1 /* 2 * Copyright (c) 1998, 2017, 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.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/safepointMechanism.inline.hpp" 39 #include "runtime/stubRoutines.hpp" 40 #include "runtime/thread.inline.hpp" 41 #include "services/threadService.hpp" 42 #include "trace/tracing.hpp" 43 #include "trace/traceMacros.hpp" 44 #include "utilities/dtrace.hpp" 45 #include "utilities/macros.hpp" 46 #include "utilities/preserveException.hpp" 47 48 #ifdef DTRACE_ENABLED 49 50 // Only bother with this argument setup if dtrace is available 51 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 52 53 54 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 55 char* bytes = NULL; \ 56 int len = 0; \ 57 jlong jtid = SharedRuntime::get_java_tid(thread); \ 58 Symbol* klassname = ((oop)obj)->klass()->name(); \ 59 if (klassname != NULL) { \ 60 bytes = (char*)klassname->bytes(); \ 61 len = klassname->utf8_length(); \ 62 } 63 64 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 65 { \ 66 if (DTraceMonitorProbes) { \ 67 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 68 HOTSPOT_MONITOR_WAIT(jtid, \ 69 (monitor), bytes, len, (millis)); \ 70 } \ 71 } 72 73 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 74 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 75 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 76 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 77 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 78 79 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 80 { \ 81 if (DTraceMonitorProbes) { \ 82 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 83 HOTSPOT_MONITOR_##probe(jtid, \ 84 (uintptr_t)(monitor), bytes, len); \ 85 } \ 86 } 87 88 #else // ndef DTRACE_ENABLED 89 90 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 91 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 92 93 #endif // ndef DTRACE_ENABLED 94 95 // Tunables ... 96 // The knob* variables are effectively final. Once set they should 97 // never be modified hence. Consider using __read_mostly with GCC. 98 99 int ObjectMonitor::Knob_ExitRelease = 0; 100 int ObjectMonitor::Knob_Verbose = 0; 101 int ObjectMonitor::Knob_VerifyInUse = 0; 102 int ObjectMonitor::Knob_VerifyMatch = 0; 103 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 104 static int Knob_LogSpins = 0; // enable jvmstat tally for spins 105 static int Knob_HandOff = 0; 106 static int Knob_ReportSettings = 0; 107 108 static int Knob_SpinBase = 0; // Floor AKA SpinMin 109 static int Knob_SpinBackOff = 0; // spin-loop backoff 110 static int Knob_CASPenalty = -1; // Penalty for failed CAS 111 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 112 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 113 static int Knob_SpinEarly = 1; 114 static int Knob_SuccEnabled = 1; // futile wake throttling 115 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 116 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 117 static int Knob_Bonus = 100; // spin success bonus 118 static int Knob_BonusB = 100; // spin success bonus 119 static int Knob_Penalty = 200; // spin failure penalty 120 static int Knob_Poverty = 1000; 121 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 122 static int Knob_FixedSpin = 0; 123 static int Knob_OState = 3; // Spinner checks thread state of _owner 124 static int Knob_UsePause = 1; 125 static int Knob_ExitPolicy = 0; 126 static int Knob_PreSpin = 10; // 20-100 likely better 127 static int Knob_ResetEvent = 0; 128 static int BackOffMask = 0; 129 130 static int Knob_FastHSSEC = 0; 131 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 132 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 133 static volatile int InitDone = 0; 134 135 // ----------------------------------------------------------------------------- 136 // Theory of operations -- Monitors lists, thread residency, etc: 137 // 138 // * A thread acquires ownership of a monitor by successfully 139 // CAS()ing the _owner field from null to non-null. 140 // 141 // * Invariant: A thread appears on at most one monitor list -- 142 // cxq, EntryList or WaitSet -- at any one time. 143 // 144 // * Contending threads "push" themselves onto the cxq with CAS 145 // and then spin/park. 146 // 147 // * After a contending thread eventually acquires the lock it must 148 // dequeue itself from either the EntryList or the cxq. 149 // 150 // * The exiting thread identifies and unparks an "heir presumptive" 151 // tentative successor thread on the EntryList. Critically, the 152 // exiting thread doesn't unlink the successor thread from the EntryList. 153 // After having been unparked, the wakee will recontend for ownership of 154 // the monitor. The successor (wakee) will either acquire the lock or 155 // re-park itself. 156 // 157 // Succession is provided for by a policy of competitive handoff. 158 // The exiting thread does _not_ grant or pass ownership to the 159 // successor thread. (This is also referred to as "handoff" succession"). 160 // Instead the exiting thread releases ownership and possibly wakes 161 // a successor, so the successor can (re)compete for ownership of the lock. 162 // If the EntryList is empty but the cxq is populated the exiting 163 // thread will drain the cxq into the EntryList. It does so by 164 // by detaching the cxq (installing null with CAS) and folding 165 // the threads from the cxq into the EntryList. The EntryList is 166 // doubly linked, while the cxq is singly linked because of the 167 // CAS-based "push" used to enqueue recently arrived threads (RATs). 168 // 169 // * Concurrency invariants: 170 // 171 // -- only the monitor owner may access or mutate the EntryList. 172 // The mutex property of the monitor itself protects the EntryList 173 // from concurrent interference. 174 // -- Only the monitor owner may detach the cxq. 175 // 176 // * The monitor entry list operations avoid locks, but strictly speaking 177 // they're not lock-free. Enter is lock-free, exit is not. 178 // For a description of 'Methods and apparatus providing non-blocking access 179 // to a resource,' see U.S. Pat. No. 7844973. 180 // 181 // * The cxq can have multiple concurrent "pushers" but only one concurrent 182 // detaching thread. This mechanism is immune from the ABA corruption. 183 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 184 // 185 // * Taken together, the cxq and the EntryList constitute or form a 186 // single logical queue of threads stalled trying to acquire the lock. 187 // We use two distinct lists to improve the odds of a constant-time 188 // dequeue operation after acquisition (in the ::enter() epilogue) and 189 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 190 // A key desideratum is to minimize queue & monitor metadata manipulation 191 // that occurs while holding the monitor lock -- that is, we want to 192 // minimize monitor lock holds times. Note that even a small amount of 193 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 194 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 195 // locks and monitor metadata. 196 // 197 // Cxq points to the set of Recently Arrived Threads attempting entry. 198 // Because we push threads onto _cxq with CAS, the RATs must take the form of 199 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 200 // the unlocking thread notices that EntryList is null but _cxq is != null. 201 // 202 // The EntryList is ordered by the prevailing queue discipline and 203 // can be organized in any convenient fashion, such as a doubly-linked list or 204 // a circular doubly-linked list. Critically, we want insert and delete operations 205 // to operate in constant-time. If we need a priority queue then something akin 206 // to Solaris' sleepq would work nicely. Viz., 207 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 208 // Queue discipline is enforced at ::exit() time, when the unlocking thread 209 // drains the cxq into the EntryList, and orders or reorders the threads on the 210 // EntryList accordingly. 211 // 212 // Barring "lock barging", this mechanism provides fair cyclic ordering, 213 // somewhat similar to an elevator-scan. 214 // 215 // * The monitor synchronization subsystem avoids the use of native 216 // synchronization primitives except for the narrow platform-specific 217 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 218 // the semantics of park-unpark. Put another way, this monitor implementation 219 // depends only on atomic operations and park-unpark. The monitor subsystem 220 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 221 // underlying OS manages the READY<->RUN transitions. 222 // 223 // * Waiting threads reside on the WaitSet list -- wait() puts 224 // the caller onto the WaitSet. 225 // 226 // * notify() or notifyAll() simply transfers threads from the WaitSet to 227 // either the EntryList or cxq. Subsequent exit() operations will 228 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 229 // it's likely the notifyee would simply impale itself on the lock held 230 // by the notifier. 231 // 232 // * An interesting alternative is to encode cxq as (List,LockByte) where 233 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 234 // variable, like _recursions, in the scheme. The threads or Events that form 235 // the list would have to be aligned in 256-byte addresses. A thread would 236 // try to acquire the lock or enqueue itself with CAS, but exiting threads 237 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 238 // Note that is is *not* word-tearing, but it does presume that full-word 239 // CAS operations are coherent with intermix with STB operations. That's true 240 // on most common processors. 241 // 242 // * See also http://blogs.sun.com/dave 243 244 245 // ----------------------------------------------------------------------------- 246 // Enter support 247 248 void ObjectMonitor::enter(TRAPS) { 249 // The following code is ordered to check the most common cases first 250 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 251 Thread * const Self = THREAD; 252 253 void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL); 254 if (cur == NULL) { 255 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 256 assert(_recursions == 0, "invariant"); 257 assert(_owner == Self, "invariant"); 258 return; 259 } 260 261 if (cur == Self) { 262 // TODO-FIXME: check for integer overflow! BUGID 6557169. 263 _recursions++; 264 return; 265 } 266 267 if (Self->is_lock_owned ((address)cur)) { 268 assert(_recursions == 0, "internal state error"); 269 _recursions = 1; 270 // Commute owner from a thread-specific on-stack BasicLockObject address to 271 // a full-fledged "Thread *". 272 _owner = Self; 273 return; 274 } 275 276 // We've encountered genuine contention. 277 assert(Self->_Stalled == 0, "invariant"); 278 Self->_Stalled = intptr_t(this); 279 280 // Try one round of spinning *before* enqueueing Self 281 // and before going through the awkward and expensive state 282 // transitions. The following spin is strictly optional ... 283 // Note that if we acquire the monitor from an initial spin 284 // we forgo posting JVMTI events and firing DTRACE probes. 285 if (Knob_SpinEarly && TrySpin (Self) > 0) { 286 assert(_owner == Self, "invariant"); 287 assert(_recursions == 0, "invariant"); 288 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 289 Self->_Stalled = 0; 290 return; 291 } 292 293 assert(_owner != Self, "invariant"); 294 assert(_succ != Self, "invariant"); 295 assert(Self->is_Java_thread(), "invariant"); 296 JavaThread * jt = (JavaThread *) Self; 297 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 298 assert(jt->thread_state() != _thread_blocked, "invariant"); 299 assert(this->object() != NULL, "invariant"); 300 assert(_count >= 0, "invariant"); 301 302 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 303 // Ensure the object-monitor relationship remains stable while there's contention. 