1 /* 2 * Copyright (c) 1998, 2019, 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_SpinLimit = 5000; // derived by an external tool - 105 106 static int Knob_Bonus = 100; // spin success bonus 107 static int Knob_BonusB = 100; // spin success bonus 108 static int Knob_Penalty = 200; // spin failure penalty 109 static int Knob_Poverty = 1000; 110 static int Knob_FixedSpin = 0; 111 static int Knob_PreSpin = 10; // 20-100 likely better 112 113 DEBUG_ONLY(static volatile bool InitDone = false;) 114 115 // ----------------------------------------------------------------------------- 116 // Theory of operations -- Monitors lists, thread residency, etc: 117 // 118 // * A thread acquires ownership of a monitor by successfully 119 // CAS()ing the _owner field from null to non-null. 120 // 121 // * Invariant: A thread appears on at most one monitor list -- 122 // cxq, EntryList or WaitSet -- at any one time. 123 // 124 // * Contending threads "push" themselves onto the cxq with CAS 125 // and then spin/park. 126 // 127 // * After a contending thread eventually acquires the lock it must 128 // dequeue itself from either the EntryList or the cxq. 129 // 130 // * The exiting thread identifies and unparks an "heir presumptive" 131 // tentative successor thread on the EntryList. Critically, the 132 // exiting thread doesn't unlink the successor thread from the EntryList. 133 // After having been unparked, the wakee will recontend for ownership of 134 // the monitor. The successor (wakee) will either acquire the lock or 135 // re-park itself. 136 // 137 // Succession is provided for by a policy of competitive handoff. 138 // The exiting thread does _not_ grant or pass ownership to the 139 // successor thread. (This is also referred to as "handoff" succession"). 140 // Instead the exiting thread releases ownership and possibly wakes 141 // a successor, so the successor can (re)compete for ownership of the lock. 142 // If the EntryList is empty but the cxq is populated the exiting 143 // thread will drain the cxq into the EntryList. It does so by 144 // by detaching the cxq (installing null with CAS) and folding 145 // the threads from the cxq into the EntryList. The EntryList is 146 // doubly linked, while the cxq is singly linked because of the 147 // CAS-based "push" used to enqueue recently arrived threads (RATs). 148 // 149 // * Concurrency invariants: 150 // 151 // -- only the monitor owner may access or mutate the EntryList. 152 // The mutex property of the monitor itself protects the EntryList 153 // from concurrent interference. 154 // -- Only the monitor owner may detach the cxq. 155 // 156 // * The monitor entry list operations avoid locks, but strictly speaking 157 // they're not lock-free. Enter is lock-free, exit is not. 158 // For a description of 'Methods and apparatus providing non-blocking access 159 // to a resource,' see U.S. Pat. No. 7844973. 160 // 161 // * The cxq can have multiple concurrent "pushers" but only one concurrent 162 // detaching thread. This mechanism is immune from the ABA corruption. 163 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 164 // 165 // * Taken together, the cxq and the EntryList constitute or form a 166 // single logical queue of threads stalled trying to acquire the lock. 167 // We use two distinct lists to improve the odds of a constant-time 168 // dequeue operation after acquisition (in the ::enter() epilogue) and 169 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 170 // A key desideratum is to minimize queue & monitor metadata manipulation 171 // that occurs while holding the monitor lock -- that is, we want to 172 // minimize monitor lock holds times. Note that even a small amount of 173 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 174 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 175 // locks and monitor metadata. 176 // 177 // Cxq points to the set of Recently Arrived Threads attempting entry. 178 // Because we push threads onto _cxq with CAS, the RATs must take the form of 179 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 180 // the unlocking thread notices that EntryList is null but _cxq is != null. 181 // 182 // The EntryList is ordered by the prevailing queue discipline and 183 // can be organized in any convenient fashion, such as a doubly-linked list or 184 // a circular doubly-linked list. Critically, we want insert and delete operations 185 // to operate in constant-time. If we need a priority queue then something akin 186 // to Solaris' sleepq would work nicely. Viz., 187 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 188 // Queue discipline is enforced at ::exit() time, when the unlocking thread 189 // drains the cxq into the EntryList, and orders or reorders the threads on the 190 // EntryList accordingly. 191 // 192 // Barring "lock barging", this mechanism provides fair cyclic ordering, 193 // somewhat similar to an elevator-scan. 194 // 195 // * The monitor synchronization subsystem avoids the use of native 196 // synchronization primitives except for the narrow platform-specific 197 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 198 // the semantics of park-unpark. Put another way, this monitor implementation 199 // depends only on atomic operations and park-unpark. The monitor subsystem 200 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 201 // underlying OS manages the READY<->RUN transitions. 202 // 203 // * Waiting threads reside on the WaitSet list -- wait() puts 204 // the caller onto the WaitSet. 205 // 206 // * notify() or notifyAll() simply transfers threads from the WaitSet to 207 // either the EntryList or cxq. Subsequent exit() operations will 208 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 209 // it's likely the notifyee would simply impale itself on the lock held 210 // by the notifier. 211 // 212 // * An interesting alternative is to encode cxq as (List,LockByte) where 213 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 214 // variable, like _recursions, in the scheme. The threads or Events that form 215 // the list would have to be aligned in 256-byte addresses. A thread would 216 // try to acquire the lock or enqueue itself with CAS, but exiting threads 217 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 218 // Note that is is *not* word-tearing, but it does presume that full-word 219 // CAS operations are coherent with intermix with STB operations. That's true 220 // on most common processors. 221 // 222 // * See also http://blogs.sun.com/dave 223 224 225 void* ObjectMonitor::operator new (size_t size) throw() { 226 return AllocateHeap(size, mtInternal); 227 } 228 void* ObjectMonitor::operator new[] (size_t size) throw() { 229 return operator new (size); 230 } 231 void ObjectMonitor::operator delete(void* p) { 232 FreeHeap(p); 233 } 234 void ObjectMonitor::operator delete[] (void *p) { 235 operator delete(p); 236 } 237 238 // ----------------------------------------------------------------------------- 239 // Enter support 240 241 void ObjectMonitor::enter(TRAPS) { 242 // The following code is ordered to check the most common cases first 243 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 244 Thread * const Self = THREAD; 245 246 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 247 if (cur == NULL) { 248 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 249 assert(_recursions == 0, "invariant"); 250 assert(_owner == Self, "invariant"); 251 return; 252 } 253 254 if (cur == Self) { 255 // TODO-FIXME: check for integer overflow! BUGID 6557169. 256 _recursions++; 257 return; 258 } 259 260 if (Self->is_lock_owned ((address)cur)) { 261 assert(_recursions == 0, "internal state error"); 262 _recursions = 1; 263 // Commute owner from a thread-specific on-stack BasicLockObject address to 264 // a full-fledged "Thread *". 265 _owner = Self; 266 return; 267 } 268 269 // We've encountered genuine contention. 270 assert(Self->_Stalled == 0, "invariant"); 271 Self->_Stalled = intptr_t(this); 272 273 // Try one round of spinning *before* enqueueing Self 274 // and before going through the awkward and expensive state 275 // transitions. The following spin is strictly optional ... 276 // Note that if we acquire the monitor from an initial spin 277 // we forgo posting JVMTI events and firing DTRACE probes. 278 if (TrySpin(Self) > 0) { 279 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 280 assert(_recursions == 0, "must be 0: recursions=" INTPTR_FORMAT, 281 _recursions); 282 assert(((oop)object())->mark() == markOopDesc::encode(this), 283 "object mark must match encoded this: mark=" INTPTR_FORMAT 284 ", encoded this=" INTPTR_FORMAT, p2i(((oop)object())->mark()), 285 p2i(markOopDesc::encode(this))); 286 Self->_Stalled = 0; 287 return; 288 } 289 290 assert(_owner != Self, "invariant"); 291 assert(_succ != Self, "invariant"); 292 assert(Self->is_Java_thread(), "invariant"); 293 JavaThread * jt = (JavaThread *) Self; 294 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 295 assert(jt->thread_state() != _thread_blocked, "invariant"); 296 assert(this->object() != NULL, "invariant"); 297 assert(_contentions >= 0, "invariant"); 298 299 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 300 // Ensure the object-monitor relationship remains stable while there's contention. 