1 /* 2 * Copyright (c) 1998, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "oops/markOop.hpp" 29 #include "oops/oop.inline.hpp" 30 #include "runtime/atomic.inline.hpp" 31 #include "runtime/handles.inline.hpp" 32 #include "runtime/interfaceSupport.hpp" 33 #include "runtime/mutexLocker.hpp" 34 #include "runtime/objectMonitor.hpp" 35 #include "runtime/objectMonitor.inline.hpp" 36 #include "runtime/orderAccess.inline.hpp" 37 #include "runtime/osThread.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #include "runtime/thread.inline.hpp" 40 #include "services/threadService.hpp" 41 #include "trace/tracing.hpp" 42 #include "trace/traceMacros.hpp" 43 #include "utilities/dtrace.hpp" 44 #include "utilities/macros.hpp" 45 #include "utilities/preserveException.hpp" 46 47 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 48 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 49 #define NOINLINE __attribute__((noinline)) 50 #else 51 #define NOINLINE 52 #endif 53 54 55 #ifdef DTRACE_ENABLED 56 57 // Only bother with this argument setup if dtrace is available 58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95 #else // ndef DTRACE_ENABLED 96 97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100 #endif // ndef DTRACE_ENABLED 101 102 // Tunables ... 103 // The knob* variables are effectively final. Once set they should 104 // never be modified hence. Consider using __read_mostly with GCC. 105 106 int ObjectMonitor::Knob_ExitRelease = 0; 107 int ObjectMonitor::Knob_Verbose = 0; 108 int ObjectMonitor::Knob_VerifyInUse = 0; 109 int ObjectMonitor::Knob_VerifyMatch = 0; 110 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 111 static int Knob_LogSpins = 0; // enable jvmstat tally for spins 112 static int Knob_HandOff = 0; 113 static int Knob_ReportSettings = 0; 114 115 static int Knob_SpinBase = 0; // Floor AKA SpinMin 116 static int Knob_SpinBackOff = 0; // spin-loop backoff 117 static int Knob_CASPenalty = -1; // Penalty for failed CAS 118 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 119 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 120 static int Knob_SpinEarly = 1; 121 static int Knob_SuccEnabled = 1; // futile wake throttling 122 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 123 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 124 static int Knob_Bonus = 100; // spin success bonus 125 static int Knob_BonusB = 100; // spin success bonus 126 static int Knob_Penalty = 200; // spin failure penalty 127 static int Knob_Poverty = 1000; 128 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 129 static int Knob_FixedSpin = 0; 130 static int Knob_OState = 3; // Spinner checks thread state of _owner 131 static int Knob_UsePause = 1; 132 static int Knob_ExitPolicy = 0; 133 static int Knob_PreSpin = 10; // 20-100 likely better 134 static int Knob_ResetEvent = 0; 135 static int BackOffMask = 0; 136 137 static int Knob_FastHSSEC = 0; 138 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 139 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 140 static volatile int InitDone = 0; 141 142 #define TrySpin TrySpin_VaryDuration 143 144 // ----------------------------------------------------------------------------- 145 // Theory of operations -- Monitors lists, thread residency, etc: 146 // 147 // * A thread acquires ownership of a monitor by successfully 148 // CAS()ing the _owner field from null to non-null. 149 // 150 // * Invariant: A thread appears on at most one monitor list -- 151 // cxq, EntryList or WaitSet -- at any one time. 152 // 153 // * Contending threads "push" themselves onto the cxq with CAS 154 // and then spin/park. 155 // 156 // * After a contending thread eventually acquires the lock it must 157 // dequeue itself from either the EntryList or the cxq. 158 // 159 // * The exiting thread identifies and unparks an "heir presumptive" 160 // tentative successor thread on the EntryList. Critically, the 161 // exiting thread doesn't unlink the successor thread from the EntryList. 162 // After having been unparked, the wakee will recontend for ownership of 163 // the monitor. The successor (wakee) will either acquire the lock or 164 // re-park itself. 165 // 166 // Succession is provided for by a policy of competitive handoff. 167 // The exiting thread does _not_ grant or pass ownership to the 168 // successor thread. (This is also referred to as "handoff" succession"). 169 // Instead the exiting thread releases ownership and possibly wakes 170 // a successor, so the successor can (re)compete for ownership of the lock. 171 // If the EntryList is empty but the cxq is populated the exiting 172 // thread will drain the cxq into the EntryList. It does so by 173 // by detaching the cxq (installing null with CAS) and folding 174 // the threads from the cxq into the EntryList. The EntryList is 175 // doubly linked, while the cxq is singly linked because of the 176 // CAS-based "push" used to enqueue recently arrived threads (RATs). 177 // 178 // * Concurrency invariants: 179 // 180 // -- only the monitor owner may access or mutate the EntryList. 181 // The mutex property of the monitor itself protects the EntryList 182 // from concurrent interference. 183 // -- Only the monitor owner may detach the cxq. 184 // 185 // * The monitor entry list operations avoid locks, but strictly speaking 186 // they're not lock-free. Enter is lock-free, exit is not. 187 // For a description of 'Methods and apparatus providing non-blocking access 188 // to a resource,' see U.S. Pat. No. 7844973. 189 // 190 // * The cxq can have multiple concurrent "pushers" but only one concurrent 191 // detaching thread. This mechanism is immune from the ABA corruption. 192 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 193 // 194 // * Taken together, the cxq and the EntryList constitute or form a 195 // single logical queue of threads stalled trying to acquire the lock. 196 // We use two distinct lists to improve the odds of a constant-time 197 // dequeue operation after acquisition (in the ::enter() epilogue) and 198 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 199 // A key desideratum is to minimize queue & monitor metadata manipulation 200 // that occurs while holding the monitor lock -- that is, we want to 201 // minimize monitor lock holds times. Note that even a small amount of 202 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 203 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 204 // locks and monitor metadata. 205 // 206 // Cxq points to the set of Recently Arrived Threads attempting entry. 207 // Because we push threads onto _cxq with CAS, the RATs must take the form of 208 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 209 // the unlocking thread notices that EntryList is null but _cxq is != null. 210 // 211 // The EntryList is ordered by the prevailing queue discipline and 212 // can be organized in any convenient fashion, such as a doubly-linked list or 213 // a circular doubly-linked list. Critically, we want insert and delete operations 214 // to operate in constant-time. If we need a priority queue then something akin 215 // to Solaris' sleepq would work nicely. Viz., 216 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 217 // Queue discipline is enforced at ::exit() time, when the unlocking thread 218 // drains the cxq into the EntryList, and orders or reorders the threads on the 219 // EntryList accordingly. 220 // 221 // Barring "lock barging", this mechanism provides fair cyclic ordering, 222 // somewhat similar to an elevator-scan. 223 // 224 // * The monitor synchronization subsystem avoids the use of native 225 // synchronization primitives except for the narrow platform-specific 226 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 227 // the semantics of park-unpark. Put another way, this monitor implementation 228 // depends only on atomic operations and park-unpark. The monitor subsystem 229 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 230 // underlying OS manages the READY<->RUN transitions. 231 // 232 // * Waiting threads reside on the WaitSet list -- wait() puts 233 // the caller onto the WaitSet. 234 // 235 // * notify() or notifyAll() simply transfers threads from the WaitSet to 236 // either the EntryList or cxq. Subsequent exit() operations will 237 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 238 // it's likely the notifyee would simply impale itself on the lock held 239 // by the notifier. 240 // 241 // * An interesting alternative is to encode cxq as (List,LockByte) where 242 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 243 // variable, like _recursions, in the scheme. The threads or Events that form 244 // the list would have to be aligned in 256-byte addresses. A thread would 245 // try to acquire the lock or enqueue itself with CAS, but exiting threads 246 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 247 // Note that is is *not* word-tearing, but it does presume that full-word 248 // CAS operations are coherent with intermix with STB operations. That's true 249 // on most common processors. 250 // 251 // * See also http://blogs.sun.com/dave 252 253 254 // ----------------------------------------------------------------------------- 255 // Enter support 256 257 void NOINLINE ObjectMonitor::enter(TRAPS) { 258 // The following code is ordered to check the most common cases first 259 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 260 Thread * const Self = THREAD; 261 262 void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL); 263 if (cur == NULL) { 264 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 265 assert(_recursions == 0, "invariant"); 266 assert(_owner == Self, "invariant"); 267 return; 268 } 269 270 if (cur == Self) { 271 // TODO-FIXME: check for integer overflow! BUGID 6557169. 272 _recursions++; 273 return; 274 } 275 276 if (Self->is_lock_owned ((address)cur)) { 277 assert(_recursions == 0, "internal state error"); 278 _recursions = 1; 279 // Commute owner from a thread-specific on-stack BasicLockObject address to 280 // a full-fledged "Thread *". 281 _owner = Self; 282 return; 283 } 284 285 // We've encountered genuine contention. 286 assert(Self->_Stalled == 0, "invariant"); 287 Self->_Stalled = intptr_t(this); 288 289 // Try one round of spinning *before* enqueueing Self 290 // and before going through the awkward and expensive state 291 // transitions. The following spin is strictly optional ... 292 // Note that if we acquire the monitor from an initial spin 293 // we forgo posting JVMTI events and firing DTRACE probes. 294 if (Knob_SpinEarly && TrySpin (Self) > 0) { 295 assert(_owner == Self, "invariant"); 296 assert(_recursions == 0, "invariant"); 297 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 298 Self->_Stalled = 0; 299 return; 300 } 301 302 assert(_owner != Self, "invariant"); 303 assert(_succ != Self, "invariant"); 304 assert(Self->is_Java_thread(), "invariant"); 305 JavaThread * jt = (JavaThread *) Self; 306 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 307 assert(jt->thread_state() != _thread_blocked, "invariant"); 308 assert(this->object() != NULL, "invariant"); 309 assert(_count >= 0, "invariant"); 310 311 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 312 // Ensure the object-monitor relationship remains stable while there's contention. 313 Atomic::inc(&_count); 314 315 EventJavaMonitorEnter event; 316 317 { // Change java thread status to indicate blocked on monitor enter. 318 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 319 320 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 321 if (JvmtiExport::should_post_monitor_contended_enter()) { 322 JvmtiExport::post_monitor_contended_enter(jt, this); 323 324 // The current thread does not yet own the monitor and does not 325 // yet appear on any queues that would get it made the successor. 