304 Atomic::inc(&_count); 305 306 EventJavaMonitorEnter event; 307 308 { // Change java thread status to indicate blocked on monitor enter. 309 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 310 311 Self->set_current_pending_monitor(this); 312 313 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 314 if (JvmtiExport::should_post_monitor_contended_enter()) { 315 JvmtiExport::post_monitor_contended_enter(jt, this); 316 317 // The current thread does not yet own the monitor and does not 318 // yet appear on any queues that would get it made the successor. 319 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 320 // handler cannot accidentally consume an unpark() meant for the 321 // ParkEvent associated with this ObjectMonitor. 322 } 323 324 OSThreadContendState osts(Self->osthread()); 325 ThreadBlockInVM tbivm(jt); 326 327 // TODO-FIXME: change the following for(;;) loop to straight-line code. 328 for (;;) { 329 jt->set_suspend_equivalent(); 330 // cleared by handle_special_suspend_equivalent_condition() 331 // or java_suspend_self() 332 333 EnterI(THREAD); 334 335 if (!ExitSuspendEquivalent(jt)) break; 336 337 // We have acquired the contended monitor, but while we were 338 // waiting another thread suspended us. We don't want to enter 339 // the monitor while suspended because that would surprise the 340 // thread that suspended us. 341 // 342 _recursions = 0; 343 _succ = NULL; 344 exit(false, Self); 345 346 jt->java_suspend_self(); 347 } 348 Self->set_current_pending_monitor(NULL); 349 350 // We cleared the pending monitor info since we've just gotten past 351 // the enter-check-for-suspend dance and we now own the monitor free 352 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 353 // destructor can go to a safepoint at the end of this block. If we 354 // do a thread dump during that safepoint, then this thread will show 355 // as having "-locked" the monitor, but the OS and java.lang.Thread 356 // states will still report that the thread is blocked trying to 357 // acquire it. 358 } 359 360 Atomic::dec(&_count); 361 assert(_count >= 0, "invariant"); 362 Self->_Stalled = 0; 363 364 // Must either set _recursions = 0 or ASSERT _recursions == 0. 365 assert(_recursions == 0, "invariant"); 366 assert(_owner == Self, "invariant"); 367 assert(_succ != Self, "invariant"); 368 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 369 370 // The thread -- now the owner -- is back in vm mode. 371 // Report the glorious news via TI,DTrace and jvmstat. 372 // The probe effect is non-trivial. All the reportage occurs 373 // while we hold the monitor, increasing the length of the critical 374 // section. Amdahl's parallel speedup law comes vividly into play. 375 // 376 // Another option might be to aggregate the events (thread local or 377 // per-monitor aggregation) and defer reporting until a more opportune 378 // time -- such as next time some thread encounters contention but has 379 // yet to acquire the lock. While spinning that thread could 380 // spinning we could increment JVMStat counters, etc. 381 382 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 383 if (JvmtiExport::should_post_monitor_contended_entered()) { 384 JvmtiExport::post_monitor_contended_entered(jt, this); 385 386 // The current thread already owns the monitor and is not going to 387 // call park() for the remainder of the monitor enter protocol. So 388 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 389 // event handler consumed an unpark() issued by the thread that 390 // just exited the monitor. 391 } 392 393 if (event.should_commit()) { 394 event.set_monitorClass(((oop)this->object())->klass()); 395 event.set_previousOwner((TYPE_THREAD)_previous_owner_tid); 396 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 397 event.commit(); 398 } 399 400 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 401 } 402 403 404 // Caveat: TryLock() is not necessarily serializing if it returns failure. 405 // Callers must compensate as needed. 406 407 int ObjectMonitor::TryLock(Thread * Self) { 408 void * own = _owner; 409 if (own != NULL) return 0; 410 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 411 // Either guarantee _recursions == 0 or set _recursions = 0. 412 assert(_recursions == 0, "invariant"); 413 assert(_owner == Self, "invariant"); 414 return 1; 415 } 416 // The lock had been free momentarily, but we lost the race to the lock. 417 // Interference -- the CAS failed. 418 // We can either return -1 or retry. 419 // Retry doesn't make as much sense because the lock was just acquired. 420 return -1; 421 } 422 423 #define MAX_RECHECK_INTERVAL 1000 424 425 void ObjectMonitor::EnterI(TRAPS) { 426 Thread * const Self = THREAD; 427 assert(Self->is_Java_thread(), "invariant"); 428 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 429 430 // Try the lock - TATAS 431 if (TryLock (Self) > 0) { 432 assert(_succ != Self, "invariant"); 433 assert(_owner == Self, "invariant"); 434 assert(_Responsible != Self, "invariant"); 435 return; 436 } 437 438 DeferredInitialize(); 439 440 // We try one round of spinning *before* enqueueing Self. 441 // 442 // If the _owner is ready but OFFPROC we could use a YieldTo() 443 // operation to donate the remainder of this thread's quantum 444 // to the owner. This has subtle but beneficial affinity 445 // effects. 446 447 if (TrySpin (Self) > 0) { 448 assert(_owner == Self, "invariant"); 449 assert(_succ != Self, "invariant"); 450 assert(_Responsible != Self, "invariant"); 451 return; 452 } 453 454 // The Spin failed -- Enqueue and park the thread ... 455 assert(_succ != Self, "invariant"); 456 assert(_owner != Self, "invariant"); 457 assert(_Responsible != Self, "invariant"); 458 459 // Enqueue "Self" on ObjectMonitor's _cxq. 460 // 461 // Node acts as a proxy for Self. 462 // As an aside, if were to ever rewrite the synchronization code mostly 463 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 464 // Java objects. This would avoid awkward lifecycle and liveness issues, 465 // as well as eliminate a subset of ABA issues. 466 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 467 468 ObjectWaiter node(Self); 469 Self->_ParkEvent->reset(); 470 node._prev = (ObjectWaiter *) 0xBAD; 471 node.TState = ObjectWaiter::TS_CXQ; 472 473 // Push "Self" onto the front of the _cxq. 474 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 475 // Note that spinning tends to reduce the rate at which threads 476 // enqueue and dequeue on EntryList|cxq. 477 ObjectWaiter * nxt; 478 for (;;) { 479 node._next = nxt = _cxq; 480 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break; 481 482 // Interference - the CAS failed because _cxq changed. Just retry. 483 // As an optional optimization we retry the lock. 484 if (TryLock (Self) > 0) { 485 assert(_succ != Self, "invariant"); 486 assert(_owner == Self, "invariant"); 487 assert(_Responsible != Self, "invariant"); 488 return; 489 } 490 } 491 492 // Check for cxq|EntryList edge transition to non-null. This indicates 493 // the onset of contention. While contention persists exiting threads 494 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 495 // operations revert to the faster 1-0 mode. This enter operation may interleave 496 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 497 // arrange for one of the contending thread to use a timed park() operations 498 // to detect and recover from the race. (Stranding is form of progress failure 499 // where the monitor is unlocked but all the contending threads remain parked). 500 // That is, at least one of the contended threads will periodically poll _owner. 501 // One of the contending threads will become the designated "Responsible" thread. 502 // The Responsible thread uses a timed park instead of a normal indefinite park 503 // operation -- it periodically wakes and checks for and recovers from potential 504 // strandings admitted by 1-0 exit operations. We need at most one Responsible 505 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 506 // be responsible for a monitor. 507 // 508 // Currently, one of the contended threads takes on the added role of "Responsible". 509 // A viable alternative would be to use a dedicated "stranding checker" thread 510 // that periodically iterated over all the threads (or active monitors) and unparked 511 // successors where there was risk of stranding. This would help eliminate the 512 // timer scalability issues we see on some platforms as we'd only have one thread 513 // -- the checker -- parked on a timer. 514 515 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 516 // Try to assume the role of responsible thread for the monitor. 517 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 518 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 519 } 520 521 // The lock might have been released while this thread was occupied queueing 522 // itself onto _cxq. To close the race and avoid "stranding" and 523 // progress-liveness failure we must resample-retry _owner before parking. 524 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 525 // In this case the ST-MEMBAR is accomplished with CAS(). 526 // 527 // TODO: Defer all thread state transitions until park-time. 528 // Since state transitions are heavy and inefficient we'd like 529 // to defer the state transitions until absolutely necessary, 530 // and in doing so avoid some transitions ... 531 532 TEVENT(Inflated enter - Contention); 533 int nWakeups = 0; 534 int recheckInterval = 1; 535 536 for (;;) { 537 538 if (TryLock(Self) > 0) break; 539 assert(_owner != Self, "invariant"); 540 541 if ((SyncFlags & 2) && _Responsible == NULL) { 542 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 543 } 544 545 // park self 546 if (_Responsible == Self || (SyncFlags & 1)) { 547 TEVENT(Inflated enter - park TIMED); 548 Self->_ParkEvent->park((jlong) recheckInterval); 549 // Increase the recheckInterval, but clamp the value. 550 recheckInterval *= 8; 551 if (recheckInterval > MAX_RECHECK_INTERVAL) { 552 recheckInterval = MAX_RECHECK_INTERVAL; 553 } 554 } else { 555 TEVENT(Inflated enter - park UNTIMED); 556 Self->_ParkEvent->park(); 557 } 558 559 if (TryLock(Self) > 0) break; 560 561 // The lock is still contested. 562 // Keep a tally of the # of futile wakeups. 563 // Note that the counter is not protected by a lock or updated by atomics. 564 // That is by design - we trade "lossy" counters which are exposed to 565 // races during updates for a lower probe effect. 566 TEVENT(Inflated enter - Futile wakeup); 567 // This PerfData object can be used in parallel with a safepoint. 568 // See the work around in PerfDataManager::destroy(). 569 OM_PERFDATA_OP(FutileWakeups, inc()); 570 ++nWakeups; 571 572 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 573 // We can defer clearing _succ until after the spin completes 574 // TrySpin() must tolerate being called with _succ == Self. 575 // Try yet another round of adaptive spinning. 576 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 577 578 // We can find that we were unpark()ed and redesignated _succ while 579 // we were spinning. That's harmless. If we iterate and call park(), 580 // park() will consume the event and return immediately and we'll 581 // just spin again. This pattern can repeat, leaving _succ to simply 582 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 583 // Alternately, we can sample fired() here, and if set, forgo spinning 584 // in the next iteration. 585 586 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 587 Self->_ParkEvent->reset(); 588 OrderAccess::fence(); 589 } 590 if (_succ == Self) _succ = NULL; 591 592 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 593 OrderAccess::fence(); 594 } 595 596 // Egress : 597 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 598 // Normally we'll find Self on the EntryList . 599 // From the perspective of the lock owner (this thread), the 600 // EntryList is stable and cxq is prepend-only. 601 // The head of cxq is volatile but the interior is stable. 