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 = try_set_owner_from(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 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 264 return; 265 } 266 267 // We've encountered genuine contention. 268 assert(Self->_Stalled == 0, "invariant"); 269 Self->_Stalled = intptr_t(this); 270 271 // Try one round of spinning *before* enqueueing Self 272 // and before going through the awkward and expensive state 273 // transitions. The following spin is strictly optional ... 274 // Note that if we acquire the monitor from an initial spin 275 // we forgo posting JVMTI events and firing DTRACE probes. 276 if (TrySpin(Self) > 0) { 277 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 278 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 279 assert(((oop)object())->mark() == markWord::encode(this), 280 "object mark must match encoded this: mark=" INTPTR_FORMAT 281 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 282 markWord::encode(this).value()); 283 Self->_Stalled = 0; 284 return; 285 } 286 287 assert(_owner != Self, "invariant"); 288 assert(_succ != Self, "invariant"); 289 assert(Self->is_Java_thread(), "invariant"); 290 JavaThread * jt = (JavaThread *) Self; 291 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 292 assert(jt->thread_state() != _thread_blocked, "invariant"); 293 assert(this->object() != NULL, "invariant"); 294 assert(_contentions >= 0, "invariant"); 295 296 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 297 // Ensure the object-monitor relationship remains stable while there's contention. 298 Atomic::inc(&_contentions); 299 300 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 301 EventJavaMonitorEnter event; 302 if (event.should_commit()) { 303 event.set_monitorClass(((oop)this->object())->klass()); 304 event.set_address((uintptr_t)(this->object_addr())); 305 } 306 307 { // Change java thread status to indicate blocked on monitor enter. 308 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 309 310 Self->set_current_pending_monitor(this); 311 312 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 313 if (JvmtiExport::should_post_monitor_contended_enter()) { 314 JvmtiExport::post_monitor_contended_enter(jt, this); 315 316 // The current thread does not yet own the monitor and does not 317 // yet appear on any queues that would get it made the successor. 318 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 319 // handler cannot accidentally consume an unpark() meant for the 320 // ParkEvent associated with this ObjectMonitor. 321 } 322 323 OSThreadContendState osts(Self->osthread()); 324 ThreadBlockInVM tbivm(jt); 325 326 // TODO-FIXME: change the following for(;;) loop to straight-line code. 327 for (;;) { 328 jt->set_suspend_equivalent(); 329 // cleared by handle_special_suspend_equivalent_condition() 330 // or java_suspend_self() 331 332 EnterI(THREAD); 333 334 if (!ExitSuspendEquivalent(jt)) break; 335 336 // We have acquired the contended monitor, but while we were 337 // waiting another thread suspended us. We don't want to enter 338 // the monitor while suspended because that would surprise the 339 // thread that suspended us. 340 // 341 _recursions = 0; 342 _succ = NULL; 343 exit(false, Self); 344 345 jt->java_suspend_self(); 346 } 347 Self->set_current_pending_monitor(NULL); 348 349 // We cleared the pending monitor info since we've just gotten past 350 // the enter-check-for-suspend dance and we now own the monitor free 351 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 352 // destructor can go to a safepoint at the end of this block. If we 353 // do a thread dump during that safepoint, then this thread will show 354 // as having "-locked" the monitor, but the OS and java.lang.Thread 355 // states will still report that the thread is blocked trying to 356 // acquire it. 357 } 358 359 Atomic::dec(&_contentions); 360 assert(_contentions >= 0, "invariant"); 361 Self->_Stalled = 0; 362 363 // Must either set _recursions = 0 or ASSERT _recursions == 0. 364 assert(_recursions == 0, "invariant"); 365 assert(_owner == Self, "invariant"); 366 assert(_succ != Self, "invariant"); 367 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 368 369 // The thread -- now the owner -- is back in vm mode. 370 // Report the glorious news via TI,DTrace and jvmstat. 371 // The probe effect is non-trivial. All the reportage occurs 372 // while we hold the monitor, increasing the length of the critical 373 // section. Amdahl's parallel speedup law comes vividly into play. 374 // 375 // Another option might be to aggregate the events (thread local or 376 // per-monitor aggregation) and defer reporting until a more opportune 377 // time -- such as next time some thread encounters contention but has 378 // yet to acquire the lock. While spinning that thread could 379 // spinning we could increment JVMStat counters, etc. 380 381 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 382 if (JvmtiExport::should_post_monitor_contended_entered()) { 383 JvmtiExport::post_monitor_contended_entered(jt, this); 384 385 // The current thread already owns the monitor and is not going to 386 // call park() for the remainder of the monitor enter protocol. So 387 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 388 // event handler consumed an unpark() issued by the thread that 389 // just exited the monitor. 390 } 391 if (event.should_commit()) { 392 event.set_previousOwner((uintptr_t)_previous_owner_tid); 393 event.commit(); 394 } 395 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 396 } 397 398 // Caveat: TryLock() is not necessarily serializing if it returns failure. 399 // Callers must compensate as needed. 400 401 int ObjectMonitor::TryLock(Thread * Self) { 402 void * own = _owner; 403 if (own != NULL) return 0; 404 if (try_set_owner_from(NULL, Self) == NULL) { 405 assert(_recursions == 0, "invariant"); 406 return 1; 407 } 408 // The lock had been free momentarily, but we lost the race to the lock. 409 // Interference -- the CAS failed. 410 // We can either return -1 or retry. 411 // Retry doesn't make as much sense because the lock was just acquired. 412 return -1; 413 } 414 415 // Convert the fields used by is_busy() to a string that can be 416 // used for diagnostic output. 417 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 418 ss->print("is_busy: contentions=%d, waiters=%d, owner=" INTPTR_FORMAT 419 ", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, _contentions, 420 _waiters, p2i(_owner), p2i(_cxq), p2i(_EntryList)); 421 return ss->base(); 422 } 423 424 #define MAX_RECHECK_INTERVAL 1000 425 426 void ObjectMonitor::EnterI(TRAPS) { 427 Thread * const Self = THREAD; 428 assert(Self->is_Java_thread(), "invariant"); 429 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 430 431 // Try the lock - TATAS 432 if (TryLock (Self) > 0) { 433 assert(_succ != Self, "invariant"); 434 assert(_owner == Self, "invariant"); 435 assert(_Responsible != Self, "invariant"); 436 return; 437 } 438 439 assert(InitDone, "Unexpectedly not initialized"); 440 441 // We try one round of spinning *before* enqueueing Self. 442 // 443 // If the _owner is ready but OFFPROC we could use a YieldTo() 444 // operation to donate the remainder of this thread's quantum 445 // to the owner. This has subtle but beneficial affinity 446 // effects. 447 448 if (TrySpin(Self) > 0) { 449 assert(_owner == Self, "invariant"); 450 assert(_succ != Self, "invariant"); 451 assert(_Responsible != Self, "invariant"); 452 return; 453 } 454 455 // The Spin failed -- Enqueue and park the thread ... 456 assert(_succ != Self, "invariant"); 457 assert(_owner != Self, "invariant"); 458 assert(_Responsible != Self, "invariant"); 459 460 // Enqueue "Self" on ObjectMonitor's _cxq. 461 // 462 // Node acts as a proxy for Self. 463 // As an aside, if were to ever rewrite the synchronization code mostly 464 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 465 // Java objects. This would avoid awkward lifecycle and liveness issues, 466 // as well as eliminate a subset of ABA issues. 