1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "jfr/support/jfrThreadId.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "memory/resourceArea.hpp" 31 #include "oops/markOop.hpp" 32 #include "oops/oop.inline.hpp" 33 #include "runtime/atomic.hpp" 34 #include "runtime/handles.inline.hpp" 35 #include "runtime/interfaceSupport.inline.hpp" 36 #include "runtime/mutexLocker.hpp" 37 #include "runtime/objectMonitor.hpp" 38 #include "runtime/objectMonitor.inline.hpp" 39 #include "runtime/orderAccess.hpp" 40 #include "runtime/osThread.hpp" 41 #include "runtime/safepointMechanism.inline.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 #include "runtime/stubRoutines.hpp" 44 #include "runtime/thread.inline.hpp" 45 #include "services/threadService.hpp" 46 #include "utilities/dtrace.hpp" 47 #include "utilities/macros.hpp" 48 #include "utilities/preserveException.hpp" 49 #if INCLUDE_JFR 50 #include "jfr/support/jfrFlush.hpp" 51 #endif 52 53 #ifdef DTRACE_ENABLED 54 55 // Only bother with this argument setup if dtrace is available 56 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 57 58 59 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 60 char* bytes = NULL; \ 61 int len = 0; \ 62 jlong jtid = SharedRuntime::get_java_tid(thread); \ 63 Symbol* klassname = ((oop)obj)->klass()->name(); \ 64 if (klassname != NULL) { \ 65 bytes = (char*)klassname->bytes(); \ 66 len = klassname->utf8_length(); \ 67 } 68 69 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 70 { \ 71 if (DTraceMonitorProbes) { \ 72 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 73 HOTSPOT_MONITOR_WAIT(jtid, \ 74 (monitor), bytes, len, (millis)); \ 75 } \ 76 } 77 78 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 79 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 80 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 81 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 82 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 83 84 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 85 { \ 86 if (DTraceMonitorProbes) { \ 87 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 88 HOTSPOT_MONITOR_##probe(jtid, \ 89 (uintptr_t)(monitor), bytes, len); \ 90 } \ 91 } 92 93 #else // ndef DTRACE_ENABLED 94 95 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 96 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 97 98 #endif // ndef DTRACE_ENABLED 99 100 // Tunables ... 101 // The knob* variables are effectively final. Once set they should 102 // never be modified hence. Consider using __read_mostly with GCC. 103 104 int ObjectMonitor::Knob_ExitRelease = 0; 105 int ObjectMonitor::Knob_InlineNotify = 1; 106 int ObjectMonitor::Knob_Verbose = 0; 107 int ObjectMonitor::Knob_VerifyInUse = 0; 108 int ObjectMonitor::Knob_VerifyMatch = 0; 109 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 110 111 static int Knob_ReportSettings = 0; 112 static int Knob_SpinBase = 0; // Floor AKA SpinMin 113 static int Knob_SpinBackOff = 0; // spin-loop backoff 114 static int Knob_CASPenalty = -1; // Penalty for failed CAS 115 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 116 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 117 static int Knob_SpinEarly = 1; 118 static int Knob_SuccEnabled = 1; // futile wake throttling 119 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 120 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 121 static int Knob_Bonus = 100; // spin success bonus 122 static int Knob_BonusB = 100; // spin success bonus 123 static int Knob_Penalty = 200; // spin failure penalty 124 static int Knob_Poverty = 1000; 125 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 126 static int Knob_FixedSpin = 0; 127 static int Knob_OState = 3; // Spinner checks thread state of _owner 128 static int Knob_UsePause = 1; 129 static int Knob_ExitPolicy = 0; 130 static int Knob_PreSpin = 10; // 20-100 likely better 131 static int Knob_ResetEvent = 0; 132 static int BackOffMask = 0; 133 134 static int Knob_FastHSSEC = 0; 135 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 136 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 137 static volatile int InitDone = 0; 138 139 // ----------------------------------------------------------------------------- 140 // Theory of operations -- Monitors lists, thread residency, etc: 141 // 142 // * A thread acquires ownership of a monitor by successfully 143 // CAS()ing the _owner field from null to non-null. 144 // 145 // * Invariant: A thread appears on at most one monitor list -- 146 // cxq, EntryList or WaitSet -- at any one time. 147 // 148 // * Contending threads "push" themselves onto the cxq with CAS 149 // and then spin/park. 150 // 151 // * After a contending thread eventually acquires the lock it must 152 // dequeue itself from either the EntryList or the cxq. 153 // 154 // * The exiting thread identifies and unparks an "heir presumptive" 155 // tentative successor thread on the EntryList. Critically, the 156 // exiting thread doesn't unlink the successor thread from the EntryList. 157 // After having been unparked, the wakee will recontend for ownership of 158 // the monitor. The successor (wakee) will either acquire the lock or 159 // re-park itself. 160 // 161 // Succession is provided for by a policy of competitive handoff. 162 // The exiting thread does _not_ grant or pass ownership to the 163 // successor thread. (This is also referred to as "handoff" succession"). 164 // Instead the exiting thread releases ownership and possibly wakes 165 // a successor, so the successor can (re)compete for ownership of the lock. 166 // If the EntryList is empty but the cxq is populated the exiting 167 // thread will drain the cxq into the EntryList. It does so by 168 // by detaching the cxq (installing null with CAS) and folding 169 // the threads from the cxq into the EntryList. The EntryList is 170 // doubly linked, while the cxq is singly linked because of the 171 // CAS-based "push" used to enqueue recently arrived threads (RATs). 172 // 173 // * Concurrency invariants: 174 // 175 // -- only the monitor owner may access or mutate the EntryList. 176 // The mutex property of the monitor itself protects the EntryList 177 // from concurrent interference. 178 // -- Only the monitor owner may detach the cxq. 179 // 180 // * The monitor entry list operations avoid locks, but strictly speaking 181 // they're not lock-free. Enter is lock-free, exit is not. 182 // For a description of 'Methods and apparatus providing non-blocking access 183 // to a resource,' see U.S. Pat. No. 7844973. 184 // 185 // * The cxq can have multiple concurrent "pushers" but only one concurrent 186 // detaching thread. This mechanism is immune from the ABA corruption. 187 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 188 // 189 // * Taken together, the cxq and the EntryList constitute or form a 190 // single logical queue of threads stalled trying to acquire the lock. 191 // We use two distinct lists to improve the odds of a constant-time 192 // dequeue operation after acquisition (in the ::enter() epilogue) and 193 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 194 // A key desideratum is to minimize queue & monitor metadata manipulation 195 // that occurs while holding the monitor lock -- that is, we want to 196 // minimize monitor lock holds times. Note that even a small amount of 197 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 198 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 199 // locks and monitor metadata. 200 // 201 // Cxq points to the set of Recently Arrived Threads attempting entry. 202 // Because we push threads onto _cxq with CAS, the RATs must take the form of 203 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 204 // the unlocking thread notices that EntryList is null but _cxq is != null. 205 // 206 // The EntryList is ordered by the prevailing queue discipline and 207 // can be organized in any convenient fashion, such as a doubly-linked list or 208 // a circular doubly-linked list. Critically, we want insert and delete operations 209 // to operate in constant-time. If we need a priority queue then something akin 210 // to Solaris' sleepq would work nicely. Viz., 211 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 212 // Queue discipline is enforced at ::exit() time, when the unlocking thread 213 // drains the cxq into the EntryList, and orders or reorders the threads on the 214 // EntryList accordingly. 215 // 216 // Barring "lock barging", this mechanism provides fair cyclic ordering, 217 // somewhat similar to an elevator-scan. 218 // 219 // * The monitor synchronization subsystem avoids the use of native 220 // synchronization primitives except for the narrow platform-specific 221 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 222 // the semantics of park-unpark. Put another way, this monitor implementation 223 // depends only on atomic operations and park-unpark. The monitor subsystem 224 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 225 // underlying OS manages the READY<->RUN transitions. 226 // 227 // * Waiting threads reside on the WaitSet list -- wait() puts 228 // the caller onto the WaitSet. 229 // 230 // * notify() or notifyAll() simply transfers threads from the WaitSet to 231 // either the EntryList or cxq. Subsequent exit() operations will 232 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 233 // it's likely the notifyee would simply impale itself on the lock held 234 // by the notifier. 235 // 236 // * An interesting alternative is to encode cxq as (List,LockByte) where 237 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 238 // variable, like _recursions, in the scheme. The threads or Events that form 239 // the list would have to be aligned in 256-byte addresses. A thread would 240 // try to acquire the lock or enqueue itself with CAS, but exiting threads 241 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 242 // Note that is is *not* word-tearing, but it does presume that full-word 243 // CAS operations are coherent with intermix with STB operations. That's true 244 // on most common processors. 245 // 246 // * See also http://blogs.sun.com/dave 247 248 249 void* ObjectMonitor::operator new (size_t size) throw() { 250 return AllocateHeap(size, mtInternal); 251 } 252 void* ObjectMonitor::operator new[] (size_t size) throw() { 253 return operator new (size); 254 } 255 void ObjectMonitor::operator delete(void* p) { 256 FreeHeap(p); 257 } 258 void ObjectMonitor::operator delete[] (void *p) { 259 operator delete(p); 260 } 261 262 // ----------------------------------------------------------------------------- 263 // Enter support 264 265 void ObjectMonitor::enter(TRAPS) { 266 // The following code is ordered to check the most common cases first 267 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 268 Thread * const Self = THREAD; 269 270 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL); 271 if (cur == NULL) { 272 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 273 assert(_recursions == 0, "invariant"); 274 assert(_owner == Self, "invariant"); 275 return; 276 } 277 278 if (cur == Self) { 279 // TODO-FIXME: check for integer overflow! BUGID 6557169. 280 _recursions++; 281 return; 282 } 283 284 if (Self->is_lock_owned ((address)cur)) { 285 assert(_recursions == 0, "internal state error"); 286 _recursions = 1; 287 // Commute owner from a thread-specific on-stack BasicLockObject address to 288 // a full-fledged "Thread *". 