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 "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "oops/markWord.hpp" 34 #include "oops/oop.inline.hpp" 35 #include "runtime/atomic.hpp" 36 #include "runtime/handles.inline.hpp" 37 #include "runtime/interfaceSupport.inline.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "runtime/objectMonitor.hpp" 40 #include "runtime/objectMonitor.inline.hpp" 41 #include "runtime/orderAccess.hpp" 42 #include "runtime/osThread.hpp" 43 #include "runtime/safepointMechanism.inline.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/stubRoutines.hpp" 46 #include "runtime/thread.inline.hpp" 47 #include "services/threadService.hpp" 48 #include "utilities/dtrace.hpp" 49 #include "utilities/macros.hpp" 50 #include "utilities/preserveException.hpp" 51 #if INCLUDE_JFR 52 #include "jfr/support/jfrFlush.hpp" 53 #endif 54 55 #ifdef DTRACE_ENABLED 56 57 // Only bother with this argument setup if dtrace is available 58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95 #else // ndef DTRACE_ENABLED 96 97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100 #endif // ndef DTRACE_ENABLED 101 102 // Tunables ... 103 // The knob* variables are effectively final. Once set they should 104 // never be modified hence. Consider using __read_mostly with GCC. 105 106 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 107 108 static int Knob_Bonus = 100; // spin success bonus 109 static int Knob_BonusB = 100; // spin success bonus 110 static int Knob_Penalty = 200; // spin failure penalty 111 static int Knob_Poverty = 1000; 112 static int Knob_FixedSpin = 0; 113 static int Knob_PreSpin = 10; // 20-100 likely better 114 115 DEBUG_ONLY(static volatile bool InitDone = false;) 116 117 // ----------------------------------------------------------------------------- 118 // Theory of operations -- Monitors lists, thread residency, etc: 119 // 120 // * A thread acquires ownership of a monitor by successfully 121 // CAS()ing the _owner field from null to non-null. 122 // 123 // * Invariant: A thread appears on at most one monitor list -- 124 // cxq, EntryList or WaitSet -- at any one time. 125 // 126 // * Contending threads "push" themselves onto the cxq with CAS 127 // and then spin/park. 128 // 129 // * After a contending thread eventually acquires the lock it must 130 // dequeue itself from either the EntryList or the cxq. 131 // 132 // * The exiting thread identifies and unparks an "heir presumptive" 133 // tentative successor thread on the EntryList. Critically, the 134 // exiting thread doesn't unlink the successor thread from the EntryList. 135 // After having been unparked, the wakee will recontend for ownership of 136 // the monitor. The successor (wakee) will either acquire the lock or 137 // re-park itself. 138 // 139 // Succession is provided for by a policy of competitive handoff. 140 // The exiting thread does _not_ grant or pass ownership to the 141 // successor thread. (This is also referred to as "handoff" succession"). 142 // Instead the exiting thread releases ownership and possibly wakes 143 // a successor, so the successor can (re)compete for ownership of the lock. 144 // If the EntryList is empty but the cxq is populated the exiting 145 // thread will drain the cxq into the EntryList. It does so by 146 // by detaching the cxq (installing null with CAS) and folding 147 // the threads from the cxq into the EntryList. The EntryList is 148 // doubly linked, while the cxq is singly linked because of the 149 // CAS-based "push" used to enqueue recently arrived threads (RATs). 150 // 151 // * Concurrency invariants: 152 // 153 // -- only the monitor owner may access or mutate the EntryList. 154 // The mutex property of the monitor itself protects the EntryList 155 // from concurrent interference. 156 // -- Only the monitor owner may detach the cxq. 157 // 158 // * The monitor entry list operations avoid locks, but strictly speaking 159 // they're not lock-free. Enter is lock-free, exit is not. 160 // For a description of 'Methods and apparatus providing non-blocking access 161 // to a resource,' see U.S. Pat. No. 7844973. 162 // 163 // * The cxq can have multiple concurrent "pushers" but only one concurrent 164 // detaching thread. This mechanism is immune from the ABA corruption. 165 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 166 // 167 // * Taken together, the cxq and the EntryList constitute or form a 168 // single logical queue of threads stalled trying to acquire the lock. 169 // We use two distinct lists to improve the odds of a constant-time 170 // dequeue operation after acquisition (in the ::enter() epilogue) and 171 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 172 // A key desideratum is to minimize queue & monitor metadata manipulation 173 // that occurs while holding the monitor lock -- that is, we want to 174 // minimize monitor lock holds times. Note that even a small amount of 175 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 176 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 177 // locks and monitor metadata. 178 // 179 // Cxq points to the set of Recently Arrived Threads attempting entry. 180 // Because we push threads onto _cxq with CAS, the RATs must take the form of 181 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 182 // the unlocking thread notices that EntryList is null but _cxq is != null. 183 // 184 // The EntryList is ordered by the prevailing queue discipline and 185 // can be organized in any convenient fashion, such as a doubly-linked list or 186 // a circular doubly-linked list. Critically, we want insert and delete operations 187 // to operate in constant-time. If we need a priority queue then something akin 188 // to Solaris' sleepq would work nicely. Viz., 189 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 190 // Queue discipline is enforced at ::exit() time, when the unlocking thread 191 // drains the cxq into the EntryList, and orders or reorders the threads on the 192 // EntryList accordingly. 193 // 194 // Barring "lock barging", this mechanism provides fair cyclic ordering, 195 // somewhat similar to an elevator-scan. 196 // 197 // * The monitor synchronization subsystem avoids the use of native 198 // synchronization primitives except for the narrow platform-specific 199 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 200 // the semantics of park-unpark. Put another way, this monitor implementation 201 // depends only on atomic operations and park-unpark. The monitor subsystem 202 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 203 // underlying OS manages the READY<->RUN transitions. 204 // 205 // * Waiting threads reside on the WaitSet list -- wait() puts 206 // the caller onto the WaitSet. 207 // 208 // * notify() or notifyAll() simply transfers threads from the WaitSet to 209 // either the EntryList or cxq. Subsequent exit() operations will 210 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 211 // it's likely the notifyee would simply impale itself on the lock held 212 // by the notifier. 213 // 214 // * An interesting alternative is to encode cxq as (List,LockByte) where 215 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 216 // variable, like _recursions, in the scheme. The threads or Events that form 217 // the list would have to be aligned in 256-byte addresses. A thread would 218 // try to acquire the lock or enqueue itself with CAS, but exiting threads 219 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 220 // Note that is is *not* word-tearing, but it does presume that full-word 221 // CAS operations are coherent with intermix with STB operations. That's true 222 // on most common processors. 223 // 224 // * See also http://blogs.sun.com/dave 225 226 227 void* ObjectMonitor::operator new (size_t size) throw() { 228 return AllocateHeap(size, mtInternal); 229 } 230 void* ObjectMonitor::operator new[] (size_t size) throw() { 231 return operator new (size); 232 } 233 void ObjectMonitor::operator delete(void* p) { 234 FreeHeap(p); 235 } 236 void ObjectMonitor::operator delete[] (void *p) { 237 operator delete(p); 238 } 239 240 // ----------------------------------------------------------------------------- 241 // Enter support 242 243 void ObjectMonitor::enter(TRAPS) { 244 // The following code is ordered to check the most common cases first 245 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 246 Thread * const Self = THREAD; 247 248 void * cur = Atomic::cmpxchg(&_owner, (void*)NULL, Self); 249 if (cur == NULL) { 250 assert(_recursions == 0, "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=" INTX_FORMAT, _recursions); 281 assert(((oop)object())->mark() == markWord::encode(this), 282 "object mark must match encoded this: mark=" INTPTR_FORMAT 283 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 284 markWord::encode(this).value()); 285 Self->_Stalled = 0; 286 return; 287 } 288 289 assert(_owner != Self, "invariant"); 290 assert(_succ != Self, "invariant"); 291 assert(Self->is_Java_thread(), "invariant"); 292 JavaThread * jt = (JavaThread *) Self; 293 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 294 assert(jt->thread_state() != _thread_blocked, "invariant"); 295 assert(this->object() != NULL, "invariant"); 296 assert(_contentions >= 0, "invariant"); 297 298 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 299 // Ensure the object-monitor relationship remains stable while there's contention. 300 Atomic::inc(&_contentions); 301 302 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 303 EventJavaMonitorEnter event; 304 if (event.should_commit()) { 305 event.set_monitorClass(((oop)this->object())->klass()); 306 event.set_address((uintptr_t)(this->object_addr())); 307 } 308 309 { // Change java thread status to indicate blocked on monitor enter. 