301 Atomic::inc(&_contentions); 302 303 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 304 EventJavaMonitorEnter event; 305 if (event.should_commit()) { 306 event.set_monitorClass(((oop)this->object())->klass()); 307 event.set_address((uintptr_t)(this->object_addr())); 308 } 309 310 { // Change java thread status to indicate blocked on monitor enter. 311 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 312 313 Self->set_current_pending_monitor(this); 314 315 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 316 if (JvmtiExport::should_post_monitor_contended_enter()) { 317 JvmtiExport::post_monitor_contended_enter(jt, this); 318 319 // The current thread does not yet own the monitor and does not 320 // yet appear on any queues that would get it made the successor. 321 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 322 // handler cannot accidentally consume an unpark() meant for the 323 // ParkEvent associated with this ObjectMonitor. 324 } 325 326 OSThreadContendState osts(Self->osthread()); 327 ThreadBlockInVM tbivm(jt); 328 329 // TODO-FIXME: change the following for(;;) loop to straight-line code. 330 for (;;) { 331 jt->set_suspend_equivalent(); 332 // cleared by handle_special_suspend_equivalent_condition() 333 // or java_suspend_self() 334 335 EnterI(THREAD); 336 337 if (!ExitSuspendEquivalent(jt)) break; 338 339 // We have acquired the contended monitor, but while we were 340 // waiting another thread suspended us. We don't want to enter 341 // the monitor while suspended because that would surprise the 342 // thread that suspended us. 343 // 344 _recursions = 0; 345 _succ = NULL; 346 exit(false, Self); 347 348 jt->java_suspend_self(); 349 } 350 Self->set_current_pending_monitor(NULL); 351 352 // We cleared the pending monitor info since we've just gotten past 353 // the enter-check-for-suspend dance and we now own the monitor free 354 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 355 // destructor can go to a safepoint at the end of this block. If we 356 // do a thread dump during that safepoint, then this thread will show 357 // as having "-locked" the monitor, but the OS and java.lang.Thread 358 // states will still report that the thread is blocked trying to 359 // acquire it. 360 } 361 362 Atomic::dec(&_contentions); 363 assert(_contentions >= 0, "invariant"); 364 Self->_Stalled = 0; 365 366 // Must either set _recursions = 0 or ASSERT _recursions == 0. 367 assert(_recursions == 0, "invariant"); 368 assert(_owner == Self, "invariant"); 369 assert(_succ != Self, "invariant"); 370 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 371 372 // The thread -- now the owner -- is back in vm mode. 373 // Report the glorious news via TI,DTrace and jvmstat. 374 // The probe effect is non-trivial. All the reportage occurs 375 // while we hold the monitor, increasing the length of the critical 376 // section. Amdahl's parallel speedup law comes vividly into play. 377 // 378 // Another option might be to aggregate the events (thread local or 379 // per-monitor aggregation) and defer reporting until a more opportune 380 // time -- such as next time some thread encounters contention but has 381 // yet to acquire the lock. While spinning that thread could 382 // spinning we could increment JVMStat counters, etc. 383 384 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 385 if (JvmtiExport::should_post_monitor_contended_entered()) { 386 JvmtiExport::post_monitor_contended_entered(jt, this); 387 388 // The current thread already owns the monitor and is not going to 389 // call park() for the remainder of the monitor enter protocol. So 390 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 391 // event handler consumed an unpark() issued by the thread that 392 // just exited the monitor. 393 } 394 if (event.should_commit()) { 395 event.set_previousOwner((uintptr_t)_previous_owner_tid); 396 event.commit(); 397 } 398 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 399 } 400 401 // Caveat: TryLock() is not necessarily serializing if it returns failure. 402 // Callers must compensate as needed. 403 404 int ObjectMonitor::TryLock(Thread * Self) { 405 void * own = _owner; 406 if (own != NULL) return 0; 407 if (Atomic::replace_if_null(Self, &_owner)) { 408 // Either guarantee _recursions == 0 or set _recursions = 0. 409 assert(_recursions == 0, "invariant"); 410 assert(_owner == Self, "invariant"); 411 return 1; 412 } 413 // The lock had been free momentarily, but we lost the race to the lock. 414 // Interference -- the CAS failed. 415 // We can either return -1 or retry. 416 // Retry doesn't make as much sense because the lock was just acquired. 417 return -1; 418 } 419 420 #define MAX_RECHECK_INTERVAL 1000 421 422 void ObjectMonitor::EnterI(TRAPS) { 423 Thread * const Self = THREAD; 424 assert(Self->is_Java_thread(), "invariant"); 425 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 426 427 // Try the lock - TATAS 428 if (TryLock (Self) > 0) { 429 assert(_succ != Self, "invariant"); 430 assert(_owner == Self, "invariant"); 431 assert(_Responsible != Self, "invariant"); 432 return; 433 } 434 435 assert(InitDone, "Unexpectedly not initialized"); 436 437 // We try one round of spinning *before* enqueueing Self. 438 // 439 // If the _owner is ready but OFFPROC we could use a YieldTo() 440 // operation to donate the remainder of this thread's quantum 441 // to the owner. This has subtle but beneficial affinity 442 // effects. 443 444 if (TrySpin(Self) > 0) { 445 assert(_owner == Self, "invariant"); 446 assert(_succ != Self, "invariant"); 447 assert(_Responsible != Self, "invariant"); 448 return; 449 } 450 451 // The Spin failed -- Enqueue and park the thread ... 452 assert(_succ != Self, "invariant"); 453 assert(_owner != Self, "invariant"); 454 assert(_Responsible != Self, "invariant"); 455 456 // Enqueue "Self" on ObjectMonitor's _cxq. 457 // 458 // Node acts as a proxy for Self. 459 // As an aside, if were to ever rewrite the synchronization code mostly 460 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 461 // Java objects. This would avoid awkward lifecycle and liveness issues, 462 // as well as eliminate a subset of ABA issues. 463 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 464 465 ObjectWaiter node(Self); 466 Self->_ParkEvent->reset(); 467 node._prev = (ObjectWaiter *) 0xBAD; 468 node.TState = ObjectWaiter::TS_CXQ; 469 470 // Push "Self" onto the front of the _cxq. 471 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 472 // Note that spinning tends to reduce the rate at which threads 473 // enqueue and dequeue on EntryList|cxq. 474 ObjectWaiter * nxt; 475 for (;;) { 476 node._next = nxt = _cxq; 477 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 478 479 // Interference - the CAS failed because _cxq changed. Just retry. 480 // As an optional optimization we retry the lock. 481 if (TryLock (Self) > 0) { 482 assert(_succ != Self, "invariant"); 483 assert(_owner == Self, "invariant"); 484 assert(_Responsible != Self, "invariant"); 485 return; 486 } 487 } 488 489 // Check for cxq|EntryList edge transition to non-null. This indicates 490 // the onset of contention. While contention persists exiting threads 491 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 492 // operations revert to the faster 1-0 mode. This enter operation may interleave 493 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 494 // arrange for one of the contending thread to use a timed park() operations 495 // to detect and recover from the race. (Stranding is form of progress failure 496 // where the monitor is unlocked but all the contending threads remain parked). 497 // That is, at least one of the contended threads will periodically poll _owner. 498 // One of the contending threads will become the designated "Responsible" thread. 499 // The Responsible thread uses a timed park instead of a normal indefinite park 500 // operation -- it periodically wakes and checks for and recovers from potential 501 // strandings admitted by 1-0 exit operations. We need at most one Responsible 502 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 503 // be responsible for a monitor. 504 // 505 // Currently, one of the contended threads takes on the added role of "Responsible". 506 // A viable alternative would be to use a dedicated "stranding checker" thread 507 // that periodically iterated over all the threads (or active monitors) and unparked 508 // successors where there was risk of stranding. This would help eliminate the 509 // timer scalability issues we see on some platforms as we'd only have one thread 510 // -- the checker -- parked on a timer. 511 512 if (nxt == NULL && _EntryList == NULL) { 513 // Try to assume the role of responsible thread for the monitor. 514 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 515 Atomic::replace_if_null(Self, &_Responsible); 516 } 517 518 // The lock might have been released while this thread was occupied queueing 519 // itself onto _cxq. To close the race and avoid "stranding" and 520 // progress-liveness failure we must resample-retry _owner before parking. 521 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 522 // In this case the ST-MEMBAR is accomplished with CAS(). 523 // 524 // TODO: Defer all thread state transitions until park-time. 525 // Since state transitions are heavy and inefficient we'd like 526 // to defer the state transitions until absolutely necessary, 527 // and in doing so avoid some transitions ... 