467 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 468 469 ObjectWaiter node(Self); 470 Self->_ParkEvent->reset(); 471 node._prev = (ObjectWaiter *) 0xBAD; 472 node.TState = ObjectWaiter::TS_CXQ; 473 474 // Push "Self" onto the front of the _cxq. 475 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 476 // Note that spinning tends to reduce the rate at which threads 477 // enqueue and dequeue on EntryList|cxq. 478 ObjectWaiter * nxt; 479 for (;;) { 480 node._next = nxt = _cxq; 481 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 482 483 // Interference - the CAS failed because _cxq changed. Just retry. 484 // As an optional optimization we retry the lock. 485 if (TryLock (Self) > 0) { 486 assert(_succ != Self, "invariant"); 487 assert(_owner == Self, "invariant"); 488 assert(_Responsible != Self, "invariant"); 489 return; 490 } 491 } 492 493 // Check for cxq|EntryList edge transition to non-null. This indicates 494 // the onset of contention. While contention persists exiting threads 495 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 496 // operations revert to the faster 1-0 mode. This enter operation may interleave 497 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 498 // arrange for one of the contending thread to use a timed park() operations 499 // to detect and recover from the race. (Stranding is form of progress failure 500 // where the monitor is unlocked but all the contending threads remain parked). 501 // That is, at least one of the contended threads will periodically poll _owner. 502 // One of the contending threads will become the designated "Responsible" thread. 503 // The Responsible thread uses a timed park instead of a normal indefinite park 504 // operation -- it periodically wakes and checks for and recovers from potential 505 // strandings admitted by 1-0 exit operations. We need at most one Responsible 506 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 507 // be responsible for a monitor. 508 // 509 // Currently, one of the contended threads takes on the added role of "Responsible". 510 // A viable alternative would be to use a dedicated "stranding checker" thread 511 // that periodically iterated over all the threads (or active monitors) and unparked 512 // successors where there was risk of stranding. This would help eliminate the 513 // timer scalability issues we see on some platforms as we'd only have one thread 514 // -- the checker -- parked on a timer. 515 516 if (nxt == NULL && _EntryList == NULL) { 517 // Try to assume the role of responsible thread for the monitor. 518 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 519 Atomic::replace_if_null(&_Responsible, Self); 520 } 521 522 // The lock might have been released while this thread was occupied queueing 523 // itself onto _cxq. To close the race and avoid "stranding" and 524 // progress-liveness failure we must resample-retry _owner before parking. 525 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 526 // In this case the ST-MEMBAR is accomplished with CAS(). 527 // 528 // TODO: Defer all thread state transitions until park-time. 529 // Since state transitions are heavy and inefficient we'd like 530 // to defer the state transitions until absolutely necessary, 531 // and in doing so avoid some transitions ... 532 533 int nWakeups = 0; 534 int recheckInterval = 1; 535 536 for (;;) { 537 538 if (TryLock(Self) > 0) break; 539 assert(_owner != Self, "invariant"); 540 541 // park self 542 if (_Responsible == Self) { 543 Self->_ParkEvent->park((jlong) recheckInterval); 544 // Increase the recheckInterval, but clamp the value. 545 recheckInterval *= 8; 546 if (recheckInterval > MAX_RECHECK_INTERVAL) { 547 recheckInterval = MAX_RECHECK_INTERVAL; 548 } 549 } else { 550 Self->_ParkEvent->park(); 551 } 552 553 if (TryLock(Self) > 0) break; 554 555 // The lock is still contested. 556 // Keep a tally of the # of futile wakeups. 557 // Note that the counter is not protected by a lock or updated by atomics. 558 // That is by design - we trade "lossy" counters which are exposed to 559 // races during updates for a lower probe effect. 560 561 // This PerfData object can be used in parallel with a safepoint. 562 // See the work around in PerfDataManager::destroy(). 563 OM_PERFDATA_OP(FutileWakeups, inc()); 564 ++nWakeups; 565 566 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 567 // We can defer clearing _succ until after the spin completes 568 // TrySpin() must tolerate being called with _succ == Self. 569 // Try yet another round of adaptive spinning. 570 if (TrySpin(Self) > 0) break; 571 572 // We can find that we were unpark()ed and redesignated _succ while 573 // we were spinning. That's harmless. If we iterate and call park(), 574 // park() will consume the event and return immediately and we'll 575 // just spin again. This pattern can repeat, leaving _succ to simply 576 // spin on a CPU. 577 578 if (_succ == Self) _succ = NULL; 579 580 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 581 OrderAccess::fence(); 582 } 583 584 // Egress : 585 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 586 // Normally we'll find Self on the EntryList . 587 // From the perspective of the lock owner (this thread), the 588 // EntryList is stable and cxq is prepend-only. 589 // The head of cxq is volatile but the interior is stable. 590 // In addition, Self.TState is stable. 591 592 assert(_owner == Self, "invariant"); 593 assert(object() != NULL, "invariant"); 594 // I'd like to write: 595 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 596 // but as we're at a safepoint that's not safe. 597 598 UnlinkAfterAcquire(Self, &node); 599 if (_succ == Self) _succ = NULL; 600 601 assert(_succ != Self, "invariant"); 602 if (_Responsible == Self) { 603 _Responsible = NULL; 604 OrderAccess::fence(); // Dekker pivot-point 605 606 // We may leave threads on cxq|EntryList without a designated 607 // "Responsible" thread. This is benign. When this thread subsequently 608 // exits the monitor it can "see" such preexisting "old" threads -- 609 // threads that arrived on the cxq|EntryList before the fence, above -- 610 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 611 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 612 // non-null and elect a new "Responsible" timer thread. 613 // 614 // This thread executes: 615 // ST Responsible=null; MEMBAR (in enter epilogue - here) 616 // LD cxq|EntryList (in subsequent exit) 617 // 618 // Entering threads in the slow/contended path execute: 619 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 620 // The (ST cxq; MEMBAR) is accomplished with CAS(). 621 // 622 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 623 // exit operation from floating above the ST Responsible=null. 624 } 625 626 // We've acquired ownership with CAS(). 627 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 628 // But since the CAS() this thread may have also stored into _succ, 629 // EntryList, cxq or Responsible. These meta-data updates must be 630 // visible __before this thread subsequently drops the lock. 631 // Consider what could occur if we didn't enforce this constraint -- 632 // STs to monitor meta-data and user-data could reorder with (become 633 // visible after) the ST in exit that drops ownership of the lock. 634 // Some other thread could then acquire the lock, but observe inconsistent 635 // or old monitor meta-data and heap data. That violates the JMM. 636 // To that end, the 1-0 exit() operation must have at least STST|LDST 637 // "release" barrier semantics. Specifically, there must be at least a 638 // STST|LDST barrier in exit() before the ST of null into _owner that drops 639 // the lock. The barrier ensures that changes to monitor meta-data and data 640 // protected by the lock will be visible before we release the lock, and 641 // therefore before some other thread (CPU) has a chance to acquire the lock. 642 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 643 // 644 // Critically, any prior STs to _succ or EntryList must be visible before 645 // the ST of null into _owner in the *subsequent* (following) corresponding 646 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 647 // execute a serializing instruction. 