310 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 311 312 Self->set_current_pending_monitor(this); 313 314 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 315 if (JvmtiExport::should_post_monitor_contended_enter()) { 316 JvmtiExport::post_monitor_contended_enter(jt, this); 317 318 // The current thread does not yet own the monitor and does not 319 // yet appear on any queues that would get it made the successor. 320 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 321 // handler cannot accidentally consume an unpark() meant for the 322 // ParkEvent associated with this ObjectMonitor. 323 } 324 325 OSThreadContendState osts(Self->osthread()); 326 ThreadBlockInVM tbivm(jt); 327 328 // TODO-FIXME: change the following for(;;) loop to straight-line code. 329 for (;;) { 330 jt->set_suspend_equivalent(); 331 // cleared by handle_special_suspend_equivalent_condition() 332 // or java_suspend_self() 333 334 EnterI(THREAD); 335 336 if (!ExitSuspendEquivalent(jt)) break; 337 338 // We have acquired the contended monitor, but while we were 339 // waiting another thread suspended us. We don't want to enter 340 // the monitor while suspended because that would surprise the 341 // thread that suspended us. 342 // 343 _recursions = 0; 344 _succ = NULL; 345 exit(false, Self); 346 347 jt->java_suspend_self(); 348 } 349 Self->set_current_pending_monitor(NULL); 350 351 // We cleared the pending monitor info since we've just gotten past 352 // the enter-check-for-suspend dance and we now own the monitor free 353 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 354 // destructor can go to a safepoint at the end of this block. If we 355 // do a thread dump during that safepoint, then this thread will show 356 // as having "-locked" the monitor, but the OS and java.lang.Thread 357 // states will still report that the thread is blocked trying to 358 // acquire it. 359 } 360 361 Atomic::dec(&_contentions); 362 assert(_contentions >= 0, "invariant"); 363 Self->_Stalled = 0; 364 365 // Must either set _recursions = 0 or ASSERT _recursions == 0. 366 assert(_recursions == 0, "invariant"); 367 assert(_owner == Self, "invariant"); 368 assert(_succ != Self, "invariant"); 369 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 370 371 // The thread -- now the owner -- is back in vm mode. 372 // Report the glorious news via TI,DTrace and jvmstat. 373 // The probe effect is non-trivial. All the reportage occurs 374 // while we hold the monitor, increasing the length of the critical 375 // section. Amdahl's parallel speedup law comes vividly into play. 376 // 377 // Another option might be to aggregate the events (thread local or 378 // per-monitor aggregation) and defer reporting until a more opportune 379 // time -- such as next time some thread encounters contention but has 380 // yet to acquire the lock. While spinning that thread could 381 // spinning we could increment JVMStat counters, etc. 382 383 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 384 if (JvmtiExport::should_post_monitor_contended_entered()) { 385 JvmtiExport::post_monitor_contended_entered(jt, this); 386 387 // The current thread already owns the monitor and is not going to 388 // call park() for the remainder of the monitor enter protocol. So 389 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 390 // event handler consumed an unpark() issued by the thread that 391 // just exited the monitor. 392 } 393 if (event.should_commit()) { 394 event.set_previousOwner((uintptr_t)_previous_owner_tid); 395 event.commit(); 396 } 397 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 398 } 399 400 // Caveat: TryLock() is not necessarily serializing if it returns failure. 401 // Callers must compensate as needed. 402 403 int ObjectMonitor::TryLock(Thread * Self) { 404 void * own = _owner; 405 if (own != NULL) return 0; 406 if (Atomic::replace_if_null(&_owner, Self)) { 407 assert(_recursions == 0, "invariant"); 408 return 1; 409 } 410 // The lock had been free momentarily, but we lost the race to the lock. 411 // Interference -- the CAS failed. 412 // We can either return -1 or retry. 413 // Retry doesn't make as much sense because the lock was just acquired. 414 return -1; 415 } 416 417 // Convert the fields used by is_busy() to a string that can be 418 // used for diagnostic output. 419 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 420 ss->print("is_busy: contentions=%d, waiters=%d, owner=" INTPTR_FORMAT 421 ", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, _contentions, 422 _waiters, p2i(_owner), p2i(_cxq), p2i(_EntryList)); 423 return ss->base(); 424 } 425 426 #define MAX_RECHECK_INTERVAL 1000 427 428 void ObjectMonitor::EnterI(TRAPS) { 429 Thread * const Self = THREAD; 430 assert(Self->is_Java_thread(), "invariant"); 431 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 432 433 // Try the lock - TATAS 434 if (TryLock (Self) > 0) { 435 assert(_succ != Self, "invariant"); 436 assert(_owner == Self, "invariant"); 437 assert(_Responsible != Self, "invariant"); 438 return; 439 } 440 441 assert(InitDone, "Unexpectedly not initialized"); 442 443 // We try one round of spinning *before* enqueueing Self. 444 // 445 // If the _owner is ready but OFFPROC we could use a YieldTo() 446 // operation to donate the remainder of this thread's quantum 447 // to the owner. This has subtle but beneficial affinity 448 // effects. 449 450 if (TrySpin(Self) > 0) { 451 assert(_owner == Self, "invariant"); 452 assert(_succ != Self, "invariant"); 453 assert(_Responsible != Self, "invariant"); 454 return; 455 } 456 457 // The Spin failed -- Enqueue and park the thread ... 458 assert(_succ != Self, "invariant"); 459 assert(_owner != Self, "invariant"); 460 assert(_Responsible != Self, "invariant"); 461 462 // Enqueue "Self" on ObjectMonitor's _cxq. 463 // 464 // Node acts as a proxy for Self. 465 // As an aside, if were to ever rewrite the synchronization code mostly 466 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 467 // Java objects. This would avoid awkward lifecycle and liveness issues, 468 // as well as eliminate a subset of ABA issues. 469 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 470 471 ObjectWaiter node(Self); 472 Self->_ParkEvent->reset(); 473 node._prev = (ObjectWaiter *) 0xBAD; 474 node.TState = ObjectWaiter::TS_CXQ; 475 476 // Push "Self" onto the front of the _cxq. 477 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 478 // Note that spinning tends to reduce the rate at which threads 479 // enqueue and dequeue on EntryList|cxq. 480 ObjectWaiter * nxt; 481 for (;;) { 482 node._next = nxt = _cxq; 483 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 484 485 // Interference - the CAS failed because _cxq changed. Just retry. 486 // As an optional optimization we retry the lock. 487 if (TryLock (Self) > 0) { 488 assert(_succ != Self, "invariant"); 489 assert(_owner == Self, "invariant"); 490 assert(_Responsible != Self, "invariant"); 491 return; 492 } 493 } 494 495 // Check for cxq|EntryList edge transition to non-null. This indicates 496 // the onset of contention. While contention persists exiting threads 497 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 498 // operations revert to the faster 1-0 mode. This enter operation may interleave 499 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 500 // arrange for one of the contending thread to use a timed park() operations 501 // to detect and recover from the race. (Stranding is form of progress failure 502 // where the monitor is unlocked but all the contending threads remain parked). 503 // That is, at least one of the contended threads will periodically poll _owner. 504 // One of the contending threads will become the designated "Responsible" thread. 505 // The Responsible thread uses a timed park instead of a normal indefinite park 506 // operation -- it periodically wakes and checks for and recovers from potential 507 // strandings admitted by 1-0 exit operations. We need at most one Responsible 508 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 509 // be responsible for a monitor. 510 // 511 // Currently, one of the contended threads takes on the added role of "Responsible". 512 // A viable alternative would be to use a dedicated "stranding checker" thread 513 // that periodically iterated over all the threads (or active monitors) and unparked 514 // successors where there was risk of stranding. This would help eliminate the 515 // timer scalability issues we see on some platforms as we'd only have one thread 516 // -- the checker -- parked on a timer. 517 518 if (nxt == NULL && _EntryList == NULL) { 519 // Try to assume the role of responsible thread for the monitor. 520 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 521 Atomic::replace_if_null(&_Responsible, Self); 522 } 523 524 // The lock might have been released while this thread was occupied queueing 525 // itself onto _cxq. To close the race and avoid "stranding" and 526 // progress-liveness failure we must resample-retry _owner before parking. 527 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 528 // In this case the ST-MEMBAR is accomplished with CAS(). 529 // 530 // TODO: Defer all thread state transitions until park-time. 531 // Since state transitions are heavy and inefficient we'd like 532 // to defer the state transitions until absolutely necessary, 533 // and in doing so avoid some transitions ... 534 535 int nWakeups = 0; 536 int recheckInterval = 1; 537 538 for (;;) { 539 540 if (TryLock(Self) > 0) break; 541 assert(_owner != Self, "invariant"); 542 543 // park self 544 if (_Responsible == Self) { 545 Self->_ParkEvent->park((jlong) recheckInterval); 546 // Increase the recheckInterval, but clamp the value. 547 recheckInterval *= 8; 548 if (recheckInterval > MAX_RECHECK_INTERVAL) { 549 recheckInterval = MAX_RECHECK_INTERVAL; 550 } 551 } else { 552 Self->_ParkEvent->park(); 553 } 554 555 if (TryLock(Self) > 0) break; 556 557 // The lock is still contested. 