289 _owner = Self; 290 return; 291 } 292 293 // We've encountered genuine contention. 294 assert(Self->_Stalled == 0, "invariant"); 295 Self->_Stalled = intptr_t(this); 296 297 // Try one round of spinning *before* enqueueing Self 298 // and before going through the awkward and expensive state 299 // transitions. The following spin is strictly optional ... 300 // Note that if we acquire the monitor from an initial spin 301 // we forgo posting JVMTI events and firing DTRACE probes. 302 if (Knob_SpinEarly && TrySpin (Self) > 0) { 303 assert(_owner == Self, "invariant"); 304 assert(_recursions == 0, "invariant"); 305 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 306 Self->_Stalled = 0; 307 return; 308 } 309 310 assert(_owner != Self, "invariant"); 311 assert(_succ != Self, "invariant"); 312 assert(Self->is_Java_thread(), "invariant"); 313 JavaThread * jt = (JavaThread *) Self; 314 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 315 assert(jt->thread_state() != _thread_blocked, "invariant"); 316 assert(this->object() != NULL, "invariant"); 317 assert(_count >= 0, "invariant"); 318 319 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 320 // Ensure the object-monitor relationship remains stable while there's contention. 321 Atomic::inc(&_count); 322 323 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 324 EventJavaMonitorEnter event; 325 if (event.should_commit()) { 326 event.set_monitorClass(((oop)this->object())->klass()); 327 event.set_address((uintptr_t)(this->object_addr())); 328 } 329 330 { // Change java thread status to indicate blocked on monitor enter. 331 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 332 333 Self->set_current_pending_monitor(this); 334 335 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 336 if (JvmtiExport::should_post_monitor_contended_enter()) { 337 JvmtiExport::post_monitor_contended_enter(jt, this); 338 339 // The current thread does not yet own the monitor and does not 340 // yet appear on any queues that would get it made the successor. 341 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 342 // handler cannot accidentally consume an unpark() meant for the 343 // ParkEvent associated with this ObjectMonitor. 344 } 345 346 OSThreadContendState osts(Self->osthread()); 347 ThreadBlockInVM tbivm(jt); 348 349 // TODO-FIXME: change the following for(;;) loop to straight-line code. 350 for (;;) { 351 jt->set_suspend_equivalent(); 352 // cleared by handle_special_suspend_equivalent_condition() 353 // or java_suspend_self() 354 355 EnterI(THREAD); 356 357 if (!ExitSuspendEquivalent(jt)) break; 358 359 // We have acquired the contended monitor, but while we were 360 // waiting another thread suspended us. We don't want to enter 361 // the monitor while suspended because that would surprise the 362 // thread that suspended us. 363 // 364 _recursions = 0; 365 _succ = NULL; 366 exit(false, Self); 367 368 jt->java_suspend_self(); 369 } 370 Self->set_current_pending_monitor(NULL); 371 372 // We cleared the pending monitor info since we've just gotten past 373 // the enter-check-for-suspend dance and we now own the monitor free 374 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 375 // destructor can go to a safepoint at the end of this block. If we 376 // do a thread dump during that safepoint, then this thread will show 377 // as having "-locked" the monitor, but the OS and java.lang.Thread 378 // states will still report that the thread is blocked trying to 379 // acquire it. 380 } 381 382 Atomic::dec(&_count); 383 assert(_count >= 0, "invariant"); 384 Self->_Stalled = 0; 385 386 // Must either set _recursions = 0 or ASSERT _recursions == 0. 387 assert(_recursions == 0, "invariant"); 388 assert(_owner == Self, "invariant"); 389 assert(_succ != Self, "invariant"); 390 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 391 392 // The thread -- now the owner -- is back in vm mode. 393 // Report the glorious news via TI,DTrace and jvmstat. 394 // The probe effect is non-trivial. All the reportage occurs 395 // while we hold the monitor, increasing the length of the critical 396 // section. Amdahl's parallel speedup law comes vividly into play. 397 // 398 // Another option might be to aggregate the events (thread local or 399 // per-monitor aggregation) and defer reporting until a more opportune 400 // time -- such as next time some thread encounters contention but has 401 // yet to acquire the lock. While spinning that thread could 402 // spinning we could increment JVMStat counters, etc. 403 404 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 405 if (JvmtiExport::should_post_monitor_contended_entered()) { 406 JvmtiExport::post_monitor_contended_entered(jt, this); 407 408 // The current thread already owns the monitor and is not going to 409 // call park() for the remainder of the monitor enter protocol. So 410 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 411 // event handler consumed an unpark() issued by the thread that 412 // just exited the monitor. 413 } 414 if (event.should_commit()) { 415 event.set_previousOwner((uintptr_t)_previous_owner_tid); 416 event.commit(); 417 } 418 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 419 } 420 421 // Caveat: TryLock() is not necessarily serializing if it returns failure. 422 // Callers must compensate as needed. 423 424 int ObjectMonitor::TryLock(Thread * Self) { 425 void * own = _owner; 426 if (own != NULL) return 0; 427 if (Atomic::replace_if_null(Self, &_owner)) { 428 // Either guarantee _recursions == 0 or set _recursions = 0. 429 assert(_recursions == 0, "invariant"); 430 assert(_owner == Self, "invariant"); 431 return 1; 432 } 433 // The lock had been free momentarily, but we lost the race to the lock. 434 // Interference -- the CAS failed. 435 // We can either return -1 or retry. 436 // Retry doesn't make as much sense because the lock was just acquired. 437 return -1; 438 } 439 440 #define MAX_RECHECK_INTERVAL 1000 441 442 void ObjectMonitor::EnterI(TRAPS) { 443 Thread * const Self = THREAD; 444 assert(Self->is_Java_thread(), "invariant"); 445 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 446 447 // Try the lock - TATAS 448 if (TryLock (Self) > 0) { 449 assert(_succ != Self, "invariant"); 450 assert(_owner == Self, "invariant"); 451 assert(_Responsible != Self, "invariant"); 452 return; 453 } 454 455 DeferredInitialize(); 456 457 // We try one round of spinning *before* enqueueing Self. 458 // 459 // If the _owner is ready but OFFPROC we could use a YieldTo() 460 // operation to donate the remainder of this thread's quantum 461 // to the owner. This has subtle but beneficial affinity 462 // effects. 463 464 if (TrySpin (Self) > 0) { 465 assert(_owner == Self, "invariant"); 466 assert(_succ != Self, "invariant"); 467 assert(_Responsible != Self, "invariant"); 468 return; 469 } 470 471 // The Spin failed -- Enqueue and park the thread ... 472 assert(_succ != Self, "invariant"); 473 assert(_owner != Self, "invariant"); 474 assert(_Responsible != Self, "invariant"); 475 476 // Enqueue "Self" on ObjectMonitor's _cxq. 477 // 478 // Node acts as a proxy for Self. 479 // As an aside, if were to ever rewrite the synchronization code mostly 480 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 481 // Java objects. This would avoid awkward lifecycle and liveness issues, 482 // as well as eliminate a subset of ABA issues. 483 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 484 485 ObjectWaiter node(Self); 486 Self->_ParkEvent->reset(); 487 node._prev = (ObjectWaiter *) 0xBAD; 488 node.TState = ObjectWaiter::TS_CXQ; 489 490 // Push "Self" onto the front of the _cxq. 491 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 492 // Note that spinning tends to reduce the rate at which threads 493 // enqueue and dequeue on EntryList|cxq. 494 ObjectWaiter * nxt; 495 for (;;) { 496 node._next = nxt = _cxq; 497 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 498 499 // Interference - the CAS failed because _cxq changed. Just retry. 500 // As an optional optimization we retry the lock. 501 if (TryLock (Self) > 0) { 502 assert(_succ != Self, "invariant"); 503 assert(_owner == Self, "invariant"); 504 assert(_Responsible != Self, "invariant"); 505 return; 506 } 507 } 508 509 // Check for cxq|EntryList edge transition to non-null. This indicates 510 // the onset of contention. While contention persists exiting threads 511 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 512 // operations revert to the faster 1-0 mode. This enter operation may interleave 513 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 514 // arrange for one of the contending thread to use a timed park() operations 515 // to detect and recover from the race. (Stranding is form of progress failure 516 // where the monitor is unlocked but all the contending threads remain parked). 517 // That is, at least one of the contended threads will periodically poll _owner. 518 // One of the contending threads will become the designated "Responsible" thread. 519 // The Responsible thread uses a timed park instead of a normal indefinite park 520 // operation -- it periodically wakes and checks for and recovers from potential 521 // strandings admitted by 1-0 exit operations. We need at most one Responsible 522 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 523 // be responsible for a monitor. 524 // 525 // Currently, one of the contended threads takes on the added role of "Responsible". 526 // A viable alternative would be to use a dedicated "stranding checker" thread 527 // that periodically iterated over all the threads (or active monitors) and unparked 528 // successors where there was risk of stranding. This would help eliminate the 529 // timer scalability issues we see on some platforms as we'd only have one thread 530 // -- the checker -- parked on a timer. 531 532 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 533 // Try to assume the role of responsible thread for the monitor. 534 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 535 Atomic::replace_if_null(Self, &_Responsible); 536 } 537 538 // The lock might have been released while this thread was occupied queueing 539 // itself onto _cxq. To close the race and avoid "stranding" and 540 // progress-liveness failure we must resample-retry _owner before parking. 541 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 542 // In this case the ST-MEMBAR is accomplished with CAS(). 543 // 544 // TODO: Defer all thread state transitions until park-time. 545 // Since state transitions are heavy and inefficient we'd like 546 // to defer the state transitions until absolutely necessary, 547 // and in doing so avoid some transitions ... 548 549 int nWakeups = 0; 550 int recheckInterval = 1; 551 552 for (;;) { 553 554 if (TryLock(Self) > 0) break; 555 assert(_owner != Self, "invariant"); 556 557 if ((SyncFlags & 2) && _Responsible == NULL) { 558 Atomic::replace_if_null(Self, &_Responsible); 559 } 560 561 // park self 562 if (_Responsible == Self || (SyncFlags & 1)) { 563 Self->_ParkEvent->park((jlong) recheckInterval); 564 // Increase the recheckInterval, but clamp the value. 565 recheckInterval *= 8; 566 if (recheckInterval > MAX_RECHECK_INTERVAL) { 567 recheckInterval = MAX_RECHECK_INTERVAL; 568 } 569 } else { 570 Self->_ParkEvent->park(); 571 } 572 573 if (TryLock(Self) > 0) break; 574 575 // The lock is still contested. 