648 649 return; 650 } 651 652 // ReenterI() is a specialized inline form of the latter half of the 653 // contended slow-path from EnterI(). We use ReenterI() only for 654 // monitor reentry in wait(). 655 // 656 // In the future we should reconcile EnterI() and ReenterI(). 657 658 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 659 assert(Self != NULL, "invariant"); 660 assert(SelfNode != NULL, "invariant"); 661 assert(SelfNode->_thread == Self, "invariant"); 662 assert(_waiters > 0, "invariant"); 663 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 664 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 665 JavaThread * jt = (JavaThread *) Self; 666 667 int nWakeups = 0; 668 for (;;) { 669 ObjectWaiter::TStates v = SelfNode->TState; 670 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 671 assert(_owner != Self, "invariant"); 672 673 if (TryLock(Self) > 0) break; 674 if (TrySpin(Self) > 0) break; 675 676 // State transition wrappers around park() ... 677 // ReenterI() wisely defers state transitions until 678 // it's clear we must park the thread. 679 { 680 OSThreadContendState osts(Self->osthread()); 681 ThreadBlockInVM tbivm(jt); 682 683 // cleared by handle_special_suspend_equivalent_condition() 684 // or java_suspend_self() 685 jt->set_suspend_equivalent(); 686 Self->_ParkEvent->park(); 687 688 // were we externally suspended while we were waiting? 689 for (;;) { 690 if (!ExitSuspendEquivalent(jt)) break; 691 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 692 jt->java_suspend_self(); 693 jt->set_suspend_equivalent(); 694 } 695 } 696 697 // Try again, but just so we distinguish between futile wakeups and 698 // successful wakeups. The following test isn't algorithmically 699 // necessary, but it helps us maintain sensible statistics. 700 if (TryLock(Self) > 0) break; 701 702 // The lock is still contested. 703 // Keep a tally of the # of futile wakeups. 704 // Note that the counter is not protected by a lock or updated by atomics. 705 // That is by design - we trade "lossy" counters which are exposed to 706 // races during updates for a lower probe effect. 707 ++nWakeups; 708 709 // Assuming this is not a spurious wakeup we'll normally 710 // find that _succ == Self. 711 if (_succ == Self) _succ = NULL; 712 713 // Invariant: after clearing _succ a contending thread 714 // *must* retry _owner before parking. 715 OrderAccess::fence(); 716 717 // This PerfData object can be used in parallel with a safepoint. 718 // See the work around in PerfDataManager::destroy(). 719 OM_PERFDATA_OP(FutileWakeups, inc()); 720 } 721 722 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 723 // Normally we'll find Self on the EntryList. 724 // Unlinking from the EntryList is constant-time and atomic-free. 725 // From the perspective of the lock owner (this thread), the 726 // EntryList is stable and cxq is prepend-only. 727 // The head of cxq is volatile but the interior is stable. 728 // In addition, Self.TState is stable. 729 730 assert(_owner == Self, "invariant"); 731 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 732 UnlinkAfterAcquire(Self, SelfNode); 733 if (_succ == Self) _succ = NULL; 734 assert(_succ != Self, "invariant"); 735 SelfNode->TState = ObjectWaiter::TS_RUN; 736 OrderAccess::fence(); // see comments at the end of EnterI() 737 } 738 739 // By convention we unlink a contending thread from EntryList|cxq immediately 740 // after the thread acquires the lock in ::enter(). Equally, we could defer 741 // unlinking the thread until ::exit()-time. 742 743 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 744 assert(_owner == Self, "invariant"); 745 assert(SelfNode->_thread == Self, "invariant"); 746 747 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 748 // Normal case: remove Self from the DLL EntryList . 749 // This is a constant-time operation. 750 ObjectWaiter * nxt = SelfNode->_next; 751 ObjectWaiter * prv = SelfNode->_prev; 752 if (nxt != NULL) nxt->_prev = prv; 753 if (prv != NULL) prv->_next = nxt; 754 if (SelfNode == _EntryList) _EntryList = nxt; 755 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 756 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 757 } else { 758 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 759 // Inopportune interleaving -- Self is still on the cxq. 760 // This usually means the enqueue of self raced an exiting thread. 761 // Normally we'll find Self near the front of the cxq, so 762 // dequeueing is typically fast. If needbe we can accelerate 763 // this with some MCS/CHL-like bidirectional list hints and advisory 764 // back-links so dequeueing from the interior will normally operate 765 // in constant-time. 766 // Dequeue Self from either the head (with CAS) or from the interior 767 // with a linear-time scan and normal non-atomic memory operations. 768 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 769 // and then unlink Self from EntryList. We have to drain eventually, 770 // so it might as well be now. 771 772 ObjectWaiter * v = _cxq; 773 assert(v != NULL, "invariant"); 774 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) { 775 // The CAS above can fail from interference IFF a "RAT" arrived. 776 // In that case Self must be in the interior and can no longer be 777 // at the head of cxq. 778 if (v == SelfNode) { 779 assert(_cxq != v, "invariant"); 780 v = _cxq; // CAS above failed - start scan at head of list 781 } 782 ObjectWaiter * p; 783 ObjectWaiter * q = NULL; 784 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 785 q = p; 786 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 787 } 788 assert(v != SelfNode, "invariant"); 789 assert(p == SelfNode, "Node not found on cxq"); 790 assert(p != _cxq, "invariant"); 791 assert(q != NULL, "invariant"); 792 assert(q->_next == p, "invariant"); 793 q->_next = p->_next; 794 } 795 } 796 797 #ifdef ASSERT 798 // Diagnostic hygiene ... 799 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 800 SelfNode->_next = (ObjectWaiter *) 0xBAD; 801 SelfNode->TState = ObjectWaiter::TS_RUN; 802 #endif 803 } 804 805 // ----------------------------------------------------------------------------- 806 // Exit support 807 // 808 // exit() 809 // ~~~~~~ 810 // Note that the collector can't reclaim the objectMonitor or deflate 811 // the object out from underneath the thread calling ::exit() as the 812 // thread calling ::exit() never transitions to a stable state. 813 // This inhibits GC, which in turn inhibits asynchronous (and 814 // inopportune) reclamation of "this". 815 // 816 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 817 // There's one exception to the claim above, however. EnterI() can call 818 // exit() to drop a lock if the acquirer has been externally suspended. 819 // In that case exit() is called with _thread_state as _thread_blocked, 820 // but the monitor's _contentions field is > 0, which inhibits reclamation. 821 // 822 // 1-0 exit 823 // ~~~~~~~~ 824 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 825 // the fast-path operators have been optimized so the common ::exit() 826 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 827 // The code emitted by fast_unlock() elides the usual MEMBAR. This 828 // greatly improves latency -- MEMBAR and CAS having considerable local 829 // latency on modern processors -- but at the cost of "stranding". Absent the 830 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 831 // ::enter() path, resulting in the entering thread being stranding 832 // and a progress-liveness failure. Stranding is extremely rare. 833 // We use timers (timed park operations) & periodic polling to detect 834 // and recover from stranding. Potentially stranded threads periodically 835 // wake up and poll the lock. See the usage of the _Responsible variable. 