1 /* 2 * Copyright (c) 1998, 2014, 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 "memory/resourceArea.hpp" 28 #include "oops/markOop.hpp" 29 #include "oops/oop.inline.hpp" 30 #include "runtime/handles.inline.hpp" 31 #include "runtime/interfaceSupport.hpp" 32 #include "runtime/mutexLocker.hpp" 33 #include "runtime/objectMonitor.hpp" 34 #include "runtime/objectMonitor.inline.hpp" 35 #include "runtime/orderAccess.inline.hpp" 36 #include "runtime/osThread.hpp" 37 #include "runtime/stubRoutines.hpp" 38 #include "runtime/thread.inline.hpp" 39 #include "services/threadService.hpp" 40 #include "trace/tracing.hpp" 41 #include "trace/traceMacros.hpp" 42 #include "utilities/dtrace.hpp" 43 #include "utilities/macros.hpp" 44 #include "utilities/preserveException.hpp" 45 #ifdef TARGET_OS_FAMILY_linux 46 # include "os_linux.inline.hpp" 47 #endif 48 #ifdef TARGET_OS_FAMILY_solaris 49 # include "os_solaris.inline.hpp" 50 #endif 51 #ifdef TARGET_OS_FAMILY_windows 52 # include "os_windows.inline.hpp" 53 #endif 54 #ifdef TARGET_OS_FAMILY_bsd 55 # include "os_bsd.inline.hpp" 56 #endif 57 58 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 59 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 60 #define ATTR __attribute__((noinline)) 61 #else 62 #define ATTR 63 #endif 64 65 66 #ifdef DTRACE_ENABLED 67 68 // Only bother with this argument setup if dtrace is available 69 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 70 71 72 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 73 char* bytes = NULL; \ 74 int len = 0; \ 75 jlong jtid = SharedRuntime::get_java_tid(thread); \ 76 Symbol* klassname = ((oop)obj)->klass()->name(); \ 77 if (klassname != NULL) { \ 78 bytes = (char*)klassname->bytes(); \ 79 len = klassname->utf8_length(); \ 80 } 81 82 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 83 { \ 84 if (DTraceMonitorProbes) { \ 85 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 86 HOTSPOT_MONITOR_WAIT(jtid, \ 87 (monitor), bytes, len, (millis)); \ 88 } \ 89 } 90 91 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 92 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 93 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 94 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 95 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 96 97 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 98 { \ 99 if (DTraceMonitorProbes) { \ 100 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 101 HOTSPOT_MONITOR_##probe(jtid, \ 102 (uintptr_t)(monitor), bytes, len); \ 103 } \ 104 } 105 106 #else // ndef DTRACE_ENABLED 107 108 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 109 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 110 111 #endif // ndef DTRACE_ENABLED 112 113 // Tunables ... 114 // The knob* variables are effectively final. Once set they should 115 // never be modified hence. Consider using __read_mostly with GCC. 116 117 int ObjectMonitor::Knob_Verbose = 0 ; 118 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 119 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 120 static int Knob_HandOff = 0 ; 121 static int Knob_ReportSettings = 0 ; 122 123 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 124 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 125 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 126 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 127 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 128 static int Knob_SpinEarly = 1 ; 129 static int Knob_SuccEnabled = 1 ; // futile wake throttling 130 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 131 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 132 static int Knob_Bonus = 100 ; // spin success bonus 133 static int Knob_BonusB = 100 ; // spin success bonus 134 static int Knob_Penalty = 200 ; // spin failure penalty 135 static int Knob_Poverty = 1000 ; 136 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 137 static int Knob_FixedSpin = 0 ; 138 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 139 static int Knob_UsePause = 1 ; 140 static int Knob_ExitPolicy = 0 ; 141 static int Knob_PreSpin = 10 ; // 20-100 likely better 142 static int Knob_ResetEvent = 0 ; 143 static int BackOffMask = 0 ; 144 145 static int Knob_FastHSSEC = 0 ; 146 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 147 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 148 static volatile int InitDone = 0 ; 149 150 #define TrySpin TrySpin_VaryDuration 151 152 // ----------------------------------------------------------------------------- 153 // Theory of operations -- Monitors lists, thread residency, etc: 154 // 155 // * A thread acquires ownership of a monitor by successfully 156 // CAS()ing the _owner field from null to non-null. 157 // 158 // * Invariant: A thread appears on at most one monitor list -- 159 // cxq, EntryList or WaitSet -- at any one time. 160 // 161 // * Contending threads "push" themselves onto the cxq with CAS 162 // and then spin/park. 163 // 164 // * After a contending thread eventually acquires the lock it must 165 // dequeue itself from either the EntryList or the cxq. 166 // 167 // * The exiting thread identifies and unparks an "heir presumptive" 168 // tentative successor thread on the EntryList. Critically, the 169 // exiting thread doesn't unlink the successor thread from the EntryList. 170 // After having been unparked, the wakee will recontend for ownership of 171 // the monitor. The successor (wakee) will either acquire the lock or 172 // re-park itself. 173 // 174 // Succession is provided for by a policy of competitive handoff. 175 // The exiting thread does _not_ grant or pass ownership to the 176 // successor thread. (This is also referred to as "handoff" succession"). 177 // Instead the exiting thread releases ownership and possibly wakes 178 // a successor, so the successor can (re)compete for ownership of the lock. 179 // If the EntryList is empty but the cxq is populated the exiting 180 // thread will drain the cxq into the EntryList. It does so by 181 // by detaching the cxq (installing null with CAS) and folding 182 // the threads from the cxq into the EntryList. The EntryList is 183 // doubly linked, while the cxq is singly linked because of the 184 // CAS-based "push" used to enqueue recently arrived threads (RATs). 185 // 186 // * Concurrency invariants: 187 // 188 // -- only the monitor owner may access or mutate the EntryList. 189 // The mutex property of the monitor itself protects the EntryList 190 // from concurrent interference. 191 // -- Only the monitor owner may detach the cxq. 192 // 193 // * The monitor entry list operations avoid locks, but strictly speaking 194 // they're not lock-free. Enter is lock-free, exit is not. 195 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 196 // 197 // * The cxq can have multiple concurrent "pushers" but only one concurrent 198 // detaching thread. This mechanism is immune from the ABA corruption. 199 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 200 // 201 // * Taken together, the cxq and the EntryList constitute or form a 202 // single logical queue of threads stalled trying to acquire the lock. 203 // We use two distinct lists to improve the odds of a constant-time 204 // dequeue operation after acquisition (in the ::enter() epilogue) and 205 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 206 // A key desideratum is to minimize queue & monitor metadata manipulation 207 // that occurs while holding the monitor lock -- that is, we want to 208 // minimize monitor lock holds times. Note that even a small amount of 209 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 210 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 211 // locks and monitor metadata. 212 // 213 // Cxq points to the the set of Recently Arrived Threads attempting entry. 214 // Because we push threads onto _cxq with CAS, the RATs must take the form of 215 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 216 // the unlocking thread notices that EntryList is null but _cxq is != null. 217 // 218 // The EntryList is ordered by the prevailing queue discipline and 219 // can be organized in any convenient fashion, such as a doubly-linked list or 220 // a circular doubly-linked list. Critically, we want insert and delete operations 221 // to operate in constant-time. If we need a priority queue then something akin 222 // to Solaris' sleepq would work nicely. Viz., 223 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 224 // Queue discipline is enforced at ::exit() time, when the unlocking thread 225 // drains the cxq into the EntryList, and orders or reorders the threads on the 226 // EntryList accordingly. 227 // 228 // Barring "lock barging", this mechanism provides fair cyclic ordering, 229 // somewhat similar to an elevator-scan. 230 // 231 // * The monitor synchronization subsystem avoids the use of native 232 // synchronization primitives except for the narrow platform-specific 233 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 234 // the semantics of park-unpark. Put another way, this monitor implementation 235 // depends only on atomic operations and park-unpark. The monitor subsystem 236 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 237 // underlying OS manages the READY<->RUN transitions. 238 // 239 // * Waiting threads reside on the WaitSet list -- wait() puts 240 // the caller onto the WaitSet. 241 // 242 // * notify() or notifyAll() simply transfers threads from the WaitSet to 243 // either the EntryList or cxq. Subsequent exit() operations will 244 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 245 // it's likely the notifyee would simply impale itself on the lock held 246 // by the notifier. 247 // 248 // * An interesting alternative is to encode cxq as (List,LockByte) where 249 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 250 // variable, like _recursions, in the scheme. The threads or Events that form 251 // the list would have to be aligned in 256-byte addresses. A thread would 252 // try to acquire the lock or enqueue itself with CAS, but exiting threads 253 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 254 // Note that is is *not* word-tearing, but it does presume that full-word 255 // CAS operations are coherent with intermix with STB operations. That's true 256 // on most common processors. 257 // 258 // * See also http://blogs.sun.com/dave 259 260 261 // ----------------------------------------------------------------------------- 262 // Enter support 263 264 bool ObjectMonitor::try_enter(Thread* THREAD) { 265 if (THREAD != _owner) { 266 if (THREAD->is_lock_owned ((address)_owner)) { 267 assert(_recursions == 0, "internal state error"); 268 _owner = THREAD ; 269 _recursions = 1 ; 270 OwnerIsThread = 1 ; 271 return true; 272 } 273 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 274 return false; 275 } 276 return true; 277 } else { 278 _recursions++; 279 return true; 280 } 281 } 282 283 void ATTR ObjectMonitor::enter(TRAPS) { 284 // The following code is ordered to check the most common cases first 285 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 286 Thread * const Self = THREAD ; 287 void * cur ; 288 289 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 290 if (cur == NULL) { 291 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 292 assert (_recursions == 0 , "invariant") ; 293 assert (_owner == Self, "invariant") ; 294 // CONSIDER: set or assert OwnerIsThread == 1 295 return ; 296 } 297 298 if (cur == Self) { 299 // TODO-FIXME: check for integer overflow! BUGID 6557169. 