576 // Keep a tally of the # of futile wakeups. 577 // Note that the counter is not protected by a lock or updated by atomics. 578 // That is by design - we trade "lossy" counters which are exposed to 579 // races during updates for a lower probe effect. 580 581 // This PerfData object can be used in parallel with a safepoint. 582 // See the work around in PerfDataManager::destroy(). 583 OM_PERFDATA_OP(FutileWakeups, inc()); 584 ++nWakeups; 585 586 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 587 // We can defer clearing _succ until after the spin completes 588 // TrySpin() must tolerate being called with _succ == Self. 589 // Try yet another round of adaptive spinning. 590 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 591 592 // We can find that we were unpark()ed and redesignated _succ while 593 // we were spinning. That's harmless. If we iterate and call park(), 594 // park() will consume the event and return immediately and we'll 595 // just spin again. This pattern can repeat, leaving _succ to simply 596 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 597 // Alternately, we can sample fired() here, and if set, forgo spinning 598 // in the next iteration. 599 600 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 601 Self->_ParkEvent->reset(); 602 OrderAccess::fence(); 603 } 604 if (_succ == Self) _succ = NULL; 605 606 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 607 OrderAccess::fence(); 608 } 609 610 // Egress : 611 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 612 // Normally we'll find Self on the EntryList . 613 // From the perspective of the lock owner (this thread), the 614 // EntryList is stable and cxq is prepend-only. 615 // The head of cxq is volatile but the interior is stable. 616 // In addition, Self.TState is stable. 617 618 assert(_owner == Self, "invariant"); 619 assert(object() != NULL, "invariant"); 620 // I'd like to write: 621 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 622 // but as we're at a safepoint that's not safe. 623 624 UnlinkAfterAcquire(Self, &node); 625 if (_succ == Self) _succ = NULL; 626 627 assert(_succ != Self, "invariant"); 628 if (_Responsible == Self) { 629 _Responsible = NULL; 630 OrderAccess::fence(); // Dekker pivot-point 631 632 // We may leave threads on cxq|EntryList without a designated 633 // "Responsible" thread. This is benign. When this thread subsequently 634 // exits the monitor it can "see" such preexisting "old" threads -- 635 // threads that arrived on the cxq|EntryList before the fence, above -- 636 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 637 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 638 // non-null and elect a new "Responsible" timer thread. 639 // 640 // This thread executes: 641 // ST Responsible=null; MEMBAR (in enter epilogue - here) 642 // LD cxq|EntryList (in subsequent exit) 643 // 644 // Entering threads in the slow/contended path execute: 645 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 646 // The (ST cxq; MEMBAR) is accomplished with CAS(). 647 // 648 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 649 // exit operation from floating above the ST Responsible=null. 650 } 651 652 // We've acquired ownership with CAS(). 653 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 654 // But since the CAS() this thread may have also stored into _succ, 655 // EntryList, cxq or Responsible. These meta-data updates must be 656 // visible __before this thread subsequently drops the lock. 657 // Consider what could occur if we didn't enforce this constraint -- 658 // STs to monitor meta-data and user-data could reorder with (become 659 // visible after) the ST in exit that drops ownership of the lock. 660 // Some other thread could then acquire the lock, but observe inconsistent 661 // or old monitor meta-data and heap data. That violates the JMM. 662 // To that end, the 1-0 exit() operation must have at least STST|LDST 663 // "release" barrier semantics. Specifically, there must be at least a 664 // STST|LDST barrier in exit() before the ST of null into _owner that drops 665 // the lock. The barrier ensures that changes to monitor meta-data and data 666 // protected by the lock will be visible before we release the lock, and 667 // therefore before some other thread (CPU) has a chance to acquire the lock. 668 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 669 // 670 // Critically, any prior STs to _succ or EntryList must be visible before 671 // the ST of null into _owner in the *subsequent* (following) corresponding 672 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 673 // execute a serializing instruction. 674 675 if (SyncFlags & 8) { 676 OrderAccess::fence(); 677 } 678 return; 679 } 680 681 // ReenterI() is a specialized inline form of the latter half of the 682 // contended slow-path from EnterI(). We use ReenterI() only for 683 // monitor reentry in wait(). 684 // 685 // In the future we should reconcile EnterI() and ReenterI(), adding 686 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 687 // loop accordingly. 688 689 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 690 assert(Self != NULL, "invariant"); 691 assert(SelfNode != NULL, "invariant"); 692 assert(SelfNode->_thread == Self, "invariant"); 693 assert(_waiters > 0, "invariant"); 694 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 695 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 696 JavaThread * jt = (JavaThread *) Self; 697 698 int nWakeups = 0; 699 for (;;) { 700 ObjectWaiter::TStates v = SelfNode->TState; 701 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 702 assert(_owner != Self, "invariant"); 703 704 if (TryLock(Self) > 0) break; 705 if (TrySpin(Self) > 0) break; 706 707 // State transition wrappers around park() ... 708 // ReenterI() wisely defers state transitions until 709 // it's clear we must park the thread. 710 { 711 OSThreadContendState osts(Self->osthread()); 712 ThreadBlockInVM tbivm(jt); 713 714 // cleared by handle_special_suspend_equivalent_condition() 715 // or java_suspend_self() 716 jt->set_suspend_equivalent(); 717 if (SyncFlags & 1) { 718 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL); 719 } else { 720 Self->_ParkEvent->park(); 721 } 722 723 // were we externally suspended while we were waiting? 724 for (;;) { 725 if (!ExitSuspendEquivalent(jt)) break; 726 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 727 jt->java_suspend_self(); 728 jt->set_suspend_equivalent(); 729 } 730 } 731 732 // Try again, but just so we distinguish between futile wakeups and 733 // successful wakeups. The following test isn't algorithmically 734 // necessary, but it helps us maintain sensible statistics. 735 if (TryLock(Self) > 0) break; 736 737 // The lock is still contested. 738 // Keep a tally of the # of futile wakeups. 739 // Note that the counter is not protected by a lock or updated by atomics. 740 // That is by design - we trade "lossy" counters which are exposed to 741 // races during updates for a lower probe effect. 742 ++nWakeups; 743 744 // Assuming this is not a spurious wakeup we'll normally 745 // find that _succ == Self. 746 if (_succ == Self) _succ = NULL; 747 748 // Invariant: after clearing _succ a contending thread 749 // *must* retry _owner before parking. 750 OrderAccess::fence(); 751 752 // This PerfData object can be used in parallel with a safepoint. 753 // See the work around in PerfDataManager::destroy(). 754 OM_PERFDATA_OP(FutileWakeups, inc()); 755 } 756 757 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 758 // Normally we'll find Self on the EntryList. 759 // Unlinking from the EntryList is constant-time and atomic-free. 760 // From the perspective of the lock owner (this thread), the 761 // EntryList is stable and cxq is prepend-only. 762 // The head of cxq is volatile but the interior is stable. 763 // In addition, Self.TState is stable. 764 765 assert(_owner == Self, "invariant"); 766 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 767 UnlinkAfterAcquire(Self, SelfNode); 768 if (_succ == Self) _succ = NULL; 769 assert(_succ != Self, "invariant"); 770 SelfNode->TState = ObjectWaiter::TS_RUN; 771 OrderAccess::fence(); // see comments at the end of EnterI() 772 } 773 774 // By convention we unlink a contending thread from EntryList|cxq immediately 775 // after the thread acquires the lock in ::enter(). Equally, we could defer 776 // unlinking the thread until ::exit()-time. 777 778 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 779 assert(_owner == Self, "invariant"); 780 assert(SelfNode->_thread == Self, "invariant"); 781 782 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 783 // Normal case: remove Self from the DLL EntryList . 784 // This is a constant-time operation. 785 ObjectWaiter * nxt = SelfNode->_next; 786 ObjectWaiter * prv = SelfNode->_prev; 787 if (nxt != NULL) nxt->_prev = prv; 788 if (prv != NULL) prv->_next = nxt; 789 if (SelfNode == _EntryList) _EntryList = nxt; 790 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 791 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 792 } else { 793 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 794 // Inopportune interleaving -- Self is still on the cxq. 795 // This usually means the enqueue of self raced an exiting thread. 796 // Normally we'll find Self near the front of the cxq, so 797 // dequeueing is typically fast. If needbe we can accelerate 798 // this with some MCS/CHL-like bidirectional list hints and advisory 799 // back-links so dequeueing from the interior will normally operate 800 // in constant-time. 801 // Dequeue Self from either the head (with CAS) or from the interior 802 // with a linear-time scan and normal non-atomic memory operations. 803 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 804 // and then unlink Self from EntryList. We have to drain eventually, 805 // so it might as well be now. 806 807 ObjectWaiter * v = _cxq; 808 assert(v != NULL, "invariant"); 809 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 810 // The CAS above can fail from interference IFF a "RAT" arrived. 811 // In that case Self must be in the interior and can no longer be 812 // at the head of cxq. 813 if (v == SelfNode) { 814 assert(_cxq != v, "invariant"); 815 v = _cxq; // CAS above failed - start scan at head of list 816 } 817 ObjectWaiter * p; 818 ObjectWaiter * q = NULL; 819 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 820 q = p; 821 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 822 } 823 assert(v != SelfNode, "invariant"); 824 assert(p == SelfNode, "Node not found on cxq"); 825 assert(p != _cxq, "invariant"); 826 assert(q != NULL, "invariant"); 827 assert(q->_next == p, "invariant"); 828 q->_next = p->_next; 829 } 830 } 831 832 #ifdef ASSERT 833 // Diagnostic hygiene ... 834 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 835 SelfNode->_next = (ObjectWaiter *) 0xBAD; 836 SelfNode->TState = ObjectWaiter::TS_RUN; 837 #endif 838 } 839 840 // ----------------------------------------------------------------------------- 841 // Exit support 842 // 843 // exit() 844 // ~~~~~~ 845 // Note that the collector can't reclaim the objectMonitor or deflate 846 // the object out from underneath the thread calling ::exit() as the 847 // thread calling ::exit() never transitions to a stable state. 