836 // 837 // The CAS() in enter provides for safety and exclusion, while the CAS or 838 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 839 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 840 // We detect and recover from stranding with timers. 841 // 842 // If a thread transiently strands it'll park until (a) another 843 // thread acquires the lock and then drops the lock, at which time the 844 // exiting thread will notice and unpark the stranded thread, or, (b) 845 // the timer expires. If the lock is high traffic then the stranding latency 846 // will be low due to (a). If the lock is low traffic then the odds of 847 // stranding are lower, although the worst-case stranding latency 848 // is longer. Critically, we don't want to put excessive load in the 849 // platform's timer subsystem. We want to minimize both the timer injection 850 // rate (timers created/sec) as well as the number of timers active at 851 // any one time. (more precisely, we want to minimize timer-seconds, which is 852 // the integral of the # of active timers at any instant over time). 853 // Both impinge on OS scalability. Given that, at most one thread parked on 854 // a monitor will use a timer. 855 // 856 // There is also the risk of a futile wake-up. If we drop the lock 857 // another thread can reacquire the lock immediately, and we can 858 // then wake a thread unnecessarily. This is benign, and we've 859 // structured the code so the windows are short and the frequency 860 // of such futile wakups is low. 861 862 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 863 Thread * const Self = THREAD; 864 if (THREAD != _owner) { 865 void* cur = _owner; 866 if (THREAD->is_lock_owned((address)cur)) { 867 assert(_recursions == 0, "invariant"); 868 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 869 _recursions = 0; 870 } else { 871 // Apparent unbalanced locking ... 872 // Naively we'd like to throw IllegalMonitorStateException. 873 // As a practical matter we can neither allocate nor throw an 874 // exception as ::exit() can be called from leaf routines. 875 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 876 // Upon deeper reflection, however, in a properly run JVM the only 877 // way we should encounter this situation is in the presence of 878 // unbalanced JNI locking. TODO: CheckJNICalls. 879 // See also: CR4414101 880 #ifdef ASSERT 881 LogStreamHandle(Error, monitorinflation) lsh; 882 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 883 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 884 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 885 print_debug_style_on(&lsh); 886 #endif 887 assert(false, "Non-balanced monitor enter/exit!"); 888 return; 889 } 890 } 891 892 if (_recursions != 0) { 893 _recursions--; // this is simple recursive enter 894 return; 895 } 896 897 // Invariant: after setting Responsible=null an thread must execute 898 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 899 _Responsible = NULL; 900 901 #if INCLUDE_JFR 902 // get the owner's thread id for the MonitorEnter event 903 // if it is enabled and the thread isn't suspended 904 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 905 _previous_owner_tid = JFR_THREAD_ID(Self); 906 } 907 #endif 908 909 for (;;) { 910 assert(THREAD == _owner, "invariant"); 911 912 // Drop the lock. 913 // release semantics: prior loads and stores from within the critical section 914 // must not float (reorder) past the following store that drops the lock. 915 // Uses a storeload to separate release_store(owner) from the 916 // successor check. The try_set_owner() below uses cmpxchg() so 917 // we get the fence down there. 918 release_clear_owner(Self); 919 OrderAccess::storeload(); 920 921 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 (try_set_owner_from(NULL, Self) != NULL) { 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 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1096 release_clear_owner(Self); 1097 OrderAccess::fence(); 1098 1099 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1100 Trigger->unpark(); 1101 1102 // Maintain stats and report events to JVMTI 1103 OM_PERFDATA_OP(Parks, inc()); 1104 } 1105 1106 1107 // ----------------------------------------------------------------------------- 1108 // Class Loader deadlock handling. 1109 // 1110 // complete_exit exits a lock returning recursion count 1111 // complete_exit/reenter operate as a wait without waiting 1112 // complete_exit requires an inflated monitor 1113 // The _owner field is not always the Thread addr even with an 1114 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1115 // thread due to contention. 1116 intx ObjectMonitor::complete_exit(TRAPS) { 1117 Thread * const Self = THREAD; 1118 assert(Self->is_Java_thread(), "Must be Java thread!"); 1119 JavaThread *jt = (JavaThread *)THREAD; 1120 1121 assert(InitDone, "Unexpectedly not initialized"); 1122 1123 if (THREAD != _owner) { 1124 void* cur = _owner; 1125 if (THREAD->is_lock_owned((address)cur)) { 1126 assert(_recursions == 0, "internal state error"); 1127 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 1128 _recursions = 0; 1129 } 1130 } 1131 1132 guarantee(Self == _owner, "complete_exit not owner"); 1133 intx save = _recursions; // record the old recursion count 1134 _recursions = 0; // set the recursion level to be 0 1135 exit(true, Self); // exit the monitor 1136 guarantee(_owner != Self, "invariant"); 1137 return save; 1138 } 1139 1140 // reenter() enters a lock and sets recursion count 1141 // complete_exit/reenter operate as a wait without waiting 1142 void ObjectMonitor::reenter(intx recursions, TRAPS) { 1143 Thread * const Self = THREAD; 1144 assert(Self->is_Java_thread(), "Must be Java thread!"); 1145 JavaThread *jt = (JavaThread *)THREAD; 1146 1147 guarantee(_owner != Self, "reenter already owner"); 1148 enter(THREAD); // enter the monitor 1149 guarantee(_recursions == 0, "reenter recursion"); 1150 _recursions = recursions; 1151 return; 1152 } 1153 1154 // Checks that the current THREAD owns this monitor and causes an 1155 // immediate return if it doesn't. We don't use the CHECK macro 1156 // because we want the IMSE to be the only exception that is thrown 1157 // from the call site when false is returned. Any other pending 1158 // exception is ignored. 1159 #define CHECK_OWNER() \ 1160 do { \ 1161 if (!check_owner(THREAD)) { \ 1162 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1163 return; \ 1164 } \ 1165 } while (false) 1166 1167 // Returns true if the specified thread owns the ObjectMonitor. 1168 // Otherwise returns false and throws IllegalMonitorStateException 1169 // (IMSE). If there is a pending exception and the specified thread 1170 // is not the owner, that exception will be replaced by the IMSE. 1171 bool ObjectMonitor::check_owner(Thread* THREAD) { 1172 if (_owner == THREAD) { 1173 return true; 1174 } 1175 void* cur = _owner; 1176 if (THREAD->is_lock_owned((address)cur)) { 1177 set_owner_from_BasicLock(cur, THREAD); // Convert from BasicLock* to Thread*. 1178 _recursions = 0; 1179 return true; 1180 } 1181 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1182 "current thread is not owner", false); 1183 } 1184 1185 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1186 ObjectMonitor* monitor, 1187 jlong notifier_tid, 1188 jlong timeout, 1189 bool timedout) { 1190 assert(event != NULL, "invariant"); 1191 assert(monitor != NULL, "invariant"); 1192 event->set_monitorClass(((oop)monitor->object())->klass()); 1193 event->set_timeout(timeout); 1194 event->set_address((uintptr_t)monitor->object_addr()); 1195 event->set_notifier(notifier_tid); 1196 event->set_timedOut(timedout); 1197 event->commit(); 1198 } 1199 1200 // ----------------------------------------------------------------------------- 1201 // Wait/Notify/NotifyAll 1202 // 1203 // Note: a subset of changes to ObjectMonitor::wait() 1204 // will need to be replicated in complete_exit 1205 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1206 Thread * const Self = THREAD; 1207 assert(Self->is_Java_thread(), "Must be Java thread!"); 1208 JavaThread *jt = (JavaThread *)THREAD; 1209 1210 assert(InitDone, "Unexpectedly not initialized"); 1211 1212 CHECK_OWNER(); // Throws IMSE if not owner. 