558 // Keep a tally of the # of futile wakeups. 559 // Note that the counter is not protected by a lock or updated by atomics. 560 // That is by design - we trade "lossy" counters which are exposed to 561 // races during updates for a lower probe effect. 562 563 // This PerfData object can be used in parallel with a safepoint. 564 // See the work around in PerfDataManager::destroy(). 565 OM_PERFDATA_OP(FutileWakeups, inc()); 566 ++nWakeups; 567 568 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 569 // We can defer clearing _succ until after the spin completes 570 // TrySpin() must tolerate being called with _succ == Self. 571 // Try yet another round of adaptive spinning. 572 if (TrySpin(Self) > 0) break; 573 574 // We can find that we were unpark()ed and redesignated _succ while 575 // we were spinning. That's harmless. If we iterate and call park(), 576 // park() will consume the event and return immediately and we'll 577 // just spin again. This pattern can repeat, leaving _succ to simply 578 // spin on a CPU. 579 580 if (_succ == Self) _succ = NULL; 581 582 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 583 OrderAccess::fence(); 584 } 585 586 // Egress : 587 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 588 // Normally we'll find Self on the EntryList . 589 // From the perspective of the lock owner (this thread), the 590 // EntryList is stable and cxq is prepend-only. 591 // The head of cxq is volatile but the interior is stable. 592 // In addition, Self.TState is stable. 593 594 assert(_owner == Self, "invariant"); 595 assert(object() != NULL, "invariant"); 596 // I'd like to write: 597 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 598 // but as we're at a safepoint that's not safe. 599 600 UnlinkAfterAcquire(Self, &node); 601 if (_succ == Self) _succ = NULL; 602 603 assert(_succ != Self, "invariant"); 604 if (_Responsible == Self) { 605 _Responsible = NULL; 606 OrderAccess::fence(); // Dekker pivot-point 607 608 // We may leave threads on cxq|EntryList without a designated 609 // "Responsible" thread. This is benign. When this thread subsequently 610 // exits the monitor it can "see" such preexisting "old" threads -- 611 // threads that arrived on the cxq|EntryList before the fence, above -- 612 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 613 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 614 // non-null and elect a new "Responsible" timer thread. 615 // 616 // This thread executes: 617 // ST Responsible=null; MEMBAR (in enter epilogue - here) 618 // LD cxq|EntryList (in subsequent exit) 619 // 620 // Entering threads in the slow/contended path execute: 621 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 622 // The (ST cxq; MEMBAR) is accomplished with CAS(). 623 // 624 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 625 // exit operation from floating above the ST Responsible=null. 626 } 627 628 // We've acquired ownership with CAS(). 629 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 630 // But since the CAS() this thread may have also stored into _succ, 631 // EntryList, cxq or Responsible. These meta-data updates must be 632 // visible __before this thread subsequently drops the lock. 633 // Consider what could occur if we didn't enforce this constraint -- 634 // STs to monitor meta-data and user-data could reorder with (become 635 // visible after) the ST in exit that drops ownership of the lock. 636 // Some other thread could then acquire the lock, but observe inconsistent 637 // or old monitor meta-data and heap data. That violates the JMM. 638 // To that end, the 1-0 exit() operation must have at least STST|LDST 639 // "release" barrier semantics. Specifically, there must be at least a 640 // STST|LDST barrier in exit() before the ST of null into _owner that drops 641 // the lock. The barrier ensures that changes to monitor meta-data and data 642 // protected by the lock will be visible before we release the lock, and 643 // therefore before some other thread (CPU) has a chance to acquire the lock. 644 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 645 // 646 // Critically, any prior STs to _succ or EntryList must be visible before 647 // the ST of null into _owner in the *subsequent* (following) corresponding 648 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 649 // execute a serializing instruction. 650 651 return; 652 } 653 654 // ReenterI() is a specialized inline form of the latter half of the 655 // contended slow-path from EnterI(). We use ReenterI() only for 656 // monitor reentry in wait(). 657 // 658 // In the future we should reconcile EnterI() and ReenterI(). 659 660 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 661 assert(Self != NULL, "invariant"); 662 assert(SelfNode != NULL, "invariant"); 663 assert(SelfNode->_thread == Self, "invariant"); 664 assert(_waiters > 0, "invariant"); 665 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 666 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 667 JavaThread * jt = (JavaThread *) Self; 668 669 int nWakeups = 0; 670 for (;;) { 671 ObjectWaiter::TStates v = SelfNode->TState; 672 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 673 assert(_owner != Self, "invariant"); 674 675 if (TryLock(Self) > 0) break; 676 if (TrySpin(Self) > 0) break; 677 678 // State transition wrappers around park() ... 679 // ReenterI() wisely defers state transitions until 680 // it's clear we must park the thread. 681 { 682 OSThreadContendState osts(Self->osthread()); 683 ThreadBlockInVM tbivm(jt); 684 685 // cleared by handle_special_suspend_equivalent_condition() 686 // or java_suspend_self() 687 jt->set_suspend_equivalent(); 688 Self->_ParkEvent->park(); 689 690 // were we externally suspended while we were waiting? 691 for (;;) { 692 if (!ExitSuspendEquivalent(jt)) break; 693 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 694 jt->java_suspend_self(); 695 jt->set_suspend_equivalent(); 696 } 697 } 698 699 // Try again, but just so we distinguish between futile wakeups and 700 // successful wakeups. The following test isn't algorithmically 701 // necessary, but it helps us maintain sensible statistics. 702 if (TryLock(Self) > 0) break; 703 704 // The lock is still contested. 705 // Keep a tally of the # of futile wakeups. 706 // Note that the counter is not protected by a lock or updated by atomics. 707 // That is by design - we trade "lossy" counters which are exposed to 708 // races during updates for a lower probe effect. 709 ++nWakeups; 710 711 // Assuming this is not a spurious wakeup we'll normally 712 // find that _succ == Self. 713 if (_succ == Self) _succ = NULL; 714 715 // Invariant: after clearing _succ a contending thread 716 // *must* retry _owner before parking. 717 OrderAccess::fence(); 718 719 // This PerfData object can be used in parallel with a safepoint. 720 // See the work around in PerfDataManager::destroy(). 721 OM_PERFDATA_OP(FutileWakeups, inc()); 722 } 723 724 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 725 // Normally we'll find Self on the EntryList. 726 // Unlinking from the EntryList is constant-time and atomic-free. 727 // From the perspective of the lock owner (this thread), the 728 // EntryList is stable and cxq is prepend-only. 729 // The head of cxq is volatile but the interior is stable. 730 // In addition, Self.TState is stable. 731 732 assert(_owner == Self, "invariant"); 733 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 734 UnlinkAfterAcquire(Self, SelfNode); 735 if (_succ == Self) _succ = NULL; 736 assert(_succ != Self, "invariant"); 737 SelfNode->TState = ObjectWaiter::TS_RUN; 738 OrderAccess::fence(); // see comments at the end of EnterI() 739 } 740 741 // By convention we unlink a contending thread from EntryList|cxq immediately 742 // after the thread acquires the lock in ::enter(). Equally, we could defer 743 // unlinking the thread until ::exit()-time. 744 745 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 746 assert(_owner == Self, "invariant"); 747 assert(SelfNode->_thread == Self, "invariant"); 748 749 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 750 // Normal case: remove Self from the DLL EntryList . 751 // This is a constant-time operation. 752 ObjectWaiter * nxt = SelfNode->_next; 753 ObjectWaiter * prv = SelfNode->_prev; 754 if (nxt != NULL) nxt->_prev = prv; 755 if (prv != NULL) prv->_next = nxt; 756 if (SelfNode == _EntryList) _EntryList = nxt; 757 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 758 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 759 } else { 760 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 761 // Inopportune interleaving -- Self is still on the cxq. 762 // This usually means the enqueue of self raced an exiting thread. 763 // Normally we'll find Self near the front of the cxq, so 764 // dequeueing is typically fast. If needbe we can accelerate 765 // this with some MCS/CHL-like bidirectional list hints and advisory 766 // back-links so dequeueing from the interior will normally operate 767 // in constant-time. 768 // Dequeue Self from either the head (with CAS) or from the interior 769 // with a linear-time scan and normal non-atomic memory operations. 770 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 771 // and then unlink Self from EntryList. We have to drain eventually, 772 // so it might as well be now. 773 774 ObjectWaiter * v = _cxq; 775 assert(v != NULL, "invariant"); 776 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) { 777 // The CAS above can fail from interference IFF a "RAT" arrived. 