602 // In addition, Self.TState is stable. 603 604 assert(_owner == Self, "invariant"); 605 assert(object() != NULL, "invariant"); 606 // I'd like to write: 607 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 608 // but as we're at a safepoint that's not safe. 609 610 UnlinkAfterAcquire(Self, &node); 611 if (_succ == Self) _succ = NULL; 612 613 assert(_succ != Self, "invariant"); 614 if (_Responsible == Self) { 615 _Responsible = NULL; 616 OrderAccess::fence(); // Dekker pivot-point 617 618 // We may leave threads on cxq|EntryList without a designated 619 // "Responsible" thread. This is benign. When this thread subsequently 620 // exits the monitor it can "see" such preexisting "old" threads -- 621 // threads that arrived on the cxq|EntryList before the fence, above -- 622 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 623 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 624 // non-null and elect a new "Responsible" timer thread. 625 // 626 // This thread executes: 627 // ST Responsible=null; MEMBAR (in enter epilogue - here) 628 // LD cxq|EntryList (in subsequent exit) 629 // 630 // Entering threads in the slow/contended path execute: 631 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 632 // The (ST cxq; MEMBAR) is accomplished with CAS(). 633 // 634 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 635 // exit operation from floating above the ST Responsible=null. 636 } 637 638 // We've acquired ownership with CAS(). 639 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 640 // But since the CAS() this thread may have also stored into _succ, 641 // EntryList, cxq or Responsible. These meta-data updates must be 642 // visible __before this thread subsequently drops the lock. 643 // Consider what could occur if we didn't enforce this constraint -- 644 // STs to monitor meta-data and user-data could reorder with (become 645 // visible after) the ST in exit that drops ownership of the lock. 646 // Some other thread could then acquire the lock, but observe inconsistent 647 // or old monitor meta-data and heap data. That violates the JMM. 648 // To that end, the 1-0 exit() operation must have at least STST|LDST 649 // "release" barrier semantics. Specifically, there must be at least a 650 // STST|LDST barrier in exit() before the ST of null into _owner that drops 651 // the lock. The barrier ensures that changes to monitor meta-data and data 652 // protected by the lock will be visible before we release the lock, and 653 // therefore before some other thread (CPU) has a chance to acquire the lock. 654 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 655 // 656 // Critically, any prior STs to _succ or EntryList must be visible before 657 // the ST of null into _owner in the *subsequent* (following) corresponding 658 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 659 // execute a serializing instruction. 660 661 if (SyncFlags & 8) { 662 OrderAccess::fence(); 663 } 664 return; 665 } 666 667 // ReenterI() is a specialized inline form of the latter half of the 668 // contended slow-path from EnterI(). We use ReenterI() only for 669 // monitor reentry in wait(). 670 // 671 // In the future we should reconcile EnterI() and ReenterI(), adding 672 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 673 // loop accordingly. 674 675 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 676 assert(Self != NULL, "invariant"); 677 assert(SelfNode != NULL, "invariant"); 678 assert(SelfNode->_thread == Self, "invariant"); 679 assert(_waiters > 0, "invariant"); 680 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 681 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 682 JavaThread * jt = (JavaThread *) Self; 683 684 int nWakeups = 0; 685 for (;;) { 686 ObjectWaiter::TStates v = SelfNode->TState; 687 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 688 assert(_owner != Self, "invariant"); 689 690 if (TryLock(Self) > 0) break; 691 if (TrySpin(Self) > 0) break; 692 693 TEVENT(Wait Reentry - parking); 694 695 // State transition wrappers around park() ... 696 // ReenterI() wisely defers state transitions until 697 // it's clear we must park the thread. 698 { 699 OSThreadContendState osts(Self->osthread()); 700 ThreadBlockInVM tbivm(jt); 701 702 // cleared by handle_special_suspend_equivalent_condition() 703 // or java_suspend_self() 704 jt->set_suspend_equivalent(); 705 if (SyncFlags & 1) { 706 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL); 707 } else { 708 Self->_ParkEvent->park(); 709 } 710 711 // were we externally suspended while we were waiting? 712 for (;;) { 713 if (!ExitSuspendEquivalent(jt)) break; 714 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 715 jt->java_suspend_self(); 716 jt->set_suspend_equivalent(); 717 } 718 } 719 720 // Try again, but just so we distinguish between futile wakeups and 721 // successful wakeups. The following test isn't algorithmically 722 // necessary, but it helps us maintain sensible statistics. 723 if (TryLock(Self) > 0) break; 724 725 // The lock is still contested. 726 // Keep a tally of the # of futile wakeups. 727 // Note that the counter is not protected by a lock or updated by atomics. 728 // That is by design - we trade "lossy" counters which are exposed to 729 // races during updates for a lower probe effect. 730 TEVENT(Wait Reentry - futile wakeup); 731 ++nWakeups; 732 733 // Assuming this is not a spurious wakeup we'll normally 734 // find that _succ == Self. 735 if (_succ == Self) _succ = NULL; 736 737 // Invariant: after clearing _succ a contending thread 738 // *must* retry _owner before parking. 739 OrderAccess::fence(); 740 741 // This PerfData object can be used in parallel with a safepoint. 742 // See the work around in PerfDataManager::destroy(). 743 OM_PERFDATA_OP(FutileWakeups, inc()); 744 } 745 746 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 747 // Normally we'll find Self on the EntryList. 748 // Unlinking from the EntryList is constant-time and atomic-free. 749 // From the perspective of the lock owner (this thread), the 750 // EntryList is stable and cxq is prepend-only. 751 // The head of cxq is volatile but the interior is stable. 752 // In addition, Self.TState is stable. 753 754 assert(_owner == Self, "invariant"); 755 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 756 UnlinkAfterAcquire(Self, SelfNode); 757 if (_succ == Self) _succ = NULL; 758 assert(_succ != Self, "invariant"); 759 SelfNode->TState = ObjectWaiter::TS_RUN; 760 OrderAccess::fence(); // see comments at the end of EnterI() 761 } 762 763 // By convention we unlink a contending thread from EntryList|cxq immediately 764 // after the thread acquires the lock in ::enter(). Equally, we could defer 765 // unlinking the thread until ::exit()-time. 766 767 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 768 assert(_owner == Self, "invariant"); 769 assert(SelfNode->_thread == Self, "invariant"); 770 771 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 772 // Normal case: remove Self from the DLL EntryList . 773 // This is a constant-time operation. 774 ObjectWaiter * nxt = SelfNode->_next; 775 ObjectWaiter * prv = SelfNode->_prev; 776 if (nxt != NULL) nxt->_prev = prv; 777 if (prv != NULL) prv->_next = nxt; 778 if (SelfNode == _EntryList) _EntryList = nxt; 779 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 780 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 781 TEVENT(Unlink from EntryList); 782 } else { 783 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 784 // Inopportune interleaving -- Self is still on the cxq. 785 // This usually means the enqueue of self raced an exiting thread. 786 // Normally we'll find Self near the front of the cxq, so 787 // dequeueing is typically fast. If needbe we can accelerate 788 // this with some MCS/CHL-like bidirectional list hints and advisory 789 // back-links so dequeueing from the interior will normally operate 790 // in constant-time. 791 // Dequeue Self from either the head (with CAS) or from the interior 792 // with a linear-time scan and normal non-atomic memory operations. 793 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 794 // and then unlink Self from EntryList. We have to drain eventually, 795 // so it might as well be now. 796 797 ObjectWaiter * v = _cxq; 798 assert(v != NULL, "invariant"); 799 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 800 // The CAS above can fail from interference IFF a "RAT" arrived. 801 // In that case Self must be in the interior and can no longer be 802 // at the head of cxq. 803 if (v == SelfNode) { 804 assert(_cxq != v, "invariant"); 805 v = _cxq; // CAS above failed - start scan at head of list 806 } 807 ObjectWaiter * p; 808 ObjectWaiter * q = NULL; 809 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 810 q = p; 811 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 812 } 813 assert(v != SelfNode, "invariant"); 814 assert(p == SelfNode, "Node not found on cxq"); 815 assert(p != _cxq, "invariant"); 816 assert(q != NULL, "invariant"); 817 assert(q->_next == p, "invariant"); 818 q->_next = p->_next; 819 } 820 TEVENT(Unlink from cxq); 821 } 822 823 #ifdef ASSERT 824 // Diagnostic hygiene ... 825 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 826 SelfNode->_next = (ObjectWaiter *) 0xBAD; 827 SelfNode->TState = ObjectWaiter::TS_RUN; 828 #endif 829 } 830 831 // ----------------------------------------------------------------------------- 832 // Exit support 833 // 834 // exit() 835 // ~~~~~~ 836 // Note that the collector can't reclaim the objectMonitor or deflate 837 // the object out from underneath the thread calling ::exit() as the 838 // thread calling ::exit() never transitions to a stable state. 839 // This inhibits GC, which in turn inhibits asynchronous (and 840 // inopportune) reclamation of "this". 841 // 842 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 843 // There's one exception to the claim above, however. EnterI() can call 844 // exit() to drop a lock if the acquirer has been externally suspended. 845 // In that case exit() is called with _thread_state as _thread_blocked, 846 // but the monitor's _count field is > 0, which inhibits reclamation. 847 // 848 // 1-0 exit 849 // ~~~~~~~~ 850 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 851 // the fast-path operators have been optimized so the common ::exit() 852 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 853 // The code emitted by fast_unlock() elides the usual MEMBAR. This 854 // greatly improves latency -- MEMBAR and CAS having considerable local 855 // latency on modern processors -- but at the cost of "stranding". Absent the 856 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 857 // ::enter() path, resulting in the entering thread being stranding 858 // and a progress-liveness failure. Stranding is extremely rare. 859 // We use timers (timed park operations) & periodic polling to detect 860 // and recover from stranding. Potentially stranded threads periodically 861 // wake up and poll the lock. See the usage of the _Responsible variable. 862 // 863 // The CAS() in enter provides for safety and exclusion, while the CAS or 864 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 865 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 866 // We detect and recover from stranding with timers. 867 // 868 // If a thread transiently strands it'll park until (a) another 869 // thread acquires the lock and then drops the lock, at which time the 870 // exiting thread will notice and unpark the stranded thread, or, (b) 871 // the timer expires. If the lock is high traffic then the stranding latency 872 // will be low due to (a). If the lock is low traffic then the odds of 873 // stranding are lower, although the worst-case stranding latency 874 // is longer. Critically, we don't want to put excessive load in the 875 // platform's timer subsystem. We want to minimize both the timer injection 876 // rate (timers created/sec) as well as the number of timers active at 877 // any one time. (more precisely, we want to minimize timer-seconds, which is 878 // the integral of the # of active timers at any instant over time). 