848 // This inhibits GC, which in turn inhibits asynchronous (and 849 // inopportune) reclamation of "this". 850 // 851 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 852 // There's one exception to the claim above, however. EnterI() can call 853 // exit() to drop a lock if the acquirer has been externally suspended. 854 // In that case exit() is called with _thread_state as _thread_blocked, 855 // but the monitor's _count field is > 0, which inhibits reclamation. 856 // 857 // 1-0 exit 858 // ~~~~~~~~ 859 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 860 // the fast-path operators have been optimized so the common ::exit() 861 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 862 // The code emitted by fast_unlock() elides the usual MEMBAR. This 863 // greatly improves latency -- MEMBAR and CAS having considerable local 864 // latency on modern processors -- but at the cost of "stranding". Absent the 865 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 866 // ::enter() path, resulting in the entering thread being stranding 867 // and a progress-liveness failure. Stranding is extremely rare. 868 // We use timers (timed park operations) & periodic polling to detect 869 // and recover from stranding. Potentially stranded threads periodically 870 // wake up and poll the lock. See the usage of the _Responsible variable. 871 // 872 // The CAS() in enter provides for safety and exclusion, while the CAS or 873 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 874 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 875 // We detect and recover from stranding with timers. 876 // 877 // If a thread transiently strands it'll park until (a) another 878 // thread acquires the lock and then drops the lock, at which time the 879 // exiting thread will notice and unpark the stranded thread, or, (b) 880 // the timer expires. If the lock is high traffic then the stranding latency 881 // will be low due to (a). If the lock is low traffic then the odds of 882 // stranding are lower, although the worst-case stranding latency 883 // is longer. Critically, we don't want to put excessive load in the 884 // platform's timer subsystem. We want to minimize both the timer injection 885 // rate (timers created/sec) as well as the number of timers active at 886 // any one time. (more precisely, we want to minimize timer-seconds, which is 887 // the integral of the # of active timers at any instant over time). 888 // Both impinge on OS scalability. Given that, at most one thread parked on 889 // a monitor will use a timer. 890 // 891 // There is also the risk of a futile wake-up. If we drop the lock 892 // another thread can reacquire the lock immediately, and we can 893 // then wake a thread unnecessarily. This is benign, and we've 894 // structured the code so the windows are short and the frequency 895 // of such futile wakups is low. 896 897 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 898 Thread * const Self = THREAD; 899 if (THREAD != _owner) { 900 if (THREAD->is_lock_owned((address) _owner)) { 901 // Transmute _owner from a BasicLock pointer to a Thread address. 902 // We don't need to hold _mutex for this transition. 903 // Non-null to Non-null is safe as long as all readers can 904 // tolerate either flavor. 905 assert(_recursions == 0, "invariant"); 906 _owner = THREAD; 907 _recursions = 0; 908 } else { 909 // Apparent unbalanced locking ... 910 // Naively we'd like to throw IllegalMonitorStateException. 911 // As a practical matter we can neither allocate nor throw an 912 // exception as ::exit() can be called from leaf routines. 913 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 914 // Upon deeper reflection, however, in a properly run JVM the only 915 // way we should encounter this situation is in the presence of 916 // unbalanced JNI locking. TODO: CheckJNICalls. 917 // See also: CR4414101 918 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 919 return; 920 } 921 } 922 923 if (_recursions != 0) { 924 _recursions--; // this is simple recursive enter 925 return; 926 } 927 928 // Invariant: after setting Responsible=null an thread must execute 929 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 930 if ((SyncFlags & 4) == 0) { 931 _Responsible = NULL; 932 } 933 934 #if INCLUDE_JFR 935 // get the owner's thread id for the MonitorEnter event 936 // if it is enabled and the thread isn't suspended 937 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 938 _previous_owner_tid = JFR_THREAD_ID(Self); 939 } 940 #endif 941 942 for (;;) { 943 assert(THREAD == _owner, "invariant"); 944 945 if (Knob_ExitPolicy == 0) { 946 // release semantics: prior loads and stores from within the critical section 947 // must not float (reorder) past the following store that drops the lock. 948 // On SPARC that requires MEMBAR #loadstore|#storestore. 949 // But of course in TSO #loadstore|#storestore is not required. 950 // I'd like to write one of the following: 951 // A. OrderAccess::release() ; _owner = NULL 952 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 953 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 954 // store into a _dummy variable. That store is not needed, but can result 955 // in massive wasteful coherency traffic on classic SMP systems. 956 // Instead, I use release_store(), which is implemented as just a simple 957 // ST on x64, x86 and SPARC. 958 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 959 OrderAccess::storeload(); // See if we need to wake a successor 960 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 961 return; 962 } 963 // Other threads are blocked trying to acquire the lock. 964 965 // Normally the exiting thread is responsible for ensuring succession, 966 // but if other successors are ready or other entering threads are spinning 967 // then this thread can simply store NULL into _owner and exit without 968 // waking a successor. The existence of spinners or ready successors 969 // guarantees proper succession (liveness). Responsibility passes to the 970 // ready or running successors. The exiting thread delegates the duty. 971 // More precisely, if a successor already exists this thread is absolved 972 // of the responsibility of waking (unparking) one. 973 // 974 // The _succ variable is critical to reducing futile wakeup frequency. 975 // _succ identifies the "heir presumptive" thread that has been made 976 // ready (unparked) but that has not yet run. We need only one such 977 // successor thread to guarantee progress. 978 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 979 // section 3.3 "Futile Wakeup Throttling" for details. 980 // 981 // Note that spinners in Enter() also set _succ non-null. 982 // In the current implementation spinners opportunistically set 983 // _succ so that exiting threads might avoid waking a successor. 984 // Another less appealing alternative would be for the exiting thread 985 // to drop the lock and then spin briefly to see if a spinner managed 986 // to acquire the lock. If so, the exiting thread could exit 987 // immediately without waking a successor, otherwise the exiting 988 // thread would need to dequeue and wake a successor. 989 // (Note that we'd need to make the post-drop spin short, but no 990 // shorter than the worst-case round-trip cache-line migration time. 991 // The dropped lock needs to become visible to the spinner, and then 992 // the acquisition of the lock by the spinner must become visible to 993 // the exiting thread). 994 995 // It appears that an heir-presumptive (successor) must be made ready. 996 // Only the current lock owner can manipulate the EntryList or 997 // drain _cxq, so we need to reacquire the lock. If we fail 998 // to reacquire the lock the responsibility for ensuring succession 999 // falls to the new owner. 1000 // 1001 if (!Atomic::replace_if_null(THREAD, &_owner)) { 1002 return; 1003 } 1004 } else { 1005 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1006 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 1007 OrderAccess::storeload(); 1008 // Ratify the previously observed values. 1009 if (_cxq == NULL || _succ != NULL) { 1010 return; 1011 } 1012 1013 // inopportune interleaving -- the exiting thread (this thread) 1014 // in the fast-exit path raced an entering thread in the slow-enter 1015 // path. 1016 // We have two choices: 1017 // A. Try to reacquire the lock. 1018 // If the CAS() fails return immediately, otherwise 1019 // we either restart/rerun the exit operation, or simply 1020 // fall-through into the code below which wakes a successor. 1021 // B. If the elements forming the EntryList|cxq are TSM 1022 // we could simply unpark() the lead thread and return 1023 // without having set _succ. 1024 if (!Atomic::replace_if_null(THREAD, &_owner)) { 1025 return; 1026 } 1027 } 1028 } 1029 1030 guarantee(_owner == THREAD, "invariant"); 1031 1032 ObjectWaiter * w = NULL; 1033 int QMode = Knob_QMode; 1034 1035 if (QMode == 2 && _cxq != NULL) { 1036 // QMode == 2 : cxq has precedence over EntryList. 1037 // Try to directly wake a successor from the cxq. 1038 // If successful, the successor will need to unlink itself from cxq. 1039 w = _cxq; 1040 assert(w != NULL, "invariant"); 1041 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1042 ExitEpilog(Self, w); 1043 return; 1044 } 1045 1046 if (QMode == 3 && _cxq != NULL) { 1047 // Aggressively drain cxq into EntryList at the first opportunity. 1048 // This policy ensure that recently-run threads live at the head of EntryList. 1049 // Drain _cxq into EntryList - bulk transfer. 1050 // First, detach _cxq. 1051 // The following loop is tantamount to: w = swap(&cxq, NULL) 1052 w = _cxq; 1053 for (;;) { 1054 assert(w != NULL, "Invariant"); 1055 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1056 if (u == w) break; 1057 w = u; 1058 } 1059 assert(w != NULL, "invariant"); 1060 1061 ObjectWaiter * q = NULL; 1062 ObjectWaiter * p; 1063 for (p = w; p != NULL; p = p->_next) { 1064 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1065 p->TState = ObjectWaiter::TS_ENTER; 1066 p->_prev = q; 1067 q = p; 1068 } 1069 1070 // Append the RATs to the EntryList 1071 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1072 ObjectWaiter * Tail; 1073 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1074 Tail = Tail->_next) 1075 /* empty */; 1076 if (Tail == NULL) { 1077 _EntryList = w; 1078 } else { 1079 Tail->_next = w; 1080 w->_prev = Tail; 1081 } 1082 1083 // Fall thru into code that tries to wake a successor from EntryList 1084 } 1085 1086 if (QMode == 4 && _cxq != NULL) { 1087 // Aggressively drain cxq into EntryList at the first opportunity. 1088 // This policy ensure that recently-run threads live at the head of EntryList. 1089 1090 // Drain _cxq into EntryList - bulk transfer. 1091 // First, detach _cxq. 1092 // The following loop is tantamount to: w = swap(&cxq, NULL) 1093 w = _cxq; 1094 for (;;) { 1095 assert(w != NULL, "Invariant"); 1096 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1097 if (u == w) break; 1098 w = u; 1099 } 1100 assert(w != NULL, "invariant"); 1101 1102 ObjectWaiter * q = NULL; 1103 ObjectWaiter * p; 1104 for (p = w; p != NULL; p = p->_next) { 1105 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1106 p->TState = ObjectWaiter::TS_ENTER; 1107 p->_prev = q; 1108 q = p; 1109 } 1110 1111 // Prepend the RATs to the EntryList 1112 if (_EntryList != NULL) { 1113 q->_next = _EntryList; 1114 _EntryList->_prev = q; 1115 } 1116 _EntryList = w; 1117 1118 // Fall thru into code that tries to wake a successor from EntryList 1119 } 1120 1121 w = _EntryList; 1122 if (w != NULL) { 1123 // I'd like to write: guarantee (w->_thread != Self). 