778 // In that case Self must be in the interior and can no longer be 779 // at the head of cxq. 780 if (v == SelfNode) { 781 assert(_cxq != v, "invariant"); 782 v = _cxq; // CAS above failed - start scan at head of list 783 } 784 ObjectWaiter * p; 785 ObjectWaiter * q = NULL; 786 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 787 q = p; 788 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 789 } 790 assert(v != SelfNode, "invariant"); 791 assert(p == SelfNode, "Node not found on cxq"); 792 assert(p != _cxq, "invariant"); 793 assert(q != NULL, "invariant"); 794 assert(q->_next == p, "invariant"); 795 q->_next = p->_next; 796 } 797 } 798 799 #ifdef ASSERT 800 // Diagnostic hygiene ... 801 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 802 SelfNode->_next = (ObjectWaiter *) 0xBAD; 803 SelfNode->TState = ObjectWaiter::TS_RUN; 804 #endif 805 } 806 807 // ----------------------------------------------------------------------------- 808 // Exit support 809 // 810 // exit() 811 // ~~~~~~ 812 // Note that the collector can't reclaim the objectMonitor or deflate 813 // the object out from underneath the thread calling ::exit() as the 814 // thread calling ::exit() never transitions to a stable state. 815 // This inhibits GC, which in turn inhibits asynchronous (and 816 // inopportune) reclamation of "this". 817 // 818 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 819 // There's one exception to the claim above, however. EnterI() can call 820 // exit() to drop a lock if the acquirer has been externally suspended. 821 // In that case exit() is called with _thread_state as _thread_blocked, 822 // but the monitor's _contentions field is > 0, which inhibits reclamation. 823 // 824 // 1-0 exit 825 // ~~~~~~~~ 826 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 827 // the fast-path operators have been optimized so the common ::exit() 828 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 829 // The code emitted by fast_unlock() elides the usual MEMBAR. This 830 // greatly improves latency -- MEMBAR and CAS having considerable local 831 // latency on modern processors -- but at the cost of "stranding". Absent the 832 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 833 // ::enter() path, resulting in the entering thread being stranding 834 // and a progress-liveness failure. Stranding is extremely rare. 835 // We use timers (timed park operations) & periodic polling to detect 836 // and recover from stranding. Potentially stranded threads periodically 837 // wake up and poll the lock. See the usage of the _Responsible variable. 838 // 839 // The CAS() in enter provides for safety and exclusion, while the CAS or 840 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 841 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 842 // We detect and recover from stranding with timers. 843 // 844 // If a thread transiently strands it'll park until (a) another 845 // thread acquires the lock and then drops the lock, at which time the 846 // exiting thread will notice and unpark the stranded thread, or, (b) 847 // the timer expires. If the lock is high traffic then the stranding latency 848 // will be low due to (a). If the lock is low traffic then the odds of 849 // stranding are lower, although the worst-case stranding latency 850 // is longer. Critically, we don't want to put excessive load in the 851 // platform's timer subsystem. We want to minimize both the timer injection 852 // rate (timers created/sec) as well as the number of timers active at 853 // any one time. (more precisely, we want to minimize timer-seconds, which is 854 // the integral of the # of active timers at any instant over time). 855 // Both impinge on OS scalability. Given that, at most one thread parked on 856 // a monitor will use a timer. 857 // 858 // There is also the risk of a futile wake-up. If we drop the lock 859 // another thread can reacquire the lock immediately, and we can 860 // then wake a thread unnecessarily. This is benign, and we've 861 // structured the code so the windows are short and the frequency 862 // of such futile wakups is low. 863 864 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 865 Thread * const Self = THREAD; 866 if (THREAD != _owner) { 867 if (THREAD->is_lock_owned((address) _owner)) { 868 // Transmute _owner from a BasicLock pointer to a Thread address. 869 // We don't need to hold _mutex for this transition. 870 // Non-null to Non-null is safe as long as all readers can 871 // tolerate either flavor. 872 assert(_recursions == 0, "invariant"); 873 _owner = THREAD; 874 _recursions = 0; 875 } else { 876 // Apparent unbalanced locking ... 877 // Naively we'd like to throw IllegalMonitorStateException. 878 // As a practical matter we can neither allocate nor throw an 879 // exception as ::exit() can be called from leaf routines. 880 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 881 // Upon deeper reflection, however, in a properly run JVM the only 882 // way we should encounter this situation is in the presence of 883 // unbalanced JNI locking. TODO: CheckJNICalls. 884 // See also: CR4414101 885 #ifdef ASSERT 886 LogStreamHandle(Error, monitorinflation) lsh; 887 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 888 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 889 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 890 print_debug_style_on(&lsh); 891 #endif 892 assert(false, "Non-balanced monitor enter/exit!"); 893 return; 894 } 895 } 896 897 if (_recursions != 0) { 898 _recursions--; // this is simple recursive enter 899 return; 900 } 901 902 // Invariant: after setting Responsible=null an thread must execute 903 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 904 _Responsible = NULL; 905 906 #if INCLUDE_JFR 907 // get the owner's thread id for the MonitorEnter event 908 // if it is enabled and the thread isn't suspended 909 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 910 _previous_owner_tid = JFR_THREAD_ID(Self); 911 } 912 #endif 913 914 for (;;) { 915 assert(THREAD == _owner, "invariant"); 916 917 // release semantics: prior loads and stores from within the critical section 918 // must not float (reorder) past the following store that drops the lock. 919 Atomic::release_store(&_owner, (void*)NULL); // drop the lock 920 OrderAccess::storeload(); // See if we need to wake a successor 921 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 922 return; 923 } 924 // Other threads are blocked trying to acquire the lock. 925 926 // Normally the exiting thread is responsible for ensuring succession, 927 // but if other successors are ready or other entering threads are spinning 928 // then this thread can simply store NULL into _owner and exit without 929 // waking a successor. The existence of spinners or ready successors 930 // guarantees proper succession (liveness). Responsibility passes to the 931 // ready or running successors. The exiting thread delegates the duty. 932 // More precisely, if a successor already exists this thread is absolved 933 // of the responsibility of waking (unparking) one. 934 // 935 // The _succ variable is critical to reducing futile wakeup frequency. 936 // _succ identifies the "heir presumptive" thread that has been made 937 // ready (unparked) but that has not yet run. We need only one such 938 // successor thread to guarantee progress. 939 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 940 // section 3.3 "Futile Wakeup Throttling" for details. 941 // 942 // Note that spinners in Enter() also set _succ non-null. 943 // In the current implementation spinners opportunistically set 944 // _succ so that exiting threads might avoid waking a successor. 945 // Another less appealing alternative would be for the exiting thread 946 // to drop the lock and then spin briefly to see if a spinner managed 947 // to acquire the lock. If so, the exiting thread could exit 948 // immediately without waking a successor, otherwise the exiting 949 // thread would need to dequeue and wake a successor. 950 // (Note that we'd need to make the post-drop spin short, but no 951 // shorter than the worst-case round-trip cache-line migration time. 952 // The dropped lock needs to become visible to the spinner, and then 953 // the acquisition of the lock by the spinner must become visible to 954 // the exiting thread). 955 956 // It appears that an heir-presumptive (successor) must be made ready. 957 // Only the current lock owner can manipulate the EntryList or 958 // drain _cxq, so we need to reacquire the lock. If we fail 959 // to reacquire the lock the responsibility for ensuring succession 960 // falls to the new owner. 961 // 962 if (!Atomic::replace_if_null(&_owner, THREAD)) { 963 return; 964 } 965 966 guarantee(_owner == THREAD, "invariant"); 967 968 ObjectWaiter * w = NULL; 969 970 w = _EntryList; 971 if (w != NULL) { 972 // I'd like to write: guarantee (w->_thread != Self). 973 // But in practice an exiting thread may find itself on the EntryList. 974 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 975 // then calls exit(). Exit release the lock by setting O._owner to NULL. 976 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 977 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 978 // release the lock "O". T2 resumes immediately after the ST of null into 979 // _owner, above. T2 notices that the EntryList is populated, so it 980 // reacquires the lock and then finds itself on the EntryList. 981 // Given all that, we have to tolerate the circumstance where "w" is 982 // associated with Self. 983 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 984 ExitEpilog(Self, w); 985 return; 986 } 987 988 // If we find that both _cxq and EntryList are null then just 989 // re-run the exit protocol from the top. 990 w = _cxq; 991 if (w == NULL) continue; 992 993 // Drain _cxq into EntryList - bulk transfer. 