326 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 327 // handler cannot accidentally consume an unpark() meant for the 328 // ParkEvent associated with this ObjectMonitor. 329 } 330 331 OSThreadContendState osts(Self->osthread()); 332 ThreadBlockInVM tbivm(jt); 333 334 Self->set_current_pending_monitor(this); 335 336 // TODO-FIXME: change the following for(;;) loop to straight-line code. 337 for (;;) { 338 jt->set_suspend_equivalent(); 339 // cleared by handle_special_suspend_equivalent_condition() 340 // or java_suspend_self() 341 342 EnterI(THREAD); 343 344 if (!ExitSuspendEquivalent(jt)) break; 345 346 // We have acquired the contended monitor, but while we were 347 // waiting another thread suspended us. We don't want to enter 348 // the monitor while suspended because that would surprise the 349 // thread that suspended us. 350 // 351 _recursions = 0; 352 _succ = NULL; 353 exit(false, Self); 354 355 jt->java_suspend_self(); 356 } 357 Self->set_current_pending_monitor(NULL); 358 359 // We cleared the pending monitor info since we've just gotten past 360 // the enter-check-for-suspend dance and we now own the monitor free 361 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 362 // destructor can go to a safepoint at the end of this block. If we 363 // do a thread dump during that safepoint, then this thread will show 364 // as having "-locked" the monitor, but the OS and java.lang.Thread 365 // states will still report that the thread is blocked trying to 366 // acquire it. 367 } 368 369 Atomic::dec(&_count); 370 assert(_count >= 0, "invariant"); 371 Self->_Stalled = 0; 372 373 // Must either set _recursions = 0 or ASSERT _recursions == 0. 374 assert(_recursions == 0, "invariant"); 375 assert(_owner == Self, "invariant"); 376 assert(_succ != Self, "invariant"); 377 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 378 379 // The thread -- now the owner -- is back in vm mode. 380 // Report the glorious news via TI,DTrace and jvmstat. 381 // The probe effect is non-trivial. All the reportage occurs 382 // while we hold the monitor, increasing the length of the critical 383 // section. Amdahl's parallel speedup law comes vividly into play. 384 // 385 // Another option might be to aggregate the events (thread local or 386 // per-monitor aggregation) and defer reporting until a more opportune 387 // time -- such as next time some thread encounters contention but has 388 // yet to acquire the lock. While spinning that thread could 389 // spinning we could increment JVMStat counters, etc. 390 391 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 392 if (JvmtiExport::should_post_monitor_contended_entered()) { 393 JvmtiExport::post_monitor_contended_entered(jt, this); 394 395 // The current thread already owns the monitor and is not going to 396 // call park() for the remainder of the monitor enter protocol. So 397 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 398 // event handler consumed an unpark() issued by the thread that 399 // just exited the monitor. 400 } 401 402 if (event.should_commit()) { 403 event.set_klass(((oop)this->object())->klass()); 404 event.set_previousOwner((TYPE_THREAD)_previous_owner_tid); 405 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 406 event.commit(); 407 } 408 409 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 410 } 411 412 413 // Caveat: TryLock() is not necessarily serializing if it returns failure. 414 // Callers must compensate as needed. 415 416 int ObjectMonitor::TryLock(Thread * Self) { 417 void * own = _owner; 418 if (own != NULL) return 0; 419 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 420 // Either guarantee _recursions == 0 or set _recursions = 0. 421 assert(_recursions == 0, "invariant"); 422 assert(_owner == Self, "invariant"); 423 return 1; 424 } 425 // The lock had been free momentarily, but we lost the race to the lock. 426 // Interference -- the CAS failed. 427 // We can either return -1 or retry. 428 // Retry doesn't make as much sense because the lock was just acquired. 429 return -1; 430 } 431 432 #define MAX_RECHECK_INTERVAL 1000 433 434 void NOINLINE ObjectMonitor::EnterI(TRAPS) { 435 Thread * const Self = THREAD; 436 assert(Self->is_Java_thread(), "invariant"); 437 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 438 439 // Try the lock - TATAS 440 if (TryLock (Self) > 0) { 441 assert(_succ != Self, "invariant"); 442 assert(_owner == Self, "invariant"); 443 assert(_Responsible != Self, "invariant"); 444 return; 445 } 446 447 DeferredInitialize(); 448 449 // We try one round of spinning *before* enqueueing Self. 450 // 451 // If the _owner is ready but OFFPROC we could use a YieldTo() 452 // operation to donate the remainder of this thread's quantum 453 // to the owner. This has subtle but beneficial affinity 454 // effects. 455 456 if (TrySpin (Self) > 0) { 457 assert(_owner == Self, "invariant"); 458 assert(_succ != Self, "invariant"); 459 assert(_Responsible != Self, "invariant"); 460 return; 461 } 462 463 // The Spin failed -- Enqueue and park the thread ... 464 assert(_succ != Self, "invariant"); 465 assert(_owner != Self, "invariant"); 466 assert(_Responsible != Self, "invariant"); 467 468 // Enqueue "Self" on ObjectMonitor's _cxq. 469 // 470 // Node acts as a proxy for Self. 471 // As an aside, if were to ever rewrite the synchronization code mostly 472 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 473 // Java objects. This would avoid awkward lifecycle and liveness issues, 474 // as well as eliminate a subset of ABA issues. 475 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 476 477 ObjectWaiter node(Self); 478 Self->_ParkEvent->reset(); 479 node._prev = (ObjectWaiter *) 0xBAD; 480 node.TState = ObjectWaiter::TS_CXQ; 481 482 // Push "Self" onto the front of the _cxq. 483 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 484 // Note that spinning tends to reduce the rate at which threads 485 // enqueue and dequeue on EntryList|cxq. 486 ObjectWaiter * nxt; 487 for (;;) { 488 node._next = nxt = _cxq; 489 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break; 490 491 // Interference - the CAS failed because _cxq changed. Just retry. 492 // As an optional optimization we retry the lock. 493 if (TryLock (Self) > 0) { 494 assert(_succ != Self, "invariant"); 495 assert(_owner == Self, "invariant"); 496 assert(_Responsible != Self, "invariant"); 497 return; 498 } 499 } 500 501 // Check for cxq|EntryList edge transition to non-null. This indicates 502 // the onset of contention. While contention persists exiting threads 503 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 504 // operations revert to the faster 1-0 mode. This enter operation may interleave 505 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 506 // arrange for one of the contending thread to use a timed park() operations 507 // to detect and recover from the race. (Stranding is form of progress failure 508 // where the monitor is unlocked but all the contending threads remain parked). 509 // That is, at least one of the contended threads will periodically poll _owner. 510 // One of the contending threads will become the designated "Responsible" thread. 511 // The Responsible thread uses a timed park instead of a normal indefinite park 512 // operation -- it periodically wakes and checks for and recovers from potential 513 // strandings admitted by 1-0 exit operations. We need at most one Responsible 514 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 515 // be responsible for a monitor. 516 // 517 // Currently, one of the contended threads takes on the added role of "Responsible". 518 // A viable alternative would be to use a dedicated "stranding checker" thread 519 // that periodically iterated over all the threads (or active monitors) and unparked 520 // successors where there was risk of stranding. This would help eliminate the 521 // timer scalability issues we see on some platforms as we'd only have one thread 522 // -- the checker -- parked on a timer. 523 524 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 525 // Try to assume the role of responsible thread for the monitor. 526 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 527 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 528 } 529 530 // The lock might have been released while this thread was occupied queueing 531 // itself onto _cxq. To close the race and avoid "stranding" and 532 // progress-liveness failure we must resample-retry _owner before parking. 533 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 534 // In this case the ST-MEMBAR is accomplished with CAS(). 535 // 536 // TODO: Defer all thread state transitions until park-time. 537 // Since state transitions are heavy and inefficient we'd like 538 // to defer the state transitions until absolutely necessary, 539 // and in doing so avoid some transitions ... 540 541 TEVENT(Inflated enter - Contention); 542 int nWakeups = 0; 543 int recheckInterval = 1; 544 545 for (;;) { 546 547 if (TryLock(Self) > 0) break; 548 assert(_owner != Self, "invariant"); 549 550 if ((SyncFlags & 2) && _Responsible == NULL) { 551 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 552 } 553 554 // park self 555 if (_Responsible == Self || (SyncFlags & 1)) { 556 TEVENT(Inflated enter - park TIMED); 557 Self->_ParkEvent->park((jlong) recheckInterval); 558 // Increase the recheckInterval, but clamp the value. 559 recheckInterval *= 8; 560 if (recheckInterval > MAX_RECHECK_INTERVAL) { 561 recheckInterval = MAX_RECHECK_INTERVAL; 562 } 563 } else { 564 TEVENT(Inflated enter - park UNTIMED); 565 Self->_ParkEvent->park(); 566 } 567 568 if (TryLock(Self) > 0) break; 569 570 // The lock is still contested. 571 // Keep a tally of the # of futile wakeups. 572 // Note that the counter is not protected by a lock or updated by atomics. 573 // That is by design - we trade "lossy" counters which are exposed to 574 // races during updates for a lower probe effect. 575 TEVENT(Inflated enter - Futile wakeup); 576 // This PerfData object can be used in parallel with a safepoint. 577 // See the work around in PerfDataManager::destroy(). 578 OM_PERFDATA_OP(FutileWakeups, inc()); 579 ++nWakeups; 580 581 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 582 // We can defer clearing _succ until after the spin completes 583 // TrySpin() must tolerate being called with _succ == Self. 584 // Try yet another round of adaptive spinning. 585 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 586 587 // We can find that we were unpark()ed and redesignated _succ while 588 // we were spinning. That's harmless. If we iterate and call park(), 589 // park() will consume the event and return immediately and we'll 590 // just spin again. This pattern can repeat, leaving _succ to simply 591 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 592 // Alternately, we can sample fired() here, and if set, forgo spinning 593 // in the next iteration. 594 595 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 596 Self->_ParkEvent->reset(); 597 OrderAccess::fence(); 598 } 599 if (_succ == Self) _succ = NULL; 600 601 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 602 OrderAccess::fence(); 603 } 604 605 // Egress : 606 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 607 // Normally we'll find Self on the EntryList . 608 // From the perspective of the lock owner (this thread), the 609 // EntryList is stable and cxq is prepend-only. 610 // The head of cxq is volatile but the interior is stable. 611 // In addition, Self.TState is stable. 612 613 assert(_owner == Self, "invariant"); 614 assert(object() != NULL, "invariant"); 615 // I'd like to write: 616 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 617 // but as we're at a safepoint that's not safe. 618 619 UnlinkAfterAcquire(Self, &node); 620 if (_succ == Self) _succ = NULL; 621 622 assert(_succ != Self, "invariant"); 623 if (_Responsible == Self) { 624 _Responsible = NULL; 625 OrderAccess::fence(); // Dekker pivot-point 626 627 // We may leave threads on cxq|EntryList without a designated 628 // "Responsible" thread. This is benign. When this thread subsequently 629 // exits the monitor it can "see" such preexisting "old" threads -- 630 // threads that arrived on the cxq|EntryList before the fence, above -- 631 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 632 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 633 // non-null and elect a new "Responsible" timer thread. 