528 529 int nWakeups = 0; 530 int recheckInterval = 1; 531 532 for (;;) { 533 534 if (TryLock(Self) > 0) break; 535 assert(_owner != Self, "invariant"); 536 537 // park self 538 if (_Responsible == Self) { 539 Self->_ParkEvent->park((jlong) recheckInterval); 540 // Increase the recheckInterval, but clamp the value. 541 recheckInterval *= 8; 542 if (recheckInterval > MAX_RECHECK_INTERVAL) { 543 recheckInterval = MAX_RECHECK_INTERVAL; 544 } 545 } else { 546 Self->_ParkEvent->park(); 547 } 548 549 if (TryLock(Self) > 0) break; 550 551 // The lock is still contested. 552 // Keep a tally of the # of futile wakeups. 553 // Note that the counter is not protected by a lock or updated by atomics. 554 // That is by design - we trade "lossy" counters which are exposed to 555 // races during updates for a lower probe effect. 556 557 // This PerfData object can be used in parallel with a safepoint. 558 // See the work around in PerfDataManager::destroy(). 559 OM_PERFDATA_OP(FutileWakeups, inc()); 560 ++nWakeups; 561 562 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 563 // We can defer clearing _succ until after the spin completes 564 // TrySpin() must tolerate being called with _succ == Self. 565 // Try yet another round of adaptive spinning. 566 if (TrySpin(Self) > 0) break; 567 568 // We can find that we were unpark()ed and redesignated _succ while 569 // we were spinning. That's harmless. If we iterate and call park(), 570 // park() will consume the event and return immediately and we'll 571 // just spin again. This pattern can repeat, leaving _succ to simply 572 // spin on a CPU. 573 574 if (_succ == Self) _succ = NULL; 575 576 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 577 OrderAccess::fence(); 578 } 579 580 // Egress : 581 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 582 // Normally we'll find Self on the EntryList . 583 // From the perspective of the lock owner (this thread), the 584 // EntryList is stable and cxq is prepend-only. 585 // The head of cxq is volatile but the interior is stable. 586 // In addition, Self.TState is stable. 587 588 assert(_owner == Self, "invariant"); 589 assert(object() != NULL, "invariant"); 590 // I'd like to write: 591 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 592 // but as we're at a safepoint that's not safe. 593 594 UnlinkAfterAcquire(Self, &node); 595 if (_succ == Self) _succ = NULL; 596 597 assert(_succ != Self, "invariant"); 598 if (_Responsible == Self) { 599 _Responsible = NULL; 600 OrderAccess::fence(); // Dekker pivot-point 601 602 // We may leave threads on cxq|EntryList without a designated 603 // "Responsible" thread. This is benign. When this thread subsequently 604 // exits the monitor it can "see" such preexisting "old" threads -- 605 // threads that arrived on the cxq|EntryList before the fence, above -- 606 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 607 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 608 // non-null and elect a new "Responsible" timer thread. 609 // 610 // This thread executes: 611 // ST Responsible=null; MEMBAR (in enter epilogue - here) 612 // LD cxq|EntryList (in subsequent exit) 613 // 614 // Entering threads in the slow/contended path execute: 615 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 616 // The (ST cxq; MEMBAR) is accomplished with CAS(). 617 // 618 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 619 // exit operation from floating above the ST Responsible=null. 620 } 621 622 // We've acquired ownership with CAS(). 623 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 624 // But since the CAS() this thread may have also stored into _succ, 625 // EntryList, cxq or Responsible. These meta-data updates must be 626 // visible __before this thread subsequently drops the lock. 627 // Consider what could occur if we didn't enforce this constraint -- 628 // STs to monitor meta-data and user-data could reorder with (become 629 // visible after) the ST in exit that drops ownership of the lock. 630 // Some other thread could then acquire the lock, but observe inconsistent 631 // or old monitor meta-data and heap data. That violates the JMM. 632 // To that end, the 1-0 exit() operation must have at least STST|LDST 633 // "release" barrier semantics. Specifically, there must be at least a 634 // STST|LDST barrier in exit() before the ST of null into _owner that drops 635 // the lock. The barrier ensures that changes to monitor meta-data and data 636 // protected by the lock will be visible before we release the lock, and 637 // therefore before some other thread (CPU) has a chance to acquire the lock. 638 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 639 // 640 // Critically, any prior STs to _succ or EntryList must be visible before 641 // the ST of null into _owner in the *subsequent* (following) corresponding 642 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 643 // execute a serializing instruction. 644 645 return; 646 } 647 648 // ReenterI() is a specialized inline form of the latter half of the 649 // contended slow-path from EnterI(). We use ReenterI() only for 650 // monitor reentry in wait(). 651 // 652 // In the future we should reconcile EnterI() and ReenterI(). 653 654 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 655 assert(Self != NULL, "invariant"); 656 assert(SelfNode != NULL, "invariant"); 657 assert(SelfNode->_thread == Self, "invariant"); 658 assert(_waiters > 0, "invariant"); 659 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 660 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 661 JavaThread * jt = (JavaThread *) Self; 662 663 int nWakeups = 0; 664 for (;;) { 665 ObjectWaiter::TStates v = SelfNode->TState; 666 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 667 assert(_owner != Self, "invariant"); 668 669 if (TryLock(Self) > 0) break; 670 if (TrySpin(Self) > 0) break; 671 672 // State transition wrappers around park() ... 673 // ReenterI() wisely defers state transitions until 674 // it's clear we must park the thread. 675 { 676 OSThreadContendState osts(Self->osthread()); 677 ThreadBlockInVM tbivm(jt); 678 679 // cleared by handle_special_suspend_equivalent_condition() 680 // or java_suspend_self() 681 jt->set_suspend_equivalent(); 682 Self->_ParkEvent->park(); 683 684 // were we externally suspended while we were waiting? 685 for (;;) { 686 if (!ExitSuspendEquivalent(jt)) break; 687 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 688 jt->java_suspend_self(); 689 jt->set_suspend_equivalent(); 690 } 691 } 692 693 // Try again, but just so we distinguish between futile wakeups and 694 // successful wakeups. The following test isn't algorithmically 695 // necessary, but it helps us maintain sensible statistics. 696 if (TryLock(Self) > 0) break; 697 698 // The lock is still contested. 699 // Keep a tally of the # of futile wakeups. 700 // Note that the counter is not protected by a lock or updated by atomics. 701 // That is by design - we trade "lossy" counters which are exposed to 702 // races during updates for a lower probe effect. 703 ++nWakeups; 704 705 // Assuming this is not a spurious wakeup we'll normally 706 // find that _succ == Self. 707 if (_succ == Self) _succ = NULL; 708 709 // Invariant: after clearing _succ a contending thread 710 // *must* retry _owner before parking. 711 OrderAccess::fence(); 712 713 // This PerfData object can be used in parallel with a safepoint. 714 // See the work around in PerfDataManager::destroy(). 715 OM_PERFDATA_OP(FutileWakeups, inc()); 716 } 717 718 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 719 // Normally we'll find Self on the EntryList. 720 // Unlinking from the EntryList is constant-time and atomic-free. 721 // From the perspective of the lock owner (this thread), the 722 // EntryList is stable and cxq is prepend-only. 723 // The head of cxq is volatile but the interior is stable. 724 // In addition, Self.TState is stable. 725 726 assert(_owner == Self, "invariant"); 727 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 728 UnlinkAfterAcquire(Self, SelfNode); 729 if (_succ == Self) _succ = NULL; 730 assert(_succ != Self, "invariant"); 731 SelfNode->TState = ObjectWaiter::TS_RUN; 732 OrderAccess::fence(); // see comments at the end of EnterI() 733 } 734 735 // By convention we unlink a contending thread from EntryList|cxq immediately 736 // after the thread acquires the lock in ::enter(). Equally, we could defer 737 // unlinking the thread until ::exit()-time. 738 739 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 740 assert(_owner == Self, "invariant"); 741 assert(SelfNode->_thread == Self, "invariant"); 742 743 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 744 // Normal case: remove Self from the DLL EntryList . 745 // This is a constant-time operation. 746 ObjectWaiter * nxt = SelfNode->_next; 747 ObjectWaiter * prv = SelfNode->_prev; 748 if (nxt != NULL) nxt->_prev = prv; 749 if (prv != NULL) prv->_next = nxt; 750 if (SelfNode == _EntryList) _EntryList = nxt; 751 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 752 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 753 } else { 754 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 755 // Inopportune interleaving -- Self is still on the cxq. 756 // This usually means the enqueue of self raced an exiting thread. 