1124 // But in practice an exiting thread may find itself on the EntryList. 1125 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1126 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1127 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1128 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1129 // release the lock "O". T2 resumes immediately after the ST of null into 1130 // _owner, above. T2 notices that the EntryList is populated, so it 1131 // reacquires the lock and then finds itself on the EntryList. 1132 // Given all that, we have to tolerate the circumstance where "w" is 1133 // associated with Self. 1134 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1135 ExitEpilog(Self, w); 1136 return; 1137 } 1138 1139 // If we find that both _cxq and EntryList are null then just 1140 // re-run the exit protocol from the top. 1141 w = _cxq; 1142 if (w == NULL) continue; 1143 1144 // Drain _cxq into EntryList - bulk transfer. 1145 // First, detach _cxq. 1146 // The following loop is tantamount to: w = swap(&cxq, NULL) 1147 for (;;) { 1148 assert(w != NULL, "Invariant"); 1149 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1150 if (u == w) break; 1151 w = u; 1152 } 1153 1154 assert(w != NULL, "invariant"); 1155 assert(_EntryList == NULL, "invariant"); 1156 1157 // Convert the LIFO SLL anchored by _cxq into a DLL. 1158 // The list reorganization step operates in O(LENGTH(w)) time. 1159 // It's critical that this step operate quickly as 1160 // "Self" still holds the outer-lock, restricting parallelism 1161 // and effectively lengthening the critical section. 1162 // Invariant: s chases t chases u. 1163 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1164 // we have faster access to the tail. 1165 1166 if (QMode == 1) { 1167 // QMode == 1 : drain cxq to EntryList, reversing order 1168 // We also reverse the order of the list. 1169 ObjectWaiter * s = NULL; 1170 ObjectWaiter * t = w; 1171 ObjectWaiter * u = NULL; 1172 while (t != NULL) { 1173 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1174 t->TState = ObjectWaiter::TS_ENTER; 1175 u = t->_next; 1176 t->_prev = u; 1177 t->_next = s; 1178 s = t; 1179 t = u; 1180 } 1181 _EntryList = s; 1182 assert(s != NULL, "invariant"); 1183 } else { 1184 // QMode == 0 or QMode == 2 1185 _EntryList = w; 1186 ObjectWaiter * q = NULL; 1187 ObjectWaiter * p; 1188 for (p = w; p != NULL; p = p->_next) { 1189 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1190 p->TState = ObjectWaiter::TS_ENTER; 1191 p->_prev = q; 1192 q = p; 1193 } 1194 } 1195 1196 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1197 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1198 1199 // See if we can abdicate to a spinner instead of waking a thread. 1200 // A primary goal of the implementation is to reduce the 1201 // context-switch rate. 1202 if (_succ != NULL) continue; 1203 1204 w = _EntryList; 1205 if (w != NULL) { 1206 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1207 ExitEpilog(Self, w); 1208 return; 1209 } 1210 } 1211 } 1212 1213 // ExitSuspendEquivalent: 1214 // A faster alternate to handle_special_suspend_equivalent_condition() 1215 // 1216 // handle_special_suspend_equivalent_condition() unconditionally 1217 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1218 // operations have high latency. Note that in ::enter() we call HSSEC 1219 // while holding the monitor, so we effectively lengthen the critical sections. 1220 // 1221 // There are a number of possible solutions: 1222 // 1223 // A. To ameliorate the problem we might also defer state transitions 1224 // to as late as possible -- just prior to parking. 1225 // Given that, we'd call HSSEC after having returned from park(), 1226 // but before attempting to acquire the monitor. This is only a 1227 // partial solution. It avoids calling HSSEC while holding the 1228 // monitor (good), but it still increases successor reacquisition latency -- 1229 // the interval between unparking a successor and the time the successor 1230 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1231 // If we use this technique we can also avoid EnterI()-exit() loop 1232 // in ::enter() where we iteratively drop the lock and then attempt 1233 // to reacquire it after suspending. 1234 // 1235 // B. In the future we might fold all the suspend bits into a 1236 // composite per-thread suspend flag and then update it with CAS(). 1237 // Alternately, a Dekker-like mechanism with multiple variables 1238 // would suffice: 1239 // ST Self->_suspend_equivalent = false 1240 // MEMBAR 1241 // LD Self_>_suspend_flags 1242 // 1243 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1244 // with a more efficient implementation, the need to use "FastHSSEC" has 1245 // decreased. - Dave 1246 1247 1248 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1249 const int Mode = Knob_FastHSSEC; 1250 if (Mode && !jSelf->is_external_suspend()) { 1251 assert(jSelf->is_suspend_equivalent(), "invariant"); 1252 jSelf->clear_suspend_equivalent(); 1253 if (2 == Mode) OrderAccess::storeload(); 1254 if (!jSelf->is_external_suspend()) return false; 1255 // We raced a suspension -- fall thru into the slow path 1256 jSelf->set_suspend_equivalent(); 1257 } 1258 return jSelf->handle_special_suspend_equivalent_condition(); 1259 } 1260 1261 1262 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1263 assert(_owner == Self, "invariant"); 1264 1265 // Exit protocol: 1266 // 1. ST _succ = wakee 1267 // 2. membar #loadstore|#storestore; 1268 // 2. ST _owner = NULL 1269 // 3. unpark(wakee) 1270 1271 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1272 ParkEvent * Trigger = Wakee->_event; 1273 1274 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1275 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1276 // out-of-scope (non-extant). 1277 Wakee = NULL; 1278 1279 // Drop the lock 1280 OrderAccess::release_store(&_owner, (void*)NULL); 1281 OrderAccess::fence(); // ST _owner vs LD in unpark() 1282 1283 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1284 Trigger->unpark(); 1285 1286 // Maintain stats and report events to JVMTI 1287 OM_PERFDATA_OP(Parks, inc()); 1288 } 1289 1290 1291 // ----------------------------------------------------------------------------- 1292 // Class Loader deadlock handling. 1293 // 1294 // complete_exit exits a lock returning recursion count 1295 // complete_exit/reenter operate as a wait without waiting 1296 // complete_exit requires an inflated monitor 1297 // The _owner field is not always the Thread addr even with an 1298 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1299 // thread due to contention. 1300 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1301 Thread * const Self = THREAD; 1302 assert(Self->is_Java_thread(), "Must be Java thread!"); 1303 JavaThread *jt = (JavaThread *)THREAD; 1304 1305 DeferredInitialize(); 1306 1307 if (THREAD != _owner) { 1308 if (THREAD->is_lock_owned ((address)_owner)) { 1309 assert(_recursions == 0, "internal state error"); 1310 _owner = THREAD; // Convert from basiclock addr to Thread addr 1311 _recursions = 0; 1312 } 1313 } 1314 1315 guarantee(Self == _owner, "complete_exit not owner"); 1316 intptr_t save = _recursions; // record the old recursion count 1317 _recursions = 0; // set the recursion level to be 0 1318 exit(true, Self); // exit the monitor 1319 guarantee(_owner != Self, "invariant"); 1320 return save; 1321 } 1322 1323 // reenter() enters a lock and sets recursion count 1324 // complete_exit/reenter operate as a wait without waiting 1325 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1326 Thread * const Self = THREAD; 1327 assert(Self->is_Java_thread(), "Must be Java thread!"); 1328 JavaThread *jt = (JavaThread *)THREAD; 1329 1330 guarantee(_owner != Self, "reenter already owner"); 1331 enter(THREAD); // enter the monitor 1332 guarantee(_recursions == 0, "reenter recursion"); 1333 _recursions = recursions; 1334 return; 1335 } 1336 1337 1338 // ----------------------------------------------------------------------------- 1339 // A macro is used below because there may already be a pending 1340 // exception which should not abort the execution of the routines 1341 // which use this (which is why we don't put this into check_slow and 1342 // call it with a CHECK argument). 1343 1344 #define CHECK_OWNER() \ 1345 do { \ 1346 if (THREAD != _owner) { \ 1347 if (THREAD->is_lock_owned((address) _owner)) { \ 1348 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1349 _recursions = 0; \ 1350 } else { \ 1351 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1352 } \ 1353 } \ 1354 } while (false) 1355 1356 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1357 // TODO-FIXME: remove check_slow() -- it's likely dead. 1358 1359 void ObjectMonitor::check_slow(TRAPS) { 1360 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1361 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1362 } 1363 1364 static int Adjust(volatile int * adr, int dx) { 1365 int v; 1366 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1367 return v; 1368 } 1369 1370 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1371 ObjectMonitor* monitor, 1372 jlong notifier_tid, 1373 jlong timeout, 1374 bool timedout) { 1375 assert(event != NULL, "invariant"); 1376 assert(monitor != NULL, "invariant"); 1377 event->set_monitorClass(((oop)monitor->object())->klass()); 1378 event->set_timeout(timeout); 1379 event->set_address((uintptr_t)monitor->object_addr()); 1380 event->set_notifier(notifier_tid); 1381 event->set_timedOut(timedout); 1382 event->commit(); 1383 } 1384 1385 // ----------------------------------------------------------------------------- 1386 // Wait/Notify/NotifyAll 1387 // 1388 // Note: a subset of changes to ObjectMonitor::wait() 1389 // will need to be replicated in complete_exit 1390 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1391 Thread * const Self = THREAD; 1392 assert(Self->is_Java_thread(), "Must be Java thread!"); 1393 JavaThread *jt = (JavaThread *)THREAD; 1394 1395 DeferredInitialize(); 1396 1397 // Throw IMSX or IEX. 1398 CHECK_OWNER(); 1399 1400 EventJavaMonitorWait event; 1401 1402 // check for a pending interrupt 1403 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1404 // post monitor waited event. Note that this is past-tense, we are done waiting. 1405 if (JvmtiExport::should_post_monitor_waited()) { 1406 // Note: 'false' parameter is passed here because the 1407 // wait was not timed out due to thread interrupt. 1408 JvmtiExport::post_monitor_waited(jt, this, false); 1409 1410 // In this short circuit of the monitor wait protocol, the 1411 // current thread never drops ownership of the monitor and 1412 // never gets added to the wait queue so the current thread 1413 // cannot be made the successor. This means that the 1414 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1415 // consume an unpark() meant for the ParkEvent associated with 1416 // this ObjectMonitor. 