300 _recursions ++ ; 301 return ; 302 } 303 304 if (Self->is_lock_owned ((address)cur)) { 305 assert (_recursions == 0, "internal state error"); 306 _recursions = 1 ; 307 // Commute owner from a thread-specific on-stack BasicLockObject address to 308 // a full-fledged "Thread *". 309 _owner = Self ; 310 OwnerIsThread = 1 ; 311 return ; 312 } 313 314 // We've encountered genuine contention. 315 assert (Self->_Stalled == 0, "invariant") ; 316 Self->_Stalled = intptr_t(this) ; 317 318 // Try one round of spinning *before* enqueueing Self 319 // and before going through the awkward and expensive state 320 // transitions. The following spin is strictly optional ... 321 // Note that if we acquire the monitor from an initial spin 322 // we forgo posting JVMTI events and firing DTRACE probes. 323 if (Knob_SpinEarly && TrySpin (Self) > 0) { 324 assert (_owner == Self , "invariant") ; 325 assert (_recursions == 0 , "invariant") ; 326 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 327 Self->_Stalled = 0 ; 328 return ; 329 } 330 331 assert (_owner != Self , "invariant") ; 332 assert (_succ != Self , "invariant") ; 333 assert (Self->is_Java_thread() , "invariant") ; 334 JavaThread * jt = (JavaThread *) Self ; 335 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 336 assert (jt->thread_state() != _thread_blocked , "invariant") ; 337 assert (this->object() != NULL , "invariant") ; 338 assert (_count >= 0, "invariant") ; 339 340 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 341 // Ensure the object-monitor relationship remains stable while there's contention. 342 Atomic::inc_ptr(&_count); 343 344 EventJavaMonitorEnter event; 345 346 { // Change java thread status to indicate blocked on monitor enter. 347 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 348 349 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 350 if (JvmtiExport::should_post_monitor_contended_enter()) { 351 JvmtiExport::post_monitor_contended_enter(jt, this); 352 353 // The current thread does not yet own the monitor and does not 354 // yet appear on any queues that would get it made the successor. 355 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 356 // handler cannot accidentally consume an unpark() meant for the 357 // ParkEvent associated with this ObjectMonitor. 358 } 359 360 OSThreadContendState osts(Self->osthread()); 361 ThreadBlockInVM tbivm(jt); 362 363 Self->set_current_pending_monitor(this); 364 365 // TODO-FIXME: change the following for(;;) loop to straight-line code. 366 for (;;) { 367 jt->set_suspend_equivalent(); 368 // cleared by handle_special_suspend_equivalent_condition() 369 // or java_suspend_self() 370 371 EnterI (THREAD) ; 372 373 if (!ExitSuspendEquivalent(jt)) break ; 374 375 // 376 // We have acquired the contended monitor, but while we were 377 // waiting another thread suspended us. We don't want to enter 378 // the monitor while suspended because that would surprise the 379 // thread that suspended us. 380 // 381 _recursions = 0 ; 382 _succ = NULL ; 383 exit (false, Self) ; 384 385 jt->java_suspend_self(); 386 } 387 Self->set_current_pending_monitor(NULL); 388 } 389 390 Atomic::dec_ptr(&_count); 391 assert (_count >= 0, "invariant") ; 392 Self->_Stalled = 0 ; 393 394 // Must either set _recursions = 0 or ASSERT _recursions == 0. 395 assert (_recursions == 0 , "invariant") ; 396 assert (_owner == Self , "invariant") ; 397 assert (_succ != Self , "invariant") ; 398 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 399 400 // The thread -- now the owner -- is back in vm mode. 401 // Report the glorious news via TI,DTrace and jvmstat. 402 // The probe effect is non-trivial. All the reportage occurs 403 // while we hold the monitor, increasing the length of the critical 404 // section. Amdahl's parallel speedup law comes vividly into play. 405 // 406 // Another option might be to aggregate the events (thread local or 407 // per-monitor aggregation) and defer reporting until a more opportune 408 // time -- such as next time some thread encounters contention but has 409 // yet to acquire the lock. While spinning that thread could 410 // spinning we could increment JVMStat counters, etc. 411 412 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 413 if (JvmtiExport::should_post_monitor_contended_entered()) { 414 JvmtiExport::post_monitor_contended_entered(jt, this); 415 416 // The current thread already owns the monitor and is not going to 417 // call park() for the remainder of the monitor enter protocol. So 418 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 419 // event handler consumed an unpark() issued by the thread that 420 // just exited the monitor. 421 } 422 423 if (event.should_commit()) { 424 event.set_klass(((oop)this->object())->klass()); 425 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 426 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 427 event.commit(); 428 } 429 430 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 431 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 432 } 433 } 434 435 436 // Caveat: TryLock() is not necessarily serializing if it returns failure. 437 // Callers must compensate as needed. 438 439 int ObjectMonitor::TryLock (Thread * Self) { 440 for (;;) { 441 void * own = _owner ; 442 if (own != NULL) return 0 ; 443 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 444 // Either guarantee _recursions == 0 or set _recursions = 0. 445 assert (_recursions == 0, "invariant") ; 446 assert (_owner == Self, "invariant") ; 447 // CONSIDER: set or assert that OwnerIsThread == 1 448 return 1 ; 449 } 450 // The lock had been free momentarily, but we lost the race to the lock. 451 // Interference -- the CAS failed. 452 // We can either return -1 or retry. 453 // Retry doesn't make as much sense because the lock was just acquired. 454 if (true) return -1 ; 455 } 456 } 457 458 void ATTR ObjectMonitor::EnterI (TRAPS) { 459 Thread * Self = THREAD ; 460 assert (Self->is_Java_thread(), "invariant") ; 461 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 462 463 // Try the lock - TATAS 464 if (TryLock (Self) > 0) { 465 assert (_succ != Self , "invariant") ; 466 assert (_owner == Self , "invariant") ; 467 assert (_Responsible != Self , "invariant") ; 468 return ; 469 } 470 471 DeferredInitialize () ; 472 473 // We try one round of spinning *before* enqueueing Self. 474 // 475 // If the _owner is ready but OFFPROC we could use a YieldTo() 476 // operation to donate the remainder of this thread's quantum 477 // to the owner. This has subtle but beneficial affinity 478 // effects. 479 480 if (TrySpin (Self) > 0) { 481 assert (_owner == Self , "invariant") ; 482 assert (_succ != Self , "invariant") ; 483 assert (_Responsible != Self , "invariant") ; 484 return ; 485 } 486 487 // The Spin failed -- Enqueue and park the thread ... 488 assert (_succ != Self , "invariant") ; 489 assert (_owner != Self , "invariant") ; 490 assert (_Responsible != Self , "invariant") ; 491 492 // Enqueue "Self" on ObjectMonitor's _cxq. 493 // 494 // Node acts as a proxy for Self. 495 // As an aside, if were to ever rewrite the synchronization code mostly 496 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 497 // Java objects. This would avoid awkward lifecycle and liveness issues, 498 // as well as eliminate a subset of ABA issues. 499 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 500 // 501 502 ObjectWaiter node(Self) ; 503 Self->_ParkEvent->reset() ; 504 node._prev = (ObjectWaiter *) 0xBAD ; 505 node.TState = ObjectWaiter::TS_CXQ ; 506 507 // Push "Self" onto the front of the _cxq. 508 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 509 // Note that spinning tends to reduce the rate at which threads 510 // enqueue and dequeue on EntryList|cxq. 511 ObjectWaiter * nxt ; 512 for (;;) { 513 node._next = nxt = _cxq ; 514 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 515 516 // Interference - the CAS failed because _cxq changed. Just retry. 517 // As an optional optimization we retry the lock. 518 if (TryLock (Self) > 0) { 519 assert (_succ != Self , "invariant") ; 520 assert (_owner == Self , "invariant") ; 521 assert (_Responsible != Self , "invariant") ; 522 return ; 523 } 524 } 525 526 // Check for cxq|EntryList edge transition to non-null. This indicates 527 // the onset of contention. While contention persists exiting threads 528 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 529 // operations revert to the faster 1-0 mode. This enter operation may interleave 530 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 531 // arrange for one of the contending thread to use a timed park() operations 532 // to detect and recover from the race. (Stranding is form of progress failure 533 // where the monitor is unlocked but all the contending threads remain parked). 534 // That is, at least one of the contended threads will periodically poll _owner. 535 // One of the contending threads will become the designated "Responsible" thread. 536 // The Responsible thread uses a timed park instead of a normal indefinite park 537 // operation -- it periodically wakes and checks for and recovers from potential 538 // strandings admitted by 1-0 exit operations. We need at most one Responsible 539 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 540 // be responsible for a monitor. 541 // 542 // Currently, one of the contended threads takes on the added role of "Responsible". 543 // A viable alternative would be to use a dedicated "stranding checker" thread 544 // that periodically iterated over all the threads (or active monitors) and unparked 545 // successors where there was risk of stranding. This would help eliminate the 546 // timer scalability issues we see on some platforms as we'd only have one thread 547 // -- the checker -- parked on a timer. 548 549 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 550 // Try to assume the role of responsible thread for the monitor. 551 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 552 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 553 } 554 555 // The lock have been released while this thread was occupied queueing 556 // itself onto _cxq. To close the race and avoid "stranding" and 557 // progress-liveness failure we must resample-retry _owner before parking. 558 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 559 // In this case the ST-MEMBAR is accomplished with CAS(). 560 // 561 // TODO: Defer all thread state transitions until park-time. 562 // Since state transitions are heavy and inefficient we'd like 563 // to defer the state transitions until absolutely necessary, 564 // and in doing so avoid some transitions ... 565 566 TEVENT (Inflated enter - Contention) ; 567 int nWakeups = 0 ; 568 int RecheckInterval = 1 ; 569 570 for (;;) { 571 572 if (TryLock (Self) > 0) break ; 573 assert (_owner != Self, "invariant") ; 574 575 if ((SyncFlags & 2) && _Responsible == NULL) { 576 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 577 } 578 579 // park self 580 if (_Responsible == Self || (SyncFlags & 1)) { 581 TEVENT (Inflated enter - park TIMED) ; 582 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 583 // Increase the RecheckInterval, but clamp the value. 584 RecheckInterval *= 8 ; 585 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 586 } else { 587 TEVENT (Inflated enter - park UNTIMED) ; 588 Self->_ParkEvent->park() ; 589 } 590 591 if (TryLock(Self) > 0) break ; 592 593 // The lock is still contested. 