879 // Both impinge on OS scalability. Given that, at most one thread parked on 880 // a monitor will use a timer. 881 // 882 // There is also the risk of a futile wake-up. If we drop the lock 883 // another thread can reacquire the lock immediately, and we can 884 // then wake a thread unnecessarily. This is benign, and we've 885 // structured the code so the windows are short and the frequency 886 // of such futile wakups is low. 887 888 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 889 Thread * const Self = THREAD; 890 if (THREAD != _owner) { 891 if (THREAD->is_lock_owned((address) _owner)) { 892 // Transmute _owner from a BasicLock pointer to a Thread address. 893 // We don't need to hold _mutex for this transition. 894 // Non-null to Non-null is safe as long as all readers can 895 // tolerate either flavor. 896 assert(_recursions == 0, "invariant"); 897 _owner = THREAD; 898 _recursions = 0; 899 } else { 900 // Apparent unbalanced locking ... 901 // Naively we'd like to throw IllegalMonitorStateException. 902 // As a practical matter we can neither allocate nor throw an 903 // exception as ::exit() can be called from leaf routines. 904 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 905 // Upon deeper reflection, however, in a properly run JVM the only 906 // way we should encounter this situation is in the presence of 907 // unbalanced JNI locking. TODO: CheckJNICalls. 908 // See also: CR4414101 909 TEVENT(Exit - Throw IMSX); 910 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 911 return; 912 } 913 } 914 915 if (_recursions != 0) { 916 _recursions--; // this is simple recursive enter 917 TEVENT(Inflated exit - recursive); 918 return; 919 } 920 921 // Invariant: after setting Responsible=null an thread must execute 922 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 923 if ((SyncFlags & 4) == 0) { 924 _Responsible = NULL; 925 } 926 927 #if INCLUDE_TRACE 928 // get the owner's thread id for the MonitorEnter event 929 // if it is enabled and the thread isn't suspended 930 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 931 _previous_owner_tid = THREAD_TRACE_ID(Self); 932 } 933 #endif 934 935 for (;;) { 936 assert(THREAD == _owner, "invariant"); 937 938 if (Knob_ExitPolicy == 0) { 939 // release semantics: prior loads and stores from within the critical section 940 // must not float (reorder) past the following store that drops the lock. 941 // On SPARC that requires MEMBAR #loadstore|#storestore. 942 // But of course in TSO #loadstore|#storestore is not required. 943 // I'd like to write one of the following: 944 // A. OrderAccess::release() ; _owner = NULL 945 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 946 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 947 // store into a _dummy variable. That store is not needed, but can result 948 // in massive wasteful coherency traffic on classic SMP systems. 949 // Instead, I use release_store(), which is implemented as just a simple 950 // ST on x64, x86 and SPARC. 951 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 952 OrderAccess::storeload(); // See if we need to wake a successor 953 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 954 TEVENT(Inflated exit - simple egress); 955 return; 956 } 957 TEVENT(Inflated exit - complex egress); 958 // Other threads are blocked trying to acquire the lock. 959 960 // Normally the exiting thread is responsible for ensuring succession, 961 // but if other successors are ready or other entering threads are spinning 962 // then this thread can simply store NULL into _owner and exit without 963 // waking a successor. The existence of spinners or ready successors 964 // guarantees proper succession (liveness). Responsibility passes to the 965 // ready or running successors. The exiting thread delegates the duty. 966 // More precisely, if a successor already exists this thread is absolved 967 // of the responsibility of waking (unparking) one. 968 // 969 // The _succ variable is critical to reducing futile wakeup frequency. 970 // _succ identifies the "heir presumptive" thread that has been made 971 // ready (unparked) but that has not yet run. We need only one such 972 // successor thread to guarantee progress. 973 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 974 // section 3.3 "Futile Wakeup Throttling" for details. 975 // 976 // Note that spinners in Enter() also set _succ non-null. 977 // In the current implementation spinners opportunistically set 978 // _succ so that exiting threads might avoid waking a successor. 979 // Another less appealing alternative would be for the exiting thread 980 // to drop the lock and then spin briefly to see if a spinner managed 981 // to acquire the lock. If so, the exiting thread could exit 982 // immediately without waking a successor, otherwise the exiting 983 // thread would need to dequeue and wake a successor. 984 // (Note that we'd need to make the post-drop spin short, but no 985 // shorter than the worst-case round-trip cache-line migration time. 986 // The dropped lock needs to become visible to the spinner, and then 987 // the acquisition of the lock by the spinner must become visible to 988 // the exiting thread). 989 990 // It appears that an heir-presumptive (successor) must be made ready. 991 // Only the current lock owner can manipulate the EntryList or 992 // drain _cxq, so we need to reacquire the lock. If we fail 993 // to reacquire the lock the responsibility for ensuring succession 994 // falls to the new owner. 995 // 996 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 997 return; 998 } 999 TEVENT(Exit - Reacquired); 1000 } else { 1001 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1002 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1003 OrderAccess::storeload(); 1004 // Ratify the previously observed values. 1005 if (_cxq == NULL || _succ != NULL) { 1006 TEVENT(Inflated exit - simple egress); 1007 return; 1008 } 1009 1010 // inopportune interleaving -- the exiting thread (this thread) 1011 // in the fast-exit path raced an entering thread in the slow-enter 1012 // path. 1013 // We have two choices: 1014 // A. Try to reacquire the lock. 1015 // If the CAS() fails return immediately, otherwise 1016 // we either restart/rerun the exit operation, or simply 1017 // fall-through into the code below which wakes a successor. 1018 // B. If the elements forming the EntryList|cxq are TSM 1019 // we could simply unpark() the lead thread and return 1020 // without having set _succ. 1021 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1022 TEVENT(Inflated exit - reacquired succeeded); 1023 return; 1024 } 1025 TEVENT(Inflated exit - reacquired failed); 1026 } else { 1027 TEVENT(Inflated exit - complex egress); 1028 } 1029 } 1030 1031 guarantee(_owner == THREAD, "invariant"); 1032 1033 ObjectWaiter * w = NULL; 1034 int QMode = Knob_QMode; 1035 1036 if (QMode == 2 && _cxq != NULL) { 1037 // QMode == 2 : cxq has precedence over EntryList. 1038 // Try to directly wake a successor from the cxq. 1039 // If successful, the successor will need to unlink itself from cxq. 1040 w = _cxq; 1041 assert(w != NULL, "invariant"); 1042 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1043 ExitEpilog(Self, w); 1044 return; 1045 } 1046 1047 if (QMode == 3 && _cxq != NULL) { 1048 // Aggressively drain cxq into EntryList at the first opportunity. 1049 // This policy ensure that recently-run threads live at the head of EntryList. 1050 // Drain _cxq into EntryList - bulk transfer. 1051 // First, detach _cxq. 1052 // The following loop is tantamount to: w = swap(&cxq, NULL) 1053 w = _cxq; 1054 for (;;) { 1055 assert(w != NULL, "Invariant"); 1056 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1057 if (u == w) break; 1058 w = u; 1059 } 1060 assert(w != NULL, "invariant"); 1061 1062 ObjectWaiter * q = NULL; 1063 ObjectWaiter * p; 1064 for (p = w; p != NULL; p = p->_next) { 1065 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1066 p->TState = ObjectWaiter::TS_ENTER; 1067 p->_prev = q; 1068 q = p; 1069 } 1070 1071 // Append the RATs to the EntryList 1072 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1073 ObjectWaiter * Tail; 1074 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1075 Tail = Tail->_next) 1076 /* empty */; 1077 if (Tail == NULL) { 1078 _EntryList = w; 1079 } else { 1080 Tail->_next = w; 1081 w->_prev = Tail; 1082 } 1083 1084 // Fall thru into code that tries to wake a successor from EntryList 1085 } 1086 1087 if (QMode == 4 && _cxq != NULL) { 1088 // Aggressively drain cxq into EntryList at the first opportunity. 1089 // This policy ensure that recently-run threads live at the head of EntryList. 1090 1091 // Drain _cxq into EntryList - bulk transfer. 1092 // First, detach _cxq. 1093 // The following loop is tantamount to: w = swap(&cxq, NULL) 1094 w = _cxq; 1095 for (;;) { 1096 assert(w != NULL, "Invariant"); 1097 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1098 if (u == w) break; 1099 w = u; 1100 } 1101 assert(w != NULL, "invariant"); 1102 1103 ObjectWaiter * q = NULL; 1104 ObjectWaiter * p; 1105 for (p = w; p != NULL; p = p->_next) { 1106 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1107 p->TState = ObjectWaiter::TS_ENTER; 1108 p->_prev = q; 1109 q = p; 1110 } 1111 1112 // Prepend the RATs to the EntryList 1113 if (_EntryList != NULL) { 1114 q->_next = _EntryList; 1115 _EntryList->_prev = q; 1116 } 1117 _EntryList = w; 1118 1119 // Fall thru into code that tries to wake a successor from EntryList 1120 } 1121 1122 w = _EntryList; 1123 if (w != NULL) { 1124 // I'd like to write: guarantee (w->_thread != Self). 1125 // But in practice an exiting thread may find itself on the EntryList. 1126 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1127 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1128 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1129 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1130 // release the lock "O". T2 resumes immediately after the ST of null into 1131 // _owner, above. T2 notices that the EntryList is populated, so it 1132 // reacquires the lock and then finds itself on the EntryList. 1133 // Given all that, we have to tolerate the circumstance where "w" is 1134 // associated with Self. 1135 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1136 ExitEpilog(Self, w); 1137 return; 1138 } 1139 1140 // If we find that both _cxq and EntryList are null then just 1141 // re-run the exit protocol from the top. 1142 w = _cxq; 1143 if (w == NULL) continue; 1144 1145 // Drain _cxq into EntryList - bulk transfer. 1146 // First, detach _cxq. 1147 // The following loop is tantamount to: w = swap(&cxq, NULL) 1148 for (;;) { 1149 assert(w != NULL, "Invariant"); 1150 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1151 if (u == w) break; 1152 w = u; 1153 } 1154 TEVENT(Inflated exit - drain cxq into EntryList); 1155 1156 assert(w != NULL, "invariant"); 1157 assert(_EntryList == NULL, "invariant"); 1158 1159 // Convert the LIFO SLL anchored by _cxq into a DLL. 1160 // The list reorganization step operates in O(LENGTH(w)) time. 1161 // It's critical that this step operate quickly as 1162 // "Self" still holds the outer-lock, restricting parallelism 1163 // and effectively lengthening the critical section. 