1417 } 1418 if (event.should_commit()) { 1419 post_monitor_wait_event(&event, this, 0, millis, false); 1420 } 1421 THROW(vmSymbols::java_lang_InterruptedException()); 1422 return; 1423 } 1424 1425 assert(Self->_Stalled == 0, "invariant"); 1426 Self->_Stalled = intptr_t(this); 1427 jt->set_current_waiting_monitor(this); 1428 1429 // create a node to be put into the queue 1430 // Critically, after we reset() the event but prior to park(), we must check 1431 // for a pending interrupt. 1432 ObjectWaiter node(Self); 1433 node.TState = ObjectWaiter::TS_WAIT; 1434 Self->_ParkEvent->reset(); 1435 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1436 1437 // Enter the waiting queue, which is a circular doubly linked list in this case 1438 // but it could be a priority queue or any data structure. 1439 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1440 // by the the owner of the monitor *except* in the case where park() 1441 // returns because of a timeout of interrupt. Contention is exceptionally rare 1442 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1443 1444 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1445 AddWaiter(&node); 1446 Thread::SpinRelease(&_WaitSetLock); 1447 1448 if ((SyncFlags & 4) == 0) { 1449 _Responsible = NULL; 1450 } 1451 intptr_t save = _recursions; // record the old recursion count 1452 _waiters++; // increment the number of waiters 1453 _recursions = 0; // set the recursion level to be 1 1454 exit(true, Self); // exit the monitor 1455 guarantee(_owner != Self, "invariant"); 1456 1457 // The thread is on the WaitSet list - now park() it. 1458 // On MP systems it's conceivable that a brief spin before we park 1459 // could be profitable. 1460 // 1461 // TODO-FIXME: change the following logic to a loop of the form 1462 // while (!timeout && !interrupted && _notified == 0) park() 1463 1464 int ret = OS_OK; 1465 int WasNotified = 0; 1466 { // State transition wrappers 1467 OSThread* osthread = Self->osthread(); 1468 OSThreadWaitState osts(osthread, true); 1469 { 1470 ThreadBlockInVM tbivm(jt); 1471 // Thread is in thread_blocked state and oop access is unsafe. 1472 jt->set_suspend_equivalent(); 1473 1474 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1475 // Intentionally empty 1476 } else if (node._notified == 0) { 1477 if (millis <= 0) { 1478 Self->_ParkEvent->park(); 1479 } else { 1480 ret = Self->_ParkEvent->park(millis); 1481 } 1482 } 1483 1484 // were we externally suspended while we were waiting? 1485 if (ExitSuspendEquivalent (jt)) { 1486 // TODO-FIXME: add -- if succ == Self then succ = null. 1487 jt->java_suspend_self(); 1488 } 1489 1490 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1491 1492 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1493 // from the WaitSet to the EntryList. 1494 // See if we need to remove Node from the WaitSet. 1495 // We use double-checked locking to avoid grabbing _WaitSetLock 1496 // if the thread is not on the wait queue. 1497 // 1498 // Note that we don't need a fence before the fetch of TState. 1499 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1500 // written by the is thread. (perhaps the fetch might even be satisfied 1501 // by a look-aside into the processor's own store buffer, although given 1502 // the length of the code path between the prior ST and this load that's 1503 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1504 // then we'll acquire the lock and then re-fetch a fresh TState value. 1505 // That is, we fail toward safety. 1506 1507 if (node.TState == ObjectWaiter::TS_WAIT) { 1508 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1509 if (node.TState == ObjectWaiter::TS_WAIT) { 1510 DequeueSpecificWaiter(&node); // unlink from WaitSet 1511 assert(node._notified == 0, "invariant"); 1512 node.TState = ObjectWaiter::TS_RUN; 1513 } 1514 Thread::SpinRelease(&_WaitSetLock); 1515 } 1516 1517 // The thread is now either on off-list (TS_RUN), 1518 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1519 // The Node's TState variable is stable from the perspective of this thread. 1520 // No other threads will asynchronously modify TState. 1521 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1522 OrderAccess::loadload(); 1523 if (_succ == Self) _succ = NULL; 1524 WasNotified = node._notified; 1525 1526 // Reentry phase -- reacquire the monitor. 1527 // re-enter contended monitor after object.wait(). 1528 // retain OBJECT_WAIT state until re-enter successfully completes 1529 // Thread state is thread_in_vm and oop access is again safe, 1530 // although the raw address of the object may have changed. 1531 // (Don't cache naked oops over safepoints, of course). 1532 1533 // post monitor waited event. Note that this is past-tense, we are done waiting. 1534 if (JvmtiExport::should_post_monitor_waited()) { 1535 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1536 1537 if (node._notified != 0 && _succ == Self) { 1538 // In this part of the monitor wait-notify-reenter protocol it 1539 // is possible (and normal) for another thread to do a fastpath 1540 // monitor enter-exit while this thread is still trying to get 1541 // to the reenter portion of the protocol. 1542 // 1543 // The ObjectMonitor was notified and the current thread is 1544 // the successor which also means that an unpark() has already 1545 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1546 // consume the unpark() that was done when the successor was 1547 // set because the same ParkEvent is shared between Java 1548 // monitors and JVM/TI RawMonitors (for now). 1549 // 1550 // We redo the unpark() to ensure forward progress, i.e., we 1551 // don't want all pending threads hanging (parked) with none 1552 // entering the unlocked monitor. 1553 node._event->unpark(); 1554 } 1555 } 1556 1557 if (event.should_commit()) { 1558 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1559 } 1560 1561 OrderAccess::fence(); 1562 1563 assert(Self->_Stalled != 0, "invariant"); 1564 Self->_Stalled = 0; 1565 1566 assert(_owner != Self, "invariant"); 1567 ObjectWaiter::TStates v = node.TState; 1568 if (v == ObjectWaiter::TS_RUN) { 1569 enter(Self); 1570 } else { 1571 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1572 ReenterI(Self, &node); 1573 node.wait_reenter_end(this); 1574 } 1575 1576 // Self has reacquired the lock. 1577 // Lifecycle - the node representing Self must not appear on any queues. 1578 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1579 // want residual elements associated with this thread left on any lists. 1580 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1581 assert(_owner == Self, "invariant"); 1582 assert(_succ != Self, "invariant"); 1583 } // OSThreadWaitState() 1584 1585 jt->set_current_waiting_monitor(NULL); 1586 1587 guarantee(_recursions == 0, "invariant"); 1588 _recursions = save; // restore the old recursion count 1589 _waiters--; // decrement the number of waiters 1590 1591 // Verify a few postconditions 1592 assert(_owner == Self, "invariant"); 1593 assert(_succ != Self, "invariant"); 1594 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1595 1596 if (SyncFlags & 32) { 1597 OrderAccess::fence(); 1598 } 1599 1600 // check if the notification happened 1601 if (!WasNotified) { 1602 // no, it could be timeout or Thread.interrupt() or both 1603 // check for interrupt event, otherwise it is timeout 1604 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1605 THROW(vmSymbols::java_lang_InterruptedException()); 1606 } 1607 } 1608 1609 // NOTE: Spurious wake up will be consider as timeout. 1610 // Monitor notify has precedence over thread interrupt. 1611 } 1612 1613 1614 // Consider: 1615 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1616 // then instead of transferring a thread from the WaitSet to the EntryList 1617 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1618 1619 void ObjectMonitor::INotify(Thread * Self) { 1620 const int policy = Knob_MoveNotifyee; 1621 1622 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1623 ObjectWaiter * iterator = DequeueWaiter(); 1624 if (iterator != NULL) { 1625 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1626 guarantee(iterator->_notified == 0, "invariant"); 1627 // Disposition - what might we do with iterator ? 1628 // a. add it directly to the EntryList - either tail (policy == 1) 1629 // or head (policy == 0). 1630 // b. push it onto the front of the _cxq (policy == 2). 1631 // For now we use (b). 1632 if (policy != 4) { 1633 iterator->TState = ObjectWaiter::TS_ENTER; 1634 } 1635 iterator->_notified = 1; 1636 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1637 1638 ObjectWaiter * list = _EntryList; 1639 if (list != NULL) { 1640 assert(list->_prev == NULL, "invariant"); 1641 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1642 assert(list != iterator, "invariant"); 1643 } 1644 1645 if (policy == 0) { // prepend to EntryList 1646 if (list == NULL) { 1647 iterator->_next = iterator->_prev = NULL; 1648 _EntryList = iterator; 1649 } else { 1650 list->_prev = iterator; 1651 iterator->_next = list; 1652 iterator->_prev = NULL; 1653 _EntryList = iterator; 1654 } 1655 } else if (policy == 1) { // append to EntryList 1656 if (list == NULL) { 1657 iterator->_next = iterator->_prev = NULL; 1658 _EntryList = iterator; 1659 } else { 1660 // CONSIDER: finding the tail currently requires a linear-time walk of 1661 // the EntryList. We can make tail access constant-time by converting to 1662 // a CDLL instead of using our current DLL. 1663 ObjectWaiter * tail; 1664 for (tail = list; tail->_next != NULL; tail = tail->_next) {} 1665 assert(tail != NULL && tail->_next == NULL, "invariant"); 1666 tail->_next = iterator; 1667 iterator->_prev = tail; 1668 iterator->_next = NULL; 1669 } 1670 } else if (policy == 2) { // prepend to cxq 1671 if (list == NULL) { 1672 iterator->_next = iterator->_prev = NULL; 1673 _EntryList = iterator; 1674 } else { 1675 iterator->TState = ObjectWaiter::TS_CXQ; 1676 for (;;) { 1677 ObjectWaiter * front = _cxq; 1678 iterator->_next = front; 1679 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1680 break; 1681 } 1682 } 1683 } 1684 } else if (policy == 3) { // append to cxq 1685 iterator->TState = ObjectWaiter::TS_CXQ; 1686 for (;;) { 1687 ObjectWaiter * tail = _cxq; 1688 if (tail == NULL) { 1689 iterator->_next = NULL; 1690 if (Atomic::replace_if_null(iterator, &_cxq)) { 1691 break; 1692 } 1693 } else { 1694 while (tail->_next != NULL) tail = tail->_next; 1695 tail->_next = iterator; 1696 iterator->_prev = tail; 1697 iterator->_next = NULL; 1698 break; 1699 } 1700 } 1701 } else { 1702 ParkEvent * ev = iterator->_event; 1703 iterator->TState = ObjectWaiter::TS_RUN; 1704 OrderAccess::fence(); 1705 ev->unpark(); 1706 } 1707 1708 // _WaitSetLock protects the wait queue, not the EntryList. We could 1709 // move the add-to-EntryList operation, above, outside the critical section 1710 // protected by _WaitSetLock. In practice that's not useful. With the 1711 // exception of wait() timeouts and interrupts the monitor owner 1712 // is the only thread that grabs _WaitSetLock. There's almost no contention 1713 // on _WaitSetLock so it's not profitable to reduce the length of the 1714 // critical section. 1715 1716 if (policy < 4) { 1717 iterator->wait_reenter_begin(this); 1718 } 1719 } 1720 Thread::SpinRelease(&_WaitSetLock); 1721 } 1722 1723 // Consider: a not-uncommon synchronization bug is to use notify() when 1724 // notifyAll() is more appropriate, potentially resulting in stranded 1725 // threads; this is one example of a lost wakeup. A useful diagnostic 1726 // option is to force all notify() operations to behave as notifyAll(). 