594 // Keep a tally of the # of futile wakeups. 595 // Note that the counter is not protected by a lock or updated by atomics. 596 // That is by design - we trade "lossy" counters which are exposed to 597 // races during updates for a lower probe effect. 598 TEVENT (Inflated enter - Futile wakeup) ; 599 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 600 ObjectMonitor::_sync_FutileWakeups->inc() ; 601 } 602 ++ nWakeups ; 603 604 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 605 // We can defer clearing _succ until after the spin completes 606 // TrySpin() must tolerate being called with _succ == Self. 607 // Try yet another round of adaptive spinning. 608 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 609 610 // We can find that we were unpark()ed and redesignated _succ while 611 // we were spinning. That's harmless. If we iterate and call park(), 612 // park() will consume the event and return immediately and we'll 613 // just spin again. This pattern can repeat, leaving _succ to simply 614 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 615 // Alternately, we can sample fired() here, and if set, forgo spinning 616 // in the next iteration. 617 618 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 619 Self->_ParkEvent->reset() ; 620 OrderAccess::fence() ; 621 } 622 if (_succ == Self) _succ = NULL ; 623 624 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 625 OrderAccess::fence() ; 626 } 627 628 // Egress : 629 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 630 // Normally we'll find Self on the EntryList . 631 // From the perspective of the lock owner (this thread), the 632 // EntryList is stable and cxq is prepend-only. 633 // The head of cxq is volatile but the interior is stable. 634 // In addition, Self.TState is stable. 635 636 assert (_owner == Self , "invariant") ; 637 assert (object() != NULL , "invariant") ; 638 // I'd like to write: 639 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 640 // but as we're at a safepoint that's not safe. 641 642 UnlinkAfterAcquire (Self, &node) ; 643 if (_succ == Self) _succ = NULL ; 644 645 assert (_succ != Self, "invariant") ; 646 if (_Responsible == Self) { 647 _Responsible = NULL ; 648 OrderAccess::fence(); // Dekker pivot-point 649 650 // We may leave threads on cxq|EntryList without a designated 651 // "Responsible" thread. This is benign. When this thread subsequently 652 // exits the monitor it can "see" such preexisting "old" threads -- 653 // threads that arrived on the cxq|EntryList before the fence, above -- 654 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 655 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 656 // non-null and elect a new "Responsible" timer thread. 657 // 658 // This thread executes: 659 // ST Responsible=null; MEMBAR (in enter epilogue - here) 660 // LD cxq|EntryList (in subsequent exit) 661 // 662 // Entering threads in the slow/contended path execute: 663 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 664 // The (ST cxq; MEMBAR) is accomplished with CAS(). 665 // 666 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 667 // exit operation from floating above the ST Responsible=null. 668 } 669 670 // We've acquired ownership with CAS(). 671 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 672 // But since the CAS() this thread may have also stored into _succ, 673 // EntryList, cxq or Responsible. These meta-data updates must be 674 // visible __before this thread subsequently drops the lock. 675 // Consider what could occur if we didn't enforce this constraint -- 676 // STs to monitor meta-data and user-data could reorder with (become 677 // visible after) the ST in exit that drops ownership of the lock. 678 // Some other thread could then acquire the lock, but observe inconsistent 679 // or old monitor meta-data and heap data. That violates the JMM. 680 // To that end, the 1-0 exit() operation must have at least STST|LDST 681 // "release" barrier semantics. Specifically, there must be at least a 682 // STST|LDST barrier in exit() before the ST of null into _owner that drops 683 // the lock. The barrier ensures that changes to monitor meta-data and data 684 // protected by the lock will be visible before we release the lock, and 685 // therefore before some other thread (CPU) has a chance to acquire the lock. 686 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 687 // 688 // Critically, any prior STs to _succ or EntryList must be visible before 689 // the ST of null into _owner in the *subsequent* (following) corresponding 690 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 691 // execute a serializing instruction. 692 693 if (SyncFlags & 8) { 694 OrderAccess::fence() ; 695 } 696 return ; 697 } 698 699 // ReenterI() is a specialized inline form of the latter half of the 700 // contended slow-path from EnterI(). We use ReenterI() only for 701 // monitor reentry in wait(). 702 // 703 // In the future we should reconcile EnterI() and ReenterI(), adding 704 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 705 // loop accordingly. 706 707 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 708 assert (Self != NULL , "invariant") ; 709 assert (SelfNode != NULL , "invariant") ; 710 assert (SelfNode->_thread == Self , "invariant") ; 711 assert (_waiters > 0 , "invariant") ; 712 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 713 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 714 JavaThread * jt = (JavaThread *) Self ; 715 716 int nWakeups = 0 ; 717 for (;;) { 718 ObjectWaiter::TStates v = SelfNode->TState ; 719 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 720 assert (_owner != Self, "invariant") ; 721 722 if (TryLock (Self) > 0) break ; 723 if (TrySpin (Self) > 0) break ; 724 725 TEVENT (Wait Reentry - parking) ; 726 727 // State transition wrappers around park() ... 728 // ReenterI() wisely defers state transitions until 729 // it's clear we must park the thread. 730 { 731 OSThreadContendState osts(Self->osthread()); 732 ThreadBlockInVM tbivm(jt); 733 734 // cleared by handle_special_suspend_equivalent_condition() 735 // or java_suspend_self() 736 jt->set_suspend_equivalent(); 737 if (SyncFlags & 1) { 738 Self->_ParkEvent->park ((jlong)1000) ; 739 } else { 740 Self->_ParkEvent->park () ; 741 } 742 743 // were we externally suspended while we were waiting? 744 for (;;) { 745 if (!ExitSuspendEquivalent (jt)) break ; 746 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 747 jt->java_suspend_self(); 748 jt->set_suspend_equivalent(); 749 } 750 } 751 752 // Try again, but just so we distinguish between futile wakeups and 753 // successful wakeups. The following test isn't algorithmically 754 // necessary, but it helps us maintain sensible statistics. 755 if (TryLock(Self) > 0) break ; 756 757 // The lock is still contested. 758 // Keep a tally of the # of futile wakeups. 759 // Note that the counter is not protected by a lock or updated by atomics. 760 // That is by design - we trade "lossy" counters which are exposed to 761 // races during updates for a lower probe effect. 762 TEVENT (Wait Reentry - futile wakeup) ; 763 ++ nWakeups ; 764 765 // Assuming this is not a spurious wakeup we'll normally 766 // find that _succ == Self. 767 if (_succ == Self) _succ = NULL ; 768 769 // Invariant: after clearing _succ a contending thread 770 // *must* retry _owner before parking. 771 OrderAccess::fence() ; 772 773 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 774 ObjectMonitor::_sync_FutileWakeups->inc() ; 775 } 776 } 777 778 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 779 // Normally we'll find Self on the EntryList. 780 // Unlinking from the EntryList is constant-time and atomic-free. 781 // From the perspective of the lock owner (this thread), the 782 // EntryList is stable and cxq is prepend-only. 783 // The head of cxq is volatile but the interior is stable. 784 // In addition, Self.TState is stable. 785 786 assert (_owner == Self, "invariant") ; 787 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 788 UnlinkAfterAcquire (Self, SelfNode) ; 789 if (_succ == Self) _succ = NULL ; 790 assert (_succ != Self, "invariant") ; 791 SelfNode->TState = ObjectWaiter::TS_RUN ; 792 OrderAccess::fence() ; // see comments at the end of EnterI() 793 } 794 795 // after the thread acquires the lock in ::enter(). Equally, we could defer 796 // unlinking the thread until ::exit()-time. 797 798 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 799 { 800 assert (_owner == Self, "invariant") ; 801 assert (SelfNode->_thread == Self, "invariant") ; 802 803 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 804 // Normal case: remove Self from the DLL EntryList . 805 // This is a constant-time operation. 806 ObjectWaiter * nxt = SelfNode->_next ; 807 ObjectWaiter * prv = SelfNode->_prev ; 808 if (nxt != NULL) nxt->_prev = prv ; 809 if (prv != NULL) prv->_next = nxt ; 810 if (SelfNode == _EntryList ) _EntryList = nxt ; 811 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 812 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 813 TEVENT (Unlink from EntryList) ; 814 } else { 815 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 816 // Inopportune interleaving -- Self is still on the cxq. 817 // This usually means the enqueue of self raced an exiting thread. 818 // Normally we'll find Self near the front of the cxq, so 819 // dequeueing is typically fast. If needbe we can accelerate 820 // this with some MCS/CHL-like bidirectional list hints and advisory 821 // back-links so dequeueing from the interior will normally operate 822 // in constant-time. 823 // Dequeue Self from either the head (with CAS) or from the interior 824 // with a linear-time scan and normal non-atomic memory operations. 825 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 826 // and then unlink Self from EntryList. We have to drain eventually, 827 // so it might as well be now. 828 829 ObjectWaiter * v = _cxq ; 830 assert (v != NULL, "invariant") ; 831 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 832 // The CAS above can fail from interference IFF a "RAT" arrived. 833 // In that case Self must be in the interior and can no longer be 834 // at the head of cxq. 835 if (v == SelfNode) { 836 assert (_cxq != v, "invariant") ; 837 v = _cxq ; // CAS above failed - start scan at head of list 838 } 839 ObjectWaiter * p ; 840 ObjectWaiter * q = NULL ; 841 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 842 q = p ; 843 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 844 } 845 assert (v != SelfNode, "invariant") ; 846 assert (p == SelfNode, "Node not found on cxq") ; 847 assert (p != _cxq, "invariant") ; 848 assert (q != NULL, "invariant") ; 849 assert (q->_next == p, "invariant") ; 850 q->_next = p->_next ; 851 } 852 TEVENT (Unlink from cxq) ; 853 } 854 855 // Diagnostic hygiene ... 856 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 857 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 858 SelfNode->TState = ObjectWaiter::TS_RUN ; 859 } 860 861 // ----------------------------------------------------------------------------- 862 // Exit support 863 // 864 // exit() 865 // ~~~~~~ 866 // Note that the collector can't reclaim the objectMonitor or deflate 867 // the object out from underneath the thread calling ::exit() as the 868 // thread calling ::exit() never transitions to a stable state. 