1164 // Invariant: s chases t chases u. 1165 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1166 // we have faster access to the tail. 1167 1168 if (QMode == 1) { 1169 // QMode == 1 : drain cxq to EntryList, reversing order 1170 // We also reverse the order of the list. 1171 ObjectWaiter * s = NULL; 1172 ObjectWaiter * t = w; 1173 ObjectWaiter * u = NULL; 1174 while (t != NULL) { 1175 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1176 t->TState = ObjectWaiter::TS_ENTER; 1177 u = t->_next; 1178 t->_prev = u; 1179 t->_next = s; 1180 s = t; 1181 t = u; 1182 } 1183 _EntryList = s; 1184 assert(s != NULL, "invariant"); 1185 } else { 1186 // QMode == 0 or QMode == 2 1187 _EntryList = w; 1188 ObjectWaiter * q = NULL; 1189 ObjectWaiter * p; 1190 for (p = w; p != NULL; p = p->_next) { 1191 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1192 p->TState = ObjectWaiter::TS_ENTER; 1193 p->_prev = q; 1194 q = p; 1195 } 1196 } 1197 1198 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1199 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1200 1201 // See if we can abdicate to a spinner instead of waking a thread. 1202 // A primary goal of the implementation is to reduce the 1203 // context-switch rate. 1204 if (_succ != NULL) continue; 1205 1206 w = _EntryList; 1207 if (w != NULL) { 1208 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1209 ExitEpilog(Self, w); 1210 return; 1211 } 1212 } 1213 } 1214 1215 // ExitSuspendEquivalent: 1216 // A faster alternate to handle_special_suspend_equivalent_condition() 1217 // 1218 // handle_special_suspend_equivalent_condition() unconditionally 1219 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1220 // operations have high latency. Note that in ::enter() we call HSSEC 1221 // while holding the monitor, so we effectively lengthen the critical sections. 1222 // 1223 // There are a number of possible solutions: 1224 // 1225 // A. To ameliorate the problem we might also defer state transitions 1226 // to as late as possible -- just prior to parking. 1227 // Given that, we'd call HSSEC after having returned from park(), 1228 // but before attempting to acquire the monitor. This is only a 1229 // partial solution. It avoids calling HSSEC while holding the 1230 // monitor (good), but it still increases successor reacquisition latency -- 1231 // the interval between unparking a successor and the time the successor 1232 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1233 // If we use this technique we can also avoid EnterI()-exit() loop 1234 // in ::enter() where we iteratively drop the lock and then attempt 1235 // to reacquire it after suspending. 1236 // 1237 // B. In the future we might fold all the suspend bits into a 1238 // composite per-thread suspend flag and then update it with CAS(). 1239 // Alternately, a Dekker-like mechanism with multiple variables 1240 // would suffice: 1241 // ST Self->_suspend_equivalent = false 1242 // MEMBAR 1243 // LD Self_>_suspend_flags 1244 // 1245 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1246 // with a more efficient implementation, the need to use "FastHSSEC" has 1247 // decreased. - Dave 1248 1249 1250 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1251 const int Mode = Knob_FastHSSEC; 1252 if (Mode && !jSelf->is_external_suspend()) { 1253 assert(jSelf->is_suspend_equivalent(), "invariant"); 1254 jSelf->clear_suspend_equivalent(); 1255 if (2 == Mode) OrderAccess::storeload(); 1256 if (!jSelf->is_external_suspend()) return false; 1257 // We raced a suspension -- fall thru into the slow path 1258 TEVENT(ExitSuspendEquivalent - raced); 1259 jSelf->set_suspend_equivalent(); 1260 } 1261 return jSelf->handle_special_suspend_equivalent_condition(); 1262 } 1263 1264 1265 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1266 assert(_owner == Self, "invariant"); 1267 1268 // Exit protocol: 1269 // 1. ST _succ = wakee 1270 // 2. membar #loadstore|#storestore; 1271 // 2. ST _owner = NULL 1272 // 3. unpark(wakee) 1273 1274 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1275 ParkEvent * Trigger = Wakee->_event; 1276 1277 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1278 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1279 // out-of-scope (non-extant). 1280 Wakee = NULL; 1281 1282 // Drop the lock 1283 OrderAccess::release_store_ptr(&_owner, NULL); 1284 OrderAccess::fence(); // ST _owner vs LD in unpark() 1285 1286 if (SafepointMechanism::poll()) { 1287 TEVENT(unpark before SAFEPOINT); 1288 } 1289 1290 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1291 Trigger->unpark(); 1292 1293 // Maintain stats and report events to JVMTI 1294 OM_PERFDATA_OP(Parks, inc()); 1295 } 1296 1297 1298 // ----------------------------------------------------------------------------- 1299 // Class Loader deadlock handling. 1300 // 1301 // complete_exit exits a lock returning recursion count 1302 // complete_exit/reenter operate as a wait without waiting 1303 // complete_exit requires an inflated monitor 1304 // The _owner field is not always the Thread addr even with an 1305 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1306 // thread due to contention. 1307 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1308 Thread * const Self = THREAD; 1309 assert(Self->is_Java_thread(), "Must be Java thread!"); 1310 JavaThread *jt = (JavaThread *)THREAD; 1311 1312 DeferredInitialize(); 1313 1314 if (THREAD != _owner) { 1315 if (THREAD->is_lock_owned ((address)_owner)) { 1316 assert(_recursions == 0, "internal state error"); 1317 _owner = THREAD; // Convert from basiclock addr to Thread addr 1318 _recursions = 0; 1319 } 1320 } 1321 1322 guarantee(Self == _owner, "complete_exit not owner"); 1323 intptr_t save = _recursions; // record the old recursion count 1324 _recursions = 0; // set the recursion level to be 0 1325 exit(true, Self); // exit the monitor 1326 guarantee(_owner != Self, "invariant"); 1327 return save; 1328 } 1329 1330 // reenter() enters a lock and sets recursion count 1331 // complete_exit/reenter operate as a wait without waiting 1332 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1333 Thread * const Self = THREAD; 1334 assert(Self->is_Java_thread(), "Must be Java thread!"); 1335 JavaThread *jt = (JavaThread *)THREAD; 1336 1337 guarantee(_owner != Self, "reenter already owner"); 1338 enter(THREAD); // enter the monitor 1339 guarantee(_recursions == 0, "reenter recursion"); 1340 _recursions = recursions; 1341 return; 1342 } 1343 1344 1345 // ----------------------------------------------------------------------------- 1346 // A macro is used below because there may already be a pending 1347 // exception which should not abort the execution of the routines 1348 // which use this (which is why we don't put this into check_slow and 1349 // call it with a CHECK argument). 1350 1351 #define CHECK_OWNER() \ 1352 do { \ 1353 if (THREAD != _owner) { \ 1354 if (THREAD->is_lock_owned((address) _owner)) { \ 1355 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1356 _recursions = 0; \ 1357 } else { \ 1358 TEVENT(Throw IMSX); \ 1359 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1360 } \ 1361 } \ 1362 } while (false) 1363 1364 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1365 // TODO-FIXME: remove check_slow() -- it's likely dead. 1366 1367 void ObjectMonitor::check_slow(TRAPS) { 1368 TEVENT(check_slow - throw IMSX); 1369 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1370 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1371 } 1372 1373 static int Adjust(volatile int * adr, int dx) { 1374 int v; 1375 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1376 return v; 1377 } 1378 1379 // helper method for posting a monitor wait event 1380 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1381 jlong notifier_tid, 1382 jlong timeout, 1383 bool timedout) { 1384 assert(event != NULL, "invariant"); 1385 event->set_monitorClass(((oop)this->object())->klass()); 1386 event->set_timeout(timeout); 1387 event->set_address((TYPE_ADDRESS)this->object_addr()); 1388 event->set_notifier(notifier_tid); 1389 event->set_timedOut(timedout); 1390 event->commit(); 1391 } 1392 1393 // ----------------------------------------------------------------------------- 1394 // Wait/Notify/NotifyAll 1395 // 1396 // Note: a subset of changes to ObjectMonitor::wait() 1397 // will need to be replicated in complete_exit 1398 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1399 Thread * const Self = THREAD; 1400 assert(Self->is_Java_thread(), "Must be Java thread!"); 1401 JavaThread *jt = (JavaThread *)THREAD; 1402 1403 DeferredInitialize(); 1404 1405 // Throw IMSX or IEX. 1406 CHECK_OWNER(); 1407 1408 EventJavaMonitorWait event; 1409 1410 // check for a pending interrupt 1411 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1412 // post monitor waited event. Note that this is past-tense, we are done waiting. 1413 if (JvmtiExport::should_post_monitor_waited()) { 1414 // Note: 'false' parameter is passed here because the 1415 // wait was not timed out due to thread interrupt. 1416 JvmtiExport::post_monitor_waited(jt, this, false); 1417 1418 // In this short circuit of the monitor wait protocol, the 1419 // current thread never drops ownership of the monitor and 1420 // never gets added to the wait queue so the current thread 1421 // cannot be made the successor. This means that the 1422 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1423 // consume an unpark() meant for the ParkEvent associated with 1424 // this ObjectMonitor. 1425 } 1426 if (event.should_commit()) { 1427 post_monitor_wait_event(&event, 0, millis, false); 1428 } 1429 TEVENT(Wait - Throw IEX); 1430 THROW(vmSymbols::java_lang_InterruptedException()); 1431 return; 1432 } 1433 1434 TEVENT(Wait); 1435 1436 assert(Self->_Stalled == 0, "invariant"); 1437 Self->_Stalled = intptr_t(this); 1438 jt->set_current_waiting_monitor(this); 1439 1440 // create a node to be put into the queue 1441 // Critically, after we reset() the event but prior to park(), we must check 1442 // for a pending interrupt. 1443 ObjectWaiter node(Self); 1444 node.TState = ObjectWaiter::TS_WAIT; 1445 Self->_ParkEvent->reset(); 1446 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1447 1448 // Enter the waiting queue, which is a circular doubly linked list in this case 1449 // but it could be a priority queue or any data structure. 1450 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1451 // by the the owner of the monitor *except* in the case where park() 1452 // returns because of a timeout of interrupt. Contention is exceptionally rare 1453 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1454 1455 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1456 AddWaiter(&node); 1457 Thread::SpinRelease(&_WaitSetLock); 1458 1459 if ((SyncFlags & 4) == 0) { 1460 _Responsible = NULL; 1461 } 1462 intptr_t save = _recursions; // record the old recursion count 1463 _waiters++; // increment the number of waiters 1464 _recursions = 0; // set the recursion level to be 1 1465 exit(true, Self); // exit the monitor 1466 guarantee(_owner != Self, "invariant"); 1467 1468 // The thread is on the WaitSet list - now park() it. 1469 // On MP systems it's conceivable that a brief spin before we park 1470 // could be profitable. 1471 // 1472 // TODO-FIXME: change the following logic to a loop of the form 1473 // while (!timeout && !interrupted && _notified == 0) park() 1474 1475 int ret = OS_OK; 1476 int WasNotified = 0; 1477 { // State transition wrappers 1478 OSThread* osthread = Self->osthread(); 1479 OSThreadWaitState osts(osthread, true); 1480 { 1481 ThreadBlockInVM tbivm(jt); 1482 // Thread is in thread_blocked state and oop access is unsafe. 1483 jt->set_suspend_equivalent(); 1484 1485 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1486 // Intentionally empty 1487 } else if (node._notified == 0) { 1488 if (millis <= 0) { 1489 Self->_ParkEvent->park(); 1490 } else { 1491 ret = Self->_ParkEvent->park(millis); 1492 } 1493 } 1494 1495 // were we externally suspended while we were waiting? 1496 if (ExitSuspendEquivalent (jt)) { 1497 // TODO-FIXME: add -- if succ == Self then succ = null. 1498 jt->java_suspend_self(); 1499 } 1500 1501 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1502 1503 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1504 // from the WaitSet to the EntryList. 