634 // 635 // This thread executes: 636 // ST Responsible=null; MEMBAR (in enter epilogue - here) 637 // LD cxq|EntryList (in subsequent exit) 638 // 639 // Entering threads in the slow/contended path execute: 640 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 641 // The (ST cxq; MEMBAR) is accomplished with CAS(). 642 // 643 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 644 // exit operation from floating above the ST Responsible=null. 645 } 646 647 // We've acquired ownership with CAS(). 648 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 649 // But since the CAS() this thread may have also stored into _succ, 650 // EntryList, cxq or Responsible. These meta-data updates must be 651 // visible __before this thread subsequently drops the lock. 652 // Consider what could occur if we didn't enforce this constraint -- 653 // STs to monitor meta-data and user-data could reorder with (become 654 // visible after) the ST in exit that drops ownership of the lock. 655 // Some other thread could then acquire the lock, but observe inconsistent 656 // or old monitor meta-data and heap data. That violates the JMM. 657 // To that end, the 1-0 exit() operation must have at least STST|LDST 658 // "release" barrier semantics. Specifically, there must be at least a 659 // STST|LDST barrier in exit() before the ST of null into _owner that drops 660 // the lock. The barrier ensures that changes to monitor meta-data and data 661 // protected by the lock will be visible before we release the lock, and 662 // therefore before some other thread (CPU) has a chance to acquire the lock. 663 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 664 // 665 // Critically, any prior STs to _succ or EntryList must be visible before 666 // the ST of null into _owner in the *subsequent* (following) corresponding 667 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 668 // execute a serializing instruction. 669 670 if (SyncFlags & 8) { 671 OrderAccess::fence(); 672 } 673 return; 674 } 675 676 // ReenterI() is a specialized inline form of the latter half of the 677 // contended slow-path from EnterI(). We use ReenterI() only for 678 // monitor reentry in wait(). 679 // 680 // In the future we should reconcile EnterI() and ReenterI(), adding 681 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 682 // loop accordingly. 683 684 void NOINLINE ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 685 assert(Self != NULL, "invariant"); 686 assert(SelfNode != NULL, "invariant"); 687 assert(SelfNode->_thread == Self, "invariant"); 688 assert(_waiters > 0, "invariant"); 689 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 690 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 691 JavaThread * jt = (JavaThread *) Self; 692 693 int nWakeups = 0; 694 for (;;) { 695 ObjectWaiter::TStates v = SelfNode->TState; 696 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 697 assert(_owner != Self, "invariant"); 698 699 if (TryLock(Self) > 0) break; 700 if (TrySpin(Self) > 0) break; 701 702 TEVENT(Wait Reentry - parking); 703 704 // State transition wrappers around park() ... 705 // ReenterI() wisely defers state transitions until 706 // it's clear we must park the thread. 707 { 708 OSThreadContendState osts(Self->osthread()); 709 ThreadBlockInVM tbivm(jt); 710 711 // cleared by handle_special_suspend_equivalent_condition() 712 // or java_suspend_self() 713 jt->set_suspend_equivalent(); 714 if (SyncFlags & 1) { 715 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL); 716 } else { 717 Self->_ParkEvent->park(); 718 } 719 720 // were we externally suspended while we were waiting? 721 for (;;) { 722 if (!ExitSuspendEquivalent(jt)) break; 723 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 724 jt->java_suspend_self(); 725 jt->set_suspend_equivalent(); 726 } 727 } 728 729 // Try again, but just so we distinguish between futile wakeups and 730 // successful wakeups. The following test isn't algorithmically 731 // necessary, but it helps us maintain sensible statistics. 732 if (TryLock(Self) > 0) break; 733 734 // The lock is still contested. 735 // Keep a tally of the # of futile wakeups. 736 // Note that the counter is not protected by a lock or updated by atomics. 737 // That is by design - we trade "lossy" counters which are exposed to 738 // races during updates for a lower probe effect. 739 TEVENT(Wait Reentry - futile wakeup); 740 ++nWakeups; 741 742 // Assuming this is not a spurious wakeup we'll normally 743 // find that _succ == Self. 744 if (_succ == Self) _succ = NULL; 745 746 // Invariant: after clearing _succ a contending thread 747 // *must* retry _owner before parking. 748 OrderAccess::fence(); 749 750 // This PerfData object can be used in parallel with a safepoint. 751 // See the work around in PerfDataManager::destroy(). 752 OM_PERFDATA_OP(FutileWakeups, inc()); 753 } 754 755 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 756 // Normally we'll find Self on the EntryList. 757 // Unlinking from the EntryList is constant-time and atomic-free. 758 // From the perspective of the lock owner (this thread), the 759 // EntryList is stable and cxq is prepend-only. 760 // The head of cxq is volatile but the interior is stable. 761 // In addition, Self.TState is stable. 762 763 assert(_owner == Self, "invariant"); 764 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 765 UnlinkAfterAcquire(Self, SelfNode); 766 if (_succ == Self) _succ = NULL; 767 assert(_succ != Self, "invariant"); 768 SelfNode->TState = ObjectWaiter::TS_RUN; 769 OrderAccess::fence(); // see comments at the end of EnterI() 770 } 771 772 // By convention we unlink a contending thread from EntryList|cxq immediately 773 // after the thread acquires the lock in ::enter(). Equally, we could defer 774 // unlinking the thread until ::exit()-time. 775 776 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 777 assert(_owner == Self, "invariant"); 778 assert(SelfNode->_thread == Self, "invariant"); 779 780 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 781 // Normal case: remove Self from the DLL EntryList . 782 // This is a constant-time operation. 783 ObjectWaiter * nxt = SelfNode->_next; 784 ObjectWaiter * prv = SelfNode->_prev; 785 if (nxt != NULL) nxt->_prev = prv; 786 if (prv != NULL) prv->_next = nxt; 787 if (SelfNode == _EntryList) _EntryList = nxt; 788 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 789 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 790 TEVENT(Unlink from EntryList); 791 } else { 792 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 793 // Inopportune interleaving -- Self is still on the cxq. 794 // This usually means the enqueue of self raced an exiting thread. 795 // Normally we'll find Self near the front of the cxq, so 796 // dequeueing is typically fast. If needbe we can accelerate 797 // this with some MCS/CHL-like bidirectional list hints and advisory 798 // back-links so dequeueing from the interior will normally operate 799 // in constant-time. 800 // Dequeue Self from either the head (with CAS) or from the interior 801 // with a linear-time scan and normal non-atomic memory operations. 802 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 803 // and then unlink Self from EntryList. We have to drain eventually, 804 // so it might as well be now. 805 806 ObjectWaiter * v = _cxq; 807 assert(v != NULL, "invariant"); 808 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 809 // The CAS above can fail from interference IFF a "RAT" arrived. 810 // In that case Self must be in the interior and can no longer be 811 // at the head of cxq. 812 if (v == SelfNode) { 813 assert(_cxq != v, "invariant"); 814 v = _cxq; // CAS above failed - start scan at head of list 815 } 816 ObjectWaiter * p; 817 ObjectWaiter * q = NULL; 818 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 819 q = p; 820 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 821 } 822 assert(v != SelfNode, "invariant"); 823 assert(p == SelfNode, "Node not found on cxq"); 824 assert(p != _cxq, "invariant"); 825 assert(q != NULL, "invariant"); 826 assert(q->_next == p, "invariant"); 827 q->_next = p->_next; 828 } 829 TEVENT(Unlink from cxq); 830 } 831 832 #ifdef ASSERT 833 // Diagnostic hygiene ... 834 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 835 SelfNode->_next = (ObjectWaiter *) 0xBAD; 836 SelfNode->TState = ObjectWaiter::TS_RUN; 837 #endif 838 } 839 840 // ----------------------------------------------------------------------------- 841 // Exit support 842 // 843 // exit() 844 // ~~~~~~ 845 // Note that the collector can't reclaim the objectMonitor or deflate 846 // the object out from underneath the thread calling ::exit() as the 847 // thread calling ::exit() never transitions to a stable state. 848 // This inhibits GC, which in turn inhibits asynchronous (and 849 // inopportune) reclamation of "this". 850 // 851 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 852 // There's one exception to the claim above, however. EnterI() can call 853 // exit() to drop a lock if the acquirer has been externally suspended. 854 // In that case exit() is called with _thread_state as _thread_blocked, 855 // but the monitor's _count field is > 0, which inhibits reclamation. 856 // 857 // 1-0 exit 858 // ~~~~~~~~ 859 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 860 // the fast-path operators have been optimized so the common ::exit() 861 // operation is 1-0. See i486.ad fast_unlock(), for instance. 862 // The code emitted by fast_unlock() elides the usual MEMBAR. This 863 // greatly improves latency -- MEMBAR and CAS having considerable local 864 // latency on modern processors -- but at the cost of "stranding". Absent the 865 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 866 // ::enter() path, resulting in the entering thread being stranding 867 // and a progress-liveness failure. Stranding is extremely rare. 868 // We use timers (timed park operations) & periodic polling to detect 869 // and recover from stranding. Potentially stranded threads periodically 870 // wake up and poll the lock. See the usage of the _Responsible variable. 871 // 872 // The CAS() in enter provides for safety and exclusion, while the CAS or 873 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 874 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 875 // We detect and recover from stranding with timers. 876 // 877 // If a thread transiently strands it'll park until (a) another 878 // thread acquires the lock and then drops the lock, at which time the 879 // exiting thread will notice and unpark the stranded thread, or, (b) 880 // the timer expires. If the lock is high traffic then the stranding latency 881 // will be low due to (a). If the lock is low traffic then the odds of 882 // stranding are lower, although the worst-case stranding latency 883 // is longer. Critically, we don't want to put excessive load in the 884 // platform's timer subsystem. We want to minimize both the timer injection 885 // rate (timers created/sec) as well as the number of timers active at 886 // any one time. (more precisely, we want to minimize timer-seconds, which is 887 // the integral of the # of active timers at any instant over time). 888 // Both impinge on OS scalability. Given that, at most one thread parked on 889 // a monitor will use a timer. 