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