1505 // See if we need to remove Node from the WaitSet. 1506 // We use double-checked locking to avoid grabbing _WaitSetLock 1507 // if the thread is not on the wait queue. 1508 // 1509 // Note that we don't need a fence before the fetch of TState. 1510 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1511 // written by the is thread. (perhaps the fetch might even be satisfied 1512 // by a look-aside into the processor's own store buffer, although given 1513 // the length of the code path between the prior ST and this load that's 1514 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1515 // then we'll acquire the lock and then re-fetch a fresh TState value. 1516 // That is, we fail toward safety. 1517 1518 if (node.TState == ObjectWaiter::TS_WAIT) { 1519 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1520 if (node.TState == ObjectWaiter::TS_WAIT) { 1521 DequeueSpecificWaiter(&node); // unlink from WaitSet 1522 assert(node._notified == 0, "invariant"); 1523 node.TState = ObjectWaiter::TS_RUN; 1524 } 1525 Thread::SpinRelease(&_WaitSetLock); 1526 } 1527 1528 // The thread is now either on off-list (TS_RUN), 1529 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1530 // The Node's TState variable is stable from the perspective of this thread. 1531 // No other threads will asynchronously modify TState. 1532 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1533 OrderAccess::loadload(); 1534 if (_succ == Self) _succ = NULL; 1535 WasNotified = node._notified; 1536 1537 // Reentry phase -- reacquire the monitor. 1538 // re-enter contended monitor after object.wait(). 1539 // retain OBJECT_WAIT state until re-enter successfully completes 1540 // Thread state is thread_in_vm and oop access is again safe, 1541 // although the raw address of the object may have changed. 1542 // (Don't cache naked oops over safepoints, of course). 1543 1544 // post monitor waited event. Note that this is past-tense, we are done waiting. 1545 if (JvmtiExport::should_post_monitor_waited()) { 1546 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1547 1548 if (node._notified != 0 && _succ == Self) { 1549 // In this part of the monitor wait-notify-reenter protocol it 1550 // is possible (and normal) for another thread to do a fastpath 1551 // monitor enter-exit while this thread is still trying to get 1552 // to the reenter portion of the protocol. 1553 // 1554 // The ObjectMonitor was notified and the current thread is 1555 // the successor which also means that an unpark() has already 1556 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1557 // consume the unpark() that was done when the successor was 1558 // set because the same ParkEvent is shared between Java 1559 // monitors and JVM/TI RawMonitors (for now). 1560 // 1561 // We redo the unpark() to ensure forward progress, i.e., we 1562 // don't want all pending threads hanging (parked) with none 1563 // entering the unlocked monitor. 1564 node._event->unpark(); 1565 } 1566 } 1567 1568 if (event.should_commit()) { 1569 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1570 } 1571 1572 OrderAccess::fence(); 1573 1574 assert(Self->_Stalled != 0, "invariant"); 1575 Self->_Stalled = 0; 1576 1577 assert(_owner != Self, "invariant"); 1578 ObjectWaiter::TStates v = node.TState; 1579 if (v == ObjectWaiter::TS_RUN) { 1580 enter(Self); 1581 } else { 1582 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1583 ReenterI(Self, &node); 1584 node.wait_reenter_end(this); 1585 } 1586 1587 // Self has reacquired the lock. 1588 // Lifecycle - the node representing Self must not appear on any queues. 1589 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1590 // want residual elements associated with this thread left on any lists. 1591 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1592 assert(_owner == Self, "invariant"); 1593 assert(_succ != Self, "invariant"); 1594 } // OSThreadWaitState() 1595 1596 jt->set_current_waiting_monitor(NULL); 1597 1598 guarantee(_recursions == 0, "invariant"); 1599 _recursions = save; // restore the old recursion count 1600 _waiters--; // decrement the number of waiters 1601 1602 // Verify a few postconditions 1603 assert(_owner == Self, "invariant"); 1604 assert(_succ != Self, "invariant"); 1605 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1606 1607 if (SyncFlags & 32) { 1608 OrderAccess::fence(); 1609 } 1610 1611 // check if the notification happened 1612 if (!WasNotified) { 1613 // no, it could be timeout or Thread.interrupt() or both 1614 // check for interrupt event, otherwise it is timeout 1615 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1616 TEVENT(Wait - throw IEX from epilog); 1617 THROW(vmSymbols::java_lang_InterruptedException()); 1618 } 1619 } 1620 1621 // NOTE: Spurious wake up will be consider as timeout. 1622 // Monitor notify has precedence over thread interrupt. 1623 } 1624 1625 1626 // Consider: 1627 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1628 // then instead of transferring a thread from the WaitSet to the EntryList 1629 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1630 1631 void ObjectMonitor::INotify(Thread * Self) { 1632 const int policy = Knob_MoveNotifyee; 1633 1634 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1635 ObjectWaiter * iterator = DequeueWaiter(); 1636 if (iterator != NULL) { 1637 TEVENT(Notify1 - Transfer); 1638 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1639 guarantee(iterator->_notified == 0, "invariant"); 1640 // Disposition - what might we do with iterator ? 1641 // a. add it directly to the EntryList - either tail (policy == 1) 1642 // or head (policy == 0). 1643 // b. push it onto the front of the _cxq (policy == 2). 1644 // For now we use (b). 1645 if (policy != 4) { 1646 iterator->TState = ObjectWaiter::TS_ENTER; 1647 } 1648 iterator->_notified = 1; 1649 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1650 1651 ObjectWaiter * list = _EntryList; 1652 if (list != NULL) { 1653 assert(list->_prev == NULL, "invariant"); 1654 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1655 assert(list != iterator, "invariant"); 1656 } 1657 1658 if (policy == 0) { // prepend to EntryList 1659 if (list == NULL) { 1660 iterator->_next = iterator->_prev = NULL; 1661 _EntryList = iterator; 1662 } else { 1663 list->_prev = iterator; 1664 iterator->_next = list; 1665 iterator->_prev = NULL; 1666 _EntryList = iterator; 1667 } 1668 } else if (policy == 1) { // append to EntryList 1669 if (list == NULL) { 1670 iterator->_next = iterator->_prev = NULL; 1671 _EntryList = iterator; 1672 } else { 1673 // CONSIDER: finding the tail currently requires a linear-time walk of 1674 // the EntryList. We can make tail access constant-time by converting to 1675 // a CDLL instead of using our current DLL. 1676 ObjectWaiter * tail; 1677 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1678 assert(tail != NULL && tail->_next == NULL, "invariant"); 1679 tail->_next = iterator; 1680 iterator->_prev = tail; 1681 iterator->_next = NULL; 1682 } 1683 } else if (policy == 2) { // prepend to cxq 1684 if (list == NULL) { 1685 iterator->_next = iterator->_prev = NULL; 1686 _EntryList = iterator; 1687 } else { 1688 iterator->TState = ObjectWaiter::TS_CXQ; 1689 for (;;) { 1690 ObjectWaiter * front = _cxq; 1691 iterator->_next = front; 1692 if (Atomic::cmpxchg_ptr(iterator, &_cxq, front) == front) { 1693 break; 1694 } 1695 } 1696 } 1697 } else if (policy == 3) { // append to cxq 1698 iterator->TState = ObjectWaiter::TS_CXQ; 1699 for (;;) { 1700 ObjectWaiter * tail = _cxq; 1701 if (tail == NULL) { 1702 iterator->_next = NULL; 1703 if (Atomic::cmpxchg_ptr(iterator, &_cxq, NULL) == NULL) { 1704 break; 1705 } 1706 } else { 1707 while (tail->_next != NULL) tail = tail->_next; 1708 tail->_next = iterator; 1709 iterator->_prev = tail; 1710 iterator->_next = NULL; 1711 break; 1712 } 1713 } 1714 } else { 1715 ParkEvent * ev = iterator->_event; 1716 iterator->TState = ObjectWaiter::TS_RUN; 1717 OrderAccess::fence(); 1718 ev->unpark(); 1719 } 1720 1721 // _WaitSetLock protects the wait queue, not the EntryList. We could 1722 // move the add-to-EntryList operation, above, outside the critical section 1723 // protected by _WaitSetLock. In practice that's not useful. With the 1724 // exception of wait() timeouts and interrupts the monitor owner 1725 // is the only thread that grabs _WaitSetLock. There's almost no contention 1726 // on _WaitSetLock so it's not profitable to reduce the length of the 1727 // critical section. 1728 1729 if (policy < 4) { 1730 iterator->wait_reenter_begin(this); 1731 } 1732 } 1733 Thread::SpinRelease(&_WaitSetLock); 1734 } 1735 1736 // Consider: a not-uncommon synchronization bug is to use notify() when 1737 // notifyAll() is more appropriate, potentially resulting in stranded 1738 // threads; this is one example of a lost wakeup. A useful diagnostic 1739 // option is to force all notify() operations to behave as notifyAll(). 1740 // 1741 // Note: We can also detect many such problems with a "minimum wait". 1742 // When the "minimum wait" is set to a small non-zero timeout value 1743 // and the program does not hang whereas it did absent "minimum wait", 1744 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses 1745 // a "minimum wait" for all park() operations; see the recheckInterval 1746 // variable and MAX_RECHECK_INTERVAL. 1747 1748 void ObjectMonitor::notify(TRAPS) { 1749 CHECK_OWNER(); 1750 if (_WaitSet == NULL) { 1751 TEVENT(Empty-Notify); 1752 return; 1753 } 1754 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1755 INotify(THREAD); 1756 OM_PERFDATA_OP(Notifications, inc(1)); 1757 } 1758 1759 1760 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1761 // from the waitset to the EntryList. This could be done more efficiently with a 1762 // single bulk transfer but in practice it's not time-critical. Beware too, 1763 // that in prepend-mode we invert the order of the waiters. Let's say that the 1764 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1765 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1766 1767 void ObjectMonitor::notifyAll(TRAPS) { 1768 CHECK_OWNER(); 1769 if (_WaitSet == NULL) { 1770 TEVENT(Empty-NotifyAll); 1771 return; 1772 } 1773 1774 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1775 int tally = 0; 1776 while (_WaitSet != NULL) { 1777 tally++; 1778 INotify(THREAD); 1779 } 1780 1781 OM_PERFDATA_OP(Notifications, inc(tally)); 1782 } 1783 1784 // ----------------------------------------------------------------------------- 1785 // Adaptive Spinning Support 1786 // 1787 // Adaptive spin-then-block - rational spinning 1788 // 1789 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1790 // algorithm. On high order SMP systems it would be better to start with 1791 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1792 // a contending thread could enqueue itself on the cxq and then spin locally 1793 // on a thread-specific variable such as its ParkEvent._Event flag. 1794 // That's left as an exercise for the reader. Note that global spinning is 1795 // not problematic on Niagara, as the L2 cache serves the interconnect and 1796 // has both low latency and massive bandwidth. 1797 // 1798 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1799 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1800 // (duration) or we can fix the count at approximately the duration of 1801 // a context switch and vary the frequency. Of course we could also 1802 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1803 // For a description of 'Adaptive spin-then-block mutual exclusion in 1804 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1805 // 1806 // This implementation varies the duration "D", where D varies with 1807 // the success rate of recent spin attempts. (D is capped at approximately 1808 // length of a round-trip context switch). The success rate for recent 1809 // spin attempts is a good predictor of the success rate of future spin 1810 // attempts. The mechanism adapts automatically to varying critical 1811 // section length (lock modality), system load and degree of parallelism. 