890 // 891 // There is also the risk of a futile wake-up. If we drop the lock 892 // another thread can reacquire the lock immediately, and we can 893 // then wake a thread unnecessarily. This is benign, and we've 894 // structured the code so the windows are short and the frequency 895 // of such futile wakups is low. 896 897 void NOINLINE ObjectMonitor::exit(bool not_suspended, TRAPS) { 898 Thread * const Self = THREAD; 899 if (THREAD != _owner) { 900 if (THREAD->is_lock_owned((address) _owner)) { 901 // Transmute _owner from a BasicLock pointer to a Thread address. 902 // We don't need to hold _mutex for this transition. 903 // Non-null to Non-null is safe as long as all readers can 904 // tolerate either flavor. 905 assert(_recursions == 0, "invariant"); 906 _owner = THREAD; 907 _recursions = 0; 908 } else { 909 // Apparent unbalanced locking ... 910 // Naively we'd like to throw IllegalMonitorStateException. 911 // As a practical matter we can neither allocate nor throw an 912 // exception as ::exit() can be called from leaf routines. 913 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 914 // Upon deeper reflection, however, in a properly run JVM the only 915 // way we should encounter this situation is in the presence of 916 // unbalanced JNI locking. TODO: CheckJNICalls. 917 // See also: CR4414101 918 TEVENT(Exit - Throw IMSX); 919 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 920 return; 921 } 922 } 923 924 if (_recursions != 0) { 925 _recursions--; // this is simple recursive enter 926 TEVENT(Inflated exit - recursive); 927 return; 928 } 929 930 // Invariant: after setting Responsible=null an thread must execute 931 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 932 if ((SyncFlags & 4) == 0) { 933 _Responsible = NULL; 934 } 935 936 #if INCLUDE_TRACE 937 // get the owner's thread id for the MonitorEnter event 938 // if it is enabled and the thread isn't suspended 939 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 940 _previous_owner_tid = THREAD_TRACE_ID(Self); 941 } 942 #endif 943 944 for (;;) { 945 assert(THREAD == _owner, "invariant"); 946 947 948 if (Knob_ExitPolicy == 0) { 949 // release semantics: prior loads and stores from within the critical section 950 // must not float (reorder) past the following store that drops the lock. 951 // On SPARC that requires MEMBAR #loadstore|#storestore. 952 // But of course in TSO #loadstore|#storestore is not required. 953 // I'd like to write one of the following: 954 // A. OrderAccess::release() ; _owner = NULL 955 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 956 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 957 // store into a _dummy variable. That store is not needed, but can result 958 // in massive wasteful coherency traffic on classic SMP systems. 959 // Instead, I use release_store(), which is implemented as just a simple 960 // ST on x64, x86 and SPARC. 961 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 962 OrderAccess::storeload(); // See if we need to wake a successor 963 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 964 TEVENT(Inflated exit - simple egress); 965 return; 966 } 967 TEVENT(Inflated exit - complex egress); 968 // Other threads are blocked trying to acquire the lock. 969 970 // Normally the exiting thread is responsible for ensuring succession, 971 // but if other successors are ready or other entering threads are spinning 972 // then this thread can simply store NULL into _owner and exit without 973 // waking a successor. The existence of spinners or ready successors 974 // guarantees proper succession (liveness). Responsibility passes to the 975 // ready or running successors. The exiting thread delegates the duty. 976 // More precisely, if a successor already exists this thread is absolved 977 // of the responsibility of waking (unparking) one. 978 // 979 // The _succ variable is critical to reducing futile wakeup frequency. 980 // _succ identifies the "heir presumptive" thread that has been made 981 // ready (unparked) but that has not yet run. We need only one such 982 // successor thread to guarantee progress. 983 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 984 // section 3.3 "Futile Wakeup Throttling" for details. 985 // 986 // Note that spinners in Enter() also set _succ non-null. 987 // In the current implementation spinners opportunistically set 988 // _succ so that exiting threads might avoid waking a successor. 989 // Another less appealing alternative would be for the exiting thread 990 // to drop the lock and then spin briefly to see if a spinner managed 991 // to acquire the lock. If so, the exiting thread could exit 992 // immediately without waking a successor, otherwise the exiting 993 // thread would need to dequeue and wake a successor. 994 // (Note that we'd need to make the post-drop spin short, but no 995 // shorter than the worst-case round-trip cache-line migration time. 996 // The dropped lock needs to become visible to the spinner, and then 997 // the acquisition of the lock by the spinner must become visible to 998 // the exiting thread). 999 1000 // It appears that an heir-presumptive (successor) must be made ready. 1001 // Only the current lock owner can manipulate the EntryList or 1002 // drain _cxq, so we need to reacquire the lock. If we fail 1003 // to reacquire the lock the responsibility for ensuring succession 1004 // falls to the new owner. 1005 // 1006 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1007 return; 1008 } 1009 TEVENT(Exit - Reacquired); 1010 } else { 1011 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1012 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1013 OrderAccess::storeload(); 1014 // Ratify the previously observed values. 1015 if (_cxq == NULL || _succ != NULL) { 1016 TEVENT(Inflated exit - simple egress); 1017 return; 1018 } 1019 1020 // inopportune interleaving -- the exiting thread (this thread) 1021 // in the fast-exit path raced an entering thread in the slow-enter 1022 // path. 1023 // We have two choices: 1024 // A. Try to reacquire the lock. 1025 // If the CAS() fails return immediately, otherwise 1026 // we either restart/rerun the exit operation, or simply 1027 // fall-through into the code below which wakes a successor. 1028 // B. If the elements forming the EntryList|cxq are TSM 1029 // we could simply unpark() the lead thread and return 1030 // without having set _succ. 1031 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1032 TEVENT(Inflated exit - reacquired succeeded); 1033 return; 1034 } 1035 TEVENT(Inflated exit - reacquired failed); 1036 } else { 1037 TEVENT(Inflated exit - complex egress); 1038 } 1039 } 1040 1041 guarantee(_owner == THREAD, "invariant"); 1042 1043 ObjectWaiter * w = NULL; 1044 int QMode = Knob_QMode; 1045 1046 if (QMode == 2 && _cxq != NULL) { 1047 // QMode == 2 : cxq has precedence over EntryList. 1048 // Try to directly wake a successor from the cxq. 1049 // If successful, the successor will need to unlink itself from cxq. 1050 w = _cxq; 1051 assert(w != NULL, "invariant"); 1052 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1053 ExitEpilog(Self, w); 1054 return; 1055 } 1056 1057 if (QMode == 3 && _cxq != NULL) { 1058 // Aggressively drain cxq into EntryList at the first opportunity. 1059 // This policy ensure that recently-run threads live at the head of EntryList. 1060 // Drain _cxq into EntryList - bulk transfer. 1061 // First, detach _cxq. 1062 // The following loop is tantamount to: w = swap(&cxq, NULL) 1063 w = _cxq; 1064 for (;;) { 1065 assert(w != NULL, "Invariant"); 1066 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1067 if (u == w) break; 1068 w = u; 1069 } 1070 assert(w != NULL, "invariant"); 1071 1072 ObjectWaiter * q = NULL; 1073 ObjectWaiter * p; 1074 for (p = w; p != NULL; p = p->_next) { 1075 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1076 p->TState = ObjectWaiter::TS_ENTER; 1077 p->_prev = q; 1078 q = p; 1079 } 1080 1081 // Append the RATs to the EntryList 1082 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1083 ObjectWaiter * Tail; 1084 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1085 Tail = Tail->_next) 1086 /* empty */; 1087 if (Tail == NULL) { 1088 _EntryList = w; 1089 } else { 1090 Tail->_next = w; 1091 w->_prev = Tail; 1092 } 1093 1094 // Fall thru into code that tries to wake a successor from EntryList 1095 } 1096 1097 if (QMode == 4 && _cxq != NULL) { 1098 // Aggressively drain cxq into EntryList at the first opportunity. 1099 // This policy ensure that recently-run threads live at the head of EntryList. 1100 1101 // Drain _cxq into EntryList - bulk transfer. 1102 // First, detach _cxq. 1103 // The following loop is tantamount to: w = swap(&cxq, NULL) 1104 w = _cxq; 1105 for (;;) { 1106 assert(w != NULL, "Invariant"); 1107 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1108 if (u == w) break; 1109 w = u; 1110 } 1111 assert(w != NULL, "invariant"); 1112 1113 ObjectWaiter * q = NULL; 1114 ObjectWaiter * p; 1115 for (p = w; p != NULL; p = p->_next) { 1116 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1117 p->TState = ObjectWaiter::TS_ENTER; 1118 p->_prev = q; 1119 q = p; 1120 } 1121 1122 // Prepend the RATs to the EntryList 1123 if (_EntryList != NULL) { 1124 q->_next = _EntryList; 1125 _EntryList->_prev = q; 1126 } 1127 _EntryList = w; 1128 1129 // Fall thru into code that tries to wake a successor from EntryList 1130 } 1131 1132 w = _EntryList; 1133 if (w != NULL) { 1134 // I'd like to write: guarantee (w->_thread != Self). 1135 // But in practice an exiting thread may find itself on the EntryList. 1136 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1137 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1138 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1139 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1140 // release the lock "O". T2 resumes immediately after the ST of null into 1141 // _owner, above. T2 notices that the EntryList is populated, so it 1142 // reacquires the lock and then finds itself on the EntryList. 1143 // Given all that, we have to tolerate the circumstance where "w" is 1144 // associated with Self. 1145 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1146 ExitEpilog(Self, w); 1147 return; 1148 } 1149 1150 // If we find that both _cxq and EntryList are null then just 1151 // re-run the exit protocol from the top. 1152 w = _cxq; 1153 if (w == NULL) continue; 1154 1155 // Drain _cxq into EntryList - bulk transfer. 1156 // First, detach _cxq. 1157 // The following loop is tantamount to: w = swap(&cxq, NULL) 1158 for (;;) { 1159 assert(w != NULL, "Invariant"); 1160 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1161 if (u == w) break; 1162 w = u; 1163 } 1164 TEVENT(Inflated exit - drain cxq into EntryList); 1165 1166 assert(w != NULL, "invariant"); 1167 assert(_EntryList == NULL, "invariant"); 1168 1169 // Convert the LIFO SLL anchored by _cxq into a DLL. 1170 // The list reorganization step operates in O(LENGTH(w)) time. 1171 // It's critical that this step operate quickly as 1172 // "Self" still holds the outer-lock, restricting parallelism 1173 // and effectively lengthening the critical section. 1174 // Invariant: s chases t chases u. 1175 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1176 // we have faster access to the tail. 1177 1178 if (QMode == 1) { 1179 // QMode == 1 : drain cxq to EntryList, reversing order 1180 // We also reverse the order of the list. 1181 ObjectWaiter * s = NULL; 1182 ObjectWaiter * t = w; 1183 ObjectWaiter * u = NULL; 1184 while (t != NULL) { 1185 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1186 t->TState = ObjectWaiter::TS_ENTER; 1187 u = t->_next; 1188 t->_prev = u; 1189 t->_next = s; 1190 s = t; 1191 t = u; 1192 } 1193 _EntryList = s; 1194 assert(s != NULL, "invariant"); 1195 } else { 1196 // QMode == 0 or QMode == 2 1197 _EntryList = w; 1198 ObjectWaiter * q = NULL; 1199 ObjectWaiter * p; 1200 for (p = w; p != NULL; p = p->_next) { 1201 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1202 p->TState = ObjectWaiter::TS_ENTER; 1203 p->_prev = q; 1204 q = p; 1205 } 1206 } 1207 1208 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1209 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1210 1211 // See if we can abdicate to a spinner instead of waking a thread. 