1213 1214 EventJavaMonitorWait event; 1215 1216 // check for a pending interrupt 1217 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1218 // post monitor waited event. Note that this is past-tense, we are done waiting. 1219 if (JvmtiExport::should_post_monitor_waited()) { 1220 // Note: 'false' parameter is passed here because the 1221 // wait was not timed out due to thread interrupt. 1222 JvmtiExport::post_monitor_waited(jt, this, false); 1223 1224 // In this short circuit of the monitor wait protocol, the 1225 // current thread never drops ownership of the monitor and 1226 // never gets added to the wait queue so the current thread 1227 // cannot be made the successor. This means that the 1228 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1229 // consume an unpark() meant for the ParkEvent associated with 1230 // this ObjectMonitor. 1231 } 1232 if (event.should_commit()) { 1233 post_monitor_wait_event(&event, this, 0, millis, false); 1234 } 1235 THROW(vmSymbols::java_lang_InterruptedException()); 1236 return; 1237 } 1238 1239 assert(Self->_Stalled == 0, "invariant"); 1240 Self->_Stalled = intptr_t(this); 1241 jt->set_current_waiting_monitor(this); 1242 1243 // create a node to be put into the queue 1244 // Critically, after we reset() the event but prior to park(), we must check 1245 // for a pending interrupt. 1246 ObjectWaiter node(Self); 1247 node.TState = ObjectWaiter::TS_WAIT; 1248 Self->_ParkEvent->reset(); 1249 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1250 1251 // Enter the waiting queue, which is a circular doubly linked list in this case 1252 // but it could be a priority queue or any data structure. 1253 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1254 // by the the owner of the monitor *except* in the case where park() 1255 // returns because of a timeout of interrupt. Contention is exceptionally rare 1256 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1257 1258 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1259 AddWaiter(&node); 1260 Thread::SpinRelease(&_WaitSetLock); 1261 1262 _Responsible = NULL; 1263 1264 intx save = _recursions; // record the old recursion count 1265 _waiters++; // increment the number of waiters 1266 _recursions = 0; // set the recursion level to be 1 1267 exit(true, Self); // exit the monitor 1268 guarantee(_owner != Self, "invariant"); 1269 1270 // The thread is on the WaitSet list - now park() it. 1271 // On MP systems it's conceivable that a brief spin before we park 1272 // could be profitable. 1273 // 1274 // TODO-FIXME: change the following logic to a loop of the form 1275 // while (!timeout && !interrupted && _notified == 0) park() 1276 1277 int ret = OS_OK; 1278 int WasNotified = 0; 1279 1280 // Need to check interrupt state whilst still _thread_in_vm 1281 bool interrupted = interruptible && jt->is_interrupted(false); 1282 1283 { // State transition wrappers 1284 OSThread* osthread = Self->osthread(); 1285 OSThreadWaitState osts(osthread, true); 1286 { 1287 ThreadBlockInVM tbivm(jt); 1288 // Thread is in thread_blocked state and oop access is unsafe. 1289 jt->set_suspend_equivalent(); 1290 1291 if (interrupted || HAS_PENDING_EXCEPTION) { 1292 // Intentionally empty 1293 } else if (node._notified == 0) { 1294 if (millis <= 0) { 1295 Self->_ParkEvent->park(); 1296 } else { 1297 ret = Self->_ParkEvent->park(millis); 1298 } 1299 } 1300 1301 // were we externally suspended while we were waiting? 1302 if (ExitSuspendEquivalent (jt)) { 1303 // TODO-FIXME: add -- if succ == Self then succ = null. 1304 jt->java_suspend_self(); 1305 } 1306 1307 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1308 1309 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1310 // from the WaitSet to the EntryList. 1311 // See if we need to remove Node from the WaitSet. 1312 // We use double-checked locking to avoid grabbing _WaitSetLock 1313 // if the thread is not on the wait queue. 1314 // 1315 // Note that we don't need a fence before the fetch of TState. 1316 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1317 // written by the is thread. (perhaps the fetch might even be satisfied 1318 // by a look-aside into the processor's own store buffer, although given 1319 // the length of the code path between the prior ST and this load that's 1320 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1321 // then we'll acquire the lock and then re-fetch a fresh TState value. 1322 // That is, we fail toward safety. 1323 1324 if (node.TState == ObjectWaiter::TS_WAIT) { 1325 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1326 if (node.TState == ObjectWaiter::TS_WAIT) { 1327 DequeueSpecificWaiter(&node); // unlink from WaitSet 1328 assert(node._notified == 0, "invariant"); 1329 node.TState = ObjectWaiter::TS_RUN; 1330 } 1331 Thread::SpinRelease(&_WaitSetLock); 1332 } 1333 1334 // The thread is now either on off-list (TS_RUN), 1335 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1336 // The Node's TState variable is stable from the perspective of this thread. 1337 // No other threads will asynchronously modify TState. 1338 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1339 OrderAccess::loadload(); 1340 if (_succ == Self) _succ = NULL; 1341 WasNotified = node._notified; 1342 1343 // Reentry phase -- reacquire the monitor. 1344 // re-enter contended monitor after object.wait(). 1345 // retain OBJECT_WAIT state until re-enter successfully completes 1346 // Thread state is thread_in_vm and oop access is again safe, 1347 // although the raw address of the object may have changed. 1348 // (Don't cache naked oops over safepoints, of course). 1349 1350 // post monitor waited event. Note that this is past-tense, we are done waiting. 1351 if (JvmtiExport::should_post_monitor_waited()) { 1352 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1353 1354 if (node._notified != 0 && _succ == Self) { 1355 // In this part of the monitor wait-notify-reenter protocol it 1356 // is possible (and normal) for another thread to do a fastpath 1357 // monitor enter-exit while this thread is still trying to get 1358 // to the reenter portion of the protocol. 1359 // 1360 // The ObjectMonitor was notified and the current thread is 1361 // the successor which also means that an unpark() has already 1362 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1363 // consume the unpark() that was done when the successor was 1364 // set because the same ParkEvent is shared between Java 1365 // monitors and JVM/TI RawMonitors (for now). 1366 // 1367 // We redo the unpark() to ensure forward progress, i.e., we 1368 // don't want all pending threads hanging (parked) with none 1369 // entering the unlocked monitor. 1370 node._event->unpark(); 1371 } 1372 } 1373 1374 if (event.should_commit()) { 1375 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1376 } 1377 1378 OrderAccess::fence(); 1379 1380 assert(Self->_Stalled != 0, "invariant"); 1381 Self->_Stalled = 0; 1382 1383 assert(_owner != Self, "invariant"); 1384 ObjectWaiter::TStates v = node.TState; 1385 if (v == ObjectWaiter::TS_RUN) { 1386 enter(Self); 1387 } else { 1388 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1389 ReenterI(Self, &node); 1390 node.wait_reenter_end(this); 1391 } 1392 1393 // Self has reacquired the lock. 1394 // Lifecycle - the node representing Self must not appear on any queues. 1395 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1396 // want residual elements associated with this thread left on any lists. 