1812 // D is maintained per-monitor in _SpinDuration and is initialized 1813 // optimistically. Spin frequency is fixed at 100%. 1814 // 1815 // Note that _SpinDuration is volatile, but we update it without locks 1816 // or atomics. The code is designed so that _SpinDuration stays within 1817 // a reasonable range even in the presence of races. The arithmetic 1818 // operations on _SpinDuration are closed over the domain of legal values, 1819 // so at worst a race will install and older but still legal value. 1820 // At the very worst this introduces some apparent non-determinism. 1821 // We might spin when we shouldn't or vice-versa, but since the spin 1822 // count are relatively short, even in the worst case, the effect is harmless. 1823 // 1824 // Care must be taken that a low "D" value does not become an 1825 // an absorbing state. Transient spinning failures -- when spinning 1826 // is overall profitable -- should not cause the system to converge 1827 // on low "D" values. We want spinning to be stable and predictable 1828 // and fairly responsive to change and at the same time we don't want 1829 // it to oscillate, become metastable, be "too" non-deterministic, 1830 // or converge on or enter undesirable stable absorbing states. 1831 // 1832 // We implement a feedback-based control system -- using past behavior 1833 // to predict future behavior. We face two issues: (a) if the 1834 // input signal is random then the spin predictor won't provide optimal 1835 // results, and (b) if the signal frequency is too high then the control 1836 // system, which has some natural response lag, will "chase" the signal. 1837 // (b) can arise from multimodal lock hold times. Transient preemption 1838 // can also result in apparent bimodal lock hold times. 1839 // Although sub-optimal, neither condition is particularly harmful, as 1840 // in the worst-case we'll spin when we shouldn't or vice-versa. 1841 // The maximum spin duration is rather short so the failure modes aren't bad. 1842 // To be conservative, I've tuned the gain in system to bias toward 1843 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1844 // "rings" or oscillates between spinning and not spinning. This happens 1845 // when spinning is just on the cusp of profitability, however, so the 1846 // situation is not dire. The state is benign -- there's no need to add 1847 // hysteresis control to damp the transition rate between spinning and 1848 // not spinning. 1849 1850 // Spinning: Fixed frequency (100%), vary duration 1851 int ObjectMonitor::TrySpin(Thread * Self) { 1852 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1853 int ctr = Knob_FixedSpin; 1854 if (ctr != 0) { 1855 while (--ctr >= 0) { 1856 if (TryLock(Self) > 0) return 1; 1857 SpinPause(); 1858 } 1859 return 0; 1860 } 1861 1862 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1863 if (TryLock(Self) > 0) { 1864 // Increase _SpinDuration ... 1865 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1866 // Raising _SpurDuration to the poverty line is key. 1867 int x = _SpinDuration; 1868 if (x < Knob_SpinLimit) { 1869 if (x < Knob_Poverty) x = Knob_Poverty; 1870 _SpinDuration = x + Knob_BonusB; 1871 } 1872 return 1; 1873 } 1874 SpinPause(); 1875 } 1876 1877 // Admission control - verify preconditions for spinning 1878 // 1879 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1880 // becoming an absorbing state. Put another way, we spin briefly to 1881 // sample, just in case the system load, parallelism, contention, or lock 1882 // modality changed. 1883 // 1884 // Consider the following alternative: 1885 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1886 // spin attempt. "Periodically" might mean after a tally of 1887 // the # of failed spin attempts (or iterations) reaches some threshold. 1888 // This takes us into the realm of 1-out-of-N spinning, where we 1889 // hold the duration constant but vary the frequency. 1890 1891 ctr = _SpinDuration; 1892 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1893 if (ctr <= 0) return 0; 1894 1895 if (Knob_SuccRestrict && _succ != NULL) return 0; 1896 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1897 TEVENT(Spin abort - notrunnable [TOP]); 1898 return 0; 1899 } 1900 1901 int MaxSpin = Knob_MaxSpinners; 1902 if (MaxSpin >= 0) { 1903 if (_Spinner > MaxSpin) { 1904 TEVENT(Spin abort -- too many spinners); 1905 return 0; 1906 } 1907 // Slightly racy, but benign ... 1908 Adjust(&_Spinner, 1); 1909 } 1910 1911 // We're good to spin ... spin ingress. 1912 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1913 // when preparing to LD...CAS _owner, etc and the CAS is likely 1914 // to succeed. 1915 int hits = 0; 1916 int msk = 0; 1917 int caspty = Knob_CASPenalty; 1918 int oxpty = Knob_OXPenalty; 1919 int sss = Knob_SpinSetSucc; 1920 if (sss && _succ == NULL) _succ = Self; 1921 Thread * prv = NULL; 1922 1923 // There are three ways to exit the following loop: 1924 // 1. A successful spin where this thread has acquired the lock. 1925 // 2. Spin failure with prejudice 1926 // 3. Spin failure without prejudice 1927 1928 while (--ctr >= 0) { 1929 1930 // Periodic polling -- Check for pending GC 1931 // Threads may spin while they're unsafe. 1932 // We don't want spinning threads to delay the JVM from reaching 1933 // a stop-the-world safepoint or to steal cycles from GC. 1934 // If we detect a pending safepoint we abort in order that 1935 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1936 // this thread, if safe, doesn't steal cycles from GC. 1937 // This is in keeping with the "no loitering in runtime" rule. 1938 // We periodically check to see if there's a safepoint pending. 1939 if ((ctr & 0xFF) == 0) { 1940 if (SafepointMechanism::poll()) { 1941 TEVENT(Spin: safepoint); 1942 goto Abort; // abrupt spin egress 1943 } 1944 if (Knob_UsePause & 1) SpinPause(); 1945 } 1946 1947 if (Knob_UsePause & 2) SpinPause(); 1948 1949 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1950 // This is useful on classic SMP systems, but is of less utility on 1951 // N1-style CMT platforms. 1952 // 1953 // Trade-off: lock acquisition latency vs coherency bandwidth. 1954 // Lock hold times are typically short. A histogram 1955 // of successful spin attempts shows that we usually acquire 1956 // the lock early in the spin. That suggests we want to 1957 // sample _owner frequently in the early phase of the spin, 1958 // but then back-off and sample less frequently as the spin 1959 // progresses. The back-off makes a good citizen on SMP big 1960 // SMP systems. Oversampling _owner can consume excessive 1961 // coherency bandwidth. Relatedly, if we _oversample _owner we 1962 // can inadvertently interfere with the the ST m->owner=null. 1963 // executed by the lock owner. 1964 if (ctr & msk) continue; 1965 ++hits; 1966 if ((hits & 0xF) == 0) { 1967 // The 0xF, above, corresponds to the exponent. 1968 // Consider: (msk+1)|msk 1969 msk = ((msk << 2)|3) & BackOffMask; 1970 } 1971 1972 // Probe _owner with TATAS 1973 // If this thread observes the monitor transition or flicker 1974 // from locked to unlocked to locked, then the odds that this 1975 // thread will acquire the lock in this spin attempt go down 1976 // considerably. The same argument applies if the CAS fails 1977 // or if we observe _owner change from one non-null value to 1978 // another non-null value. In such cases we might abort 1979 // the spin without prejudice or apply a "penalty" to the 1980 // spin count-down variable "ctr", reducing it by 100, say. 1981 1982 Thread * ox = (Thread *) _owner; 1983 if (ox == NULL) { 1984 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 1985 if (ox == NULL) { 1986 // The CAS succeeded -- this thread acquired ownership 1987 // Take care of some bookkeeping to exit spin state. 1988 if (sss && _succ == Self) { 1989 _succ = NULL; 1990 } 1991 if (MaxSpin > 0) Adjust(&_Spinner, -1); 1992 1993 // Increase _SpinDuration : 1994 // The spin was successful (profitable) so we tend toward 1995 // longer spin attempts in the future. 1996 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1997 // If we acquired the lock early in the spin cycle it 1998 // makes sense to increase _SpinDuration proportionally. 1999 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2000 int x = _SpinDuration; 2001 if (x < Knob_SpinLimit) { 2002 if (x < Knob_Poverty) x = Knob_Poverty; 2003 _SpinDuration = x + Knob_Bonus; 2004 } 2005 return 1; 2006 } 2007 2008 // The CAS failed ... we can take any of the following actions: 2009 // * penalize: ctr -= Knob_CASPenalty 2010 // * exit spin with prejudice -- goto Abort; 2011 // * exit spin without prejudice. 2012 // * Since CAS is high-latency, retry again immediately. 2013 prv = ox; 2014 TEVENT(Spin: cas failed); 2015 if (caspty == -2) break; 2016 if (caspty == -1) goto Abort; 2017 ctr -= caspty; 2018 continue; 2019 } 2020 2021 // Did lock ownership change hands ? 2022 if (ox != prv && prv != NULL) { 2023 TEVENT(spin: Owner changed) 2024 if (oxpty == -2) break; 2025 if (oxpty == -1) goto Abort; 2026 ctr -= oxpty; 2027 } 2028 prv = ox; 2029 2030 // Abort the spin if the owner is not executing. 2031 // The owner must be executing in order to drop the lock. 2032 // Spinning while the owner is OFFPROC is idiocy. 2033 // Consider: ctr -= RunnablePenalty ; 2034 if (Knob_OState && NotRunnable (Self, ox)) { 2035 TEVENT(Spin abort - notrunnable); 2036 goto Abort; 2037 } 2038 if (sss && _succ == NULL) _succ = Self; 2039 } 2040 2041 // Spin failed with prejudice -- reduce _SpinDuration. 2042 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2043 // AIMD is globally stable. 2044 TEVENT(Spin failure); 2045 { 2046 int x = _SpinDuration; 2047 if (x > 0) { 2048 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2049 // This is globally sample and tends to damp the response. 2050 x -= Knob_Penalty; 2051 if (x < 0) x = 0; 2052 _SpinDuration = x; 2053 } 2054 } 2055 2056 Abort: 2057 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2058 if (sss && _succ == Self) { 2059 _succ = NULL; 2060 // Invariant: after setting succ=null a contending thread 2061 // must recheck-retry _owner before parking. This usually happens 2062 // in the normal usage of TrySpin(), but it's safest 2063 // to make TrySpin() as foolproof as possible. 2064 OrderAccess::fence(); 2065 if (TryLock(Self) > 0) return 1; 2066 } 2067 return 0; 2068 } 2069 2070 // NotRunnable() -- informed spinning 2071 // 2072 // Don't bother spinning if the owner is not eligible to drop the lock. 2073 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2074 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2075 // The thread must be runnable in order to drop the lock in timely fashion. 2076 // If the _owner is not runnable then spinning will not likely be 2077 // successful (profitable). 2078 // 2079 // Beware -- the thread referenced by _owner could have died 2080 // so a simply fetch from _owner->_thread_state might trap. 2081 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2082 // Because of the lifecycle issues the schedctl and _thread_state values 2083 // observed by NotRunnable() might be garbage. NotRunnable must 2084 // tolerate this and consider the observed _thread_state value 2085 // as advisory. 2086 // 2087 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2088 // a thread pointer. 2089 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2090 // with the LSB of _owner. Another option would be to probablistically probe 2091 // the putative _owner->TypeTag value. 2092 // 2093 // Checking _thread_state isn't perfect. Even if the thread is 2094 // in_java it might be blocked on a page-fault or have been preempted 2095 // and sitting on a ready/dispatch queue. _thread state in conjunction 2096 // with schedctl.sc_state gives us a good picture of what the 2097 // thread is doing, however. 