757 // Normally we'll find Self near the front of the cxq, so 758 // dequeueing is typically fast. If needbe we can accelerate 759 // this with some MCS/CHL-like bidirectional list hints and advisory 760 // back-links so dequeueing from the interior will normally operate 761 // in constant-time. 762 // Dequeue Self from either the head (with CAS) or from the interior 763 // with a linear-time scan and normal non-atomic memory operations. 764 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 765 // and then unlink Self from EntryList. We have to drain eventually, 766 // so it might as well be now. 767 768 ObjectWaiter * v = _cxq; 769 assert(v != NULL, "invariant"); 770 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 771 // The CAS above can fail from interference IFF a "RAT" arrived. 772 // In that case Self must be in the interior and can no longer be 773 // at the head of cxq. 774 if (v == SelfNode) { 775 assert(_cxq != v, "invariant"); 776 v = _cxq; // CAS above failed - start scan at head of list 777 } 778 ObjectWaiter * p; 779 ObjectWaiter * q = NULL; 780 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 781 q = p; 782 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 783 } 784 assert(v != SelfNode, "invariant"); 785 assert(p == SelfNode, "Node not found on cxq"); 786 assert(p != _cxq, "invariant"); 787 assert(q != NULL, "invariant"); 788 assert(q->_next == p, "invariant"); 789 q->_next = p->_next; 790 } 791 } 792 793 #ifdef ASSERT 794 // Diagnostic hygiene ... 795 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 796 SelfNode->_next = (ObjectWaiter *) 0xBAD; 797 SelfNode->TState = ObjectWaiter::TS_RUN; 798 #endif 799 } 800 801 // ----------------------------------------------------------------------------- 802 // Exit support 803 // 804 // exit() 805 // ~~~~~~ 806 // Note that the collector can't reclaim the objectMonitor or deflate 807 // the object out from underneath the thread calling ::exit() as the 808 // thread calling ::exit() never transitions to a stable state. 809 // This inhibits GC, which in turn inhibits asynchronous (and 810 // inopportune) reclamation of "this". 811 // 812 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 813 // There's one exception to the claim above, however. EnterI() can call 814 // exit() to drop a lock if the acquirer has been externally suspended. 815 // In that case exit() is called with _thread_state as _thread_blocked, 816 // but the monitor's _contentions field is > 0, which inhibits reclamation. 817 // 818 // 1-0 exit 819 // ~~~~~~~~ 820 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 821 // the fast-path operators have been optimized so the common ::exit() 822 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 823 // The code emitted by fast_unlock() elides the usual MEMBAR. This 824 // greatly improves latency -- MEMBAR and CAS having considerable local 825 // latency on modern processors -- but at the cost of "stranding". Absent the 826 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 827 // ::enter() path, resulting in the entering thread being stranding 828 // and a progress-liveness failure. Stranding is extremely rare. 829 // We use timers (timed park operations) & periodic polling to detect 830 // and recover from stranding. Potentially stranded threads periodically 831 // wake up and poll the lock. See the usage of the _Responsible variable. 832 // 833 // The CAS() in enter provides for safety and exclusion, while the CAS or 834 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 835 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 836 // We detect and recover from stranding with timers. 837 // 838 // If a thread transiently strands it'll park until (a) another 839 // thread acquires the lock and then drops the lock, at which time the 840 // exiting thread will notice and unpark the stranded thread, or, (b) 841 // the timer expires. If the lock is high traffic then the stranding latency 842 // will be low due to (a). If the lock is low traffic then the odds of 843 // stranding are lower, although the worst-case stranding latency 844 // is longer. Critically, we don't want to put excessive load in the 845 // platform's timer subsystem. We want to minimize both the timer injection 846 // rate (timers created/sec) as well as the number of timers active at 847 // any one time. (more precisely, we want to minimize timer-seconds, which is 848 // the integral of the # of active timers at any instant over time). 849 // Both impinge on OS scalability. Given that, at most one thread parked on 850 // a monitor will use a timer. 851 // 852 // There is also the risk of a futile wake-up. If we drop the lock 853 // another thread can reacquire the lock immediately, and we can 854 // then wake a thread unnecessarily. This is benign, and we've 855 // structured the code so the windows are short and the frequency 856 // of such futile wakups is low. 857 858 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 859 Thread * const Self = THREAD; 860 if (THREAD != _owner) { 861 if (THREAD->is_lock_owned((address) _owner)) { 862 // Transmute _owner from a BasicLock pointer to a Thread address. 863 // We don't need to hold _mutex for this transition. 864 // Non-null to Non-null is safe as long as all readers can 865 // tolerate either flavor. 866 assert(_recursions == 0, "invariant"); 867 _owner = THREAD; 868 _recursions = 0; 869 } else { 870 // Apparent unbalanced locking ... 871 // Naively we'd like to throw IllegalMonitorStateException. 872 // As a practical matter we can neither allocate nor throw an 873 // exception as ::exit() can be called from leaf routines. 874 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 875 // Upon deeper reflection, however, in a properly run JVM the only 876 // way we should encounter this situation is in the presence of 877 // unbalanced JNI locking. TODO: CheckJNICalls. 878 // See also: CR4414101 879 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 880 return; 881 } 882 } 883 884 if (_recursions != 0) { 885 _recursions--; // this is simple recursive enter 886 return; 887 } 888 889 // Invariant: after setting Responsible=null an thread must execute 890 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 891 _Responsible = NULL; 892 893 #if INCLUDE_JFR 894 // get the owner's thread id for the MonitorEnter event 895 // if it is enabled and the thread isn't suspended 896 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 897 _previous_owner_tid = JFR_THREAD_ID(Self); 898 } 899 #endif 900 901 for (;;) { 902 assert(THREAD == _owner, "invariant"); 903 904 // release semantics: prior loads and stores from within the critical section 905 // must not float (reorder) past the following store that drops the lock. 906 // On SPARC that requires MEMBAR #loadstore|#storestore. 907 // But of course in TSO #loadstore|#storestore is not required. 908 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 909 OrderAccess::storeload(); // See if we need to wake a successor 910 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 911 return; 912 } 913 // Other threads are blocked trying to acquire the lock. 914 915 // Normally the exiting thread is responsible for ensuring succession, 916 // but if other successors are ready or other entering threads are spinning 917 // then this thread can simply store NULL into _owner and exit without 918 // waking a successor. The existence of spinners or ready successors 919 // guarantees proper succession (liveness). Responsibility passes to the 920 // ready or running successors. The exiting thread delegates the duty. 921 // More precisely, if a successor already exists this thread is absolved 922 // of the responsibility of waking (unparking) one. 923 // 924 // The _succ variable is critical to reducing futile wakeup frequency. 925 // _succ identifies the "heir presumptive" thread that has been made 926 // ready (unparked) but that has not yet run. We need only one such 927 // successor thread to guarantee progress. 928 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 929 // section 3.3 "Futile Wakeup Throttling" for details. 930 // 931 // Note that spinners in Enter() also set _succ non-null. 932 // In the current implementation spinners opportunistically set 933 // _succ so that exiting threads might avoid waking a successor. 934 // Another less appealing alternative would be for the exiting thread 935 // to drop the lock and then spin briefly to see if a spinner managed 936 // to acquire the lock. If so, the exiting thread could exit 937 // immediately without waking a successor, otherwise the exiting 938 // thread would need to dequeue and wake a successor. 939 // (Note that we'd need to make the post-drop spin short, but no 940 // shorter than the worst-case round-trip cache-line migration time. 941 // The dropped lock needs to become visible to the spinner, and then 942 // the acquisition of the lock by the spinner must become visible to 943 // the exiting thread). 944 945 // It appears that an heir-presumptive (successor) must be made ready. 946 // Only the current lock owner can manipulate the EntryList or 947 // drain _cxq, so we need to reacquire the lock. If we fail 948 // to reacquire the lock the responsibility for ensuring succession 949 // falls to the new owner. 950 // 951 if (!Atomic::replace_if_null(THREAD, &_owner)) { 952 return; 953 } 954 955 guarantee(_owner == THREAD, "invariant"); 956 957 ObjectWaiter * w = NULL; 958 959 w = _EntryList; 960 if (w != NULL) { 961 // I'd like to write: guarantee (w->_thread != Self). 