869 // This inhibits GC, which in turn inhibits asynchronous (and 870 // inopportune) reclamation of "this". 871 // 872 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 873 // There's one exception to the claim above, however. EnterI() can call 874 // exit() to drop a lock if the acquirer has been externally suspended. 875 // In that case exit() is called with _thread_state as _thread_blocked, 876 // but the monitor's _count field is > 0, which inhibits reclamation. 877 // 878 // 1-0 exit 879 // ~~~~~~~~ 880 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 881 // the fast-path operators have been optimized so the common ::exit() 882 // operation is 1-0. See i486.ad fast_unlock(), for instance. 883 // The code emitted by fast_unlock() elides the usual MEMBAR. This 884 // greatly improves latency -- MEMBAR and CAS having considerable local 885 // latency on modern processors -- but at the cost of "stranding". Absent the 886 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 887 // ::enter() path, resulting in the entering thread being stranding 888 // and a progress-liveness failure. Stranding is extremely rare. 889 // We use timers (timed park operations) & periodic polling to detect 890 // and recover from stranding. Potentially stranded threads periodically 891 // wake up and poll the lock. See the usage of the _Responsible variable. 892 // 893 // The CAS() in enter provides for safety and exclusion, while the CAS or 894 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 895 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 896 // We detect and recover from stranding with timers. 897 // 898 // If a thread transiently strands it'll park until (a) another 899 // thread acquires the lock and then drops the lock, at which time the 900 // exiting thread will notice and unpark the stranded thread, or, (b) 901 // the timer expires. If the lock is high traffic then the stranding latency 902 // will be low due to (a). If the lock is low traffic then the odds of 903 // stranding are lower, although the worst-case stranding latency 904 // is longer. Critically, we don't want to put excessive load in the 905 // platform's timer subsystem. We want to minimize both the timer injection 906 // rate (timers created/sec) as well as the number of timers active at 907 // any one time. (more precisely, we want to minimize timer-seconds, which is 908 // the integral of the # of active timers at any instant over time). 909 // Both impinge on OS scalability. Given that, at most one thread parked on 910 // a monitor will use a timer. 911 912 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 913 Thread * Self = THREAD ; 914 if (THREAD != _owner) { 915 if (THREAD->is_lock_owned((address) _owner)) { 916 // Transmute _owner from a BasicLock pointer to a Thread address. 917 // We don't need to hold _mutex for this transition. 918 // Non-null to Non-null is safe as long as all readers can 919 // tolerate either flavor. 920 assert (_recursions == 0, "invariant") ; 921 _owner = THREAD ; 922 _recursions = 0 ; 923 OwnerIsThread = 1 ; 924 } else { 925 // NOTE: we need to handle unbalanced monitor enter/exit 926 // in native code by throwing an exception. 927 // TODO: Throw an IllegalMonitorStateException ? 928 TEVENT (Exit - Throw IMSX) ; 929 assert(false, "Non-balanced monitor enter/exit!"); 930 if (false) { 931 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 932 } 933 return; 934 } 935 } 936 937 if (_recursions != 0) { 938 _recursions--; // this is simple recursive enter 939 TEVENT (Inflated exit - recursive) ; 940 return ; 941 } 942 943 // Invariant: after setting Responsible=null an thread must execute 944 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 945 if ((SyncFlags & 4) == 0) { 946 _Responsible = NULL ; 947 } 948 949 #if INCLUDE_TRACE 950 // get the owner's thread id for the MonitorEnter event 951 // if it is enabled and the thread isn't suspended 952 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 953 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 954 } 955 #endif 956 957 for (;;) { 958 assert (THREAD == _owner, "invariant") ; 959 960 961 if (Knob_ExitPolicy == 0) { 962 // release semantics: prior loads and stores from within the critical section 963 // must not float (reorder) past the following store that drops the lock. 964 // On SPARC that requires MEMBAR #loadstore|#storestore. 965 // But of course in TSO #loadstore|#storestore is not required. 966 // I'd like to write one of the following: 967 // A. OrderAccess::release() ; _owner = NULL 968 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 969 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 970 // store into a _dummy variable. That store is not needed, but can result 971 // in massive wasteful coherency traffic on classic SMP systems. 972 // Instead, I use release_store(), which is implemented as just a simple 973 // ST on x64, x86 and SPARC. 974 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 975 OrderAccess::storeload() ; // See if we need to wake a successor 976 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 977 TEVENT (Inflated exit - simple egress) ; 978 return ; 979 } 980 TEVENT (Inflated exit - complex egress) ; 981 982 // Normally the exiting thread is responsible for ensuring succession, 983 // but if other successors are ready or other entering threads are spinning 984 // then this thread can simply store NULL into _owner and exit without 985 // waking a successor. The existence of spinners or ready successors 986 // guarantees proper succession (liveness). Responsibility passes to the 987 // ready or running successors. The exiting thread delegates the duty. 988 // More precisely, if a successor already exists this thread is absolved 989 // of the responsibility of waking (unparking) one. 990 // 991 // The _succ variable is critical to reducing futile wakeup frequency. 992 // _succ identifies the "heir presumptive" thread that has been made 993 // ready (unparked) but that has not yet run. We need only one such 994 // successor thread to guarantee progress. 995 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 996 // section 3.3 "Futile Wakeup Throttling" for details. 997 // 998 // Note that spinners in Enter() also set _succ non-null. 999 // In the current implementation spinners opportunistically set 1000 // _succ so that exiting threads might avoid waking a successor. 1001 // Another less appealing alternative would be for the exiting thread 1002 // to drop the lock and then spin briefly to see if a spinner managed 1003 // to acquire the lock. If so, the exiting thread could exit 1004 // immediately without waking a successor, otherwise the exiting 1005 // thread would need to dequeue and wake a successor. 1006 // (Note that we'd need to make the post-drop spin short, but no 1007 // shorter than the worst-case round-trip cache-line migration time. 1008 // The dropped lock needs to become visible to the spinner, and then 1009 // the acquisition of the lock by the spinner must become visible to 1010 // the exiting thread). 1011 // 1012 1013 // It appears that an heir-presumptive (successor) must be made ready. 1014 // Only the current lock owner can manipulate the EntryList or 1015 // drain _cxq, so we need to reacquire the lock. If we fail 1016 // to reacquire the lock the responsibility for ensuring succession 1017 // falls to the new owner. 1018 // 1019 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1020 return ; 1021 } 1022 TEVENT (Exit - Reacquired) ; 1023 } else { 1024 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1025 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1026 OrderAccess::storeload() ; 1027 // Ratify the previously observed values. 1028 if (_cxq == NULL || _succ != NULL) { 1029 TEVENT (Inflated exit - simple egress) ; 1030 return ; 1031 } 1032 1033 // inopportune interleaving -- the exiting thread (this thread) 1034 // in the fast-exit path raced an entering thread in the slow-enter 1035 // path. 1036 // We have two choices: 1037 // A. Try to reacquire the lock. 1038 // If the CAS() fails return immediately, otherwise 1039 // we either restart/rerun the exit operation, or simply 1040 // fall-through into the code below which wakes a successor. 1041 // B. If the elements forming the EntryList|cxq are TSM 1042 // we could simply unpark() the lead thread and return 1043 // without having set _succ. 1044 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1045 TEVENT (Inflated exit - reacquired succeeded) ; 1046 return ; 1047 } 1048 TEVENT (Inflated exit - reacquired failed) ; 1049 } else { 1050 TEVENT (Inflated exit - complex egress) ; 1051 } 1052 } 1053 1054 guarantee (_owner == THREAD, "invariant") ; 1055 1056 ObjectWaiter * w = NULL ; 1057 int QMode = Knob_QMode ; 1058 1059 if (QMode == 2 && _cxq != NULL) { 1060 // QMode == 2 : cxq has precedence over EntryList. 1061 // Try to directly wake a successor from the cxq. 1062 // If successful, the successor will need to unlink itself from cxq. 1063 w = _cxq ; 1064 assert (w != NULL, "invariant") ; 1065 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1066 ExitEpilog (Self, w) ; 1067 return ; 1068 } 1069 1070 if (QMode == 3 && _cxq != NULL) { 1071 // Aggressively drain cxq into EntryList at the first opportunity. 1072 // This policy ensure that recently-run threads live at the head of EntryList. 1073 // Drain _cxq into EntryList - bulk transfer. 1074 // First, detach _cxq. 1075 // The following loop is tantamount to: w = swap (&cxq, NULL) 1076 w = _cxq ; 1077 for (;;) { 1078 assert (w != NULL, "Invariant") ; 1079 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1080 if (u == w) break ; 1081 w = u ; 1082 } 1083 assert (w != NULL , "invariant") ; 1084 1085 ObjectWaiter * q = NULL ; 1086 ObjectWaiter * p ; 1087 for (p = w ; p != NULL ; p = p->_next) { 1088 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1089 p->TState = ObjectWaiter::TS_ENTER ; 1090 p->_prev = q ; 1091 q = p ; 1092 } 1093 1094 // Append the RATs to the EntryList 1095 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1096 ObjectWaiter * Tail ; 1097 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1098 if (Tail == NULL) { 1099 _EntryList = w ; 1100 } else { 1101 Tail->_next = w ; 1102 w->_prev = Tail ; 1103 } 1104 1105 // Fall thru into code that tries to wake a successor from EntryList 1106 } 1107 1108 if (QMode == 4 && _cxq != NULL) { 1109 // Aggressively drain cxq into EntryList at the first opportunity. 1110 // This policy ensure that recently-run threads live at the head of EntryList. 1111 1112 // Drain _cxq into EntryList - bulk transfer. 1113 // First, detach _cxq. 1114 // The following loop is tantamount to: w = swap (&cxq, NULL) 1115 w = _cxq ; 1116 for (;;) { 1117 assert (w != NULL, "Invariant") ; 1118 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1119 if (u == w) break ; 1120 w = u ; 1121 } 1122 assert (w != NULL , "invariant") ; 1123 1124 ObjectWaiter * q = NULL ; 1125 ObjectWaiter * p ; 1126 for (p = w ; p != NULL ; p = p->_next) { 1127 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1128 p->TState = ObjectWaiter::TS_ENTER ; 1129 p->_prev = q ; 1130 q = p ; 1131 } 1132 1133 // Prepend the RATs to the EntryList 1134 if (_EntryList != NULL) { 1135 q->_next = _EntryList ; 1136 _EntryList->_prev = q ; 1137 } 1138 _EntryList = w ; 1139 1140 // Fall thru into code that tries to wake a successor from EntryList 1141 } 1142 1143 w = _EntryList ; 1144 if (w != NULL) { 1145 // I'd like to write: guarantee (w->_thread != Self). 