1397 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1398 assert(_owner == Self, "invariant"); 1399 assert(_succ != Self, "invariant"); 1400 } // OSThreadWaitState() 1401 1402 jt->set_current_waiting_monitor(NULL); 1403 1404 guarantee(_recursions == 0, "invariant"); 1405 _recursions = save; // restore the old recursion count 1406 _waiters--; // decrement the number of waiters 1407 1408 // Verify a few postconditions 1409 assert(_owner == Self, "invariant"); 1410 assert(_succ != Self, "invariant"); 1411 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1412 1413 // check if the notification happened 1414 if (!WasNotified) { 1415 // no, it could be timeout or Thread.interrupt() or both 1416 // check for interrupt event, otherwise it is timeout 1417 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1418 THROW(vmSymbols::java_lang_InterruptedException()); 1419 } 1420 } 1421 1422 // NOTE: Spurious wake up will be consider as timeout. 1423 // Monitor notify has precedence over thread interrupt. 1424 } 1425 1426 1427 // Consider: 1428 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1429 // then instead of transferring a thread from the WaitSet to the EntryList 1430 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1431 1432 void ObjectMonitor::INotify(Thread * Self) { 1433 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1434 ObjectWaiter * iterator = DequeueWaiter(); 1435 if (iterator != NULL) { 1436 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1437 guarantee(iterator->_notified == 0, "invariant"); 1438 // Disposition - what might we do with iterator ? 1439 // a. add it directly to the EntryList - either tail (policy == 1) 1440 // or head (policy == 0). 1441 // b. push it onto the front of the _cxq (policy == 2). 1442 // For now we use (b). 1443 1444 iterator->TState = ObjectWaiter::TS_ENTER; 1445 1446 iterator->_notified = 1; 1447 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1448 1449 ObjectWaiter * list = _EntryList; 1450 if (list != NULL) { 1451 assert(list->_prev == NULL, "invariant"); 1452 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1453 assert(list != iterator, "invariant"); 1454 } 1455 1456 // prepend to cxq 1457 if (list == NULL) { 1458 iterator->_next = iterator->_prev = NULL; 1459 _EntryList = iterator; 1460 } else { 1461 iterator->TState = ObjectWaiter::TS_CXQ; 1462 for (;;) { 1463 ObjectWaiter * front = _cxq; 1464 iterator->_next = front; 1465 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1466 break; 1467 } 1468 } 1469 } 1470 1471 // _WaitSetLock protects the wait queue, not the EntryList. We could 1472 // move the add-to-EntryList operation, above, outside the critical section 1473 // protected by _WaitSetLock. In practice that's not useful. With the 1474 // exception of wait() timeouts and interrupts the monitor owner 1475 // is the only thread that grabs _WaitSetLock. There's almost no contention 1476 // on _WaitSetLock so it's not profitable to reduce the length of the 1477 // critical section. 1478 1479 iterator->wait_reenter_begin(this); 1480 } 1481 Thread::SpinRelease(&_WaitSetLock); 1482 } 1483 1484 // Consider: a not-uncommon synchronization bug is to use notify() when 1485 // notifyAll() is more appropriate, potentially resulting in stranded 1486 // threads; this is one example of a lost wakeup. A useful diagnostic 1487 // option is to force all notify() operations to behave as notifyAll(). 1488 // 1489 // Note: We can also detect many such problems with a "minimum wait". 1490 // When the "minimum wait" is set to a small non-zero timeout value 1491 // and the program does not hang whereas it did absent "minimum wait", 1492 // that suggests a lost wakeup bug. 1493 1494 void ObjectMonitor::notify(TRAPS) { 1495 CHECK_OWNER(); // Throws IMSE if not owner. 1496 if (_WaitSet == NULL) { 1497 return; 1498 } 1499 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1500 INotify(THREAD); 1501 OM_PERFDATA_OP(Notifications, inc(1)); 1502 } 1503 1504 1505 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1506 // from the waitset to the EntryList. This could be done more efficiently with a 1507 // single bulk transfer but in practice it's not time-critical. Beware too, 1508 // that in prepend-mode we invert the order of the waiters. Let's say that the 1509 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1510 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1511 1512 void ObjectMonitor::notifyAll(TRAPS) { 1513 CHECK_OWNER(); // Throws IMSE if not owner. 1514 if (_WaitSet == NULL) { 1515 return; 1516 } 1517 1518 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1519 int tally = 0; 1520 while (_WaitSet != NULL) { 1521 tally++; 1522 INotify(THREAD); 1523 } 1524 1525 OM_PERFDATA_OP(Notifications, inc(tally)); 1526 } 1527 1528 // ----------------------------------------------------------------------------- 1529 // Adaptive Spinning Support 1530 // 1531 // Adaptive spin-then-block - rational spinning 1532 // 1533 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1534 // algorithm. On high order SMP systems it would be better to start with 1535 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1536 // a contending thread could enqueue itself on the cxq and then spin locally 1537 // on a thread-specific variable such as its ParkEvent._Event flag. 1538 // That's left as an exercise for the reader. Note that global spinning is 1539 // not problematic on Niagara, as the L2 cache serves the interconnect and 1540 // has both low latency and massive bandwidth. 1541 // 1542 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1543 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1544 // (duration) or we can fix the count at approximately the duration of 1545 // a context switch and vary the frequency. Of course we could also 1546 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1547 // For a description of 'Adaptive spin-then-block mutual exclusion in 1548 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1549 // 1550 // This implementation varies the duration "D", where D varies with 1551 // the success rate of recent spin attempts. (D is capped at approximately 1552 // length of a round-trip context switch). The success rate for recent 1553 // spin attempts is a good predictor of the success rate of future spin 1554 // attempts. The mechanism adapts automatically to varying critical 1555 // section length (lock modality), system load and degree of parallelism. 1556 // D is maintained per-monitor in _SpinDuration and is initialized 1557 // optimistically. Spin frequency is fixed at 100%. 1558 // 1559 // Note that _SpinDuration is volatile, but we update it without locks 1560 // or atomics. The code is designed so that _SpinDuration stays within 1561 // a reasonable range even in the presence of races. The arithmetic 1562 // operations on _SpinDuration are closed over the domain of legal values, 1563 // so at worst a race will install and older but still legal value. 1564 // At the very worst this introduces some apparent non-determinism. 1565 // We might spin when we shouldn't or vice-versa, but since the spin 1566 // count are relatively short, even in the worst case, the effect is harmless. 1567 // 1568 // Care must be taken that a low "D" value does not become an 1569 // an absorbing state. Transient spinning failures -- when spinning 1570 // is overall profitable -- should not cause the system to converge 1571 // on low "D" values. We want spinning to be stable and predictable 1572 // and fairly responsive to change and at the same time we don't want 1573 // it to oscillate, become metastable, be "too" non-deterministic, 1574 // or converge on or enter undesirable stable absorbing states. 1575 // 1576 // We implement a feedback-based control system -- using past behavior 1577 // to predict future behavior. We face two issues: (a) if the 1578 // input signal is random then the spin predictor won't provide optimal 1579 // results, and (b) if the signal frequency is too high then the control 1580 // system, which has some natural response lag, will "chase" the signal. 