1212 // A primary goal of the implementation is to reduce the 1213 // context-switch rate. 1214 if (_succ != NULL) continue; 1215 1216 w = _EntryList; 1217 if (w != NULL) { 1218 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1219 ExitEpilog(Self, w); 1220 return; 1221 } 1222 } 1223 } 1224 1225 // ExitSuspendEquivalent: 1226 // A faster alternate to handle_special_suspend_equivalent_condition() 1227 // 1228 // handle_special_suspend_equivalent_condition() unconditionally 1229 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1230 // operations have high latency. Note that in ::enter() we call HSSEC 1231 // while holding the monitor, so we effectively lengthen the critical sections. 1232 // 1233 // There are a number of possible solutions: 1234 // 1235 // A. To ameliorate the problem we might also defer state transitions 1236 // to as late as possible -- just prior to parking. 1237 // Given that, we'd call HSSEC after having returned from park(), 1238 // but before attempting to acquire the monitor. This is only a 1239 // partial solution. It avoids calling HSSEC while holding the 1240 // monitor (good), but it still increases successor reacquisition latency -- 1241 // the interval between unparking a successor and the time the successor 1242 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1243 // If we use this technique we can also avoid EnterI()-exit() loop 1244 // in ::enter() where we iteratively drop the lock and then attempt 1245 // to reacquire it after suspending. 1246 // 1247 // B. In the future we might fold all the suspend bits into a 1248 // composite per-thread suspend flag and then update it with CAS(). 1249 // Alternately, a Dekker-like mechanism with multiple variables 1250 // would suffice: 1251 // ST Self->_suspend_equivalent = false 1252 // MEMBAR 1253 // LD Self_>_suspend_flags 1254 // 1255 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1256 // with a more efficient implementation, the need to use "FastHSSEC" has 1257 // decreased. - Dave 1258 1259 1260 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1261 const int Mode = Knob_FastHSSEC; 1262 if (Mode && !jSelf->is_external_suspend()) { 1263 assert(jSelf->is_suspend_equivalent(), "invariant"); 1264 jSelf->clear_suspend_equivalent(); 1265 if (2 == Mode) OrderAccess::storeload(); 1266 if (!jSelf->is_external_suspend()) return false; 1267 // We raced a suspension -- fall thru into the slow path 1268 TEVENT(ExitSuspendEquivalent - raced); 1269 jSelf->set_suspend_equivalent(); 1270 } 1271 return jSelf->handle_special_suspend_equivalent_condition(); 1272 } 1273 1274 1275 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1276 assert(_owner == Self, "invariant"); 1277 1278 // Exit protocol: 1279 // 1. ST _succ = wakee 1280 // 2. membar #loadstore|#storestore; 1281 // 2. ST _owner = NULL 1282 // 3. unpark(wakee) 1283 1284 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1285 ParkEvent * Trigger = Wakee->_event; 1286 1287 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1288 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1289 // out-of-scope (non-extant). 1290 Wakee = NULL; 1291 1292 // Drop the lock 1293 OrderAccess::release_store_ptr(&_owner, NULL); 1294 OrderAccess::fence(); // ST _owner vs LD in unpark() 1295 1296 if (SafepointSynchronize::do_call_back()) { 1297 TEVENT(unpark before SAFEPOINT); 1298 } 1299 1300 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1301 Trigger->unpark(); 1302 1303 // Maintain stats and report events to JVMTI 1304 OM_PERFDATA_OP(Parks, inc()); 1305 } 1306 1307 1308 // ----------------------------------------------------------------------------- 1309 // Class Loader deadlock handling. 1310 // 1311 // complete_exit exits a lock returning recursion count 1312 // complete_exit/reenter operate as a wait without waiting 1313 // complete_exit requires an inflated monitor 1314 // The _owner field is not always the Thread addr even with an 1315 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1316 // thread due to contention. 1317 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1318 Thread * const Self = THREAD; 1319 assert(Self->is_Java_thread(), "Must be Java thread!"); 1320 JavaThread *jt = (JavaThread *)THREAD; 1321 1322 DeferredInitialize(); 1323 1324 if (THREAD != _owner) { 1325 if (THREAD->is_lock_owned ((address)_owner)) { 1326 assert(_recursions == 0, "internal state error"); 1327 _owner = THREAD; // Convert from basiclock addr to Thread addr 1328 _recursions = 0; 1329 } 1330 } 1331 1332 guarantee(Self == _owner, "complete_exit not owner"); 1333 intptr_t save = _recursions; // record the old recursion count 1334 _recursions = 0; // set the recursion level to be 0 1335 exit(true, Self); // exit the monitor 1336 guarantee(_owner != Self, "invariant"); 1337 return save; 1338 } 1339 1340 // reenter() enters a lock and sets recursion count 1341 // complete_exit/reenter operate as a wait without waiting 1342 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1343 Thread * const Self = THREAD; 1344 assert(Self->is_Java_thread(), "Must be Java thread!"); 1345 JavaThread *jt = (JavaThread *)THREAD; 1346 1347 guarantee(_owner != Self, "reenter already owner"); 1348 enter(THREAD); // enter the monitor 1349 guarantee(_recursions == 0, "reenter recursion"); 1350 _recursions = recursions; 1351 return; 1352 } 1353 1354 1355 // ----------------------------------------------------------------------------- 1356 // A macro is used below because there may already be a pending 1357 // exception which should not abort the execution of the routines 1358 // which use this (which is why we don't put this into check_slow and 1359 // call it with a CHECK argument). 1360 1361 #define CHECK_OWNER() \ 1362 do { \ 1363 if (THREAD != _owner) { \ 1364 if (THREAD->is_lock_owned((address) _owner)) { \ 1365 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1366 _recursions = 0; \ 1367 } else { \ 1368 TEVENT(Throw IMSX); \ 1369 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1370 } \ 1371 } \ 1372 } while (false) 1373 1374 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1375 // TODO-FIXME: remove check_slow() -- it's likely dead. 1376 1377 void ObjectMonitor::check_slow(TRAPS) { 1378 TEVENT(check_slow - throw IMSX); 1379 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1380 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1381 } 1382 1383 static int Adjust(volatile int * adr, int dx) { 1384 int v; 1385 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1386 return v; 1387 } 1388 1389 // helper method for posting a monitor wait event 1390 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1391 jlong notifier_tid, 1392 jlong timeout, 1393 bool timedout) { 1394 assert(event != NULL, "invariant"); 1395 event->set_klass(((oop)this->object())->klass()); 1396 event->set_timeout(timeout); 1397 event->set_address((TYPE_ADDRESS)this->object_addr()); 1398 event->set_notifier(notifier_tid); 1399 event->set_timedOut(timedout); 1400 event->commit(); 1401 } 1402 1403 // ----------------------------------------------------------------------------- 1404 // Wait/Notify/NotifyAll 1405 // 1406 // Note: a subset of changes to ObjectMonitor::wait() 1407 // will need to be replicated in complete_exit 1408 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1409 Thread * const Self = THREAD; 1410 assert(Self->is_Java_thread(), "Must be Java thread!"); 1411 JavaThread *jt = (JavaThread *)THREAD; 1412 1413 DeferredInitialize(); 1414 1415 // Throw IMSX or IEX. 1416 CHECK_OWNER(); 1417 1418 EventJavaMonitorWait event; 1419 1420 // check for a pending interrupt 1421 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1422 // post monitor waited event. Note that this is past-tense, we are done waiting. 1423 if (JvmtiExport::should_post_monitor_waited()) { 1424 // Note: 'false' parameter is passed here because the 1425 // wait was not timed out due to thread interrupt. 1426 JvmtiExport::post_monitor_waited(jt, this, false); 1427 1428 // In this short circuit of the monitor wait protocol, the 1429 // current thread never drops ownership of the monitor and 1430 // never gets added to the wait queue so the current thread 1431 // cannot be made the successor. This means that the 1432 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1433 // consume an unpark() meant for the ParkEvent associated with 1434 // this ObjectMonitor. 1435 } 1436 if (event.should_commit()) { 1437 post_monitor_wait_event(&event, 0, millis, false); 1438 } 1439 TEVENT(Wait - Throw IEX); 1440 THROW(vmSymbols::java_lang_InterruptedException()); 1441 return; 1442 } 1443 1444 TEVENT(Wait); 1445 1446 assert(Self->_Stalled == 0, "invariant"); 1447 Self->_Stalled = intptr_t(this); 1448 jt->set_current_waiting_monitor(this); 1449 1450 // create a node to be put into the queue 1451 // Critically, after we reset() the event but prior to park(), we must check 1452 // for a pending interrupt. 1453 ObjectWaiter node(Self); 1454 node.TState = ObjectWaiter::TS_WAIT; 1455 Self->_ParkEvent->reset(); 1456 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1457 1458 // Enter the waiting queue, which is a circular doubly linked list in this case 1459 // but it could be a priority queue or any data structure. 1460 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1461 // by the the owner of the monitor *except* in the case where park() 1462 // returns because of a timeout of interrupt. Contention is exceptionally rare 1463 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1464 1465 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1466 AddWaiter(&node); 1467 Thread::SpinRelease(&_WaitSetLock); 1468 1469 if ((SyncFlags & 4) == 0) { 1470 _Responsible = NULL; 1471 } 1472 intptr_t save = _recursions; // record the old recursion count 1473 _waiters++; // increment the number of waiters 1474 _recursions = 0; // set the recursion level to be 1 1475 exit(true, Self); // exit the monitor 1476 guarantee(_owner != Self, "invariant"); 1477 1478 // The thread is on the WaitSet list - now park() it. 1479 // On MP systems it's conceivable that a brief spin before we park 1480 // could be profitable. 1481 // 1482 // TODO-FIXME: change the following logic to a loop of the form 1483 // while (!timeout && !interrupted && _notified == 0) park() 1484 1485 int ret = OS_OK; 1486 int WasNotified = 0; 1487 { // State transition wrappers 1488 OSThread* osthread = Self->osthread(); 1489 OSThreadWaitState osts(osthread, true); 1490 { 1491 ThreadBlockInVM tbivm(jt); 1492 // Thread is in thread_blocked state and oop access is unsafe. 1493 jt->set_suspend_equivalent(); 1494 1495 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1496 // Intentionally empty 1497 } else if (node._notified == 0) { 1498 if (millis <= 0) { 1499 Self->_ParkEvent->park(); 1500 } else { 1501 ret = Self->_ParkEvent->park(millis); 1502 } 1503 } 1504 1505 // were we externally suspended while we were waiting? 1506 if (ExitSuspendEquivalent (jt)) { 1507 // TODO-FIXME: add -- if succ == Self then succ = null. 1508 jt->java_suspend_self(); 1509 } 1510 1511 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1512 1513 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1514 // from the WaitSet to the EntryList. 1515 // See if we need to remove Node from the WaitSet. 1516 // We use double-checked locking to avoid grabbing _WaitSetLock 1517 // if the thread is not on the wait queue. 1518 // 1519 // Note that we don't need a fence before the fetch of TState. 