1727 // 1728 // Note: We can also detect many such problems with a "minimum wait". 1729 // When the "minimum wait" is set to a small non-zero timeout value 1730 // and the program does not hang whereas it did absent "minimum wait", 1731 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses 1732 // a "minimum wait" for all park() operations; see the recheckInterval 1733 // variable and MAX_RECHECK_INTERVAL. 1734 1735 void ObjectMonitor::notify(TRAPS) { 1736 CHECK_OWNER(); 1737 if (_WaitSet == NULL) { 1738 return; 1739 } 1740 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1741 INotify(THREAD); 1742 OM_PERFDATA_OP(Notifications, inc(1)); 1743 } 1744 1745 1746 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1747 // from the waitset to the EntryList. This could be done more efficiently with a 1748 // single bulk transfer but in practice it's not time-critical. Beware too, 1749 // that in prepend-mode we invert the order of the waiters. Let's say that the 1750 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1751 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1752 1753 void ObjectMonitor::notifyAll(TRAPS) { 1754 CHECK_OWNER(); 1755 if (_WaitSet == NULL) { 1756 return; 1757 } 1758 1759 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1760 int tally = 0; 1761 while (_WaitSet != NULL) { 1762 tally++; 1763 INotify(THREAD); 1764 } 1765 1766 OM_PERFDATA_OP(Notifications, inc(tally)); 1767 } 1768 1769 // ----------------------------------------------------------------------------- 1770 // Adaptive Spinning Support 1771 // 1772 // Adaptive spin-then-block - rational spinning 1773 // 1774 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1775 // algorithm. On high order SMP systems it would be better to start with 1776 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1777 // a contending thread could enqueue itself on the cxq and then spin locally 1778 // on a thread-specific variable such as its ParkEvent._Event flag. 1779 // That's left as an exercise for the reader. Note that global spinning is 1780 // not problematic on Niagara, as the L2 cache serves the interconnect and 1781 // has both low latency and massive bandwidth. 1782 // 1783 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1784 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1785 // (duration) or we can fix the count at approximately the duration of 1786 // a context switch and vary the frequency. Of course we could also 1787 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1788 // For a description of 'Adaptive spin-then-block mutual exclusion in 1789 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1790 // 1791 // This implementation varies the duration "D", where D varies with 1792 // the success rate of recent spin attempts. (D is capped at approximately 1793 // length of a round-trip context switch). The success rate for recent 1794 // spin attempts is a good predictor of the success rate of future spin 1795 // attempts. The mechanism adapts automatically to varying critical 1796 // section length (lock modality), system load and degree of parallelism. 1797 // D is maintained per-monitor in _SpinDuration and is initialized 1798 // optimistically. Spin frequency is fixed at 100%. 1799 // 1800 // Note that _SpinDuration is volatile, but we update it without locks 1801 // or atomics. The code is designed so that _SpinDuration stays within 1802 // a reasonable range even in the presence of races. The arithmetic 1803 // operations on _SpinDuration are closed over the domain of legal values, 1804 // so at worst a race will install and older but still legal value. 1805 // At the very worst this introduces some apparent non-determinism. 1806 // We might spin when we shouldn't or vice-versa, but since the spin 1807 // count are relatively short, even in the worst case, the effect is harmless. 1808 // 1809 // Care must be taken that a low "D" value does not become an 1810 // an absorbing state. Transient spinning failures -- when spinning 1811 // is overall profitable -- should not cause the system to converge 1812 // on low "D" values. We want spinning to be stable and predictable 1813 // and fairly responsive to change and at the same time we don't want 1814 // it to oscillate, become metastable, be "too" non-deterministic, 1815 // or converge on or enter undesirable stable absorbing states. 1816 // 1817 // We implement a feedback-based control system -- using past behavior 1818 // to predict future behavior. We face two issues: (a) if the 1819 // input signal is random then the spin predictor won't provide optimal 1820 // results, and (b) if the signal frequency is too high then the control 1821 // system, which has some natural response lag, will "chase" the signal. 1822 // (b) can arise from multimodal lock hold times. Transient preemption 1823 // can also result in apparent bimodal lock hold times. 1824 // Although sub-optimal, neither condition is particularly harmful, as 1825 // in the worst-case we'll spin when we shouldn't or vice-versa. 1826 // The maximum spin duration is rather short so the failure modes aren't bad. 1827 // To be conservative, I've tuned the gain in system to bias toward 1828 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1829 // "rings" or oscillates between spinning and not spinning. This happens 1830 // when spinning is just on the cusp of profitability, however, so the 1831 // situation is not dire. The state is benign -- there's no need to add 1832 // hysteresis control to damp the transition rate between spinning and 1833 // not spinning. 1834 1835 // Spinning: Fixed frequency (100%), vary duration 1836 int ObjectMonitor::TrySpin(Thread * Self) { 1837 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1838 int ctr = Knob_FixedSpin; 1839 if (ctr != 0) { 1840 while (--ctr >= 0) { 1841 if (TryLock(Self) > 0) return 1; 1842 SpinPause(); 1843 } 1844 return 0; 1845 } 1846 1847 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1848 if (TryLock(Self) > 0) { 1849 // Increase _SpinDuration ... 1850 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1851 // Raising _SpurDuration to the poverty line is key. 1852 int x = _SpinDuration; 1853 if (x < Knob_SpinLimit) { 1854 if (x < Knob_Poverty) x = Knob_Poverty; 1855 _SpinDuration = x + Knob_BonusB; 1856 } 1857 return 1; 1858 } 1859 SpinPause(); 1860 } 1861 1862 // Admission control - verify preconditions for spinning 1863 // 1864 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1865 // becoming an absorbing state. Put another way, we spin briefly to 1866 // sample, just in case the system load, parallelism, contention, or lock 1867 // modality changed. 1868 // 1869 // Consider the following alternative: 1870 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1871 // spin attempt. "Periodically" might mean after a tally of 1872 // the # of failed spin attempts (or iterations) reaches some threshold. 1873 // This takes us into the realm of 1-out-of-N spinning, where we 1874 // hold the duration constant but vary the frequency. 1875 1876 ctr = _SpinDuration; 1877 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 1878 if (ctr <= 0) return 0; 1879 1880 if (Knob_SuccRestrict && _succ != NULL) return 0; 1881 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1882 return 0; 1883 } 1884 1885 int MaxSpin = Knob_MaxSpinners; 1886 if (MaxSpin >= 0) { 1887 if (_Spinner > MaxSpin) { 1888 return 0; 1889 } 1890 // Slightly racy, but benign ... 1891 Adjust(&_Spinner, 1); 1892 } 1893 1894 // We're good to spin ... spin ingress. 1895 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1896 // when preparing to LD...CAS _owner, etc and the CAS is likely 1897 // to succeed. 1898 int hits = 0; 1899 int msk = 0; 1900 int caspty = Knob_CASPenalty; 1901 int oxpty = Knob_OXPenalty; 1902 int sss = Knob_SpinSetSucc; 1903 if (sss && _succ == NULL) _succ = Self; 1904 Thread * prv = NULL; 1905 1906 // There are three ways to exit the following loop: 1907 // 1. A successful spin where this thread has acquired the lock. 1908 // 2. Spin failure with prejudice 1909 // 3. Spin failure without prejudice 1910 1911 while (--ctr >= 0) { 1912 1913 // Periodic polling -- Check for pending GC 1914 // Threads may spin while they're unsafe. 1915 // We don't want spinning threads to delay the JVM from reaching 1916 // a stop-the-world safepoint or to steal cycles from GC. 1917 // If we detect a pending safepoint we abort in order that 1918 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1919 // this thread, if safe, doesn't steal cycles from GC. 1920 // This is in keeping with the "no loitering in runtime" rule. 1921 // We periodically check to see if there's a safepoint pending. 1922 if ((ctr & 0xFF) == 0) { 1923 if (SafepointMechanism::poll(Self)) { 1924 goto Abort; // abrupt spin egress 1925 } 1926 if (Knob_UsePause & 1) SpinPause(); 1927 } 1928 1929 if (Knob_UsePause & 2) SpinPause(); 1930 1931 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 1932 // This is useful on classic SMP systems, but is of less utility on 1933 // N1-style CMT platforms. 1934 // 1935 // Trade-off: lock acquisition latency vs coherency bandwidth. 1936 // Lock hold times are typically short. A histogram 1937 // of successful spin attempts shows that we usually acquire 1938 // the lock early in the spin. That suggests we want to 1939 // sample _owner frequently in the early phase of the spin, 1940 // but then back-off and sample less frequently as the spin 1941 // progresses. The back-off makes a good citizen on SMP big 1942 // SMP systems. Oversampling _owner can consume excessive 1943 // coherency bandwidth. Relatedly, if we _oversample _owner we 1944 // can inadvertently interfere with the the ST m->owner=null. 1945 // executed by the lock owner. 1946 if (ctr & msk) continue; 1947 ++hits; 1948 if ((hits & 0xF) == 0) { 1949 // The 0xF, above, corresponds to the exponent. 1950 // Consider: (msk+1)|msk 1951 msk = ((msk << 2)|3) & BackOffMask; 1952 } 1953 1954 // Probe _owner with TATAS 1955 // If this thread observes the monitor transition or flicker 1956 // from locked to unlocked to locked, then the odds that this 1957 // thread will acquire the lock in this spin attempt go down 1958 // considerably. The same argument applies if the CAS fails 1959 // or if we observe _owner change from one non-null value to 1960 // another non-null value. In such cases we might abort 1961 // the spin without prejudice or apply a "penalty" to the 1962 // spin count-down variable "ctr", reducing it by 100, say. 1963 1964 Thread * ox = (Thread *) _owner; 1965 if (ox == NULL) { 1966 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL); 1967 if (ox == NULL) { 1968 // The CAS succeeded -- this thread acquired ownership 1969 // Take care of some bookkeeping to exit spin state. 1970 if (sss && _succ == Self) { 1971 _succ = NULL; 1972 } 1973 if (MaxSpin > 0) Adjust(&_Spinner, -1); 1974 1975 // Increase _SpinDuration : 1976 // The spin was successful (profitable) so we tend toward 1977 // longer spin attempts in the future. 1978 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1979 // If we acquired the lock early in the spin cycle it 1980 // makes sense to increase _SpinDuration proportionally. 