1146 // But in practice an exiting thread may find itself on the EntryList. 1147 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1148 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1149 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1150 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1151 // release the lock "O". T2 resumes immediately after the ST of null into 1152 // _owner, above. T2 notices that the EntryList is populated, so it 1153 // reacquires the lock and then finds itself on the EntryList. 1154 // Given all that, we have to tolerate the circumstance where "w" is 1155 // associated with Self. 1156 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1157 ExitEpilog (Self, w) ; 1158 return ; 1159 } 1160 1161 // If we find that both _cxq and EntryList are null then just 1162 // re-run the exit protocol from the top. 1163 w = _cxq ; 1164 if (w == NULL) continue ; 1165 1166 // Drain _cxq into EntryList - bulk transfer. 1167 // First, detach _cxq. 1168 // The following loop is tantamount to: w = swap (&cxq, NULL) 1169 for (;;) { 1170 assert (w != NULL, "Invariant") ; 1171 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1172 if (u == w) break ; 1173 w = u ; 1174 } 1175 TEVENT (Inflated exit - drain cxq into EntryList) ; 1176 1177 assert (w != NULL , "invariant") ; 1178 assert (_EntryList == NULL , "invariant") ; 1179 1180 // Convert the LIFO SLL anchored by _cxq into a DLL. 1181 // The list reorganization step operates in O(LENGTH(w)) time. 1182 // It's critical that this step operate quickly as 1183 // "Self" still holds the outer-lock, restricting parallelism 1184 // and effectively lengthening the critical section. 1185 // Invariant: s chases t chases u. 1186 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1187 // we have faster access to the tail. 1188 1189 if (QMode == 1) { 1190 // QMode == 1 : drain cxq to EntryList, reversing order 1191 // We also reverse the order of the list. 1192 ObjectWaiter * s = NULL ; 1193 ObjectWaiter * t = w ; 1194 ObjectWaiter * u = NULL ; 1195 while (t != NULL) { 1196 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1197 t->TState = ObjectWaiter::TS_ENTER ; 1198 u = t->_next ; 1199 t->_prev = u ; 1200 t->_next = s ; 1201 s = t; 1202 t = u ; 1203 } 1204 _EntryList = s ; 1205 assert (s != NULL, "invariant") ; 1206 } else { 1207 // QMode == 0 or QMode == 2 1208 _EntryList = w ; 1209 ObjectWaiter * q = NULL ; 1210 ObjectWaiter * p ; 1211 for (p = w ; p != NULL ; p = p->_next) { 1212 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1213 p->TState = ObjectWaiter::TS_ENTER ; 1214 p->_prev = q ; 1215 q = p ; 1216 } 1217 } 1218 1219 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1220 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1221 1222 // See if we can abdicate to a spinner instead of waking a thread. 1223 // A primary goal of the implementation is to reduce the 1224 // context-switch rate. 1225 if (_succ != NULL) continue; 1226 1227 w = _EntryList ; 1228 if (w != NULL) { 1229 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1230 ExitEpilog (Self, w) ; 1231 return ; 1232 } 1233 } 1234 } 1235 1236 // ExitSuspendEquivalent: 1237 // A faster alternate to handle_special_suspend_equivalent_condition() 1238 // 1239 // handle_special_suspend_equivalent_condition() unconditionally 1240 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1241 // operations have high latency. Note that in ::enter() we call HSSEC 1242 // while holding the monitor, so we effectively lengthen the critical sections. 1243 // 1244 // There are a number of possible solutions: 1245 // 1246 // A. To ameliorate the problem we might also defer state transitions 1247 // to as late as possible -- just prior to parking. 1248 // Given that, we'd call HSSEC after having returned from park(), 1249 // but before attempting to acquire the monitor. This is only a 1250 // partial solution. It avoids calling HSSEC while holding the 1251 // monitor (good), but it still increases successor reacquisition latency -- 1252 // the interval between unparking a successor and the time the successor 1253 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1254 // If we use this technique we can also avoid EnterI()-exit() loop 1255 // in ::enter() where we iteratively drop the lock and then attempt 1256 // to reacquire it after suspending. 1257 // 1258 // B. In the future we might fold all the suspend bits into a 1259 // composite per-thread suspend flag and then update it with CAS(). 1260 // Alternately, a Dekker-like mechanism with multiple variables 1261 // would suffice: 1262 // ST Self->_suspend_equivalent = false 1263 // MEMBAR 1264 // LD Self_>_suspend_flags 1265 // 1266 1267 1268 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1269 int Mode = Knob_FastHSSEC ; 1270 if (Mode && !jSelf->is_external_suspend()) { 1271 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1272 jSelf->clear_suspend_equivalent() ; 1273 if (2 == Mode) OrderAccess::storeload() ; 1274 if (!jSelf->is_external_suspend()) return false ; 1275 // We raced a suspension -- fall thru into the slow path 1276 TEVENT (ExitSuspendEquivalent - raced) ; 1277 jSelf->set_suspend_equivalent() ; 1278 } 1279 return jSelf->handle_special_suspend_equivalent_condition() ; 1280 } 1281 1282 1283 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1284 assert (_owner == Self, "invariant") ; 1285 1286 // Exit protocol: 1287 // 1. ST _succ = wakee 1288 // 2. membar #loadstore|#storestore; 1289 // 2. ST _owner = NULL 1290 // 3. unpark(wakee) 1291 1292 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1293 ParkEvent * Trigger = Wakee->_event ; 1294 1295 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1296 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1297 // out-of-scope (non-extant). 1298 Wakee = NULL ; 1299 1300 // Drop the lock 1301 OrderAccess::release_store_ptr (&_owner, NULL) ; 1302 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1303 1304 if (SafepointSynchronize::do_call_back()) { 1305 TEVENT (unpark before SAFEPOINT) ; 1306 } 1307 1308 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1309 Trigger->unpark() ; 1310 1311 // Maintain stats and report events to JVMTI 1312 if (ObjectMonitor::_sync_Parks != NULL) { 1313 ObjectMonitor::_sync_Parks->inc() ; 1314 } 1315 } 1316 1317 1318 // ----------------------------------------------------------------------------- 1319 // Class Loader deadlock handling. 1320 // 1321 // complete_exit exits a lock returning recursion count 1322 // complete_exit/reenter operate as a wait without waiting 1323 // complete_exit requires an inflated monitor 1324 // The _owner field is not always the Thread addr even with an 1325 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1326 // thread due to contention. 1327 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1328 Thread * const Self = THREAD; 1329 assert(Self->is_Java_thread(), "Must be Java thread!"); 1330 JavaThread *jt = (JavaThread *)THREAD; 1331 1332 DeferredInitialize(); 1333 1334 if (THREAD != _owner) { 1335 if (THREAD->is_lock_owned ((address)_owner)) { 1336 assert(_recursions == 0, "internal state error"); 1337 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1338 _recursions = 0 ; 1339 OwnerIsThread = 1 ; 1340 } 1341 } 1342 1343 guarantee(Self == _owner, "complete_exit not owner"); 1344 intptr_t save = _recursions; // record the old recursion count 1345 _recursions = 0; // set the recursion level to be 0 1346 exit (true, Self) ; // exit the monitor 1347 guarantee (_owner != Self, "invariant"); 1348 return save; 1349 } 1350 1351 // reenter() enters a lock and sets recursion count 1352 // complete_exit/reenter operate as a wait without waiting 1353 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1354 Thread * const Self = THREAD; 1355 assert(Self->is_Java_thread(), "Must be Java thread!"); 1356 JavaThread *jt = (JavaThread *)THREAD; 1357 1358 guarantee(_owner != Self, "reenter already owner"); 1359 enter (THREAD); // enter the monitor 1360 guarantee (_recursions == 0, "reenter recursion"); 1361 _recursions = recursions; 1362 return; 1363 } 1364 1365 1366 // ----------------------------------------------------------------------------- 1367 // A macro is used below because there may already be a pending 1368 // exception which should not abort the execution of the routines 1369 // which use this (which is why we don't put this into check_slow and 1370 // call it with a CHECK argument). 1371 1372 #define CHECK_OWNER() \ 1373 do { \ 1374 if (THREAD != _owner) { \ 1375 if (THREAD->is_lock_owned((address) _owner)) { \ 1376 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1377 _recursions = 0; \ 1378 OwnerIsThread = 1 ; \ 1379 } else { \ 1380 TEVENT (Throw IMSX) ; \ 1381 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1382 } \ 1383 } \ 1384 } while (false) 1385 1386 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1387 // TODO-FIXME: remove check_slow() -- it's likely dead. 1388 1389 void ObjectMonitor::check_slow(TRAPS) { 1390 TEVENT (check_slow - throw IMSX) ; 1391 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1392 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1393 } 1394 1395 static int Adjust (volatile int * adr, int dx) { 1396 int v ; 1397 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1398 return v ; 1399 } 1400 1401 // helper method for posting a monitor wait event 1402 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1403 jlong notifier_tid, 1404 jlong timeout, 1405 bool timedout) { 1406 event->set_klass(((oop)this->object())->klass()); 1407 event->set_timeout((TYPE_ULONG)timeout); 1408 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1409 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1410 event->set_timedOut((TYPE_BOOLEAN)timedout); 1411 event->commit(); 1412 } 1413 1414 // ----------------------------------------------------------------------------- 1415 // Wait/Notify/NotifyAll 1416 // 1417 // Note: a subset of changes to ObjectMonitor::wait() 1418 // will need to be replicated in complete_exit above 1419 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1420 Thread * const Self = THREAD ; 1421 assert(Self->is_Java_thread(), "Must be Java thread!"); 1422 JavaThread *jt = (JavaThread *)THREAD; 1423 1424 DeferredInitialize () ; 1425 1426 // Throw IMSX or IEX. 1427 CHECK_OWNER(); 1428 1429 EventJavaMonitorWait event; 1430 1431 // check for a pending interrupt 1432 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1433 // post monitor waited event. Note that this is past-tense, we are done waiting. 1434 if (JvmtiExport::should_post_monitor_waited()) { 1435 // Note: 'false' parameter is passed here because the 1436 // wait was not timed out due to thread interrupt. 1437 JvmtiExport::post_monitor_waited(jt, this, false); 1438 1439 // In this short circuit of the monitor wait protocol, the 1440 // current thread never drops ownership of the monitor and 1441 // never gets added to the wait queue so the current thread 1442 // cannot be made the successor. This means that the 1443 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1444 // consume an unpark() meant for the ParkEvent associated with 1445 // this ObjectMonitor. 1446 } 1447 if (event.should_commit()) { 1448 post_monitor_wait_event(&event, 0, millis, false); 1449 } 1450 TEVENT (Wait - Throw IEX) ; 1451 THROW(vmSymbols::java_lang_InterruptedException()); 1452 return ; 1453 } 1454 1455 TEVENT (Wait) ; 1456 1457 assert (Self->_Stalled == 0, "invariant") ; 1458 Self->_Stalled = intptr_t(this) ; 1459 jt->set_current_waiting_monitor(this); 1460 1461 // create a node to be put into the queue 1462 // Critically, after we reset() the event but prior to park(), we must check 1463 // for a pending interrupt. 