1520 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1521 // written by the is thread. (perhaps the fetch might even be satisfied 1522 // by a look-aside into the processor's own store buffer, although given 1523 // the length of the code path between the prior ST and this load that's 1524 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1525 // then we'll acquire the lock and then re-fetch a fresh TState value. 1526 // That is, we fail toward safety. 1527 1528 if (node.TState == ObjectWaiter::TS_WAIT) { 1529 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1530 if (node.TState == ObjectWaiter::TS_WAIT) { 1531 DequeueSpecificWaiter(&node); // unlink from WaitSet 1532 assert(node._notified == 0, "invariant"); 1533 node.TState = ObjectWaiter::TS_RUN; 1534 } 1535 Thread::SpinRelease(&_WaitSetLock); 1536 } 1537 1538 // The thread is now either on off-list (TS_RUN), 1539 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1540 // The Node's TState variable is stable from the perspective of this thread. 1541 // No other threads will asynchronously modify TState. 1542 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1543 OrderAccess::loadload(); 1544 if (_succ == Self) _succ = NULL; 1545 WasNotified = node._notified; 1546 1547 // Reentry phase -- reacquire the monitor. 1548 // re-enter contended monitor after object.wait(). 1549 // retain OBJECT_WAIT state until re-enter successfully completes 1550 // Thread state is thread_in_vm and oop access is again safe, 1551 // although the raw address of the object may have changed. 1552 // (Don't cache naked oops over safepoints, of course). 1553 1554 // post monitor waited event. Note that this is past-tense, we are done waiting. 1555 if (JvmtiExport::should_post_monitor_waited()) { 1556 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1557 1558 if (node._notified != 0 && _succ == Self) { 1559 // In this part of the monitor wait-notify-reenter protocol it 1560 // is possible (and normal) for another thread to do a fastpath 1561 // monitor enter-exit while this thread is still trying to get 1562 // to the reenter portion of the protocol. 1563 // 1564 // The ObjectMonitor was notified and the current thread is 1565 // the successor which also means that an unpark() has already 1566 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1567 // consume the unpark() that was done when the successor was 1568 // set because the same ParkEvent is shared between Java 1569 // monitors and JVM/TI RawMonitors (for now). 1570 // 1571 // We redo the unpark() to ensure forward progress, i.e., we 1572 // don't want all pending threads hanging (parked) with none 1573 // entering the unlocked monitor. 1574 node._event->unpark(); 1575 } 1576 } 1577 1578 if (event.should_commit()) { 1579 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1580 } 1581 1582 OrderAccess::fence(); 1583 1584 assert(Self->_Stalled != 0, "invariant"); 1585 Self->_Stalled = 0; 1586 1587 assert(_owner != Self, "invariant"); 1588 ObjectWaiter::TStates v = node.TState; 1589 if (v == ObjectWaiter::TS_RUN) { 1590 enter(Self); 1591 } else { 1592 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1593 ReenterI(Self, &node); 1594 node.wait_reenter_end(this); 1595 } 1596 1597 // Self has reacquired the lock. 1598 // Lifecycle - the node representing Self must not appear on any queues. 1599 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1600 // want residual elements associated with this thread left on any lists. 1601 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1602 assert(_owner == Self, "invariant"); 1603 assert(_succ != Self, "invariant"); 1604 } // OSThreadWaitState() 1605 1606 jt->set_current_waiting_monitor(NULL); 1607 1608 guarantee(_recursions == 0, "invariant"); 1609 _recursions = save; // restore the old recursion count 1610 _waiters--; // decrement the number of waiters 1611 1612 // Verify a few postconditions 1613 assert(_owner == Self, "invariant"); 1614 assert(_succ != Self, "invariant"); 1615 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1616 1617 if (SyncFlags & 32) { 1618 OrderAccess::fence(); 1619 } 1620 1621 // check if the notification happened 1622 if (!WasNotified) { 1623 // no, it could be timeout or Thread.interrupt() or both 1624 // check for interrupt event, otherwise it is timeout 1625 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1626 TEVENT(Wait - throw IEX from epilog); 1627 THROW(vmSymbols::java_lang_InterruptedException()); 1628 } 1629 } 1630 1631 // NOTE: Spurious wake up will be consider as timeout. 1632 // Monitor notify has precedence over thread interrupt. 1633 } 1634 1635 1636 // Consider: 1637 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1638 // then instead of transferring a thread from the WaitSet to the EntryList 1639 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1640 1641 void ObjectMonitor::INotify(Thread * Self) { 1642 const int policy = Knob_MoveNotifyee; 1643 1644 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1645 ObjectWaiter * iterator = DequeueWaiter(); 1646 if (iterator != NULL) { 1647 TEVENT(Notify1 - Transfer); 1648 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1649 guarantee(iterator->_notified == 0, "invariant"); 1650 // Disposition - what might we do with iterator ? 1651 // a. add it directly to the EntryList - either tail (policy == 1) 1652 // or head (policy == 0). 1653 // b. push it onto the front of the _cxq (policy == 2). 1654 // For now we use (b). 1655 if (policy != 4) { 1656 iterator->TState = ObjectWaiter::TS_ENTER; 1657 } 1658 iterator->_notified = 1; 1659 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1660 1661 ObjectWaiter * list = _EntryList; 1662 if (list != NULL) { 1663 assert(list->_prev == NULL, "invariant"); 1664 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1665 assert(list != iterator, "invariant"); 1666 } 1667 1668 if (policy == 0) { // prepend to EntryList 1669 if (list == NULL) { 1670 iterator->_next = iterator->_prev = NULL; 1671 _EntryList = iterator; 1672 } else { 1673 list->_prev = iterator; 1674 iterator->_next = list; 1675 iterator->_prev = NULL; 1676 _EntryList = iterator; 1677 } 1678 } else if (policy == 1) { // append to EntryList 1679 if (list == NULL) { 1680 iterator->_next = iterator->_prev = NULL; 1681 _EntryList = iterator; 1682 } else { 1683 // CONSIDER: finding the tail currently requires a linear-time walk of 1684 // the EntryList. We can make tail access constant-time by converting to 1685 // a CDLL instead of using our current DLL. 1686 ObjectWaiter * tail; 1687 for (tail = list; tail->_next != NULL; tail = tail->_next) /* empty */; 1688 assert(tail != NULL && tail->_next == NULL, "invariant"); 1689 tail->_next = iterator; 1690 iterator->_prev = tail; 1691 iterator->_next = NULL; 1692 } 1693 } else if (policy == 2) { // prepend to cxq 1694 if (list == NULL) { 1695 iterator->_next = iterator->_prev = NULL; 1696 _EntryList = iterator; 1697 } else { 1698 iterator->TState = ObjectWaiter::TS_CXQ; 1699 for (;;) { 1700 ObjectWaiter * front = _cxq; 1701 iterator->_next = front; 1702 if (Atomic::cmpxchg_ptr(iterator, &_cxq, front) == front) { 1703 break; 1704 } 1705 } 1706 } 1707 } else if (policy == 3) { // append to cxq 1708 iterator->TState = ObjectWaiter::TS_CXQ; 1709 for (;;) { 1710 ObjectWaiter * tail = _cxq; 1711 if (tail == NULL) { 1712 iterator->_next = NULL; 1713 if (Atomic::cmpxchg_ptr(iterator, &_cxq, NULL) == NULL) { 1714 break; 1715 } 1716 } else { 1717 while (tail->_next != NULL) tail = tail->_next; 1718 tail->_next = iterator; 1719 iterator->_prev = tail; 1720 iterator->_next = NULL; 1721 break; 1722 } 1723 } 1724 } else { 1725 ParkEvent * ev = iterator->_event; 1726 iterator->TState = ObjectWaiter::TS_RUN; 1727 OrderAccess::fence(); 1728 ev->unpark(); 1729 } 1730 1731 // _WaitSetLock protects the wait queue, not the EntryList. We could 1732 // move the add-to-EntryList operation, above, outside the critical section 1733 // protected by _WaitSetLock. In practice that's not useful. With the 1734 // exception of wait() timeouts and interrupts the monitor owner 1735 // is the only thread that grabs _WaitSetLock. There's almost no contention 1736 // on _WaitSetLock so it's not profitable to reduce the length of the 1737 // critical section. 1738 1739 if (policy < 4) { 1740 iterator->wait_reenter_begin(this); 1741 } 1742 } 1743 Thread::SpinRelease(&_WaitSetLock); 1744 } 1745 1746 // Consider: a not-uncommon synchronization bug is to use notify() when 1747 // notifyAll() is more appropriate, potentially resulting in stranded 1748 // threads; this is one example of a lost wakeup. A useful diagnostic 1749 // option is to force all notify() operations to behave as notifyAll(). 1750 // 1751 // Note: We can also detect many such problems with a "minimum wait". 1752 // When the "minimum wait" is set to a small non-zero timeout value 1753 // and the program does not hang whereas it did absent "minimum wait", 1754 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses 1755 // a "minimum wait" for all park() operations; see the recheckInterval 1756 // variable and MAX_RECHECK_INTERVAL. 1757 1758 void ObjectMonitor::notify(TRAPS) { 1759 CHECK_OWNER(); 1760 if (_WaitSet == NULL) { 1761 TEVENT(Empty-Notify); 1762 return; 1763 } 1764 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1765 INotify(THREAD); 1766 OM_PERFDATA_OP(Notifications, inc(1)); 1767 } 1768 1769 1770 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1771 // from the waitset to the EntryList. This could be done more efficiently with a 1772 // single bulk transfer but in practice it's not time-critical. Beware too, 1773 // that in prepend-mode we invert the order of the waiters. Let's say that the 1774 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1775 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1776 1777 void ObjectMonitor::notifyAll(TRAPS) { 1778 CHECK_OWNER(); 1779 if (_WaitSet == NULL) { 1780 TEVENT(Empty-NotifyAll); 1781 return; 1782 } 1783 1784 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1785 int tally = 0; 1786 while (_WaitSet != NULL) { 1787 tally++; 1788 INotify(THREAD); 1789 } 1790 1791 OM_PERFDATA_OP(Notifications, inc(tally)); 1792 } 1793 1794 // ----------------------------------------------------------------------------- 1795 // Adaptive Spinning Support 1796 // 1797 // Adaptive spin-then-block - rational spinning 1798 // 1799 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1800 // algorithm. On high order SMP systems it would be better to start with 1801 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1802 // a contending thread could enqueue itself on the cxq and then spin locally 1803 // on a thread-specific variable such as its ParkEvent._Event flag. 1804 // That's left as an exercise for the reader. Note that global spinning is 1805 // not problematic on Niagara, as the L2 cache serves the interconnect and 1806 // has both low latency and massive bandwidth. 1807 // 1808 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1809 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1810 // (duration) or we can fix the count at approximately the duration of 1811 // a context switch and vary the frequency. Of course we could also 1812 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1813 // For a description of 'Adaptive spin-then-block mutual exclusion in 1814 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1815 // 1816 // This implementation varies the duration "D", where D varies with 1817 // the success rate of recent spin attempts. (D is capped at approximately 1818 // length of a round-trip context switch). The success rate for recent 1819 // spin attempts is a good predictor of the success rate of future spin 1820 // attempts. The mechanism adapts automatically to varying critical 1821 // section length (lock modality), system load and degree of parallelism. 1822 // D is maintained per-monitor in _SpinDuration and is initialized 1823 // optimistically. Spin frequency is fixed at 100%. 