962 // But in practice an exiting thread may find itself on the EntryList. 963 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 964 // then calls exit(). Exit release the lock by setting O._owner to NULL. 965 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 966 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 967 // release the lock "O". T2 resumes immediately after the ST of null into 968 // _owner, above. T2 notices that the EntryList is populated, so it 969 // reacquires the lock and then finds itself on the EntryList. 970 // Given all that, we have to tolerate the circumstance where "w" is 971 // associated with Self. 972 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 973 ExitEpilog(Self, w); 974 return; 975 } 976 977 // If we find that both _cxq and EntryList are null then just 978 // re-run the exit protocol from the top. 979 w = _cxq; 980 if (w == NULL) continue; 981 982 // Drain _cxq into EntryList - bulk transfer. 983 // First, detach _cxq. 984 // The following loop is tantamount to: w = swap(&cxq, NULL) 985 for (;;) { 986 assert(w != NULL, "Invariant"); 987 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 988 if (u == w) break; 989 w = u; 990 } 991 992 assert(w != NULL, "invariant"); 993 assert(_EntryList == NULL, "invariant"); 994 995 // Convert the LIFO SLL anchored by _cxq into a DLL. 996 // The list reorganization step operates in O(LENGTH(w)) time. 997 // It's critical that this step operate quickly as 998 // "Self" still holds the outer-lock, restricting parallelism 999 // and effectively lengthening the critical section. 1000 // Invariant: s chases t chases u. 1001 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1002 // we have faster access to the tail. 1003 1004 _EntryList = w; 1005 ObjectWaiter * q = NULL; 1006 ObjectWaiter * p; 1007 for (p = w; p != NULL; p = p->_next) { 1008 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1009 p->TState = ObjectWaiter::TS_ENTER; 1010 p->_prev = q; 1011 q = p; 1012 } 1013 1014 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1015 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1016 1017 // See if we can abdicate to a spinner instead of waking a thread. 1018 // A primary goal of the implementation is to reduce the 1019 // context-switch rate. 1020 if (_succ != NULL) continue; 1021 1022 w = _EntryList; 1023 if (w != NULL) { 1024 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1025 ExitEpilog(Self, w); 1026 return; 1027 } 1028 } 1029 } 1030 1031 // ExitSuspendEquivalent: 1032 // A faster alternate to handle_special_suspend_equivalent_condition() 1033 // 1034 // handle_special_suspend_equivalent_condition() unconditionally 1035 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1036 // operations have high latency. Note that in ::enter() we call HSSEC 1037 // while holding the monitor, so we effectively lengthen the critical sections. 1038 // 1039 // There are a number of possible solutions: 1040 // 1041 // A. To ameliorate the problem we might also defer state transitions 1042 // to as late as possible -- just prior to parking. 1043 // Given that, we'd call HSSEC after having returned from park(), 1044 // but before attempting to acquire the monitor. This is only a 1045 // partial solution. It avoids calling HSSEC while holding the 1046 // monitor (good), but it still increases successor reacquisition latency -- 1047 // the interval between unparking a successor and the time the successor 1048 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1049 // If we use this technique we can also avoid EnterI()-exit() loop 1050 // in ::enter() where we iteratively drop the lock and then attempt 1051 // to reacquire it after suspending. 1052 // 1053 // B. In the future we might fold all the suspend bits into a 1054 // composite per-thread suspend flag and then update it with CAS(). 1055 // Alternately, a Dekker-like mechanism with multiple variables 1056 // would suffice: 1057 // ST Self->_suspend_equivalent = false 1058 // MEMBAR 1059 // LD Self_>_suspend_flags 1060 1061 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1062 return jSelf->handle_special_suspend_equivalent_condition(); 1063 } 1064 1065 1066 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1067 assert(_owner == Self, "invariant"); 1068 1069 // Exit protocol: 1070 // 1. ST _succ = wakee 1071 // 2. membar #loadstore|#storestore; 1072 // 2. ST _owner = NULL 1073 // 3. unpark(wakee) 1074 1075 _succ = Wakee->_thread; 1076 ParkEvent * Trigger = Wakee->_event; 1077 1078 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1079 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1080 // out-of-scope (non-extant). 1081 Wakee = NULL; 1082 1083 // Drop the lock 1084 OrderAccess::release_store(&_owner, (void*)NULL); 1085 OrderAccess::fence(); // ST _owner vs LD in unpark() 1086 1087 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1088 Trigger->unpark(); 1089 1090 // Maintain stats and report events to JVMTI 1091 OM_PERFDATA_OP(Parks, inc()); 1092 } 1093 1094 1095 // ----------------------------------------------------------------------------- 1096 // Class Loader deadlock handling. 1097 // 1098 // complete_exit exits a lock returning recursion count 1099 // complete_exit/reenter operate as a wait without waiting 1100 // complete_exit requires an inflated monitor 1101 // The _owner field is not always the Thread addr even with an 1102 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1103 // thread due to contention. 1104 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1105 Thread * const Self = THREAD; 1106 assert(Self->is_Java_thread(), "Must be Java thread!"); 1107 JavaThread *jt = (JavaThread *)THREAD; 1108 1109 assert(InitDone, "Unexpectedly not initialized"); 1110 1111 if (THREAD != _owner) { 1112 if (THREAD->is_lock_owned ((address)_owner)) { 1113 assert(_recursions == 0, "internal state error"); 1114 _owner = THREAD; // Convert from basiclock addr to Thread addr 1115 _recursions = 0; 1116 } 1117 } 1118 1119 guarantee(Self == _owner, "complete_exit not owner"); 1120 intptr_t save = _recursions; // record the old recursion count 1121 _recursions = 0; // set the recursion level to be 0 1122 exit(true, Self); // exit the monitor 1123 guarantee(_owner != Self, "invariant"); 1124 return save; 1125 } 1126 1127 // reenter() enters a lock and sets recursion count 1128 // complete_exit/reenter operate as a wait without waiting 1129 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1130 Thread * const Self = THREAD; 1131 assert(Self->is_Java_thread(), "Must be Java thread!"); 1132 JavaThread *jt = (JavaThread *)THREAD; 1133 1134 guarantee(_owner != Self, "reenter already owner"); 1135 enter(THREAD); // enter the monitor 1136 guarantee(_recursions == 0, "reenter recursion"); 1137 _recursions = recursions; 1138 return; 1139 } 1140 1141 1142 // ----------------------------------------------------------------------------- 1143 // A macro is used below because there may already be a pending 1144 // exception which should not abort the execution of the routines 1145 // which use this (which is why we don't put this into check_slow and 1146 // call it with a CHECK argument). 1147 1148 #define CHECK_OWNER() \ 1149 do { \ 1150 if (THREAD != _owner) { \ 1151 if (THREAD->is_lock_owned((address) _owner)) { \ 1152 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1153 _recursions = 0; \ 1154 } else { \ 1155 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1156 } \ 1157 } \ 1158 } while (false) 1159 1160 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1161 // TODO-FIXME: remove check_slow() -- it's likely dead. 1162 1163 void ObjectMonitor::check_slow(TRAPS) { 1164 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1165 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1166 } 1167 1168 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1169 ObjectMonitor* monitor, 1170 jlong notifier_tid, 1171 jlong timeout, 1172 bool timedout) { 1173 assert(event != NULL, "invariant"); 1174 assert(monitor != NULL, "invariant"); 1175 event->set_monitorClass(((oop)monitor->object())->klass()); 1176 event->set_timeout(timeout); 1177 event->set_address((uintptr_t)monitor->object_addr()); 1178 event->set_notifier(notifier_tid); 1179 event->set_timedOut(timedout); 1180 event->commit(); 1181 } 1182 1183 // ----------------------------------------------------------------------------- 1184 // Wait/Notify/NotifyAll 1185 // 1186 // Note: a subset of changes to ObjectMonitor::wait() 1187 // will need to be replicated in complete_exit 1188 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1189 Thread * const Self = THREAD; 1190 assert(Self->is_Java_thread(), "Must be Java thread!"); 1191 JavaThread *jt = (JavaThread *)THREAD; 1192 1193 assert(InitDone, "Unexpectedly not initialized"); 1194 1195 // Throw IMSX or IEX. 1196 CHECK_OWNER(); 1197 1198 EventJavaMonitorWait event; 1199 1200 // check for a pending interrupt 1201 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1202 // post monitor waited event. Note that this is past-tense, we are done waiting. 1203 if (JvmtiExport::should_post_monitor_waited()) { 1204 // Note: 'false' parameter is passed here because the 1205 // wait was not timed out due to thread interrupt. 