1464 ObjectWaiter node(Self); 1465 node.TState = ObjectWaiter::TS_WAIT ; 1466 Self->_ParkEvent->reset() ; 1467 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1468 1469 // Enter the waiting queue, which is a circular doubly linked list in this case 1470 // but it could be a priority queue or any data structure. 1471 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1472 // by the the owner of the monitor *except* in the case where park() 1473 // returns because of a timeout of interrupt. Contention is exceptionally rare 1474 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1475 1476 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1477 AddWaiter (&node) ; 1478 Thread::SpinRelease (&_WaitSetLock) ; 1479 1480 if ((SyncFlags & 4) == 0) { 1481 _Responsible = NULL ; 1482 } 1483 intptr_t save = _recursions; // record the old recursion count 1484 _waiters++; // increment the number of waiters 1485 _recursions = 0; // set the recursion level to be 1 1486 exit (true, Self) ; // exit the monitor 1487 guarantee (_owner != Self, "invariant") ; 1488 1489 // The thread is on the WaitSet list - now park() it. 1490 // On MP systems it's conceivable that a brief spin before we park 1491 // could be profitable. 1492 // 1493 // TODO-FIXME: change the following logic to a loop of the form 1494 // while (!timeout && !interrupted && _notified == 0) park() 1495 1496 int ret = OS_OK ; 1497 int WasNotified = 0 ; 1498 { // State transition wrappers 1499 OSThread* osthread = Self->osthread(); 1500 OSThreadWaitState osts(osthread, true); 1501 { 1502 ThreadBlockInVM tbivm(jt); 1503 // Thread is in thread_blocked state and oop access is unsafe. 1504 jt->set_suspend_equivalent(); 1505 1506 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1507 // Intentionally empty 1508 } else 1509 if (node._notified == 0) { 1510 if (millis <= 0) { 1511 Self->_ParkEvent->park () ; 1512 } else { 1513 ret = Self->_ParkEvent->park (millis) ; 1514 } 1515 } 1516 1517 // were we externally suspended while we were waiting? 1518 if (ExitSuspendEquivalent (jt)) { 1519 // TODO-FIXME: add -- if succ == Self then succ = null. 1520 jt->java_suspend_self(); 1521 } 1522 1523 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1524 1525 1526 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1527 // from the WaitSet to the EntryList. 1528 // See if we need to remove Node from the WaitSet. 1529 // We use double-checked locking to avoid grabbing _WaitSetLock 1530 // if the thread is not on the wait queue. 1531 // 1532 // Note that we don't need a fence before the fetch of TState. 1533 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1534 // written by the is thread. (perhaps the fetch might even be satisfied 1535 // by a look-aside into the processor's own store buffer, although given 1536 // the length of the code path between the prior ST and this load that's 1537 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1538 // then we'll acquire the lock and then re-fetch a fresh TState value. 1539 // That is, we fail toward safety. 1540 1541 if (node.TState == ObjectWaiter::TS_WAIT) { 1542 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1543 if (node.TState == ObjectWaiter::TS_WAIT) { 1544 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1545 assert(node._notified == 0, "invariant"); 1546 node.TState = ObjectWaiter::TS_RUN ; 1547 } 1548 Thread::SpinRelease (&_WaitSetLock) ; 1549 } 1550 1551 // The thread is now either on off-list (TS_RUN), 1552 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1553 // The Node's TState variable is stable from the perspective of this thread. 1554 // No other threads will asynchronously modify TState. 1555 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1556 OrderAccess::loadload() ; 1557 if (_succ == Self) _succ = NULL ; 1558 WasNotified = node._notified ; 1559 1560 // Reentry phase -- reacquire the monitor. 1561 // re-enter contended monitor after object.wait(). 1562 // retain OBJECT_WAIT state until re-enter successfully completes 1563 // Thread state is thread_in_vm and oop access is again safe, 1564 // although the raw address of the object may have changed. 1565 // (Don't cache naked oops over safepoints, of course). 1566 1567 // post monitor waited event. Note that this is past-tense, we are done waiting. 1568 if (JvmtiExport::should_post_monitor_waited()) { 1569 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1570 1571 if (node._notified != 0 && _succ == Self) { 1572 // In this part of the monitor wait-notify-reenter protocol it 1573 // is possible (and normal) for another thread to do a fastpath 1574 // monitor enter-exit while this thread is still trying to get 1575 // to the reenter portion of the protocol. 1576 // 1577 // The ObjectMonitor was notified and the current thread is 1578 // the successor which also means that an unpark() has already 1579 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1580 // consume the unpark() that was done when the successor was 1581 // set because the same ParkEvent is shared between Java 1582 // monitors and JVM/TI RawMonitors (for now). 1583 // 1584 // We redo the unpark() to ensure forward progress, i.e., we 1585 // don't want all pending threads hanging (parked) with none 1586 // entering the unlocked monitor. 1587 node._event->unpark(); 1588 } 1589 } 1590 1591 if (event.should_commit()) { 1592 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1593 } 1594 1595 OrderAccess::fence() ; 1596 1597 assert (Self->_Stalled != 0, "invariant") ; 1598 Self->_Stalled = 0 ; 1599 1600 assert (_owner != Self, "invariant") ; 1601 ObjectWaiter::TStates v = node.TState ; 1602 if (v == ObjectWaiter::TS_RUN) { 1603 enter (Self) ; 1604 } else { 1605 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1606 ReenterI (Self, &node) ; 1607 node.wait_reenter_end(this); 1608 } 1609 1610 // Self has reacquired the lock. 1611 // Lifecycle - the node representing Self must not appear on any queues. 1612 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1613 // want residual elements associated with this thread left on any lists. 1614 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1615 assert (_owner == Self, "invariant") ; 1616 assert (_succ != Self , "invariant") ; 1617 } // OSThreadWaitState() 1618 1619 jt->set_current_waiting_monitor(NULL); 1620 1621 guarantee (_recursions == 0, "invariant") ; 1622 _recursions = save; // restore the old recursion count 1623 _waiters--; // decrement the number of waiters 1624 1625 // Verify a few postconditions 1626 assert (_owner == Self , "invariant") ; 1627 assert (_succ != Self , "invariant") ; 1628 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1629 1630 if (SyncFlags & 32) { 1631 OrderAccess::fence() ; 1632 } 1633 1634 // check if the notification happened 1635 if (!WasNotified) { 1636 // no, it could be timeout or Thread.interrupt() or both 1637 // check for interrupt event, otherwise it is timeout 1638 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1639 TEVENT (Wait - throw IEX from epilog) ; 1640 THROW(vmSymbols::java_lang_InterruptedException()); 1641 } 1642 } 1643 1644 // NOTE: Spurious wake up will be consider as timeout. 1645 // Monitor notify has precedence over thread interrupt. 1646 } 1647 1648 1649 // Consider: 1650 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1651 // then instead of transferring a thread from the WaitSet to the EntryList 1652 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1653 1654 void ObjectMonitor::notify(TRAPS) { 1655 CHECK_OWNER(); 1656 if (_WaitSet == NULL) { 1657 TEVENT (Empty-Notify) ; 1658 return ; 1659 } 1660 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1661 1662 int Policy = Knob_MoveNotifyee ; 1663 1664 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1665 ObjectWaiter * iterator = DequeueWaiter() ; 1666 if (iterator != NULL) { 1667 TEVENT (Notify1 - Transfer) ; 1668 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1669 guarantee (iterator->_notified == 0, "invariant") ; 1670 if (Policy != 4) { 1671 iterator->TState = ObjectWaiter::TS_ENTER ; 1672 } 1673 iterator->_notified = 1 ; 1674 Thread * Self = THREAD; 1675 iterator->_notifier_tid = Self->osthread()->thread_id(); 1676 1677 ObjectWaiter * List = _EntryList ; 1678 if (List != NULL) { 1679 assert (List->_prev == NULL, "invariant") ; 1680 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1681 assert (List != iterator, "invariant") ; 1682 } 1683 1684 if (Policy == 0) { // prepend to EntryList 1685 if (List == NULL) { 1686 iterator->_next = iterator->_prev = NULL ; 1687 _EntryList = iterator ; 1688 } else { 1689 List->_prev = iterator ; 1690 iterator->_next = List ; 1691 iterator->_prev = NULL ; 1692 _EntryList = iterator ; 1693 } 1694 } else 1695 if (Policy == 1) { // append to EntryList 1696 if (List == NULL) { 1697 iterator->_next = iterator->_prev = NULL ; 1698 _EntryList = iterator ; 1699 } else { 1700 // CONSIDER: finding the tail currently requires a linear-time walk of 1701 // the EntryList. We can make tail access constant-time by converting to 1702 // a CDLL instead of using our current DLL. 1703 ObjectWaiter * Tail ; 1704 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1705 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1706 Tail->_next = iterator ; 1707 iterator->_prev = Tail ; 1708 iterator->_next = NULL ; 1709 } 1710 } else 1711 if (Policy == 2) { // prepend to cxq 1712 // prepend to cxq 1713 if (List == NULL) { 1714 iterator->_next = iterator->_prev = NULL ; 1715 _EntryList = iterator ; 1716 } else { 1717 iterator->TState = ObjectWaiter::TS_CXQ ; 1718 for (;;) { 1719 ObjectWaiter * Front = _cxq ; 1720 iterator->_next = Front ; 1721 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1722 break ; 1723 } 1724 } 1725 } 1726 } else 1727 if (Policy == 3) { // append to cxq 1728 iterator->TState = ObjectWaiter::TS_CXQ ; 1729 for (;;) { 1730 ObjectWaiter * Tail ; 1731 Tail = _cxq ; 1732 if (Tail == NULL) { 1733 iterator->_next = NULL ; 1734 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1735 break ; 1736 } 1737 } else { 1738 while (Tail->_next != NULL) Tail = Tail->_next ; 1739 Tail->_next = iterator ; 1740 iterator->_prev = Tail ; 1741 iterator->_next = NULL ; 1742 break ; 1743 } 1744 } 1745 } else { 1746 ParkEvent * ev = iterator->_event ; 1747 iterator->TState = ObjectWaiter::TS_RUN ; 1748 OrderAccess::fence() ; 1749 ev->unpark() ; 1750 } 1751 1752 if (Policy < 4) { 1753 iterator->wait_reenter_begin(this); 1754 } 1755 1756 // _WaitSetLock protects the wait queue, not the EntryList. We could 1757 // move the add-to-EntryList operation, above, outside the critical section 1758 // protected by _WaitSetLock. In practice that's not useful. With the 1759 // exception of wait() timeouts and interrupts the monitor owner 1760 // is the only thread that grabs _WaitSetLock. There's almost no contention 1761 // on _WaitSetLock so it's not profitable to reduce the length of the 1762 // critical section. 