1824 // 1825 // Note that _SpinDuration is volatile, but we update it without locks 1826 // or atomics. The code is designed so that _SpinDuration stays within 1827 // a reasonable range even in the presence of races. The arithmetic 1828 // operations on _SpinDuration are closed over the domain of legal values, 1829 // so at worst a race will install and older but still legal value. 1830 // At the very worst this introduces some apparent non-determinism. 1831 // We might spin when we shouldn't or vice-versa, but since the spin 1832 // count are relatively short, even in the worst case, the effect is harmless. 1833 // 1834 // Care must be taken that a low "D" value does not become an 1835 // an absorbing state. Transient spinning failures -- when spinning 1836 // is overall profitable -- should not cause the system to converge 1837 // on low "D" values. We want spinning to be stable and predictable 1838 // and fairly responsive to change and at the same time we don't want 1839 // it to oscillate, become metastable, be "too" non-deterministic, 1840 // or converge on or enter undesirable stable absorbing states. 1841 // 1842 // We implement a feedback-based control system -- using past behavior 1843 // to predict future behavior. We face two issues: (a) if the 1844 // input signal is random then the spin predictor won't provide optimal 1845 // results, and (b) if the signal frequency is too high then the control 1846 // system, which has some natural response lag, will "chase" the signal. 1847 // (b) can arise from multimodal lock hold times. Transient preemption 1848 // can also result in apparent bimodal lock hold times. 1849 // Although sub-optimal, neither condition is particularly harmful, as 1850 // in the worst-case we'll spin when we shouldn't or vice-versa. 1851 // The maximum spin duration is rather short so the failure modes aren't bad. 1852 // To be conservative, I've tuned the gain in system to bias toward 1853 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1854 // "rings" or oscillates between spinning and not spinning. This happens 1855 // when spinning is just on the cusp of profitability, however, so the 1856 // situation is not dire. The state is benign -- there's no need to add 1857 // hysteresis control to damp the transition rate between spinning and 1858 // not spinning. 1859 1860 intptr_t ObjectMonitor::SpinCallbackArgument = 0; 1861 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL; 1862 1863 // Spinning: Fixed frequency (100%), vary duration 1864 1865 1866 int ObjectMonitor::TrySpin_VaryDuration(Thread * Self) { 1867 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1868 int ctr = Knob_FixedSpin; 1869 if (ctr != 0) { 1870 while (--ctr >= 0) { 1871 if (TryLock(Self) > 0) return 1; 1872 SpinPause(); 1873 } 1874 return 0; 1875 } 1876 1877 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1878 if (TryLock(Self) > 0) { 1879 // Increase _SpinDuration ... 1880 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1881 // Raising _SpurDuration to the poverty line is key. 1882 int x = _SpinDuration; 1883 if (x < Knob_SpinLimit) { 1884 if (x < Knob_Poverty) x = Knob_Poverty; 1885 _SpinDuration = x + Knob_BonusB; 1886 } 1887 return 1; 1888 } 1889 SpinPause(); 1890 } 1891 1892 // Admission control - verify preconditions for spinning 1893 // 1894 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1895 // becoming an absorbing state. Put another way, we spin briefly to 1896 // sample, just in case the system load, parallelism, contention, or lock 1897 // modality changed. 1898 // 1899 // Consider the following alternative: 1900 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1901 // spin attempt. "Periodically" might mean after a tally of 1902 // the # of failed spin attempts (or iterations) reaches some threshold. 1903 // This takes us into the realm of 1-out-of-N spinning, where we 1904 // hold the duration constant but vary the frequency. 1905 1906 ctr = _SpinDuration; 1907 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1908 if (ctr <= 0) return 0; 1909 1910 if (Knob_SuccRestrict && _succ != NULL) return 0; 1911 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1912 TEVENT(Spin abort - notrunnable [TOP]); 1913 return 0; 1914 } 1915 1916 int MaxSpin = Knob_MaxSpinners; 1917 if (MaxSpin >= 0) { 1918 if (_Spinner > MaxSpin) { 1919 TEVENT(Spin abort -- too many spinners); 1920 return 0; 1921 } 1922 // Slightly racy, but benign ... 1923 Adjust(&_Spinner, 1); 1924 } 1925 1926 // We're good to spin ... spin ingress. 1927 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1928 // when preparing to LD...CAS _owner, etc and the CAS is likely 1929 // to succeed. 1930 int hits = 0; 1931 int msk = 0; 1932 int caspty = Knob_CASPenalty; 1933 int oxpty = Knob_OXPenalty; 1934 int sss = Knob_SpinSetSucc; 1935 if (sss && _succ == NULL) _succ = Self; 1936 Thread * prv = NULL; 1937 1938 // There are three ways to exit the following loop: 1939 // 1. A successful spin where this thread has acquired the lock. 1940 // 2. Spin failure with prejudice 1941 // 3. Spin failure without prejudice 1942 1943 while (--ctr >= 0) { 1944 1945 // Periodic polling -- Check for pending GC 1946 // Threads may spin while they're unsafe. 1947 // We don't want spinning threads to delay the JVM from reaching 1948 // a stop-the-world safepoint or to steal cycles from GC. 1949 // If we detect a pending safepoint we abort in order that 1950 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1951 // this thread, if safe, doesn't steal cycles from GC. 1952 // This is in keeping with the "no loitering in runtime" rule. 1953 // We periodically check to see if there's a safepoint pending. 1954 if ((ctr & 0xFF) == 0) { 1955 if (SafepointSynchronize::do_call_back()) { 1956 TEVENT(Spin: safepoint); 1957 goto Abort; // abrupt spin egress 1958 } 1959 if (Knob_UsePause & 1) SpinPause(); 1960 1961 int (*scb)(intptr_t,int) = SpinCallbackFunction; 1962 if (hits > 50 && scb != NULL) { 1963 int abend = (*scb)(SpinCallbackArgument, 0); 1964 } 1965 } 1966 1967 if (Knob_UsePause & 2) SpinPause(); 1968 1969 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1970 // This is useful on classic SMP systems, but is of less utility on 1971 // N1-style CMT platforms. 1972 // 1973 // Trade-off: lock acquisition latency vs coherency bandwidth. 1974 // Lock hold times are typically short. A histogram 1975 // of successful spin attempts shows that we usually acquire 1976 // the lock early in the spin. That suggests we want to 1977 // sample _owner frequently in the early phase of the spin, 1978 // but then back-off and sample less frequently as the spin 1979 // progresses. The back-off makes a good citizen on SMP big 1980 // SMP systems. Oversampling _owner can consume excessive 1981 // coherency bandwidth. Relatedly, if we _oversample _owner we 1982 // can inadvertently interfere with the the ST m->owner=null. 1983 // executed by the lock owner. 1984 if (ctr & msk) continue; 1985 ++hits; 1986 if ((hits & 0xF) == 0) { 1987 // The 0xF, above, corresponds to the exponent. 1988 // Consider: (msk+1)|msk 1989 msk = ((msk << 2)|3) & BackOffMask; 1990 } 1991 1992 // Probe _owner with TATAS 1993 // If this thread observes the monitor transition or flicker 1994 // from locked to unlocked to locked, then the odds that this 1995 // thread will acquire the lock in this spin attempt go down 1996 // considerably. The same argument applies if the CAS fails 1997 // or if we observe _owner change from one non-null value to 1998 // another non-null value. In such cases we might abort 1999 // the spin without prejudice or apply a "penalty" to the 2000 // spin count-down variable "ctr", reducing it by 100, say. 2001 2002 Thread * ox = (Thread *) _owner; 2003 if (ox == NULL) { 2004 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 2005 if (ox == NULL) { 2006 // The CAS succeeded -- this thread acquired ownership 2007 // Take care of some bookkeeping to exit spin state. 2008 if (sss && _succ == Self) { 2009 _succ = NULL; 2010 } 2011 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2012 2013 // Increase _SpinDuration : 2014 // The spin was successful (profitable) so we tend toward 2015 // longer spin attempts in the future. 2016 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2017 // If we acquired the lock early in the spin cycle it 2018 // makes sense to increase _SpinDuration proportionally. 2019 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2020 int x = _SpinDuration; 2021 if (x < Knob_SpinLimit) { 2022 if (x < Knob_Poverty) x = Knob_Poverty; 2023 _SpinDuration = x + Knob_Bonus; 2024 } 2025 return 1; 2026 } 2027 2028 // The CAS failed ... we can take any of the following actions: 2029 // * penalize: ctr -= Knob_CASPenalty 2030 // * exit spin with prejudice -- goto Abort; 2031 // * exit spin without prejudice. 2032 // * Since CAS is high-latency, retry again immediately. 2033 prv = ox; 2034 TEVENT(Spin: cas failed); 2035 if (caspty == -2) break; 2036 if (caspty == -1) goto Abort; 2037 ctr -= caspty; 2038 continue; 2039 } 2040 2041 // Did lock ownership change hands ? 2042 if (ox != prv && prv != NULL) { 2043 TEVENT(spin: Owner changed) 2044 if (oxpty == -2) break; 2045 if (oxpty == -1) goto Abort; 2046 ctr -= oxpty; 2047 } 2048 prv = ox; 2049 2050 // Abort the spin if the owner is not executing. 2051 // The owner must be executing in order to drop the lock. 2052 // Spinning while the owner is OFFPROC is idiocy. 2053 // Consider: ctr -= RunnablePenalty ; 2054 if (Knob_OState && NotRunnable (Self, ox)) { 2055 TEVENT(Spin abort - notrunnable); 2056 goto Abort; 2057 } 2058 if (sss && _succ == NULL) _succ = Self; 2059 } 2060 2061 // Spin failed with prejudice -- reduce _SpinDuration. 2062 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2063 // AIMD is globally stable. 2064 TEVENT(Spin failure); 2065 { 2066 int x = _SpinDuration; 2067 if (x > 0) { 2068 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2069 // This is globally sample and tends to damp the response. 2070 x -= Knob_Penalty; 2071 if (x < 0) x = 0; 2072 _SpinDuration = x; 2073 } 2074 } 2075 2076 Abort: 2077 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2078 if (sss && _succ == Self) { 2079 _succ = NULL; 2080 // Invariant: after setting succ=null a contending thread 2081 // must recheck-retry _owner before parking. This usually happens 2082 // in the normal usage of TrySpin(), but it's safest 2083 // to make TrySpin() as foolproof as possible. 2084 OrderAccess::fence(); 2085 if (TryLock(Self) > 0) return 1; 2086 } 2087 return 0; 2088 } 2089 2090 // NotRunnable() -- informed spinning 2091 // 2092 // Don't bother spinning if the owner is not eligible to drop the lock. 2093 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2094 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2095 // The thread must be runnable in order to drop the lock in timely fashion. 2096 // If the _owner is not runnable then spinning will not likely be 2097 // successful (profitable). 2098 // 2099 // Beware -- the thread referenced by _owner could have died 2100 // so a simply fetch from _owner->_thread_state might trap. 2101 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2102 // Because of the lifecycle issues the schedctl and _thread_state values 2103 // observed by NotRunnable() might be garbage. NotRunnable must 2104 // tolerate this and consider the observed _thread_state value 2105 // as advisory. 2106 // 2107 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2108 // a thread pointer. 2109 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2110 // with the LSB of _owner. Another option would be to probablistically probe 2111 // the putative _owner->TypeTag value. 2112 // 2113 // Checking _thread_state isn't perfect. Even if the thread is 2114 // in_java it might be blocked on a page-fault or have been preempted 2115 // and sitting on a ready/dispatch queue. _thread state in conjunction 2116 // with schedctl.sc_state gives us a good picture of what the 2117 // thread is doing, however. 2118 // 2119 // TODO: check schedctl.sc_state. 