994 // First, detach _cxq. 995 // The following loop is tantamount to: w = swap(&cxq, NULL) 996 for (;;) { 997 assert(w != NULL, "Invariant"); 998 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL); 999 if (u == w) break; 1000 w = u; 1001 } 1002 1003 assert(w != NULL, "invariant"); 1004 assert(_EntryList == NULL, "invariant"); 1005 1006 // Convert the LIFO SLL anchored by _cxq into a DLL. 1007 // The list reorganization step operates in O(LENGTH(w)) time. 1008 // It's critical that this step operate quickly as 1009 // "Self" still holds the outer-lock, restricting parallelism 1010 // and effectively lengthening the critical section. 1011 // Invariant: s chases t chases u. 1012 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1013 // we have faster access to the tail. 1014 1015 _EntryList = w; 1016 ObjectWaiter * q = NULL; 1017 ObjectWaiter * p; 1018 for (p = w; p != NULL; p = p->_next) { 1019 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1020 p->TState = ObjectWaiter::TS_ENTER; 1021 p->_prev = q; 1022 q = p; 1023 } 1024 1025 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1026 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1027 1028 // See if we can abdicate to a spinner instead of waking a thread. 1029 // A primary goal of the implementation is to reduce the 1030 // context-switch rate. 1031 if (_succ != NULL) continue; 1032 1033 w = _EntryList; 1034 if (w != NULL) { 1035 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1036 ExitEpilog(Self, w); 1037 return; 1038 } 1039 } 1040 } 1041 1042 // ExitSuspendEquivalent: 1043 // A faster alternate to handle_special_suspend_equivalent_condition() 1044 // 1045 // handle_special_suspend_equivalent_condition() unconditionally 1046 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1047 // operations have high latency. Note that in ::enter() we call HSSEC 1048 // while holding the monitor, so we effectively lengthen the critical sections. 1049 // 1050 // There are a number of possible solutions: 1051 // 1052 // A. To ameliorate the problem we might also defer state transitions 1053 // to as late as possible -- just prior to parking. 1054 // Given that, we'd call HSSEC after having returned from park(), 1055 // but before attempting to acquire the monitor. This is only a 1056 // partial solution. It avoids calling HSSEC while holding the 1057 // monitor (good), but it still increases successor reacquisition latency -- 1058 // the interval between unparking a successor and the time the successor 1059 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1060 // If we use this technique we can also avoid EnterI()-exit() loop 1061 // in ::enter() where we iteratively drop the lock and then attempt 1062 // to reacquire it after suspending. 1063 // 1064 // B. In the future we might fold all the suspend bits into a 1065 // composite per-thread suspend flag and then update it with CAS(). 1066 // Alternately, a Dekker-like mechanism with multiple variables 1067 // would suffice: 1068 // ST Self->_suspend_equivalent = false 1069 // MEMBAR 1070 // LD Self_>_suspend_flags 1071 1072 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1073 return jSelf->handle_special_suspend_equivalent_condition(); 1074 } 1075 1076 1077 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1078 assert(_owner == Self, "invariant"); 1079 1080 // Exit protocol: 1081 // 1. ST _succ = wakee 1082 // 2. membar #loadstore|#storestore; 1083 // 2. ST _owner = NULL 1084 // 3. unpark(wakee) 1085 1086 _succ = Wakee->_thread; 1087 ParkEvent * Trigger = Wakee->_event; 1088 1089 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1090 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1091 // out-of-scope (non-extant). 1092 Wakee = NULL; 1093 1094 // Drop the lock 1095 Atomic::release_store(&_owner, (void*)NULL); 1096 OrderAccess::fence(); // ST _owner vs LD in unpark() 1097 1098 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1099 Trigger->unpark(); 1100 1101 // Maintain stats and report events to JVMTI 1102 OM_PERFDATA_OP(Parks, inc()); 1103 } 1104 1105 1106 // ----------------------------------------------------------------------------- 1107 // Class Loader deadlock handling. 1108 // 1109 // complete_exit exits a lock returning recursion count 1110 // complete_exit/reenter operate as a wait without waiting 1111 // complete_exit requires an inflated monitor 1112 // The _owner field is not always the Thread addr even with an 1113 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1114 // thread due to contention. 1115 intx ObjectMonitor::complete_exit(TRAPS) { 1116 Thread * const Self = THREAD; 1117 assert(Self->is_Java_thread(), "Must be Java thread!"); 1118 JavaThread *jt = (JavaThread *)THREAD; 1119 1120 assert(InitDone, "Unexpectedly not initialized"); 1121 1122 if (THREAD != _owner) { 1123 if (THREAD->is_lock_owned ((address)_owner)) { 1124 assert(_recursions == 0, "internal state error"); 1125 _owner = THREAD; // Convert from basiclock addr to Thread addr 1126 _recursions = 0; 1127 } 1128 } 1129 1130 guarantee(Self == _owner, "complete_exit not owner"); 1131 intx save = _recursions; // record the old recursion count 1132 _recursions = 0; // set the recursion level to be 0 1133 exit(true, Self); // exit the monitor 1134 guarantee(_owner != Self, "invariant"); 1135 return save; 1136 } 1137 1138 // reenter() enters a lock and sets recursion count 1139 // complete_exit/reenter operate as a wait without waiting 1140 void ObjectMonitor::reenter(intx recursions, TRAPS) { 1141 Thread * const Self = THREAD; 1142 assert(Self->is_Java_thread(), "Must be Java thread!"); 1143 JavaThread *jt = (JavaThread *)THREAD; 1144 1145 guarantee(_owner != Self, "reenter already owner"); 1146 enter(THREAD); // enter the monitor 1147 guarantee(_recursions == 0, "reenter recursion"); 1148 _recursions = recursions; 1149 return; 1150 } 1151 1152 // Checks that the current THREAD owns this monitor and causes an 1153 // immediate return if it doesn't. We don't use the CHECK macro 1154 // because we want the IMSE to be the only exception that is thrown 1155 // from the call site when false is returned. Any other pending 1156 // exception is ignored. 1157 #define CHECK_OWNER() \ 1158 do { \ 1159 if (!check_owner(THREAD)) { \ 1160 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1161 return; \ 1162 } \ 1163 } while (false) 1164 1165 // Returns true if the specified thread owns the ObjectMonitor. 1166 // Otherwise returns false and throws IllegalMonitorStateException 1167 // (IMSE). If there is a pending exception and the specified thread 1168 // is not the owner, that exception will be replaced by the IMSE. 1169 bool ObjectMonitor::check_owner(Thread* THREAD) { 1170 if (_owner == THREAD) { 1171 return true; 1172 } 1173 if (THREAD->is_lock_owned((address)_owner)) { 1174 _owner = THREAD; // convert from BasicLock addr to Thread addr 1175 _recursions = 0; 1176 return true; 1177 } 1178 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1179 "current thread is not owner", false); 1180 } 1181 1182 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1183 ObjectMonitor* monitor, 1184 jlong notifier_tid, 1185 jlong timeout, 1186 bool timedout) { 1187 assert(event != NULL, "invariant"); 1188 assert(monitor != NULL, "invariant"); 1189 event->set_monitorClass(((oop)monitor->object())->klass()); 1190 event->set_timeout(timeout); 1191 event->set_address((uintptr_t)monitor->object_addr()); 1192 event->set_notifier(notifier_tid); 1193 event->set_timedOut(timedout); 1194 event->commit(); 1195 } 1196 1197 // ----------------------------------------------------------------------------- 1198 // Wait/Notify/NotifyAll 1199 // 1200 // Note: a subset of changes to ObjectMonitor::wait() 1201 // will need to be replicated in complete_exit 1202 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1203 Thread * const Self = THREAD; 1204 assert(Self->is_Java_thread(), "Must be Java thread!"); 1205 JavaThread *jt = (JavaThread *)THREAD; 1206 1207 assert(InitDone, "Unexpectedly not initialized"); 1208 1209 CHECK_OWNER(); // Throws IMSE if not owner. 1210 1211 EventJavaMonitorWait event; 1212 1213 // check for a pending interrupt 1214 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1215 // post monitor waited event. Note that this is past-tense, we are done waiting. 1216 if (JvmtiExport::should_post_monitor_waited()) { 1217 // Note: 'false' parameter is passed here because the 1218 // wait was not timed out due to thread interrupt. 1219 JvmtiExport::post_monitor_waited(jt, this, false); 1220 1221 // In this short circuit of the monitor wait protocol, the 1222 // current thread never drops ownership of the monitor and 1223 // never gets added to the wait queue so the current thread 1224 // cannot be made the successor. This means that the 1225 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1226 // consume an unpark() meant for the ParkEvent associated with 1227 // this ObjectMonitor. 1228 } 1229 if (event.should_commit()) { 1230 post_monitor_wait_event(&event, this, 0, millis, false); 1231 } 1232 THROW(vmSymbols::java_lang_InterruptedException()); 1233 return; 1234 } 1235 1236 assert(Self->_Stalled == 0, "invariant"); 1237 Self->_Stalled = intptr_t(this); 1238 jt->set_current_waiting_monitor(this); 1239 1240 // create a node to be put into the queue 1241 // Critically, after we reset() the event but prior to park(), we must check 1242 // for a pending interrupt. 1243 ObjectWaiter node(Self); 1244 node.TState = ObjectWaiter::TS_WAIT; 1245 Self->_ParkEvent->reset(); 1246 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1247 1248 // Enter the waiting queue, which is a circular doubly linked list in this case 1249 // but it could be a priority queue or any data structure. 