1763 } 1764 1765 Thread::SpinRelease (&_WaitSetLock) ; 1766 1767 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1768 ObjectMonitor::_sync_Notifications->inc() ; 1769 } 1770 } 1771 1772 1773 void ObjectMonitor::notifyAll(TRAPS) { 1774 CHECK_OWNER(); 1775 ObjectWaiter* iterator; 1776 if (_WaitSet == NULL) { 1777 TEVENT (Empty-NotifyAll) ; 1778 return ; 1779 } 1780 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1781 1782 int Policy = Knob_MoveNotifyee ; 1783 int Tally = 0 ; 1784 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1785 1786 for (;;) { 1787 iterator = DequeueWaiter () ; 1788 if (iterator == NULL) break ; 1789 TEVENT (NotifyAll - Transfer1) ; 1790 ++Tally ; 1791 1792 // Disposition - what might we do with iterator ? 1793 // a. add it directly to the EntryList - either tail or head. 1794 // b. push it onto the front of the _cxq. 1795 // For now we use (a). 1796 1797 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1798 guarantee (iterator->_notified == 0, "invariant") ; 1799 iterator->_notified = 1 ; 1800 Thread * Self = THREAD; 1801 iterator->_notifier_tid = Self->osthread()->thread_id(); 1802 if (Policy != 4) { 1803 iterator->TState = ObjectWaiter::TS_ENTER ; 1804 } 1805 1806 ObjectWaiter * List = _EntryList ; 1807 if (List != NULL) { 1808 assert (List->_prev == NULL, "invariant") ; 1809 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1810 assert (List != iterator, "invariant") ; 1811 } 1812 1813 if (Policy == 0) { // prepend to EntryList 1814 if (List == NULL) { 1815 iterator->_next = iterator->_prev = NULL ; 1816 _EntryList = iterator ; 1817 } else { 1818 List->_prev = iterator ; 1819 iterator->_next = List ; 1820 iterator->_prev = NULL ; 1821 _EntryList = iterator ; 1822 } 1823 } else 1824 if (Policy == 1) { // append to EntryList 1825 if (List == NULL) { 1826 iterator->_next = iterator->_prev = NULL ; 1827 _EntryList = iterator ; 1828 } else { 1829 // CONSIDER: finding the tail currently requires a linear-time walk of 1830 // the EntryList. We can make tail access constant-time by converting to 1831 // a CDLL instead of using our current DLL. 1832 ObjectWaiter * Tail ; 1833 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1834 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1835 Tail->_next = iterator ; 1836 iterator->_prev = Tail ; 1837 iterator->_next = NULL ; 1838 } 1839 } else 1840 if (Policy == 2) { // prepend to cxq 1841 // prepend to cxq 1842 iterator->TState = ObjectWaiter::TS_CXQ ; 1843 for (;;) { 1844 ObjectWaiter * Front = _cxq ; 1845 iterator->_next = Front ; 1846 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1847 break ; 1848 } 1849 } 1850 } else 1851 if (Policy == 3) { // append to cxq 1852 iterator->TState = ObjectWaiter::TS_CXQ ; 1853 for (;;) { 1854 ObjectWaiter * Tail ; 1855 Tail = _cxq ; 1856 if (Tail == NULL) { 1857 iterator->_next = NULL ; 1858 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1859 break ; 1860 } 1861 } else { 1862 while (Tail->_next != NULL) Tail = Tail->_next ; 1863 Tail->_next = iterator ; 1864 iterator->_prev = Tail ; 1865 iterator->_next = NULL ; 1866 break ; 1867 } 1868 } 1869 } else { 1870 ParkEvent * ev = iterator->_event ; 1871 iterator->TState = ObjectWaiter::TS_RUN ; 1872 OrderAccess::fence() ; 1873 ev->unpark() ; 1874 } 1875 1876 if (Policy < 4) { 1877 iterator->wait_reenter_begin(this); 1878 } 1879 1880 // _WaitSetLock protects the wait queue, not the EntryList. We could 1881 // move the add-to-EntryList operation, above, outside the critical section 1882 // protected by _WaitSetLock. In practice that's not useful. With the 1883 // exception of wait() timeouts and interrupts the monitor owner 1884 // is the only thread that grabs _WaitSetLock. There's almost no contention 1885 // on _WaitSetLock so it's not profitable to reduce the length of the 1886 // critical section. 1887 } 1888 1889 Thread::SpinRelease (&_WaitSetLock) ; 1890 1891 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1892 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1893 } 1894 } 1895 1896 // ----------------------------------------------------------------------------- 1897 // Adaptive Spinning Support 1898 // 1899 // Adaptive spin-then-block - rational spinning 1900 // 1901 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1902 // algorithm. On high order SMP systems it would be better to start with 1903 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1904 // a contending thread could enqueue itself on the cxq and then spin locally 1905 // on a thread-specific variable such as its ParkEvent._Event flag. 1906 // That's left as an exercise for the reader. Note that global spinning is 1907 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1908 // low latency and massive bandwidth. 1909 // 1910 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1911 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1912 // (duration) or we can fix the count at approximately the duration of 1913 // a context switch and vary the frequency. Of course we could also 1914 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1915 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1916 // 1917 // This implementation varies the duration "D", where D varies with 1918 // the success rate of recent spin attempts. (D is capped at approximately 1919 // length of a round-trip context switch). The success rate for recent 1920 // spin attempts is a good predictor of the success rate of future spin 1921 // attempts. The mechanism adapts automatically to varying critical 1922 // section length (lock modality), system load and degree of parallelism. 1923 // D is maintained per-monitor in _SpinDuration and is initialized 1924 // optimistically. Spin frequency is fixed at 100%. 1925 // 1926 // Note that _SpinDuration is volatile, but we update it without locks 1927 // or atomics. The code is designed so that _SpinDuration stays within 1928 // a reasonable range even in the presence of races. The arithmetic 1929 // operations on _SpinDuration are closed over the domain of legal values, 1930 // so at worst a race will install and older but still legal value. 1931 // At the very worst this introduces some apparent non-determinism. 1932 // We might spin when we shouldn't or vice-versa, but since the spin 1933 // count are relatively short, even in the worst case, the effect is harmless. 1934 // 1935 // Care must be taken that a low "D" value does not become an 1936 // an absorbing state. Transient spinning failures -- when spinning 1937 // is overall profitable -- should not cause the system to converge 1938 // on low "D" values. We want spinning to be stable and predictable 1939 // and fairly responsive to change and at the same time we don't want 1940 // it to oscillate, become metastable, be "too" non-deterministic, 1941 // or converge on or enter undesirable stable absorbing states. 1942 // 1943 // We implement a feedback-based control system -- using past behavior 1944 // to predict future behavior. We face two issues: (a) if the 1945 // input signal is random then the spin predictor won't provide optimal 1946 // results, and (b) if the signal frequency is too high then the control 1947 // system, which has some natural response lag, will "chase" the signal. 1948 // (b) can arise from multimodal lock hold times. Transient preemption 1949 // can also result in apparent bimodal lock hold times. 1950 // Although sub-optimal, neither condition is particularly harmful, as 1951 // in the worst-case we'll spin when we shouldn't or vice-versa. 1952 // The maximum spin duration is rather short so the failure modes aren't bad. 1953 // To be conservative, I've tuned the gain in system to bias toward 1954 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1955 // "rings" or oscillates between spinning and not spinning. This happens 1956 // when spinning is just on the cusp of profitability, however, so the 1957 // situation is not dire. The state is benign -- there's no need to add 1958 // hysteresis control to damp the transition rate between spinning and 1959 // not spinning. 1960 // 1961 1962 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1963 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1964 1965 // Spinning: Fixed frequency (100%), vary duration 1966 1967 1968 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1969 1970 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1971 int ctr = Knob_FixedSpin ; 1972 if (ctr != 0) { 1973 while (--ctr >= 0) { 1974 if (TryLock (Self) > 0) return 1 ; 1975 SpinPause () ; 1976 } 1977 return 0 ; 1978 } 1979 1980 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 1981 if (TryLock(Self) > 0) { 1982 // Increase _SpinDuration ... 1983 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1984 // Raising _SpurDuration to the poverty line is key. 1985 int x = _SpinDuration ; 1986 if (x < Knob_SpinLimit) { 1987 if (x < Knob_Poverty) x = Knob_Poverty ; 1988 _SpinDuration = x + Knob_BonusB ; 1989 } 1990 return 1 ; 1991 } 1992 SpinPause () ; 1993 } 1994 1995 // Admission control - verify preconditions for spinning 1996 // 1997 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1998 // becoming an absorbing state. Put another way, we spin briefly to 1999 // sample, just in case the system load, parallelism, contention, or lock 2000 // modality changed. 2001 // 2002 // Consider the following alternative: 2003 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2004 // spin attempt. "Periodically" might mean after a tally of 2005 // the # of failed spin attempts (or iterations) reaches some threshold. 2006 // This takes us into the realm of 1-out-of-N spinning, where we 2007 // hold the duration constant but vary the frequency. 2008 2009 ctr = _SpinDuration ; 2010 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2011 if (ctr <= 0) return 0 ; 2012 2013 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2014 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2015 TEVENT (Spin abort - notrunnable [TOP]); 2016 return 0 ; 2017 } 2018 2019 int MaxSpin = Knob_MaxSpinners ; 2020 if (MaxSpin >= 0) { 2021 if (_Spinner > MaxSpin) { 2022 TEVENT (Spin abort -- too many spinners) ; 2023 return 0 ; 2024 } 2025 // Slightly racy, but benign ... 2026 Adjust (&_Spinner, 1) ; 2027 } 2028 2029 // We're good to spin ... spin ingress. 2030 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2031 // when preparing to LD...CAS _owner, etc and the CAS is likely 2032 // to succeed. 2033 int hits = 0 ; 2034 int msk = 0 ; 2035 int caspty = Knob_CASPenalty ; 2036 int oxpty = Knob_OXPenalty ; 2037 int sss = Knob_SpinSetSucc ; 2038 if (sss && _succ == NULL ) _succ = Self ; 2039 Thread * prv = NULL ; 2040 2041 // There are three ways to exit the following loop: 2042 // 1. A successful spin where this thread has acquired the lock. 2043 // 2. Spin failure with prejudice 2044 // 3. Spin failure without prejudice 2045 2046 while (--ctr >= 0) { 2047 2048 // Periodic polling -- Check for pending GC 2049 // Threads may spin while they're unsafe. 2050 // We don't want spinning threads to delay the JVM from reaching 2051 // a stop-the-world safepoint or to steal cycles from GC. 2052 // If we detect a pending safepoint we abort in order that 2053 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2054 // this thread, if safe, doesn't steal cycles from GC. 2055 // This is in keeping with the "no loitering in runtime" rule. 2056 // We periodically check to see if there's a safepoint pending. 2057 if ((ctr & 0xFF) == 0) { 2058 if (SafepointSynchronize::do_call_back()) { 2059 TEVENT (Spin: safepoint) ; 2060 goto Abort ; // abrupt spin egress 2061 } 2062 if (Knob_UsePause & 1) SpinPause () ; 2063 2064 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2065 if (hits > 50 && scb != NULL) { 2066 int abend = (*scb)(SpinCallbackArgument, 0) ; 2067 } 2068 } 2069 2070 if (Knob_UsePause & 2) SpinPause() ; 2071 2072 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2073 // This is useful on classic SMP systems, but is of less utility on 2074 // N1-style CMT platforms. 