2120 // We'll need to use SafeFetch32() to read from the schedctl block. 2121 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2122 // 2123 // The return value from NotRunnable() is *advisory* -- the 2124 // result is based on sampling and is not necessarily coherent. 2125 // The caller must tolerate false-negative and false-positive errors. 2126 // Spinning, in general, is probabilistic anyway. 2127 2128 2129 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2130 // Check ox->TypeTag == 2BAD. 2131 if (ox == NULL) return 0; 2132 2133 // Avoid transitive spinning ... 2134 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2135 // Immediately after T1 acquires L it's possible that T2, also 2136 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2137 // This occurs transiently after T1 acquired L but before 2138 // T1 managed to clear T1.Stalled. T2 does not need to abort 2139 // its spin in this circumstance. 2140 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2141 2142 if (BlockedOn == 1) return 1; 2143 if (BlockedOn != 0) { 2144 return BlockedOn != intptr_t(this) && _owner == ox; 2145 } 2146 2147 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2148 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2149 // consider also: jst != _thread_in_Java -- but that's overspecific. 2150 return jst == _thread_blocked || jst == _thread_in_native; 2151 } 2152 2153 2154 // ----------------------------------------------------------------------------- 2155 // WaitSet management ... 2156 2157 ObjectWaiter::ObjectWaiter(Thread* thread) { 2158 _next = NULL; 2159 _prev = NULL; 2160 _notified = 0; 2161 TState = TS_RUN; 2162 _thread = thread; 2163 _event = thread->_ParkEvent; 2164 _active = false; 2165 assert(_event != NULL, "invariant"); 2166 } 2167 2168 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2169 JavaThread *jt = (JavaThread *)this->_thread; 2170 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2171 } 2172 2173 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2174 JavaThread *jt = (JavaThread *)this->_thread; 2175 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2176 } 2177 2178 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2179 assert(node != NULL, "should not add NULL node"); 2180 assert(node->_prev == NULL, "node already in list"); 2181 assert(node->_next == NULL, "node already in list"); 2182 // put node at end of queue (circular doubly linked list) 2183 if (_WaitSet == NULL) { 2184 _WaitSet = node; 2185 node->_prev = node; 2186 node->_next = node; 2187 } else { 2188 ObjectWaiter* head = _WaitSet; 2189 ObjectWaiter* tail = head->_prev; 2190 assert(tail->_next == head, "invariant check"); 2191 tail->_next = node; 2192 head->_prev = node; 2193 node->_next = head; 2194 node->_prev = tail; 2195 } 2196 } 2197 2198 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2199 // dequeue the very first waiter 2200 ObjectWaiter* waiter = _WaitSet; 2201 if (waiter) { 2202 DequeueSpecificWaiter(waiter); 2203 } 2204 return waiter; 2205 } 2206 2207 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2208 assert(node != NULL, "should not dequeue NULL node"); 2209 assert(node->_prev != NULL, "node already removed from list"); 2210 assert(node->_next != NULL, "node already removed from list"); 2211 // when the waiter has woken up because of interrupt, 2212 // timeout or other spurious wake-up, dequeue the 2213 // waiter from waiting list 2214 ObjectWaiter* next = node->_next; 2215 if (next == node) { 2216 assert(node->_prev == node, "invariant check"); 2217 _WaitSet = NULL; 2218 } else { 2219 ObjectWaiter* prev = node->_prev; 2220 assert(prev->_next == node, "invariant check"); 2221 assert(next->_prev == node, "invariant check"); 2222 next->_prev = prev; 2223 prev->_next = next; 2224 if (_WaitSet == node) { 2225 _WaitSet = next; 2226 } 2227 } 2228 node->_next = NULL; 2229 node->_prev = NULL; 2230 } 2231 2232 // ----------------------------------------------------------------------------- 2233 // PerfData support 2234 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2235 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2236 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2237 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2238 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2239 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2240 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2241 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2242 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2243 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2244 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2245 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2246 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2247 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2248 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2249 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2250 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2251 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2252 2253 // One-shot global initialization for the sync subsystem. 2254 // We could also defer initialization and initialize on-demand 2255 // the first time we call inflate(). Initialization would 2256 // be protected - like so many things - by the MonitorCache_lock. 2257 2258 void ObjectMonitor::Initialize() { 2259 static int InitializationCompleted = 0; 2260 assert(InitializationCompleted == 0, "invariant"); 2261 InitializationCompleted = 1; 2262 if (UsePerfData) { 2263 EXCEPTION_MARK; 2264 #define NEWPERFCOUNTER(n) \ 2265 { \ 2266 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2267 CHECK); \ 2268 } 2269 #define NEWPERFVARIABLE(n) \ 2270 { \ 2271 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2272 CHECK); \ 2273 } 2274 NEWPERFCOUNTER(_sync_Inflations); 2275 NEWPERFCOUNTER(_sync_Deflations); 2276 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2277 NEWPERFCOUNTER(_sync_FutileWakeups); 2278 NEWPERFCOUNTER(_sync_Parks); 2279 NEWPERFCOUNTER(_sync_EmptyNotifications); 2280 NEWPERFCOUNTER(_sync_Notifications); 2281 NEWPERFCOUNTER(_sync_SlowEnter); 2282 NEWPERFCOUNTER(_sync_SlowExit); 2283 NEWPERFCOUNTER(_sync_SlowNotify); 2284 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2285 NEWPERFCOUNTER(_sync_FailedSpins); 2286 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2287 NEWPERFCOUNTER(_sync_PrivateA); 2288 NEWPERFCOUNTER(_sync_PrivateB); 2289 NEWPERFCOUNTER(_sync_MonInCirculation); 2290 NEWPERFCOUNTER(_sync_MonScavenged); 2291 NEWPERFVARIABLE(_sync_MonExtant); 2292 #undef NEWPERFCOUNTER 2293 #undef NEWPERFVARIABLE 2294 } 2295 } 2296 2297 static char * kvGet(char * kvList, const char * Key) { 2298 if (kvList == NULL) return NULL; 2299 size_t n = strlen(Key); 2300 char * Search; 2301 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2302 if (strncmp (Search, Key, n) == 0) { 2303 if (Search[n] == '=') return Search + n + 1; 2304 if (Search[n] == 0) return(char *) "1"; 2305 } 2306 } 2307 return NULL; 2308 } 2309 2310 static int kvGetInt(char * kvList, const char * Key, int Default) { 2311 char * v = kvGet(kvList, Key); 2312 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2313 if (Knob_ReportSettings && v != NULL) { 2314 tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ; 2315 tty->flush(); 2316 } 2317 return rslt; 2318 } 2319 2320 void ObjectMonitor::DeferredInitialize() { 2321 if (InitDone > 0) return; 2322 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2323 while (InitDone != 1) /* empty */; 2324 return; 2325 } 2326 2327 // One-shot global initialization ... 2328 // The initialization is idempotent, so we don't need locks. 2329 // In the future consider doing this via os::init_2(). 2330 // SyncKnobs consist of <Key>=<Value> pairs in the style 2331 // of environment variables. Start by converting ':' to NUL. 2332 2333 if (SyncKnobs == NULL) SyncKnobs = ""; 2334 2335 size_t sz = strlen(SyncKnobs); 2336 char * knobs = (char *) malloc(sz + 2); 2337 if (knobs == NULL) { 2338 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2339 guarantee(0, "invariant"); 2340 } 2341 strcpy(knobs, SyncKnobs); 2342 knobs[sz+1] = 0; 2343 for (char * p = knobs; *p; p++) { 2344 if (*p == ':') *p = 0; 2345 } 2346 2347 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2348 SETKNOB(ReportSettings); 2349 SETKNOB(ExitRelease); 2350 SETKNOB(Verbose); 2351 SETKNOB(VerifyInUse); 2352 SETKNOB(VerifyMatch); 2353 SETKNOB(FixedSpin); 2354 SETKNOB(SpinLimit); 2355 SETKNOB(SpinBase); 2356 SETKNOB(SpinBackOff); 2357 SETKNOB(CASPenalty); 2358 SETKNOB(OXPenalty); 2359 SETKNOB(LogSpins); 2360 SETKNOB(SpinSetSucc); 2361 SETKNOB(SuccEnabled); 2362 SETKNOB(SuccRestrict); 2363 SETKNOB(Penalty); 2364 SETKNOB(Bonus); 2365 SETKNOB(BonusB); 2366 SETKNOB(Poverty); 2367 SETKNOB(SpinAfterFutile); 2368 SETKNOB(UsePause); 2369 SETKNOB(SpinEarly); 2370 SETKNOB(OState); 2371 SETKNOB(MaxSpinners); 2372 SETKNOB(PreSpin); 2373 SETKNOB(ExitPolicy); 2374 SETKNOB(QMode); 2375 SETKNOB(ResetEvent); 2376 SETKNOB(MoveNotifyee); 2377 SETKNOB(FastHSSEC); 2378 #undef SETKNOB 2379 2380 if (Knob_Verbose) { 2381 sanity_checks(); 2382 } 2383 2384 if (os::is_MP()) { 2385 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2386 if (Knob_ReportSettings) { 2387 tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask); 2388 } 2389 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2390 } else { 2391 Knob_SpinLimit = 0; 2392 Knob_SpinBase = 0; 2393 Knob_PreSpin = 0; 2394 Knob_FixedSpin = -1; 2395 } 2396 2397 if (Knob_LogSpins == 0) { 2398 ObjectMonitor::_sync_FailedSpins = NULL; 2399 } 2400 2401 free(knobs); 2402 OrderAccess::fence(); 2403 InitDone = 1; 2404 } 2405 2406 void ObjectMonitor::sanity_checks() { 2407 int error_cnt = 0; 2408 int warning_cnt = 0; 2409 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2410 2411 if (verbose) { 2412 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2413 sizeof(ObjectMonitor)); 2414 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT, 2415 sizeof(PaddedEnd<ObjectMonitor>)); 2416 } 2417 2418 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2419 if (verbose) { 2420 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2421 } 2422 2423 ObjectMonitor dummy; 2424 u_char *addr_begin = (u_char*)&dummy; 2425 u_char *addr_header = (u_char*)&dummy._header; 2426 u_char *addr_owner = (u_char*)&dummy._owner; 2427 2428 uint offset_header = (uint)(addr_header - addr_begin); 2429 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2430 2431 uint offset_owner = (uint)(addr_owner - addr_begin); 2432 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2433 2434 if ((uint)(addr_header - addr_begin) != 0) { 2435 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2436 error_cnt++; 2437 } 2438 2439 if (cache_line_size != 0) { 2440 // We were able to determine the L1 data cache line size so 2441 // do some cache line specific sanity checks 2442 2443 if ((offset_owner - offset_header) < cache_line_size) { 2444 tty->print_cr("WARNING: the _header and _owner fields are closer " 2445 "than a cache line which permits false sharing."); 2446 warning_cnt++; 2447 } 2448 2449 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) { 2450 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a " 2451 "multiple of a cache line which permits false sharing."); 2452 warning_cnt++; 2453 } 2454 } 2455 2456 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2457 &warning_cnt); 2458 2459 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2460 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2461 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2462 } 2463 2464 guarantee(error_cnt == 0, 2465 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2466 } 2467 2468 #ifndef PRODUCT 2469 void ObjectMonitor::verify() { 2470 } 2471 2472 void ObjectMonitor::print() { 2473 } 2474 2475 void ObjectMonitor_test() { 2476 ObjectMonitor::sanity_checks(); 2477 } 2478 #endif