1250 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1251 // by the the owner of the monitor *except* in the case where park() 1252 // returns because of a timeout of interrupt. Contention is exceptionally rare 1253 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1254 1255 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1256 AddWaiter(&node); 1257 Thread::SpinRelease(&_WaitSetLock); 1258 1259 _Responsible = NULL; 1260 1261 intx save = _recursions; // record the old recursion count 1262 _waiters++; // increment the number of waiters 1263 _recursions = 0; // set the recursion level to be 1 1264 exit(true, Self); // exit the monitor 1265 guarantee(_owner != Self, "invariant"); 1266 1267 // The thread is on the WaitSet list - now park() it. 1268 // On MP systems it's conceivable that a brief spin before we park 1269 // could be profitable. 1270 // 1271 // TODO-FIXME: change the following logic to a loop of the form 1272 // while (!timeout && !interrupted && _notified == 0) park() 1273 1274 int ret = OS_OK; 1275 int WasNotified = 0; 1276 1277 // Need to check interrupt state whilst still _thread_in_vm 1278 bool interrupted = interruptible && jt->is_interrupted(false); 1279 1280 { // State transition wrappers 1281 OSThread* osthread = Self->osthread(); 1282 OSThreadWaitState osts(osthread, true); 1283 { 1284 ThreadBlockInVM tbivm(jt); 1285 // Thread is in thread_blocked state and oop access is unsafe. 1286 jt->set_suspend_equivalent(); 1287 1288 if (interrupted || HAS_PENDING_EXCEPTION) { 1289 // Intentionally empty 1290 } else if (node._notified == 0) { 1291 if (millis <= 0) { 1292 Self->_ParkEvent->park(); 1293 } else { 1294 ret = Self->_ParkEvent->park(millis); 1295 } 1296 } 1297 1298 // were we externally suspended while we were waiting? 1299 if (ExitSuspendEquivalent (jt)) { 1300 // TODO-FIXME: add -- if succ == Self then succ = null. 1301 jt->java_suspend_self(); 1302 } 1303 1304 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1305 1306 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1307 // from the WaitSet to the EntryList. 1308 // See if we need to remove Node from the WaitSet. 1309 // We use double-checked locking to avoid grabbing _WaitSetLock 1310 // if the thread is not on the wait queue. 1311 // 1312 // Note that we don't need a fence before the fetch of TState. 1313 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1314 // written by the is thread. (perhaps the fetch might even be satisfied 1315 // by a look-aside into the processor's own store buffer, although given 1316 // the length of the code path between the prior ST and this load that's 1317 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1318 // then we'll acquire the lock and then re-fetch a fresh TState value. 1319 // That is, we fail toward safety. 1320 1321 if (node.TState == ObjectWaiter::TS_WAIT) { 1322 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1323 if (node.TState == ObjectWaiter::TS_WAIT) { 1324 DequeueSpecificWaiter(&node); // unlink from WaitSet 1325 assert(node._notified == 0, "invariant"); 1326 node.TState = ObjectWaiter::TS_RUN; 1327 } 1328 Thread::SpinRelease(&_WaitSetLock); 1329 } 1330 1331 // The thread is now either on off-list (TS_RUN), 1332 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1333 // The Node's TState variable is stable from the perspective of this thread. 1334 // No other threads will asynchronously modify TState. 1335 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1336 OrderAccess::loadload(); 1337 if (_succ == Self) _succ = NULL; 1338 WasNotified = node._notified; 1339 1340 // Reentry phase -- reacquire the monitor. 1341 // re-enter contended monitor after object.wait(). 1342 // retain OBJECT_WAIT state until re-enter successfully completes 1343 // Thread state is thread_in_vm and oop access is again safe, 1344 // although the raw address of the object may have changed. 1345 // (Don't cache naked oops over safepoints, of course). 1346 1347 // post monitor waited event. Note that this is past-tense, we are done waiting. 1348 if (JvmtiExport::should_post_monitor_waited()) { 1349 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1350 1351 if (node._notified != 0 && _succ == Self) { 1352 // In this part of the monitor wait-notify-reenter protocol it 1353 // is possible (and normal) for another thread to do a fastpath 1354 // monitor enter-exit while this thread is still trying to get 1355 // to the reenter portion of the protocol. 1356 // 1357 // The ObjectMonitor was notified and the current thread is 1358 // the successor which also means that an unpark() has already 1359 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1360 // consume the unpark() that was done when the successor was 1361 // set because the same ParkEvent is shared between Java 1362 // monitors and JVM/TI RawMonitors (for now). 1363 // 1364 // We redo the unpark() to ensure forward progress, i.e., we 1365 // don't want all pending threads hanging (parked) with none 1366 // entering the unlocked monitor. 1367 node._event->unpark(); 1368 } 1369 } 1370 1371 if (event.should_commit()) { 1372 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1373 } 1374 1375 OrderAccess::fence(); 1376 1377 assert(Self->_Stalled != 0, "invariant"); 1378 Self->_Stalled = 0; 1379 1380 assert(_owner != Self, "invariant"); 1381 ObjectWaiter::TStates v = node.TState; 1382 if (v == ObjectWaiter::TS_RUN) { 1383 enter(Self); 1384 } else { 1385 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1386 ReenterI(Self, &node); 1387 node.wait_reenter_end(this); 1388 } 1389 1390 // Self has reacquired the lock. 1391 // Lifecycle - the node representing Self must not appear on any queues. 1392 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1393 // want residual elements associated with this thread left on any lists. 1394 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1395 assert(_owner == Self, "invariant"); 1396 assert(_succ != Self, "invariant"); 1397 } // OSThreadWaitState() 1398 1399 jt->set_current_waiting_monitor(NULL); 1400 1401 guarantee(_recursions == 0, "invariant"); 1402 _recursions = save; // restore the old recursion count 1403 _waiters--; // decrement the number of waiters 1404 1405 // Verify a few postconditions 1406 assert(_owner == Self, "invariant"); 1407 assert(_succ != Self, "invariant"); 1408 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1409 1410 // check if the notification happened 1411 if (!WasNotified) { 1412 // no, it could be timeout or Thread.interrupt() or both 1413 // check for interrupt event, otherwise it is timeout 1414 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1415 THROW(vmSymbols::java_lang_InterruptedException()); 1416 } 1417 } 1418 1419 // NOTE: Spurious wake up will be consider as timeout. 1420 // Monitor notify has precedence over thread interrupt. 1421 } 1422 1423 1424 // Consider: 1425 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1426 // then instead of transferring a thread from the WaitSet to the EntryList 1427 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1428 1429 void ObjectMonitor::INotify(Thread * Self) { 1430 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1431 ObjectWaiter * iterator = DequeueWaiter(); 1432 if (iterator != NULL) { 1433 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1434 guarantee(iterator->_notified == 0, "invariant"); 1435 // Disposition - what might we do with iterator ? 1436 // a. add it directly to the EntryList - either tail (policy == 1) 1437 // or head (policy == 0). 1438 // b. push it onto the front of the _cxq (policy == 2). 1439 // For now we use (b). 1440 1441 iterator->TState = ObjectWaiter::TS_ENTER; 1442 1443 iterator->_notified = 1; 1444 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1445 1446 ObjectWaiter * list = _EntryList; 1447 if (list != NULL) { 1448 assert(list->_prev == NULL, "invariant"); 1449 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1450 assert(list != iterator, "invariant"); 1451 } 1452 1453 // prepend to cxq 1454 if (list == NULL) { 1455 iterator->_next = iterator->_prev = NULL; 1456 _EntryList = iterator; 1457 } else { 1458 iterator->TState = ObjectWaiter::TS_CXQ; 1459 for (;;) { 1460 ObjectWaiter * front = _cxq; 1461 iterator->_next = front; 1462 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1463 break; 1464 } 1465 } 1466 } 1467 1468 // _WaitSetLock protects the wait queue, not the EntryList. We could 1469 // move the add-to-EntryList operation, above, outside the critical section 1470 // protected by _WaitSetLock. In practice that's not useful. With the 1471 // exception of wait() timeouts and interrupts the monitor owner 1472 // is the only thread that grabs _WaitSetLock. There's almost no contention 1473 // on _WaitSetLock so it's not profitable to reduce the length of the 1474 // critical section. 1475 1476 iterator->wait_reenter_begin(this); 1477 } 1478 Thread::SpinRelease(&_WaitSetLock); 1479 } 1480 1481 // Consider: a not-uncommon synchronization bug is to use notify() when 1482 // notifyAll() is more appropriate, potentially resulting in stranded 1483 // threads; this is one example of a lost wakeup. A useful diagnostic 1484 // option is to force all notify() operations to behave as notifyAll(). 1485 // 1486 // Note: We can also detect many such problems with a "minimum wait". 1487 // When the "minimum wait" is set to a small non-zero timeout value 1488 // and the program does not hang whereas it did absent "minimum wait", 1489 // that suggests a lost wakeup bug. 1490 1491 void ObjectMonitor::notify(TRAPS) { 1492 CHECK_OWNER(); // Throws IMSE if not owner. 1493 if (_WaitSet == NULL) { 1494 return; 1495 } 1496 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1497 INotify(THREAD); 1498 OM_PERFDATA_OP(Notifications, inc(1)); 1499 } 1500 1501 1502 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1503 // from the waitset to the EntryList. This could be done more efficiently with a 1504 // single bulk transfer but in practice it's not time-critical. Beware too, 1505 // that in prepend-mode we invert the order of the waiters. Let's say that the 1506 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1507 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1508 1509 void ObjectMonitor::notifyAll(TRAPS) { 1510 CHECK_OWNER(); // Throws IMSE if not owner. 