1981 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1982 int x = _SpinDuration; 1983 if (x < Knob_SpinLimit) { 1984 if (x < Knob_Poverty) x = Knob_Poverty; 1985 _SpinDuration = x + Knob_Bonus; 1986 } 1987 return 1; 1988 } 1989 1990 // The CAS failed ... we can take any of the following actions: 1991 // * penalize: ctr -= Knob_CASPenalty 1992 // * exit spin with prejudice -- goto Abort; 1993 // * exit spin without prejudice. 1994 // * Since CAS is high-latency, retry again immediately. 1995 prv = ox; 1996 if (caspty == -2) break; 1997 if (caspty == -1) goto Abort; 1998 ctr -= caspty; 1999 continue; 2000 } 2001 2002 // Did lock ownership change hands ? 2003 if (ox != prv && prv != NULL) { 2004 if (oxpty == -2) break; 2005 if (oxpty == -1) goto Abort; 2006 ctr -= oxpty; 2007 } 2008 prv = ox; 2009 2010 // Abort the spin if the owner is not executing. 2011 // The owner must be executing in order to drop the lock. 2012 // Spinning while the owner is OFFPROC is idiocy. 2013 // Consider: ctr -= RunnablePenalty ; 2014 if (Knob_OState && NotRunnable (Self, ox)) { 2015 goto Abort; 2016 } 2017 if (sss && _succ == NULL) _succ = Self; 2018 } 2019 2020 // Spin failed with prejudice -- reduce _SpinDuration. 2021 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2022 // AIMD is globally stable. 2023 { 2024 int x = _SpinDuration; 2025 if (x > 0) { 2026 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2027 // This is globally sample and tends to damp the response. 2028 x -= Knob_Penalty; 2029 if (x < 0) x = 0; 2030 _SpinDuration = x; 2031 } 2032 } 2033 2034 Abort: 2035 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2036 if (sss && _succ == Self) { 2037 _succ = NULL; 2038 // Invariant: after setting succ=null a contending thread 2039 // must recheck-retry _owner before parking. This usually happens 2040 // in the normal usage of TrySpin(), but it's safest 2041 // to make TrySpin() as foolproof as possible. 2042 OrderAccess::fence(); 2043 if (TryLock(Self) > 0) return 1; 2044 } 2045 return 0; 2046 } 2047 2048 // NotRunnable() -- informed spinning 2049 // 2050 // Don't bother spinning if the owner is not eligible to drop the lock. 2051 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2052 // The thread must be runnable in order to drop the lock in timely fashion. 2053 // If the _owner is not runnable then spinning will not likely be 2054 // successful (profitable). 2055 // 2056 // Beware -- the thread referenced by _owner could have died 2057 // so a simply fetch from _owner->_thread_state might trap. 2058 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2059 // Because of the lifecycle issues, the _thread_state values 2060 // observed by NotRunnable() might be garbage. NotRunnable must 2061 // tolerate this and consider the observed _thread_state value 2062 // as advisory. 2063 // 2064 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2065 // a thread pointer. 2066 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2067 // with the LSB of _owner. Another option would be to probabilistically probe 2068 // the putative _owner->TypeTag value. 2069 // 2070 // Checking _thread_state isn't perfect. Even if the thread is 2071 // in_java it might be blocked on a page-fault or have been preempted 2072 // and sitting on a ready/dispatch queue. 2073 // 2074 // The return value from NotRunnable() is *advisory* -- the 2075 // result is based on sampling and is not necessarily coherent. 2076 // The caller must tolerate false-negative and false-positive errors. 2077 // Spinning, in general, is probabilistic anyway. 2078 2079 2080 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2081 // Check ox->TypeTag == 2BAD. 2082 if (ox == NULL) return 0; 2083 2084 // Avoid transitive spinning ... 2085 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2086 // Immediately after T1 acquires L it's possible that T2, also 2087 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2088 // This occurs transiently after T1 acquired L but before 2089 // T1 managed to clear T1.Stalled. T2 does not need to abort 2090 // its spin in this circumstance. 2091 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2092 2093 if (BlockedOn == 1) return 1; 2094 if (BlockedOn != 0) { 2095 return BlockedOn != intptr_t(this) && _owner == ox; 2096 } 2097 2098 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2099 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2100 // consider also: jst != _thread_in_Java -- but that's overspecific. 2101 return jst == _thread_blocked || jst == _thread_in_native; 2102 } 2103 2104 2105 // ----------------------------------------------------------------------------- 2106 // WaitSet management ... 2107 2108 ObjectWaiter::ObjectWaiter(Thread* thread) { 2109 _next = NULL; 2110 _prev = NULL; 2111 _notified = 0; 2112 _notifier_tid = 0; 2113 TState = TS_RUN; 2114 _thread = thread; 2115 _event = thread->_ParkEvent; 2116 _active = false; 2117 assert(_event != NULL, "invariant"); 2118 } 2119 2120 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2121 JavaThread *jt = (JavaThread *)this->_thread; 2122 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2123 } 2124 2125 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2126 JavaThread *jt = (JavaThread *)this->_thread; 2127 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2128 } 2129 2130 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2131 assert(node != NULL, "should not add NULL node"); 2132 assert(node->_prev == NULL, "node already in list"); 2133 assert(node->_next == NULL, "node already in list"); 2134 // put node at end of queue (circular doubly linked list) 2135 if (_WaitSet == NULL) { 2136 _WaitSet = node; 2137 node->_prev = node; 2138 node->_next = node; 2139 } else { 2140 ObjectWaiter* head = _WaitSet; 2141 ObjectWaiter* tail = head->_prev; 2142 assert(tail->_next == head, "invariant check"); 2143 tail->_next = node; 2144 head->_prev = node; 2145 node->_next = head; 2146 node->_prev = tail; 2147 } 2148 } 2149 2150 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2151 // dequeue the very first waiter 2152 ObjectWaiter* waiter = _WaitSet; 2153 if (waiter) { 2154 DequeueSpecificWaiter(waiter); 2155 } 2156 return waiter; 2157 } 2158 2159 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2160 assert(node != NULL, "should not dequeue NULL node"); 2161 assert(node->_prev != NULL, "node already removed from list"); 2162 assert(node->_next != NULL, "node already removed from list"); 2163 // when the waiter has woken up because of interrupt, 2164 // timeout or other spurious wake-up, dequeue the 2165 // waiter from waiting list 2166 ObjectWaiter* next = node->_next; 2167 if (next == node) { 2168 assert(node->_prev == node, "invariant check"); 2169 _WaitSet = NULL; 2170 } else { 2171 ObjectWaiter* prev = node->_prev; 2172 assert(prev->_next == node, "invariant check"); 2173 assert(next->_prev == node, "invariant check"); 2174 next->_prev = prev; 2175 prev->_next = next; 2176 if (_WaitSet == node) { 2177 _WaitSet = next; 2178 } 2179 } 2180 node->_next = NULL; 2181 node->_prev = NULL; 2182 } 2183 2184 // ----------------------------------------------------------------------------- 2185 // PerfData support 2186 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2187 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2188 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2189 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2190 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2191 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2192 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2193 2194 // One-shot global initialization for the sync subsystem. 2195 // We could also defer initialization and initialize on-demand 2196 // the first time we call inflate(). Initialization would 2197 // be protected - like so many things - by the MonitorCache_lock. 2198 2199 void ObjectMonitor::Initialize() { 2200 static int InitializationCompleted = 0; 2201 assert(InitializationCompleted == 0, "invariant"); 2202 InitializationCompleted = 1; 2203 if (UsePerfData) { 2204 EXCEPTION_MARK; 2205 #define NEWPERFCOUNTER(n) \ 2206 { \ 2207 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2208 CHECK); \ 2209 } 2210 #define NEWPERFVARIABLE(n) \ 2211 { \ 2212 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2213 CHECK); \ 2214 } 2215 NEWPERFCOUNTER(_sync_Inflations); 2216 NEWPERFCOUNTER(_sync_Deflations); 2217 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2218 NEWPERFCOUNTER(_sync_FutileWakeups); 2219 NEWPERFCOUNTER(_sync_Parks); 2220 NEWPERFCOUNTER(_sync_Notifications); 2221 NEWPERFVARIABLE(_sync_MonExtant); 2222 #undef NEWPERFCOUNTER 2223 #undef NEWPERFVARIABLE 2224 } 2225 } 2226 2227 static char * kvGet(char * kvList, const char * Key) { 2228 if (kvList == NULL) return NULL; 2229 size_t n = strlen(Key); 2230 char * Search; 2231 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2232 if (strncmp (Search, Key, n) == 0) { 2233 if (Search[n] == '=') return Search + n + 1; 2234 if (Search[n] == 0) return(char *) "1"; 2235 } 2236 } 2237 return NULL; 2238 } 2239 2240 static int kvGetInt(char * kvList, const char * Key, int Default) { 2241 char * v = kvGet(kvList, Key); 2242 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2243 if (Knob_ReportSettings && v != NULL) { 2244 tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ; 2245 tty->flush(); 2246 } 2247 return rslt; 2248 } 2249 2250 void ObjectMonitor::DeferredInitialize() { 2251 if (InitDone > 0) return; 2252 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2253 while (InitDone != 1) /* empty */; 2254 return; 2255 } 2256 2257 // One-shot global initialization ... 2258 // The initialization is idempotent, so we don't need locks. 2259 // In the future consider doing this via os::init_2(). 2260 // SyncKnobs consist of <Key>=<Value> pairs in the style 2261 // of environment variables. Start by converting ':' to NUL. 2262 2263 if (SyncKnobs == NULL) SyncKnobs = ""; 2264 2265 size_t sz = strlen(SyncKnobs); 2266 char * knobs = (char *) os::malloc(sz + 2, mtInternal); 2267 if (knobs == NULL) { 2268 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2269 guarantee(0, "invariant"); 2270 } 2271 strcpy(knobs, SyncKnobs); 2272 knobs[sz+1] = 0; 2273 for (char * p = knobs; *p; p++) { 2274 if (*p == ':') *p = 0; 2275 } 2276 2277 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2278 SETKNOB(ReportSettings); 2279 SETKNOB(ExitRelease); 2280 SETKNOB(InlineNotify); 2281 SETKNOB(Verbose); 2282 SETKNOB(VerifyInUse); 2283 SETKNOB(VerifyMatch); 2284 SETKNOB(FixedSpin); 2285 SETKNOB(SpinLimit); 2286 SETKNOB(SpinBase); 2287 SETKNOB(SpinBackOff); 2288 SETKNOB(CASPenalty); 2289 SETKNOB(OXPenalty); 2290 SETKNOB(SpinSetSucc); 2291 SETKNOB(SuccEnabled); 2292 SETKNOB(SuccRestrict); 2293 SETKNOB(Penalty); 2294 SETKNOB(Bonus); 2295 SETKNOB(BonusB); 2296 SETKNOB(Poverty); 2297 SETKNOB(SpinAfterFutile); 2298 SETKNOB(UsePause); 2299 SETKNOB(SpinEarly); 2300 SETKNOB(OState); 2301 SETKNOB(MaxSpinners); 2302 SETKNOB(PreSpin); 2303 SETKNOB(ExitPolicy); 2304 SETKNOB(QMode); 2305 SETKNOB(ResetEvent); 2306 SETKNOB(MoveNotifyee); 2307 SETKNOB(FastHSSEC); 2308 #undef SETKNOB 2309 2310 if (os::is_MP()) { 2311 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2312 if (Knob_ReportSettings) { 2313 tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask); 2314 } 2315 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2316 } else { 2317 Knob_SpinLimit = 0; 2318 Knob_SpinBase = 0; 2319 Knob_PreSpin = 0; 2320 Knob_FixedSpin = -1; 2321 } 2322 2323 os::free(knobs); 2324 OrderAccess::fence(); 2325 InitDone = 1; 2326 } 2327