2098 // 2099 // TODO: check schedctl.sc_state. 2100 // We'll need to use SafeFetch32() to read from the schedctl block. 2101 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2102 // 2103 // The return value from NotRunnable() is *advisory* -- the 2104 // result is based on sampling and is not necessarily coherent. 2105 // The caller must tolerate false-negative and false-positive errors. 2106 // Spinning, in general, is probabilistic anyway. 2107 2108 2109 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2110 // Check ox->TypeTag == 2BAD. 2111 if (ox == NULL) return 0; 2112 2113 // Avoid transitive spinning ... 2114 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2115 // Immediately after T1 acquires L it's possible that T2, also 2116 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2117 // This occurs transiently after T1 acquired L but before 2118 // T1 managed to clear T1.Stalled. T2 does not need to abort 2119 // its spin in this circumstance. 2120 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2121 2122 if (BlockedOn == 1) return 1; 2123 if (BlockedOn != 0) { 2124 return BlockedOn != intptr_t(this) && _owner == ox; 2125 } 2126 2127 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2128 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2129 // consider also: jst != _thread_in_Java -- but that's overspecific. 2130 return jst == _thread_blocked || jst == _thread_in_native; 2131 } 2132 2133 2134 // ----------------------------------------------------------------------------- 2135 // WaitSet management ... 2136 2137 ObjectWaiter::ObjectWaiter(Thread* thread) { 2138 _next = NULL; 2139 _prev = NULL; 2140 _notified = 0; 2141 TState = TS_RUN; 2142 _thread = thread; 2143 _event = thread->_ParkEvent; 2144 _active = false; 2145 assert(_event != NULL, "invariant"); 2146 } 2147 2148 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2149 JavaThread *jt = (JavaThread *)this->_thread; 2150 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2151 } 2152 2153 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2154 JavaThread *jt = (JavaThread *)this->_thread; 2155 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2156 } 2157 2158 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2159 assert(node != NULL, "should not add NULL node"); 2160 assert(node->_prev == NULL, "node already in list"); 2161 assert(node->_next == NULL, "node already in list"); 2162 // put node at end of queue (circular doubly linked list) 2163 if (_WaitSet == NULL) { 2164 _WaitSet = node; 2165 node->_prev = node; 2166 node->_next = node; 2167 } else { 2168 ObjectWaiter* head = _WaitSet; 2169 ObjectWaiter* tail = head->_prev; 2170 assert(tail->_next == head, "invariant check"); 2171 tail->_next = node; 2172 head->_prev = node; 2173 node->_next = head; 2174 node->_prev = tail; 2175 } 2176 } 2177 2178 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2179 // dequeue the very first waiter 2180 ObjectWaiter* waiter = _WaitSet; 2181 if (waiter) { 2182 DequeueSpecificWaiter(waiter); 2183 } 2184 return waiter; 2185 } 2186 2187 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2188 assert(node != NULL, "should not dequeue NULL node"); 2189 assert(node->_prev != NULL, "node already removed from list"); 2190 assert(node->_next != NULL, "node already removed from list"); 2191 // when the waiter has woken up because of interrupt, 2192 // timeout or other spurious wake-up, dequeue the 2193 // waiter from waiting list 2194 ObjectWaiter* next = node->_next; 2195 if (next == node) { 2196 assert(node->_prev == node, "invariant check"); 2197 _WaitSet = NULL; 2198 } else { 2199 ObjectWaiter* prev = node->_prev; 2200 assert(prev->_next == node, "invariant check"); 2201 assert(next->_prev == node, "invariant check"); 2202 next->_prev = prev; 2203 prev->_next = next; 2204 if (_WaitSet == node) { 2205 _WaitSet = next; 2206 } 2207 } 2208 node->_next = NULL; 2209 node->_prev = NULL; 2210 } 2211 2212 // ----------------------------------------------------------------------------- 2213 // PerfData support 2214 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2215 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2216 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2217 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2218 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2219 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2220 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2221 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2222 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2223 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2224 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2225 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2226 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2227 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2228 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2229 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2230 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2231 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2232 2233 // One-shot global initialization for the sync subsystem. 2234 // We could also defer initialization and initialize on-demand 2235 // the first time we call inflate(). Initialization would 2236 // be protected - like so many things - by the MonitorCache_lock. 2237 2238 void ObjectMonitor::Initialize() { 2239 static int InitializationCompleted = 0; 2240 assert(InitializationCompleted == 0, "invariant"); 2241 InitializationCompleted = 1; 2242 if (UsePerfData) { 2243 EXCEPTION_MARK; 2244 #define NEWPERFCOUNTER(n) \ 2245 { \ 2246 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2247 CHECK); \ 2248 } 2249 #define NEWPERFVARIABLE(n) \ 2250 { \ 2251 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2252 CHECK); \ 2253 } 2254 NEWPERFCOUNTER(_sync_Inflations); 2255 NEWPERFCOUNTER(_sync_Deflations); 2256 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2257 NEWPERFCOUNTER(_sync_FutileWakeups); 2258 NEWPERFCOUNTER(_sync_Parks); 2259 NEWPERFCOUNTER(_sync_EmptyNotifications); 2260 NEWPERFCOUNTER(_sync_Notifications); 2261 NEWPERFCOUNTER(_sync_SlowEnter); 2262 NEWPERFCOUNTER(_sync_SlowExit); 2263 NEWPERFCOUNTER(_sync_SlowNotify); 2264 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2265 NEWPERFCOUNTER(_sync_FailedSpins); 2266 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2267 NEWPERFCOUNTER(_sync_PrivateA); 2268 NEWPERFCOUNTER(_sync_PrivateB); 2269 NEWPERFCOUNTER(_sync_MonInCirculation); 2270 NEWPERFCOUNTER(_sync_MonScavenged); 2271 NEWPERFVARIABLE(_sync_MonExtant); 2272 #undef NEWPERFCOUNTER 2273 #undef NEWPERFVARIABLE 2274 } 2275 } 2276 2277 static char * kvGet(char * kvList, const char * Key) { 2278 if (kvList == NULL) return NULL; 2279 size_t n = strlen(Key); 2280 char * Search; 2281 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2282 if (strncmp (Search, Key, n) == 0) { 2283 if (Search[n] == '=') return Search + n + 1; 2284 if (Search[n] == 0) return(char *) "1"; 2285 } 2286 } 2287 return NULL; 2288 } 2289 2290 static int kvGetInt(char * kvList, const char * Key, int Default) { 2291 char * v = kvGet(kvList, Key); 2292 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2293 if (Knob_ReportSettings && v != NULL) { 2294 tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ; 2295 tty->flush(); 2296 } 2297 return rslt; 2298 } 2299 2300 void ObjectMonitor::DeferredInitialize() { 2301 if (InitDone > 0) return; 2302 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2303 while (InitDone != 1) /* empty */; 2304 return; 2305 } 2306 2307 // One-shot global initialization ... 2308 // The initialization is idempotent, so we don't need locks. 2309 // In the future consider doing this via os::init_2(). 2310 // SyncKnobs consist of <Key>=<Value> pairs in the style 2311 // of environment variables. Start by converting ':' to NUL. 2312 2313 if (SyncKnobs == NULL) SyncKnobs = ""; 2314 2315 size_t sz = strlen(SyncKnobs); 2316 char * knobs = (char *) malloc(sz + 2); 2317 if (knobs == NULL) { 2318 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2319 guarantee(0, "invariant"); 2320 } 2321 strcpy(knobs, SyncKnobs); 2322 knobs[sz+1] = 0; 2323 for (char * p = knobs; *p; p++) { 2324 if (*p == ':') *p = 0; 2325 } 2326 2327 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2328 SETKNOB(ReportSettings); 2329 SETKNOB(ExitRelease); 2330 SETKNOB(Verbose); 2331 SETKNOB(VerifyInUse); 2332 SETKNOB(VerifyMatch); 2333 SETKNOB(FixedSpin); 2334 SETKNOB(SpinLimit); 2335 SETKNOB(SpinBase); 2336 SETKNOB(SpinBackOff); 2337 SETKNOB(CASPenalty); 2338 SETKNOB(OXPenalty); 2339 SETKNOB(LogSpins); 2340 SETKNOB(SpinSetSucc); 2341 SETKNOB(SuccEnabled); 2342 SETKNOB(SuccRestrict); 2343 SETKNOB(Penalty); 2344 SETKNOB(Bonus); 2345 SETKNOB(BonusB); 2346 SETKNOB(Poverty); 2347 SETKNOB(SpinAfterFutile); 2348 SETKNOB(UsePause); 2349 SETKNOB(SpinEarly); 2350 SETKNOB(OState); 2351 SETKNOB(MaxSpinners); 2352 SETKNOB(PreSpin); 2353 SETKNOB(ExitPolicy); 2354 SETKNOB(QMode); 2355 SETKNOB(ResetEvent); 2356 SETKNOB(MoveNotifyee); 2357 SETKNOB(FastHSSEC); 2358 #undef SETKNOB 2359 2360 if (Knob_Verbose) { 2361 sanity_checks(); 2362 } 2363 2364 if (os::is_MP()) { 2365 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2366 if (Knob_ReportSettings) { 2367 tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask); 2368 } 2369 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2370 } else { 2371 Knob_SpinLimit = 0; 2372 Knob_SpinBase = 0; 2373 Knob_PreSpin = 0; 2374 Knob_FixedSpin = -1; 2375 } 2376 2377 if (Knob_LogSpins == 0) { 2378 ObjectMonitor::_sync_FailedSpins = NULL; 2379 } 2380 2381 free(knobs); 2382 OrderAccess::fence(); 2383 InitDone = 1; 2384 } 2385 2386 void ObjectMonitor::sanity_checks() { 2387 int error_cnt = 0; 2388 int warning_cnt = 0; 2389 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2390 2391 if (verbose) { 2392 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2393 sizeof(ObjectMonitor)); 2394 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT, 2395 sizeof(PaddedEnd<ObjectMonitor>)); 2396 } 2397 2398 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2399 if (verbose) { 2400 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2401 } 2402 2403 ObjectMonitor dummy; 2404 u_char *addr_begin = (u_char*)&dummy; 2405 u_char *addr_header = (u_char*)&dummy._header; 2406 u_char *addr_owner = (u_char*)&dummy._owner; 2407 2408 uint offset_header = (uint)(addr_header - addr_begin); 2409 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2410 2411 uint offset_owner = (uint)(addr_owner - addr_begin); 2412 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2413 2414 if ((uint)(addr_header - addr_begin) != 0) { 2415 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2416 error_cnt++; 2417 } 2418 2419 if (cache_line_size != 0) { 2420 // We were able to determine the L1 data cache line size so 2421 // do some cache line specific sanity checks 2422 2423 if ((offset_owner - offset_header) < cache_line_size) { 2424 tty->print_cr("WARNING: the _header and _owner fields are closer " 2425 "than a cache line which permits false sharing."); 2426 warning_cnt++; 2427 } 2428 2429 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) { 2430 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a " 2431 "multiple of a cache line which permits false sharing."); 2432 warning_cnt++; 2433 } 2434 } 2435 2436 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2437 &warning_cnt); 2438 2439 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2440 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2441 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2442 } 2443 2444 guarantee(error_cnt == 0, 2445 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2446 } 2447 2448 #ifndef PRODUCT 2449 void ObjectMonitor_test() { 2450 ObjectMonitor::sanity_checks(); 2451 } 2452 #endif