1511 if (_WaitSet == NULL) { 1512 return; 1513 } 1514 1515 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1516 int tally = 0; 1517 while (_WaitSet != NULL) { 1518 tally++; 1519 INotify(THREAD); 1520 } 1521 1522 OM_PERFDATA_OP(Notifications, inc(tally)); 1523 } 1524 1525 // ----------------------------------------------------------------------------- 1526 // Adaptive Spinning Support 1527 // 1528 // Adaptive spin-then-block - rational spinning 1529 // 1530 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1531 // algorithm. On high order SMP systems it would be better to start with 1532 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1533 // a contending thread could enqueue itself on the cxq and then spin locally 1534 // on a thread-specific variable such as its ParkEvent._Event flag. 1535 // That's left as an exercise for the reader. Note that global spinning is 1536 // not problematic on Niagara, as the L2 cache serves the interconnect and 1537 // has both low latency and massive bandwidth. 1538 // 1539 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1540 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1541 // (duration) or we can fix the count at approximately the duration of 1542 // a context switch and vary the frequency. Of course we could also 1543 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1544 // For a description of 'Adaptive spin-then-block mutual exclusion in 1545 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1546 // 1547 // This implementation varies the duration "D", where D varies with 1548 // the success rate of recent spin attempts. (D is capped at approximately 1549 // length of a round-trip context switch). The success rate for recent 1550 // spin attempts is a good predictor of the success rate of future spin 1551 // attempts. The mechanism adapts automatically to varying critical 1552 // section length (lock modality), system load and degree of parallelism. 1553 // D is maintained per-monitor in _SpinDuration and is initialized 1554 // optimistically. Spin frequency is fixed at 100%. 1555 // 1556 // Note that _SpinDuration is volatile, but we update it without locks 1557 // or atomics. The code is designed so that _SpinDuration stays within 1558 // a reasonable range even in the presence of races. The arithmetic 1559 // operations on _SpinDuration are closed over the domain of legal values, 1560 // so at worst a race will install and older but still legal value. 1561 // At the very worst this introduces some apparent non-determinism. 1562 // We might spin when we shouldn't or vice-versa, but since the spin 1563 // count are relatively short, even in the worst case, the effect is harmless. 1564 // 1565 // Care must be taken that a low "D" value does not become an 1566 // an absorbing state. Transient spinning failures -- when spinning 1567 // is overall profitable -- should not cause the system to converge 1568 // on low "D" values. We want spinning to be stable and predictable 1569 // and fairly responsive to change and at the same time we don't want 1570 // it to oscillate, become metastable, be "too" non-deterministic, 1571 // or converge on or enter undesirable stable absorbing states. 1572 // 1573 // We implement a feedback-based control system -- using past behavior 1574 // to predict future behavior. We face two issues: (a) if the 1575 // input signal is random then the spin predictor won't provide optimal 1576 // results, and (b) if the signal frequency is too high then the control 1577 // system, which has some natural response lag, will "chase" the signal. 1578 // (b) can arise from multimodal lock hold times. Transient preemption 1579 // can also result in apparent bimodal lock hold times. 1580 // Although sub-optimal, neither condition is particularly harmful, as 1581 // in the worst-case we'll spin when we shouldn't or vice-versa. 1582 // The maximum spin duration is rather short so the failure modes aren't bad. 1583 // To be conservative, I've tuned the gain in system to bias toward 1584 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1585 // "rings" or oscillates between spinning and not spinning. This happens 1586 // when spinning is just on the cusp of profitability, however, so the 1587 // situation is not dire. The state is benign -- there's no need to add 1588 // hysteresis control to damp the transition rate between spinning and 1589 // not spinning. 1590 1591 // Spinning: Fixed frequency (100%), vary duration 1592 int ObjectMonitor::TrySpin(Thread * Self) { 1593 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1594 int ctr = Knob_FixedSpin; 1595 if (ctr != 0) { 1596 while (--ctr >= 0) { 1597 if (TryLock(Self) > 0) return 1; 1598 SpinPause(); 1599 } 1600 return 0; 1601 } 1602 1603 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1604 if (TryLock(Self) > 0) { 1605 // Increase _SpinDuration ... 1606 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1607 // Raising _SpurDuration to the poverty line is key. 1608 int x = _SpinDuration; 1609 if (x < Knob_SpinLimit) { 1610 if (x < Knob_Poverty) x = Knob_Poverty; 1611 _SpinDuration = x + Knob_BonusB; 1612 } 1613 return 1; 1614 } 1615 SpinPause(); 1616 } 1617 1618 // Admission control - verify preconditions for spinning 1619 // 1620 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1621 // becoming an absorbing state. Put another way, we spin briefly to 1622 // sample, just in case the system load, parallelism, contention, or lock 1623 // modality changed. 1624 // 1625 // Consider the following alternative: 1626 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1627 // spin attempt. "Periodically" might mean after a tally of 1628 // the # of failed spin attempts (or iterations) reaches some threshold. 1629 // This takes us into the realm of 1-out-of-N spinning, where we 1630 // hold the duration constant but vary the frequency. 1631 1632 ctr = _SpinDuration; 1633 if (ctr <= 0) return 0; 1634 1635 if (NotRunnable(Self, (Thread *) _owner)) { 1636 return 0; 1637 } 1638 1639 // We're good to spin ... spin ingress. 1640 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1641 // when preparing to LD...CAS _owner, etc and the CAS is likely 1642 // to succeed. 1643 if (_succ == NULL) { 1644 _succ = Self; 1645 } 1646 Thread * prv = NULL; 1647 1648 // There are three ways to exit the following loop: 1649 // 1. A successful spin where this thread has acquired the lock. 1650 // 2. Spin failure with prejudice 1651 // 3. Spin failure without prejudice 1652 1653 while (--ctr >= 0) { 1654 1655 // Periodic polling -- Check for pending GC 1656 // Threads may spin while they're unsafe. 1657 // We don't want spinning threads to delay the JVM from reaching 1658 // a stop-the-world safepoint or to steal cycles from GC. 1659 // If we detect a pending safepoint we abort in order that 1660 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1661 // this thread, if safe, doesn't steal cycles from GC. 1662 // This is in keeping with the "no loitering in runtime" rule. 1663 // We periodically check to see if there's a safepoint pending. 1664 if ((ctr & 0xFF) == 0) { 1665 if (SafepointMechanism::should_block(Self)) { 1666 goto Abort; // abrupt spin egress 1667 } 1668 SpinPause(); 1669 } 1670 1671 // Probe _owner with TATAS 1672 // If this thread observes the monitor transition or flicker 1673 // from locked to unlocked to locked, then the odds that this 1674 // thread will acquire the lock in this spin attempt go down 1675 // considerably. The same argument applies if the CAS fails 1676 // or if we observe _owner change from one non-null value to 1677 // another non-null value. In such cases we might abort 1678 // the spin without prejudice or apply a "penalty" to the 1679 // spin count-down variable "ctr", reducing it by 100, say. 1680 1681 Thread * ox = (Thread *) _owner; 1682 if (ox == NULL) { 1683 ox = (Thread*)Atomic::cmpxchg(&_owner, (void*)NULL, Self); 1684 if (ox == NULL) { 1685 // The CAS succeeded -- this thread acquired ownership 1686 // Take care of some bookkeeping to exit spin state. 1687 if (_succ == Self) { 1688 _succ = NULL; 1689 } 1690 1691 // Increase _SpinDuration : 1692 // The spin was successful (profitable) so we tend toward 1693 // longer spin attempts in the future. 1694 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1695 // If we acquired the lock early in the spin cycle it 1696 // makes sense to increase _SpinDuration proportionally. 1697 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1698 int x = _SpinDuration; 1699 if (x < Knob_SpinLimit) { 1700 if (x < Knob_Poverty) x = Knob_Poverty; 1701 _SpinDuration = x + Knob_Bonus; 1702 } 1703 return 1; 1704 } 1705 1706 // The CAS failed ... we can take any of the following actions: 1707 // * penalize: ctr -= CASPenalty 1708 // * exit spin with prejudice -- goto Abort; 1709 // * exit spin without prejudice. 1710 // * Since CAS is high-latency, retry again immediately. 1711 prv = ox; 1712 goto Abort; 1713 } 1714 1715 // Did lock ownership change hands ? 1716 if (ox != prv && prv != NULL) { 1717 goto Abort; 1718 } 1719 prv = ox; 1720 1721 // Abort the spin if the owner is not executing. 1722 // The owner must be executing in order to drop the lock. 1723 // Spinning while the owner is OFFPROC is idiocy. 1724 // Consider: ctr -= RunnablePenalty ; 1725 if (NotRunnable(Self, ox)) { 1726 goto Abort; 1727 } 1728 if (_succ == NULL) { 1729 _succ = Self; 1730 } 1731 } 1732 1733 // Spin failed with prejudice -- reduce _SpinDuration. 1734 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1735 // AIMD is globally stable. 