1206 JvmtiExport::post_monitor_waited(jt, this, false); 1207 1208 // In this short circuit of the monitor wait protocol, the 1209 // current thread never drops ownership of the monitor and 1210 // never gets added to the wait queue so the current thread 1211 // cannot be made the successor. This means that the 1212 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1213 // consume an unpark() meant for the ParkEvent associated with 1214 // this ObjectMonitor. 1215 } 1216 if (event.should_commit()) { 1217 post_monitor_wait_event(&event, this, 0, millis, false); 1218 } 1219 THROW(vmSymbols::java_lang_InterruptedException()); 1220 return; 1221 } 1222 1223 assert(Self->_Stalled == 0, "invariant"); 1224 Self->_Stalled = intptr_t(this); 1225 jt->set_current_waiting_monitor(this); 1226 1227 // create a node to be put into the queue 1228 // Critically, after we reset() the event but prior to park(), we must check 1229 // for a pending interrupt. 1230 ObjectWaiter node(Self); 1231 node.TState = ObjectWaiter::TS_WAIT; 1232 Self->_ParkEvent->reset(); 1233 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1234 1235 // Enter the waiting queue, which is a circular doubly linked list in this case 1236 // but it could be a priority queue or any data structure. 1237 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1238 // by the the owner of the monitor *except* in the case where park() 1239 // returns because of a timeout of interrupt. Contention is exceptionally rare 1240 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1241 1242 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1243 AddWaiter(&node); 1244 Thread::SpinRelease(&_WaitSetLock); 1245 1246 _Responsible = NULL; 1247 1248 intptr_t save = _recursions; // record the old recursion count 1249 _waiters++; // increment the number of waiters 1250 _recursions = 0; // set the recursion level to be 1 1251 exit(true, Self); // exit the monitor 1252 guarantee(_owner != Self, "invariant"); 1253 1254 // The thread is on the WaitSet list - now park() it. 1255 // On MP systems it's conceivable that a brief spin before we park 1256 // could be profitable. 1257 // 1258 // TODO-FIXME: change the following logic to a loop of the form 1259 // while (!timeout && !interrupted && _notified == 0) park() 1260 1261 int ret = OS_OK; 1262 int WasNotified = 0; 1263 { // State transition wrappers 1264 OSThread* osthread = Self->osthread(); 1265 OSThreadWaitState osts(osthread, true); 1266 { 1267 ThreadBlockInVM tbivm(jt); 1268 // Thread is in thread_blocked state and oop access is unsafe. 1269 jt->set_suspend_equivalent(); 1270 1271 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1272 // Intentionally empty 1273 } else if (node._notified == 0) { 1274 if (millis <= 0) { 1275 Self->_ParkEvent->park(); 1276 } else { 1277 ret = Self->_ParkEvent->park(millis); 1278 } 1279 } 1280 1281 // were we externally suspended while we were waiting? 1282 if (ExitSuspendEquivalent (jt)) { 1283 // TODO-FIXME: add -- if succ == Self then succ = null. 1284 jt->java_suspend_self(); 1285 } 1286 1287 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1288 1289 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1290 // from the WaitSet to the EntryList. 1291 // See if we need to remove Node from the WaitSet. 1292 // We use double-checked locking to avoid grabbing _WaitSetLock 1293 // if the thread is not on the wait queue. 1294 // 1295 // Note that we don't need a fence before the fetch of TState. 1296 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1297 // written by the is thread. (perhaps the fetch might even be satisfied 1298 // by a look-aside into the processor's own store buffer, although given 1299 // the length of the code path between the prior ST and this load that's 1300 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1301 // then we'll acquire the lock and then re-fetch a fresh TState value. 1302 // That is, we fail toward safety. 1303 1304 if (node.TState == ObjectWaiter::TS_WAIT) { 1305 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1306 if (node.TState == ObjectWaiter::TS_WAIT) { 1307 DequeueSpecificWaiter(&node); // unlink from WaitSet 1308 assert(node._notified == 0, "invariant"); 1309 node.TState = ObjectWaiter::TS_RUN; 1310 } 1311 Thread::SpinRelease(&_WaitSetLock); 1312 } 1313 1314 // The thread is now either on off-list (TS_RUN), 1315 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1316 // The Node's TState variable is stable from the perspective of this thread. 1317 // No other threads will asynchronously modify TState. 1318 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1319 OrderAccess::loadload(); 1320 if (_succ == Self) _succ = NULL; 1321 WasNotified = node._notified; 1322 1323 // Reentry phase -- reacquire the monitor. 1324 // re-enter contended monitor after object.wait(). 1325 // retain OBJECT_WAIT state until re-enter successfully completes 1326 // Thread state is thread_in_vm and oop access is again safe, 1327 // although the raw address of the object may have changed. 1328 // (Don't cache naked oops over safepoints, of course). 1329 1330 // post monitor waited event. Note that this is past-tense, we are done waiting. 1331 if (JvmtiExport::should_post_monitor_waited()) { 1332 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1333 1334 if (node._notified != 0 && _succ == Self) { 1335 // In this part of the monitor wait-notify-reenter protocol it 1336 // is possible (and normal) for another thread to do a fastpath 1337 // monitor enter-exit while this thread is still trying to get 1338 // to the reenter portion of the protocol. 1339 // 1340 // The ObjectMonitor was notified and the current thread is 1341 // the successor which also means that an unpark() has already 1342 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1343 // consume the unpark() that was done when the successor was 1344 // set because the same ParkEvent is shared between Java 1345 // monitors and JVM/TI RawMonitors (for now). 1346 // 1347 // We redo the unpark() to ensure forward progress, i.e., we 1348 // don't want all pending threads hanging (parked) with none 1349 // entering the unlocked monitor. 1350 node._event->unpark(); 1351 } 1352 } 1353 1354 if (event.should_commit()) { 1355 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1356 } 1357 1358 OrderAccess::fence(); 1359 1360 assert(Self->_Stalled != 0, "invariant"); 1361 Self->_Stalled = 0; 1362 1363 assert(_owner != Self, "invariant"); 1364 ObjectWaiter::TStates v = node.TState; 1365 if (v == ObjectWaiter::TS_RUN) { 1366 enter(Self); 1367 } else { 1368 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1369 ReenterI(Self, &node); 1370 node.wait_reenter_end(this); 1371 } 1372 1373 // Self has reacquired the lock. 1374 // Lifecycle - the node representing Self must not appear on any queues. 1375 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1376 // want residual elements associated with this thread left on any lists. 1377 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1378 assert(_owner == Self, "invariant"); 1379 assert(_succ != Self, "invariant"); 1380 } // OSThreadWaitState() 1381 1382 jt->set_current_waiting_monitor(NULL); 1383 1384 guarantee(_recursions == 0, "invariant"); 1385 _recursions = save; // restore the old recursion count 1386 _waiters--; // decrement the number of waiters 1387 1388 // Verify a few postconditions 1389 assert(_owner == Self, "invariant"); 1390 assert(_succ != Self, "invariant"); 1391 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1392 1393 // check if the notification happened 1394 if (!WasNotified) { 1395 // no, it could be timeout or Thread.interrupt() or both 1396 // check for interrupt event, otherwise it is timeout 1397 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1398 THROW(vmSymbols::java_lang_InterruptedException()); 1399 } 1400 } 1401 1402 // NOTE: Spurious wake up will be consider as timeout. 1403 // Monitor notify has precedence over thread interrupt. 1404 } 1405 1406 1407 // Consider: 1408 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1409 // then instead of transferring a thread from the WaitSet to the EntryList 1410 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1411 1412 void ObjectMonitor::INotify(Thread * Self) { 1413 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1414 ObjectWaiter * iterator = DequeueWaiter(); 1415 if (iterator != NULL) { 1416 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1417 guarantee(iterator->_notified == 0, "invariant"); 1418 // Disposition - what might we do with iterator ? 1419 // a. add it directly to the EntryList - either tail (policy == 1) 1420 // or head (policy == 0). 1421 // b. push it onto the front of the _cxq (policy == 2). 1422 // For now we use (b). 1423 1424 iterator->TState = ObjectWaiter::TS_ENTER; 1425 1426 iterator->_notified = 1; 1427 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1428 1429 ObjectWaiter * list = _EntryList; 1430 if (list != NULL) { 1431 assert(list->_prev == NULL, "invariant"); 1432 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1433 assert(list != iterator, "invariant"); 1434 } 1435 1436 // prepend to cxq 1437 if (list == NULL) { 1438 iterator->_next = iterator->_prev = NULL; 1439 _EntryList = iterator; 1440 } else { 1441 iterator->TState = ObjectWaiter::TS_CXQ; 1442 for (;;) { 1443 ObjectWaiter * front = _cxq; 1444 iterator->_next = front; 1445 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1446 break; 1447 } 1448 } 1449 } 1450 1451 // _WaitSetLock protects the wait queue, not the EntryList. We could 1452 // move the add-to-EntryList operation, above, outside the critical section 1453 // protected by _WaitSetLock. In practice that's not useful. With the 1454 // exception of wait() timeouts and interrupts the monitor owner 1455 // is the only thread that grabs _WaitSetLock. There's almost no contention 1456 // on _WaitSetLock so it's not profitable to reduce the length of the 1457 // critical section. 