n) \ 2262 { \ 2263 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2264 CHECK); \ 2265 } 2266 NEWPERFCOUNTER(_sync_Inflations); 2267 NEWPERFCOUNTER(_sync_Deflations); 2268 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2269 NEWPERFCOUNTER(_sync_FutileWakeups); 2270 NEWPERFCOUNTER(_sync_Parks); 2271 NEWPERFCOUNTER(_sync_EmptyNotifications); 2272 NEWPERFCOUNTER(_sync_Notifications); 2273 NEWPERFCOUNTER(_sync_SlowEnter); 2274 NEWPERFCOUNTER(_sync_SlowExit); 2275 NEWPERFCOUNTER(_sync_SlowNotify); 2276 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2277 NEWPERFCOUNTER(_sync_FailedSpins); 2278 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2279 NEWPERFCOUNTER(_sync_PrivateA); 2280 NEWPERFCOUNTER(_sync_PrivateB); 2281 NEWPERFCOUNTER(_sync_MonInCirculation); 2282 NEWPERFCOUNTER(_sync_MonScavenged); 2283 NEWPERFVARIABLE(_sync_MonExtant); 2284 #undef NEWPERFCOUNTER 2285 #undef NEWPERFVARIABLE 2286 } 2287 } 2288 2289 static char * kvGet(char * kvList, const char * Key) { 2290 if (kvList == NULL) return NULL; 2291 size_t n = strlen(Key); 2292 char * Search; 2293 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2294 if (strncmp (Search, Key, n) == 0) { 2295 if (Search[n] == '=') return Search + n + 1; 2296 if (Search[n] == 0) return(char *) "1"; 2297 } 2298 } 2299 return NULL; 2300 } 2301 2302 static int kvGetInt(char * kvList, const char * Key, int Default) { 2303 char * v = kvGet(kvList, Key); 2304 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2305 if (Knob_ReportSettings && v != NULL) { 2306 tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ; 2307 tty->flush(); 2308 } 2309 return rslt; 2310 } 2311 2312 void ObjectMonitor::DeferredInitialize() { 2313 if (InitDone > 0) return; 2314 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2315 while (InitDone != 1) /* empty */; 2316 return; 2317 } 2318 2319 // One-shot global initialization ... 2320 // The initialization is idempotent, so we don't need locks. 2321 // In the future consider doing this via os::init_2(). 2322 // SyncKnobs consist of <Key>=<Value> pairs in the style 2323 // of environment variables. Start by converting ':' to NUL. 2324 2325 if (SyncKnobs == NULL) SyncKnobs = ""; 2326 2327 size_t sz = strlen(SyncKnobs); 2328 char * knobs = (char *) malloc(sz + 2); 2329 if (knobs == NULL) { 2330 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2331 guarantee(0, "invariant"); 2332 } 2333 strcpy(knobs, SyncKnobs); 2334 knobs[sz+1] = 0; 2335 for (char * p = knobs; *p; p++) { 2336 if (*p == ':') *p = 0; 2337 } 2338 2339 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2340 SETKNOB(ReportSettings); 2341 SETKNOB(ExitRelease); 2342 SETKNOB(Verbose); 2343 SETKNOB(VerifyInUse); 2344 SETKNOB(VerifyMatch); 2345 SETKNOB(FixedSpin); 2346 SETKNOB(SpinLimit); 2347 SETKNOB(SpinBase); 2348 SETKNOB(SpinBackOff); 2349 SETKNOB(CASPenalty); 2350 SETKNOB(OXPenalty); 2351 SETKNOB(LogSpins); 2352 SETKNOB(SpinSetSucc); 2353 SETKNOB(SuccEnabled); 2354 SETKNOB(SuccRestrict); 2355 SETKNOB(Penalty); 2356 SETKNOB(Bonus); 2357 SETKNOB(BonusB); 2358 SETKNOB(Poverty); 2359 SETKNOB(SpinAfterFutile); 2360 SETKNOB(UsePause); 2361 SETKNOB(SpinEarly); 2362 SETKNOB(OState); 2363 SETKNOB(MaxSpinners); 2364 SETKNOB(PreSpin); 2365 SETKNOB(ExitPolicy); 2366 SETKNOB(QMode); 2367 SETKNOB(ResetEvent); 2368 SETKNOB(MoveNotifyee); 2369 SETKNOB(FastHSSEC); 2370 #undef SETKNOB 2371 2372 if (Knob_Verbose) { 2373 sanity_checks(); 2374 } 2375 2376 if (os::is_MP()) { 2377 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2378 if (Knob_ReportSettings) { 2379 tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask); 2380 } 2381 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2382 } else { 2383 Knob_SpinLimit = 0; 2384 Knob_SpinBase = 0; 2385 Knob_PreSpin = 0; 2386 Knob_FixedSpin = -1; 2387 } 2388 2389 if (Knob_LogSpins == 0) { 2390 ObjectMonitor::_sync_FailedSpins = NULL; 2391 } 2392 2393 free(knobs); 2394 OrderAccess::fence(); 2395 InitDone = 1; 2396 } 2397 2398 void ObjectMonitor::sanity_checks() { 2399 int error_cnt = 0; 2400 int warning_cnt = 0; 2401 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2402 2403 if (verbose) { 2404 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2405 sizeof(ObjectMonitor)); 2406 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT, 2407 sizeof(PaddedEnd<ObjectMonitor>)); 2408 } 2409 2410 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2411 if (verbose) { 2412 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2413 } 2414 2415 ObjectMonitor dummy; 2416 u_char *addr_begin = (u_char*)&dummy; 2417 u_char *addr_header = (u_char*)&dummy._header; 2418 u_char *addr_owner = (u_char*)&dummy._owner; 2419 2420 uint offset_header = (uint)(addr_header - addr_begin); 2421 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2422 2423 uint offset_owner = (uint)(addr_owner - addr_begin); 2424 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2425 2426 if ((uint)(addr_header - addr_begin) != 0) { 2427 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2428 error_cnt++; 2429 } 2430 2431 if (cache_line_size != 0) { 2432 // We were able to determine the L1 data cache line size so 2433 // do some cache line specific sanity checks 2434 2435 if ((offset_owner - offset_header) < cache_line_size) { 2436 tty->print_cr("WARNING: the _header and _owner fields are closer " 2437 "than a cache line which permits false sharing."); 2438 warning_cnt++; 2439 } 2440 2441 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) { 2442 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a " 2443 "multiple of a cache line which permits false sharing."); 2444 warning_cnt++; 2445 } 2446 } 2447 2448 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2449 &warning_cnt); 2450 2451 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2452 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2453 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2454 } 2455 2456 guarantee(error_cnt == 0, 2457 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2458 } 2459 2460 #ifndef PRODUCT 2461 void ObjectMonitor::verify() { 2462 } 2463 2464 void ObjectMonitor::print() { 2465 } 2466 2467 void ObjectMonitor_test() { 2468 ObjectMonitor::sanity_checks(); 2469 } 2470 #endif