1736 { 1737 int x = _SpinDuration; 1738 if (x > 0) { 1739 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1740 // This is globally sample and tends to damp the response. 1741 x -= Knob_Penalty; 1742 if (x < 0) x = 0; 1743 _SpinDuration = x; 1744 } 1745 } 1746 1747 Abort: 1748 if (_succ == Self) { 1749 _succ = NULL; 1750 // Invariant: after setting succ=null a contending thread 1751 // must recheck-retry _owner before parking. This usually happens 1752 // in the normal usage of TrySpin(), but it's safest 1753 // to make TrySpin() as foolproof as possible. 1754 OrderAccess::fence(); 1755 if (TryLock(Self) > 0) return 1; 1756 } 1757 return 0; 1758 } 1759 1760 // NotRunnable() -- informed spinning 1761 // 1762 // Don't bother spinning if the owner is not eligible to drop the lock. 1763 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1764 // The thread must be runnable in order to drop the lock in timely fashion. 1765 // If the _owner is not runnable then spinning will not likely be 1766 // successful (profitable). 1767 // 1768 // Beware -- the thread referenced by _owner could have died 1769 // so a simply fetch from _owner->_thread_state might trap. 1770 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1771 // Because of the lifecycle issues, the _thread_state values 1772 // observed by NotRunnable() might be garbage. NotRunnable must 1773 // tolerate this and consider the observed _thread_state value 1774 // as advisory. 1775 // 1776 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1777 // a thread pointer. 1778 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1779 // with the LSB of _owner. Another option would be to probabilistically probe 1780 // the putative _owner->TypeTag value. 1781 // 1782 // Checking _thread_state isn't perfect. Even if the thread is 1783 // in_java it might be blocked on a page-fault or have been preempted 1784 // and sitting on a ready/dispatch queue. 1785 // 1786 // The return value from NotRunnable() is *advisory* -- the 1787 // result is based on sampling and is not necessarily coherent. 1788 // The caller must tolerate false-negative and false-positive errors. 1789 // Spinning, in general, is probabilistic anyway. 1790 1791 1792 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1793 // Check ox->TypeTag == 2BAD. 1794 if (ox == NULL) return 0; 1795 1796 // Avoid transitive spinning ... 1797 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1798 // Immediately after T1 acquires L it's possible that T2, also 1799 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1800 // This occurs transiently after T1 acquired L but before 1801 // T1 managed to clear T1.Stalled. T2 does not need to abort 1802 // its spin in this circumstance. 1803 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1804 1805 if (BlockedOn == 1) return 1; 1806 if (BlockedOn != 0) { 1807 return BlockedOn != intptr_t(this) && _owner == ox; 1808 } 1809 1810 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1811 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1812 // consider also: jst != _thread_in_Java -- but that's overspecific. 1813 return jst == _thread_blocked || jst == _thread_in_native; 1814 } 1815 1816 1817 // ----------------------------------------------------------------------------- 1818 // WaitSet management ... 1819 1820 ObjectWaiter::ObjectWaiter(Thread* thread) { 1821 _next = NULL; 1822 _prev = NULL; 1823 _notified = 0; 1824 _notifier_tid = 0; 1825 TState = TS_RUN; 1826 _thread = thread; 1827 _event = thread->_ParkEvent; 1828 _active = false; 1829 assert(_event != NULL, "invariant"); 1830 } 1831 1832 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1833 JavaThread *jt = (JavaThread *)this->_thread; 1834 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1835 } 1836 1837 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1838 JavaThread *jt = (JavaThread *)this->_thread; 1839 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1840 } 1841 1842 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1843 assert(node != NULL, "should not add NULL node"); 1844 assert(node->_prev == NULL, "node already in list"); 1845 assert(node->_next == NULL, "node already in list"); 1846 // put node at end of queue (circular doubly linked list) 1847 if (_WaitSet == NULL) { 1848 _WaitSet = node; 1849 node->_prev = node; 1850 node->_next = node; 1851 } else { 1852 ObjectWaiter* head = _WaitSet; 1853 ObjectWaiter* tail = head->_prev; 1854 assert(tail->_next == head, "invariant check"); 1855 tail->_next = node; 1856 head->_prev = node; 1857 node->_next = head; 1858 node->_prev = tail; 1859 } 1860 } 1861 1862 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1863 // dequeue the very first waiter 1864 ObjectWaiter* waiter = _WaitSet; 1865 if (waiter) { 1866 DequeueSpecificWaiter(waiter); 1867 } 1868 return waiter; 1869 } 1870 1871 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1872 assert(node != NULL, "should not dequeue NULL node"); 1873 assert(node->_prev != NULL, "node already removed from list"); 1874 assert(node->_next != NULL, "node already removed from list"); 1875 // when the waiter has woken up because of interrupt, 1876 // timeout or other spurious wake-up, dequeue the 1877 // waiter from waiting list 1878 ObjectWaiter* next = node->_next; 1879 if (next == node) { 1880 assert(node->_prev == node, "invariant check"); 1881 _WaitSet = NULL; 1882 } else { 1883 ObjectWaiter* prev = node->_prev; 1884 assert(prev->_next == node, "invariant check"); 1885 assert(next->_prev == node, "invariant check"); 1886 next->_prev = prev; 1887 prev->_next = next; 1888 if (_WaitSet == node) { 1889 _WaitSet = next; 1890 } 1891 } 1892 node->_next = NULL; 1893 node->_prev = NULL; 1894 } 1895 1896 // ----------------------------------------------------------------------------- 1897 // PerfData support 1898 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 1899 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 1900 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 1901 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 1902 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 1903 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 1904 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 1905 1906 // One-shot global initialization for the sync subsystem. 1907 // We could also defer initialization and initialize on-demand 1908 // the first time we call ObjectSynchronizer::inflate(). 1909 // Initialization would be protected - like so many things - by 1910 // the MonitorCache_lock. 1911 1912 void ObjectMonitor::Initialize() { 1913 assert(!InitDone, "invariant"); 1914 1915 if (!os::is_MP()) { 1916 Knob_SpinLimit = 0; 1917 Knob_PreSpin = 0; 1918 Knob_FixedSpin = -1; 1919 } 1920 1921 if (UsePerfData) { 1922 EXCEPTION_MARK; 1923 #define NEWPERFCOUNTER(n) \ 1924 { \ 1925 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 1926 CHECK); \ 1927 } 1928 #define NEWPERFVARIABLE(n) \ 1929 { \ 1930 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 1931 CHECK); \ 1932 } 1933 NEWPERFCOUNTER(_sync_Inflations); 1934 NEWPERFCOUNTER(_sync_Deflations); 1935 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 1936 NEWPERFCOUNTER(_sync_FutileWakeups); 1937 NEWPERFCOUNTER(_sync_Parks); 1938 NEWPERFCOUNTER(_sync_Notifications); 1939 NEWPERFVARIABLE(_sync_MonExtant); 1940 #undef NEWPERFCOUNTER 1941 #undef NEWPERFVARIABLE 1942 } 1943 1944 DEBUG_ONLY(InitDone = true;) 1945 } 1946 1947 void ObjectMonitor::print_on(outputStream* st) const { 1948 // The minimal things to print for markWord printing, more can be added for debugging and logging. 1949 st->print("{contentions=0x%08x,waiters=0x%08x" 1950 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 1951 contentions(), waiters(), recursions(), 1952 p2i(owner())); 1953 } 1954 void ObjectMonitor::print() const { print_on(tty); } 1955 1956 #ifdef ASSERT 1957 // Print the ObjectMonitor like a debugger would: 1958 // 1959 // (ObjectMonitor) 0x00007fdfb6012e40 = { 1960 // _header = 0x0000000000000001 1961 // _object = 0x000000070ff45fd0 1962 // _next_om = 0x0000000000000000 1963 // _pad_buf0 = { 1964 // [0] = '\0' 1965 // ... 1966 // [103] = '\0' 1967 // } 1968 // _owner = 0x0000000000000000 1969 // _previous_owner_tid = 0 1970 // _recursions = 0 1971 // _EntryList = 0x0000000000000000 1972 // _cxq = 0x0000000000000000 1973 // _succ = 0x0000000000000000 1974 // _Responsible = 0x0000000000000000 1975 // _Spinner = 0 1976 // _SpinDuration = 5000 1977 // _contentions = 0 1978 // _WaitSet = 0x0000700009756248 1979 // _waiters = 1 1980 // _WaitSetLock = 0 1981 // } 1982 // 1983 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 1984 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 1985 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 1986 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 1987 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(_next_om)); 1988 st->print_cr(" _pad_buf0 = {"); 1989 st->print_cr(" [0] = '\\0'"); 1990 st->print_cr(" ..."); 1991 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 1992 st->print_cr(" }"); 1993 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 1994 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 1995 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 1996 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 1997 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 1998 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 1999 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2000 st->print_cr(" _Spinner = %d", _Spinner); 2001 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2002 st->print_cr(" _contentions = %d", _contentions); 2003 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2004 st->print_cr(" _waiters = %d", _waiters); 2005 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2006 st->print_cr("}"); 2007 } 2008 #endif