1458 1459 iterator->wait_reenter_begin(this); 1460 } 1461 Thread::SpinRelease(&_WaitSetLock); 1462 } 1463 1464 // Consider: a not-uncommon synchronization bug is to use notify() when 1465 // notifyAll() is more appropriate, potentially resulting in stranded 1466 // threads; this is one example of a lost wakeup. A useful diagnostic 1467 // option is to force all notify() operations to behave as notifyAll(). 1468 // 1469 // Note: We can also detect many such problems with a "minimum wait". 1470 // When the "minimum wait" is set to a small non-zero timeout value 1471 // and the program does not hang whereas it did absent "minimum wait", 1472 // that suggests a lost wakeup bug. 1473 1474 void ObjectMonitor::notify(TRAPS) { 1475 CHECK_OWNER(); 1476 if (_WaitSet == NULL) { 1477 return; 1478 } 1479 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1480 INotify(THREAD); 1481 OM_PERFDATA_OP(Notifications, inc(1)); 1482 } 1483 1484 1485 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1486 // from the waitset to the EntryList. This could be done more efficiently with a 1487 // single bulk transfer but in practice it's not time-critical. Beware too, 1488 // that in prepend-mode we invert the order of the waiters. Let's say that the 1489 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1490 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1491 1492 void ObjectMonitor::notifyAll(TRAPS) { 1493 CHECK_OWNER(); 1494 if (_WaitSet == NULL) { 1495 return; 1496 } 1497 1498 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1499 int tally = 0; 1500 while (_WaitSet != NULL) { 1501 tally++; 1502 INotify(THREAD); 1503 } 1504 1505 OM_PERFDATA_OP(Notifications, inc(tally)); 1506 } 1507 1508 // ----------------------------------------------------------------------------- 1509 // Adaptive Spinning Support 1510 // 1511 // Adaptive spin-then-block - rational spinning 1512 // 1513 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1514 // algorithm. On high order SMP systems it would be better to start with 1515 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1516 // a contending thread could enqueue itself on the cxq and then spin locally 1517 // on a thread-specific variable such as its ParkEvent._Event flag. 1518 // That's left as an exercise for the reader. Note that global spinning is 1519 // not problematic on Niagara, as the L2 cache serves the interconnect and 1520 // has both low latency and massive bandwidth. 1521 // 1522 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1523 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1524 // (duration) or we can fix the count at approximately the duration of 1525 // a context switch and vary the frequency. Of course we could also 1526 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1527 // For a description of 'Adaptive spin-then-block mutual exclusion in 1528 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1529 // 1530 // This implementation varies the duration "D", where D varies with 1531 // the success rate of recent spin attempts. (D is capped at approximately 1532 // length of a round-trip context switch). The success rate for recent 1533 // spin attempts is a good predictor of the success rate of future spin 1534 // attempts. The mechanism adapts automatically to varying critical 1535 // section length (lock modality), system load and degree of parallelism. 1536 // D is maintained per-monitor in _SpinDuration and is initialized 1537 // optimistically. Spin frequency is fixed at 100%. 1538 // 1539 // Note that _SpinDuration is volatile, but we update it without locks 1540 // or atomics. The code is designed so that _SpinDuration stays within 1541 // a reasonable range even in the presence of races. The arithmetic 1542 // operations on _SpinDuration are closed over the domain of legal values, 1543 // so at worst a race will install and older but still legal value. 1544 // At the very worst this introduces some apparent non-determinism. 1545 // We might spin when we shouldn't or vice-versa, but since the spin 1546 // count are relatively short, even in the worst case, the effect is harmless. 1547 // 1548 // Care must be taken that a low "D" value does not become an 1549 // an absorbing state. Transient spinning failures -- when spinning 1550 // is overall profitable -- should not cause the system to converge 1551 // on low "D" values. We want spinning to be stable and predictable 1552 // and fairly responsive to change and at the same time we don't want 1553 // it to oscillate, become metastable, be "too" non-deterministic, 1554 // or converge on or enter undesirable stable absorbing states. 1555 // 1556 // We implement a feedback-based control system -- using past behavior 1557 // to predict future behavior. We face two issues: (a) if the 1558 // input signal is random then the spin predictor won't provide optimal 1559 // results, and (b) if the signal frequency is too high then the control 1560 // system, which has some natural response lag, will "chase" the signal. 1561 // (b) can arise from multimodal lock hold times. Transient preemption 1562 // can also result in apparent bimodal lock hold times. 1563 // Although sub-optimal, neither condition is particularly harmful, as 1564 // in the worst-case we'll spin when we shouldn't or vice-versa. 1565 // The maximum spin duration is rather short so the failure modes aren't bad. 1566 // To be conservative, I've tuned the gain in system to bias toward 1567 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1568 // "rings" or oscillates between spinning and not spinning. This happens 1569 // when spinning is just on the cusp of profitability, however, so the 1570 // situation is not dire. The state is benign -- there's no need to add 1571 // hysteresis control to damp the transition rate between spinning and 1572 // not spinning. 1573 1574 // Spinning: Fixed frequency (100%), vary duration 1575 int ObjectMonitor::TrySpin(Thread * Self) { 1576 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1577 int ctr = Knob_FixedSpin; 1578 if (ctr != 0) { 1579 while (--ctr >= 0) { 1580 if (TryLock(Self) > 0) return 1; 1581 SpinPause(); 1582 } 1583 return 0; 1584 } 1585 1586 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1587 if (TryLock(Self) > 0) { 1588 // Increase _SpinDuration ... 1589 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1590 // Raising _SpurDuration to the poverty line is key. 1591 int x = _SpinDuration; 1592 if (x < Knob_SpinLimit) { 1593 if (x < Knob_Poverty) x = Knob_Poverty; 1594 _SpinDuration = x + Knob_BonusB; 1595 } 1596 return 1; 1597 } 1598 SpinPause(); 1599 } 1600 1601 // Admission control - verify preconditions for spinning 1602 // 1603 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1604 // becoming an absorbing state. Put another way, we spin briefly to 1605 // sample, just in case the system load, parallelism, contention, or lock 1606 // modality changed. 1607 // 1608 // Consider the following alternative: 1609 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1610 // spin attempt. "Periodically" might mean after a tally of 1611 // the # of failed spin attempts (or iterations) reaches some threshold. 1612 // This takes us into the realm of 1-out-of-N spinning, where we 1613 // hold the duration constant but vary the frequency. 1614 1615 ctr = _SpinDuration; 1616 if (ctr <= 0) return 0; 1617 1618 if (NotRunnable(Self, (Thread *) _owner)) { 1619 return 0; 1620 } 1621 1622 // We're good to spin ... spin ingress. 1623 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1624 // when preparing to LD...CAS _owner, etc and the CAS is likely 1625 // to succeed. 1626 if (_succ == NULL) { 1627 _succ = Self; 1628 } 1629 Thread * prv = NULL; 1630 1631 // There are three ways to exit the following loop: 1632 // 1. A successful spin where this thread has acquired the lock. 1633 // 2. Spin failure with prejudice 1634 // 3. Spin failure without prejudice 1635 1636 while (--ctr >= 0) { 1637 1638 // Periodic polling -- Check for pending GC 1639 // Threads may spin while they're unsafe. 1640 // We don't want spinning threads to delay the JVM from reaching 1641 // a stop-the-world safepoint or to steal cycles from GC. 1642 // If we detect a pending safepoint we abort in order that 1643 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1644 // this thread, if safe, doesn't steal cycles from GC. 1645 // This is in keeping with the "no loitering in runtime" rule. 1646 // We periodically check to see if there's a safepoint pending. 1647 if ((ctr & 0xFF) == 0) { 1648 if (SafepointMechanism::should_block(Self)) { 1649 goto Abort; // abrupt spin egress 1650 } 1651 SpinPause(); 1652 } 1653 1654 // Probe _owner with TATAS 1655 // If this thread observes the monitor transition or flicker 1656 // from locked to unlocked to locked, then the odds that this 1657 // thread will acquire the lock in this spin attempt go down 1658 // considerably. The same argument applies if the CAS fails 1659 // or if we observe _owner change from one non-null value to 1660 // another non-null value. In such cases we might abort 1661 // the spin without prejudice or apply a "penalty" to the 1662 // spin count-down variable "ctr", reducing it by 100, say. 1663 1664 Thread * ox = (Thread *) _owner; 1665 if (ox == NULL) { 1666 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1667 if (ox == NULL) { 1668 // The CAS succeeded -- this thread acquired ownership 1669 // Take care of some bookkeeping to exit spin state. 