1581 // (b) can arise from multimodal lock hold times. Transient preemption 1582 // can also result in apparent bimodal lock hold times. 1583 // Although sub-optimal, neither condition is particularly harmful, as 1584 // in the worst-case we'll spin when we shouldn't or vice-versa. 1585 // The maximum spin duration is rather short so the failure modes aren't bad. 1586 // To be conservative, I've tuned the gain in system to bias toward 1587 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1588 // "rings" or oscillates between spinning and not spinning. This happens 1589 // when spinning is just on the cusp of profitability, however, so the 1590 // situation is not dire. The state is benign -- there's no need to add 1591 // hysteresis control to damp the transition rate between spinning and 1592 // not spinning. 1593 1594 // Spinning: Fixed frequency (100%), vary duration 1595 int ObjectMonitor::TrySpin(Thread * Self) { 1596 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1597 int ctr = Knob_FixedSpin; 1598 if (ctr != 0) { 1599 while (--ctr >= 0) { 1600 if (TryLock(Self) > 0) return 1; 1601 SpinPause(); 1602 } 1603 return 0; 1604 } 1605 1606 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1607 if (TryLock(Self) > 0) { 1608 // Increase _SpinDuration ... 1609 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1610 // Raising _SpurDuration to the poverty line is key. 1611 int x = _SpinDuration; 1612 if (x < Knob_SpinLimit) { 1613 if (x < Knob_Poverty) x = Knob_Poverty; 1614 _SpinDuration = x + Knob_BonusB; 1615 } 1616 return 1; 1617 } 1618 SpinPause(); 1619 } 1620 1621 // Admission control - verify preconditions for spinning 1622 // 1623 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1624 // becoming an absorbing state. Put another way, we spin briefly to 1625 // sample, just in case the system load, parallelism, contention, or lock 1626 // modality changed. 1627 // 1628 // Consider the following alternative: 1629 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1630 // spin attempt. "Periodically" might mean after a tally of 1631 // the # of failed spin attempts (or iterations) reaches some threshold. 1632 // This takes us into the realm of 1-out-of-N spinning, where we 1633 // hold the duration constant but vary the frequency. 1634 1635 ctr = _SpinDuration; 1636 if (ctr <= 0) return 0; 1637 1638 if (NotRunnable(Self, (Thread *) _owner)) { 1639 return 0; 1640 } 1641 1642 // We're good to spin ... spin ingress. 1643 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1644 // when preparing to LD...CAS _owner, etc and the CAS is likely 1645 // to succeed. 1646 if (_succ == NULL) { 1647 _succ = Self; 1648 } 1649 Thread * prv = NULL; 1650 1651 // There are three ways to exit the following loop: 1652 // 1. A successful spin where this thread has acquired the lock. 1653 // 2. Spin failure with prejudice 1654 // 3. Spin failure without prejudice 1655 1656 while (--ctr >= 0) { 1657 1658 // Periodic polling -- Check for pending GC 1659 // Threads may spin while they're unsafe. 1660 // We don't want spinning threads to delay the JVM from reaching 1661 // a stop-the-world safepoint or to steal cycles from GC. 1662 // If we detect a pending safepoint we abort in order that 1663 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1664 // this thread, if safe, doesn't steal cycles from GC. 1665 // This is in keeping with the "no loitering in runtime" rule. 1666 // We periodically check to see if there's a safepoint pending. 1667 if ((ctr & 0xFF) == 0) { 1668 if (SafepointMechanism::should_block(Self)) { 1669 goto Abort; // abrupt spin egress 1670 } 1671 SpinPause(); 1672 } 1673 1674 // Probe _owner with TATAS 1675 // If this thread observes the monitor transition or flicker 1676 // from locked to unlocked to locked, then the odds that this 1677 // thread will acquire the lock in this spin attempt go down 1678 // considerably. The same argument applies if the CAS fails 1679 // or if we observe _owner change from one non-null value to 1680 // another non-null value. In such cases we might abort 1681 // the spin without prejudice or apply a "penalty" to the 1682 // spin count-down variable "ctr", reducing it by 100, say. 1683 1684 Thread * ox = (Thread *) _owner; 1685 if (ox == NULL) { 1686 ox = (Thread*)try_set_owner_from(NULL, Self); 1687 if (ox == NULL) { 1688 // The CAS succeeded -- this thread acquired ownership 1689 // Take care of some bookkeeping to exit spin state. 1690 if (_succ == Self) { 1691 _succ = NULL; 1692 } 1693 1694 // Increase _SpinDuration : 1695 // The spin was successful (profitable) so we tend toward 1696 // longer spin attempts in the future. 1697 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1698 // If we acquired the lock early in the spin cycle it 1699 // makes sense to increase _SpinDuration proportionally. 1700 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1701 int x = _SpinDuration; 1702 if (x < Knob_SpinLimit) { 1703 if (x < Knob_Poverty) x = Knob_Poverty; 1704 _SpinDuration = x + Knob_Bonus; 1705 } 1706 return 1; 1707 } 1708 1709 // The CAS failed ... we can take any of the following actions: 1710 // * penalize: ctr -= CASPenalty 1711 // * exit spin with prejudice -- goto Abort; 1712 // * exit spin without prejudice. 1713 // * Since CAS is high-latency, retry again immediately. 1714 prv = ox; 1715 goto Abort; 1716 } 1717 1718 // Did lock ownership change hands ? 1719 if (ox != prv && prv != NULL) { 1720 goto Abort; 1721 } 1722 prv = ox; 1723 1724 // Abort the spin if the owner is not executing. 1725 // The owner must be executing in order to drop the lock. 1726 // Spinning while the owner is OFFPROC is idiocy. 1727 // Consider: ctr -= RunnablePenalty ; 1728 if (NotRunnable(Self, ox)) { 1729 goto Abort; 1730 } 1731 if (_succ == NULL) { 1732 _succ = Self; 1733 } 1734 } 1735 1736 // Spin failed with prejudice -- reduce _SpinDuration. 1737 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1738 // AIMD is globally stable. 1739 { 1740 int x = _SpinDuration; 1741 if (x > 0) { 1742 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1743 // This is globally sample and tends to damp the response. 1744 x -= Knob_Penalty; 1745 if (x < 0) x = 0; 1746 _SpinDuration = x; 1747 } 1748 } 1749 1750 Abort: 1751 if (_succ == Self) { 1752 _succ = NULL; 1753 // Invariant: after setting succ=null a contending thread 1754 // must recheck-retry _owner before parking. This usually happens 1755 // in the normal usage of TrySpin(), but it's safest 1756 // to make TrySpin() as foolproof as possible. 1757 OrderAccess::fence(); 1758 if (TryLock(Self) > 0) return 1; 1759 } 1760 return 0; 1761 } 1762 1763 // NotRunnable() -- informed spinning 1764 // 1765 // Don't bother spinning if the owner is not eligible to drop the lock. 1766 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1767 // The thread must be runnable in order to drop the lock in timely fashion. 1768 // If the _owner is not runnable then spinning will not likely be 1769 // successful (profitable). 1770 // 1771 // Beware -- the thread referenced by _owner could have died 1772 // so a simply fetch from _owner->_thread_state might trap. 1773 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1774 // Because of the lifecycle issues, the _thread_state values 1775 // observed by NotRunnable() might be garbage. NotRunnable must 1776 // tolerate this and consider the observed _thread_state value 1777 // as advisory. 1778 // 1779 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1780 // a thread pointer. 1781 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1782 // with the LSB of _owner. Another option would be to probabilistically probe 1783 // the putative _owner->TypeTag value. 1784 // 1785 // Checking _thread_state isn't perfect. Even if the thread is 1786 // in_java it might be blocked on a page-fault or have been preempted 1787 // and sitting on a ready/dispatch queue. 