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