1670 if (_succ == Self) { 1671 _succ = NULL; 1672 } 1673 1674 // Increase _SpinDuration : 1675 // The spin was successful (profitable) so we tend toward 1676 // longer spin attempts in the future. 1677 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1678 // If we acquired the lock early in the spin cycle it 1679 // makes sense to increase _SpinDuration proportionally. 1680 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1681 int x = _SpinDuration; 1682 if (x < Knob_SpinLimit) { 1683 if (x < Knob_Poverty) x = Knob_Poverty; 1684 _SpinDuration = x + Knob_Bonus; 1685 } 1686 return 1; 1687 } 1688 1689 // The CAS failed ... we can take any of the following actions: 1690 // * penalize: ctr -= CASPenalty 1691 // * exit spin with prejudice -- goto Abort; 1692 // * exit spin without prejudice. 1693 // * Since CAS is high-latency, retry again immediately. 1694 prv = ox; 1695 goto Abort; 1696 } 1697 1698 // Did lock ownership change hands ? 1699 if (ox != prv && prv != NULL) { 1700 goto Abort; 1701 } 1702 prv = ox; 1703 1704 // Abort the spin if the owner is not executing. 1705 // The owner must be executing in order to drop the lock. 1706 // Spinning while the owner is OFFPROC is idiocy. 1707 // Consider: ctr -= RunnablePenalty ; 1708 if (NotRunnable(Self, ox)) { 1709 goto Abort; 1710 } 1711 if (_succ == NULL) { 1712 _succ = Self; 1713 } 1714 } 1715 1716 // Spin failed with prejudice -- reduce _SpinDuration. 1717 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1718 // AIMD is globally stable. 1719 { 1720 int x = _SpinDuration; 1721 if (x > 0) { 1722 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1723 // This is globally sample and tends to damp the response. 1724 x -= Knob_Penalty; 1725 if (x < 0) x = 0; 1726 _SpinDuration = x; 1727 } 1728 } 1729 1730 Abort: 1731 if (_succ == Self) { 1732 _succ = NULL; 1733 // Invariant: after setting succ=null a contending thread 1734 // must recheck-retry _owner before parking. This usually happens 1735 // in the normal usage of TrySpin(), but it's safest 1736 // to make TrySpin() as foolproof as possible. 1737 OrderAccess::fence(); 1738 if (TryLock(Self) > 0) return 1; 1739 } 1740 return 0; 1741 } 1742 1743 // NotRunnable() -- informed spinning 1744 // 1745 // Don't bother spinning if the owner is not eligible to drop the lock. 1746 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1747 // The thread must be runnable in order to drop the lock in timely fashion. 1748 // If the _owner is not runnable then spinning will not likely be 1749 // successful (profitable). 1750 // 1751 // Beware -- the thread referenced by _owner could have died 1752 // so a simply fetch from _owner->_thread_state might trap. 1753 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1754 // Because of the lifecycle issues, the _thread_state values 1755 // observed by NotRunnable() might be garbage. NotRunnable must 1756 // tolerate this and consider the observed _thread_state value 1757 // as advisory. 1758 // 1759 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1760 // a thread pointer. 1761 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1762 // with the LSB of _owner. Another option would be to probabilistically probe 1763 // the putative _owner->TypeTag value. 1764 // 1765 // Checking _thread_state isn't perfect. Even if the thread is 1766 // in_java it might be blocked on a page-fault or have been preempted 1767 // and sitting on a ready/dispatch queue. 1768 // 1769 // The return value from NotRunnable() is *advisory* -- the 1770 // result is based on sampling and is not necessarily coherent. 1771 // The caller must tolerate false-negative and false-positive errors. 1772 // Spinning, in general, is probabilistic anyway. 1773 1774 1775 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1776 // Check ox->TypeTag == 2BAD. 1777 if (ox == NULL) return 0; 1778 1779 // Avoid transitive spinning ... 1780 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1781 // Immediately after T1 acquires L it's possible that T2, also 1782 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1783 // This occurs transiently after T1 acquired L but before 1784 // T1 managed to clear T1.Stalled. T2 does not need to abort 1785 // its spin in this circumstance. 1786 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1787 1788 if (BlockedOn == 1) return 1; 1789 if (BlockedOn != 0) { 1790 return BlockedOn != intptr_t(this) && _owner == ox; 1791 } 1792 1793 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1794 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1795 // consider also: jst != _thread_in_Java -- but that's overspecific. 1796 return jst == _thread_blocked || jst == _thread_in_native; 1797 } 1798 1799 1800 // ----------------------------------------------------------------------------- 1801 // WaitSet management ... 1802 1803 ObjectWaiter::ObjectWaiter(Thread* thread) { 1804 _next = NULL; 1805 _prev = NULL; 1806 _notified = 0; 1807 _notifier_tid = 0; 1808 TState = TS_RUN; 1809 _thread = thread; 1810 _event = thread->_ParkEvent; 1811 _active = false; 1812 assert(_event != NULL, "invariant"); 1813 } 1814 1815 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1816 JavaThread *jt = (JavaThread *)this->_thread; 1817 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1818 } 1819 1820 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1821 JavaThread *jt = (JavaThread *)this->_thread; 1822 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1823 } 1824 1825 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1826 assert(node != NULL, "should not add NULL node"); 1827 assert(node->_prev == NULL, "node already in list"); 1828 assert(node->_next == NULL, "node already in list"); 1829 // put node at end of queue (circular doubly linked list) 1830 if (_WaitSet == NULL) { 1831 _WaitSet = node; 1832 node->_prev = node; 1833 node->_next = node; 1834 } else { 1835 ObjectWaiter* head = _WaitSet; 1836 ObjectWaiter* tail = head->_prev; 1837 assert(tail->_next == head, "invariant check"); 1838 tail->_next = node; 1839 head->_prev = node; 1840 node->_next = head; 1841 node->_prev = tail; 1842 } 1843 } 1844 1845 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1846 // dequeue the very first waiter 1847 ObjectWaiter* waiter = _WaitSet; 1848 if (waiter) { 1849 DequeueSpecificWaiter(waiter); 1850 } 1851 return waiter; 1852 } 1853 1854 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1855 assert(node != NULL, "should not dequeue NULL node"); 1856 assert(node->_prev != NULL, "node already removed from list"); 1857 assert(node->_next != NULL, "node already removed from list"); 1858 // when the waiter has woken up because of interrupt, 1859 // timeout or other spurious wake-up, dequeue the 1860 // waiter from waiting list 1861 ObjectWaiter* next = node->_next; 1862 if (next == node) { 1863 assert(node->_prev == node, "invariant check"); 1864 _WaitSet = NULL; 1865 } else { 1866 ObjectWaiter* prev = node->_prev; 1867 assert(prev->_next == node, "invariant check"); 1868 assert(next->_prev == node, "invariant check"); 1869 next->_prev = prev; 1870 prev->_next = next; 1871 if (_WaitSet == node) { 1872 _WaitSet = next; 1873 } 1874 } 1875 node->_next = NULL; 1876 node->_prev = NULL; 1877 } 1878 1879 // ----------------------------------------------------------------------------- 1880 // PerfData support 1881 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 1882 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 1883 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 1884 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 1885 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 1886 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 1887 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 1888 1889 // One-shot global initialization for the sync subsystem. 1890 // We could also defer initialization and initialize on-demand 1891 // the first time we call ObjectSynchronizer::inflate(). 1892 // Initialization would be protected - like so many things - by 1893 // the MonitorCache_lock. 1894 1895 void ObjectMonitor::Initialize() { 1896 assert(!InitDone, "invariant"); 1897 1898 if (!os::is_MP()) { 1899 Knob_SpinLimit = 0; 1900 Knob_PreSpin = 0; 1901 Knob_FixedSpin = -1; 1902 } 1903 1904 if (UsePerfData) { 1905 EXCEPTION_MARK; 1906 #define NEWPERFCOUNTER(n) \ 1907 { \ 1908 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 1909 CHECK); \ 1910 } 1911 #define NEWPERFVARIABLE(n) \ 1912 { \ 1913 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 1914 CHECK); \ 1915 } 1916 NEWPERFCOUNTER(_sync_Inflations); 1917 NEWPERFCOUNTER(_sync_Deflations); 1918 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 1919 NEWPERFCOUNTER(_sync_FutileWakeups); 1920 NEWPERFCOUNTER(_sync_Parks); 1921 NEWPERFCOUNTER(_sync_Notifications); 1922 NEWPERFVARIABLE(_sync_MonExtant); 1923 #undef NEWPERFCOUNTER 1924 #undef NEWPERFVARIABLE 1925 } 1926 1927 DEBUG_ONLY(InitDone = true;) 1928 }