1788 // 1789 // The return value from NotRunnable() is *advisory* -- the 1790 // result is based on sampling and is not necessarily coherent. 1791 // The caller must tolerate false-negative and false-positive errors. 1792 // Spinning, in general, is probabilistic anyway. 1793 1794 1795 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1796 // Check ox->TypeTag == 2BAD. 1797 if (ox == NULL) return 0; 1798 1799 // Avoid transitive spinning ... 1800 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1801 // Immediately after T1 acquires L it's possible that T2, also 1802 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1803 // This occurs transiently after T1 acquired L but before 1804 // T1 managed to clear T1.Stalled. T2 does not need to abort 1805 // its spin in this circumstance. 1806 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1807 1808 if (BlockedOn == 1) return 1; 1809 if (BlockedOn != 0) { 1810 return BlockedOn != intptr_t(this) && _owner == ox; 1811 } 1812 1813 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1814 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1815 // consider also: jst != _thread_in_Java -- but that's overspecific. 1816 return jst == _thread_blocked || jst == _thread_in_native; 1817 } 1818 1819 1820 // ----------------------------------------------------------------------------- 1821 // WaitSet management ... 1822 1823 ObjectWaiter::ObjectWaiter(Thread* thread) { 1824 _next = NULL; 1825 _prev = NULL; 1826 _notified = 0; 1827 _notifier_tid = 0; 1828 TState = TS_RUN; 1829 _thread = thread; 1830 _event = thread->_ParkEvent; 1831 _active = false; 1832 assert(_event != NULL, "invariant"); 1833 } 1834 1835 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1836 JavaThread *jt = (JavaThread *)this->_thread; 1837 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1838 } 1839 1840 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1841 JavaThread *jt = (JavaThread *)this->_thread; 1842 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1843 } 1844 1845 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1846 assert(node != NULL, "should not add NULL node"); 1847 assert(node->_prev == NULL, "node already in list"); 1848 assert(node->_next == NULL, "node already in list"); 1849 // put node at end of queue (circular doubly linked list) 1850 if (_WaitSet == NULL) { 1851 _WaitSet = node; 1852 node->_prev = node; 1853 node->_next = node; 1854 } else { 1855 ObjectWaiter* head = _WaitSet; 1856 ObjectWaiter* tail = head->_prev; 1857 assert(tail->_next == head, "invariant check"); 1858 tail->_next = node; 1859 head->_prev = node; 1860 node->_next = head; 1861 node->_prev = tail; 1862 } 1863 } 1864 1865 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1866 // dequeue the very first waiter 1867 ObjectWaiter* waiter = _WaitSet; 1868 if (waiter) { 1869 DequeueSpecificWaiter(waiter); 1870 } 1871 return waiter; 1872 } 1873 1874 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1875 assert(node != NULL, "should not dequeue NULL node"); 1876 assert(node->_prev != NULL, "node already removed from list"); 1877 assert(node->_next != NULL, "node already removed from list"); 1878 // when the waiter has woken up because of interrupt, 1879 // timeout or other spurious wake-up, dequeue the 1880 // waiter from waiting list 1881 ObjectWaiter* next = node->_next; 1882 if (next == node) { 1883 assert(node->_prev == node, "invariant check"); 1884 _WaitSet = NULL; 1885 } else { 1886 ObjectWaiter* prev = node->_prev; 1887 assert(prev->_next == node, "invariant check"); 1888 assert(next->_prev == node, "invariant check"); 1889 next->_prev = prev; 1890 prev->_next = next; 1891 if (_WaitSet == node) { 1892 _WaitSet = next; 1893 } 1894 } 1895 node->_next = NULL; 1896 node->_prev = NULL; 1897 } 1898 1899 // ----------------------------------------------------------------------------- 1900 // PerfData support 1901 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 1902 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 1903 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 1904 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 1905 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 1906 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 1907 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 1908 1909 // One-shot global initialization for the sync subsystem. 1910 // We could also defer initialization and initialize on-demand 1911 // the first time we call ObjectSynchronizer::inflate(). 1912 // Initialization would be protected - like so many things - by 1913 // the MonitorCache_lock. 1914 1915 void ObjectMonitor::Initialize() { 1916 assert(!InitDone, "invariant"); 1917 1918 if (!os::is_MP()) { 1919 Knob_SpinLimit = 0; 1920 Knob_PreSpin = 0; 1921 Knob_FixedSpin = -1; 1922 } 1923 1924 if (UsePerfData) { 1925 EXCEPTION_MARK; 1926 #define NEWPERFCOUNTER(n) \ 1927 { \ 1928 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 1929 CHECK); \ 1930 } 1931 #define NEWPERFVARIABLE(n) \ 1932 { \ 1933 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 1934 CHECK); \ 1935 } 1936 NEWPERFCOUNTER(_sync_Inflations); 1937 NEWPERFCOUNTER(_sync_Deflations); 1938 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 1939 NEWPERFCOUNTER(_sync_FutileWakeups); 1940 NEWPERFCOUNTER(_sync_Parks); 1941 NEWPERFCOUNTER(_sync_Notifications); 1942 NEWPERFVARIABLE(_sync_MonExtant); 1943 #undef NEWPERFCOUNTER 1944 #undef NEWPERFVARIABLE 1945 } 1946 1947 DEBUG_ONLY(InitDone = true;) 1948 } 1949 1950 void ObjectMonitor::print_on(outputStream* st) const { 1951 // The minimal things to print for markWord printing, more can be added for debugging and logging. 1952 st->print("{contentions=0x%08x,waiters=0x%08x" 1953 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 1954 contentions(), waiters(), recursions(), 1955 p2i(owner())); 1956 } 1957 void ObjectMonitor::print() const { print_on(tty); } 1958 1959 #ifdef ASSERT 1960 // Print the ObjectMonitor like a debugger would: 1961 // 1962 // (ObjectMonitor) 0x00007fdfb6012e40 = { 1963 // _header = 0x0000000000000001 1964 // _object = 0x000000070ff45fd0 1965 // _next_om = 0x0000000000000000 1966 // _pad_buf0 = { 1967 // [0] = '\0' 1968 // ... 1969 // [103] = '\0' 1970 // } 1971 // _owner = 0x0000000000000000 1972 // _previous_owner_tid = 0 1973 // _recursions = 0 1974 // _EntryList = 0x0000000000000000 1975 // _cxq = 0x0000000000000000 1976 // _succ = 0x0000000000000000 1977 // _Responsible = 0x0000000000000000 1978 // _Spinner = 0 1979 // _SpinDuration = 5000 1980 // _contentions = 0 1981 // _WaitSet = 0x0000700009756248 1982 // _waiters = 1 1983 // _WaitSetLock = 0 1984 // } 1985 // 1986 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 1987 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 1988 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 1989 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 1990 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(_next_om)); 1991 st->print_cr(" _pad_buf0 = {"); 1992 st->print_cr(" [0] = '\\0'"); 1993 st->print_cr(" ..."); 1994 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 1995 st->print_cr(" }"); 1996 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 1997 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 1998 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 1999 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2000 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2001 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2002 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2003 st->print_cr(" _Spinner = %d", _Spinner); 2004 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2005 st->print_cr(" _contentions = %d", _contentions); 2006 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2007 st->print_cr(" _waiters = %d", _waiters); 2008 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2009 st->print_cr("}"); 2010 } 2011 #endif