2075 // 2076 // Trade-off: lock acquisition latency vs coherency bandwidth. 2077 // Lock hold times are typically short. A histogram 2078 // of successful spin attempts shows that we usually acquire 2079 // the lock early in the spin. That suggests we want to 2080 // sample _owner frequently in the early phase of the spin, 2081 // but then back-off and sample less frequently as the spin 2082 // progresses. The back-off makes a good citizen on SMP big 2083 // SMP systems. Oversampling _owner can consume excessive 2084 // coherency bandwidth. Relatedly, if we _oversample _owner we 2085 // can inadvertently interfere with the the ST m->owner=null. 2086 // executed by the lock owner. 2087 if (ctr & msk) continue ; 2088 ++hits ; 2089 if ((hits & 0xF) == 0) { 2090 // The 0xF, above, corresponds to the exponent. 2091 // Consider: (msk+1)|msk 2092 msk = ((msk << 2)|3) & BackOffMask ; 2093 } 2094 2095 // Probe _owner with TATAS 2096 // If this thread observes the monitor transition or flicker 2097 // from locked to unlocked to locked, then the odds that this 2098 // thread will acquire the lock in this spin attempt go down 2099 // considerably. The same argument applies if the CAS fails 2100 // or if we observe _owner change from one non-null value to 2101 // another non-null value. In such cases we might abort 2102 // the spin without prejudice or apply a "penalty" to the 2103 // spin count-down variable "ctr", reducing it by 100, say. 2104 2105 Thread * ox = (Thread *) _owner ; 2106 if (ox == NULL) { 2107 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2108 if (ox == NULL) { 2109 // The CAS succeeded -- this thread acquired ownership 2110 // Take care of some bookkeeping to exit spin state. 2111 if (sss && _succ == Self) { 2112 _succ = NULL ; 2113 } 2114 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2115 2116 // Increase _SpinDuration : 2117 // The spin was successful (profitable) so we tend toward 2118 // longer spin attempts in the future. 2119 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2120 // If we acquired the lock early in the spin cycle it 2121 // makes sense to increase _SpinDuration proportionally. 2122 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2123 int x = _SpinDuration ; 2124 if (x < Knob_SpinLimit) { 2125 if (x < Knob_Poverty) x = Knob_Poverty ; 2126 _SpinDuration = x + Knob_Bonus ; 2127 } 2128 return 1 ; 2129 } 2130 2131 // The CAS failed ... we can take any of the following actions: 2132 // * penalize: ctr -= Knob_CASPenalty 2133 // * exit spin with prejudice -- goto Abort; 2134 // * exit spin without prejudice. 2135 // * Since CAS is high-latency, retry again immediately. 2136 prv = ox ; 2137 TEVENT (Spin: cas failed) ; 2138 if (caspty == -2) break ; 2139 if (caspty == -1) goto Abort ; 2140 ctr -= caspty ; 2141 continue ; 2142 } 2143 2144 // Did lock ownership change hands ? 2145 if (ox != prv && prv != NULL ) { 2146 TEVENT (spin: Owner changed) 2147 if (oxpty == -2) break ; 2148 if (oxpty == -1) goto Abort ; 2149 ctr -= oxpty ; 2150 } 2151 prv = ox ; 2152 2153 // Abort the spin if the owner is not executing. 2154 // The owner must be executing in order to drop the lock. 2155 // Spinning while the owner is OFFPROC is idiocy. 2156 // Consider: ctr -= RunnablePenalty ; 2157 if (Knob_OState && NotRunnable (Self, ox)) { 2158 TEVENT (Spin abort - notrunnable); 2159 goto Abort ; 2160 } 2161 if (sss && _succ == NULL ) _succ = Self ; 2162 } 2163 2164 // Spin failed with prejudice -- reduce _SpinDuration. 2165 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2166 // AIMD is globally stable. 2167 TEVENT (Spin failure) ; 2168 { 2169 int x = _SpinDuration ; 2170 if (x > 0) { 2171 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2172 // This is globally sample and tends to damp the response. 2173 x -= Knob_Penalty ; 2174 if (x < 0) x = 0 ; 2175 _SpinDuration = x ; 2176 } 2177 } 2178 2179 Abort: 2180 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2181 if (sss && _succ == Self) { 2182 _succ = NULL ; 2183 // Invariant: after setting succ=null a contending thread 2184 // must recheck-retry _owner before parking. This usually happens 2185 // in the normal usage of TrySpin(), but it's safest 2186 // to make TrySpin() as foolproof as possible. 2187 OrderAccess::fence() ; 2188 if (TryLock(Self) > 0) return 1 ; 2189 } 2190 return 0 ; 2191 } 2192 2193 // NotRunnable() -- informed spinning 2194 // 2195 // Don't bother spinning if the owner is not eligible to drop the lock. 2196 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2197 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2198 // The thread must be runnable in order to drop the lock in timely fashion. 2199 // If the _owner is not runnable then spinning will not likely be 2200 // successful (profitable). 2201 // 2202 // Beware -- the thread referenced by _owner could have died 2203 // so a simply fetch from _owner->_thread_state might trap. 2204 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2205 // Because of the lifecycle issues the schedctl and _thread_state values 2206 // observed by NotRunnable() might be garbage. NotRunnable must 2207 // tolerate this and consider the observed _thread_state value 2208 // as advisory. 2209 // 2210 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2211 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2212 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2213 // with the LSB of _owner. Another option would be to probablistically probe 2214 // the putative _owner->TypeTag value. 2215 // 2216 // Checking _thread_state isn't perfect. Even if the thread is 2217 // in_java it might be blocked on a page-fault or have been preempted 2218 // and sitting on a ready/dispatch queue. _thread state in conjunction 2219 // with schedctl.sc_state gives us a good picture of what the 2220 // thread is doing, however. 2221 // 2222 // TODO: check schedctl.sc_state. 2223 // We'll need to use SafeFetch32() to read from the schedctl block. 2224 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2225 // 2226 // The return value from NotRunnable() is *advisory* -- the 2227 // result is based on sampling and is not necessarily coherent. 2228 // The caller must tolerate false-negative and false-positive errors. 2229 // Spinning, in general, is probabilistic anyway. 2230 2231 2232 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2233 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2234 if (!OwnerIsThread) return 0 ; 2235 2236 if (ox == NULL) return 0 ; 2237 2238 // Avoid transitive spinning ... 2239 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2240 // Immediately after T1 acquires L it's possible that T2, also 2241 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2242 // This occurs transiently after T1 acquired L but before 2243 // T1 managed to clear T1.Stalled. T2 does not need to abort 2244 // its spin in this circumstance. 2245 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2246 2247 if (BlockedOn == 1) return 1 ; 2248 if (BlockedOn != 0) { 2249 return BlockedOn != intptr_t(this) && _owner == ox ; 2250 } 2251 2252 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2253 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2254 // consider also: jst != _thread_in_Java -- but that's overspecific. 2255 return jst == _thread_blocked || jst == _thread_in_native ; 2256 } 2257 2258 2259 // ----------------------------------------------------------------------------- 2260 // WaitSet management ... 2261 2262 ObjectWaiter::ObjectWaiter(Thread* thread) { 2263 _next = NULL; 2264 _prev = NULL; 2265 _notified = 0; 2266 TState = TS_RUN ; 2267 _thread = thread; 2268 _event = thread->_ParkEvent ; 2269 _active = false; 2270 assert (_event != NULL, "invariant") ; 2271 } 2272 2273 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2274 JavaThread *jt = (JavaThread *)this->_thread; 2275 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2276 } 2277 2278 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2279 JavaThread *jt = (JavaThread *)this->_thread; 2280 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2281 } 2282 2283 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2284 assert(node != NULL, "should not dequeue NULL node"); 2285 assert(node->_prev == NULL, "node already in list"); 2286 assert(node->_next == NULL, "node already in list"); 2287 // put node at end of queue (circular doubly linked list) 2288 if (_WaitSet == NULL) { 2289 _WaitSet = node; 2290 node->_prev = node; 2291 node->_next = node; 2292 } else { 2293 ObjectWaiter* head = _WaitSet ; 2294 ObjectWaiter* tail = head->_prev; 2295 assert(tail->_next == head, "invariant check"); 2296 tail->_next = node; 2297 head->_prev = node; 2298 node->_next = head; 2299 node->_prev = tail; 2300 } 2301 } 2302 2303 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2304 // dequeue the very first waiter 2305 ObjectWaiter* waiter = _WaitSet; 2306 if (waiter) { 2307 DequeueSpecificWaiter(waiter); 2308 } 2309 return waiter; 2310 } 2311 2312 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2313 assert(node != NULL, "should not dequeue NULL node"); 2314 assert(node->_prev != NULL, "node already removed from list"); 2315 assert(node->_next != NULL, "node already removed from list"); 2316 // when the waiter has woken up because of interrupt, 2317 // timeout or other spurious wake-up, dequeue the 2318 // waiter from waiting list 2319 ObjectWaiter* next = node->_next; 2320 if (next == node) { 2321 assert(node->_prev == node, "invariant check"); 2322 _WaitSet = NULL; 2323 } else { 2324 ObjectWaiter* prev = node->_prev; 2325 assert(prev->_next == node, "invariant check"); 2326 assert(next->_prev == node, "invariant check"); 2327 next->_prev = prev; 2328 prev->_next = next; 2329 if (_WaitSet == node) { 2330 _WaitSet = next; 2331 } 2332 } 2333 node->_next = NULL; 2334 node->_prev = NULL; 2335 } 2336 2337 // ----------------------------------------------------------------------------- 2338 // PerfData support 2339 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2340 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2341 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2342 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2343 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2344 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2345 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2346 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2347 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2348 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2349 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2350 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2351 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2352 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2353 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2354 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2355 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2356 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2357 2358 // One-shot global initialization for the sync subsystem. 2359 // We could also defer initialization and initialize on-demand 2360 // the first time we call inflate(). Initialization would 2361 // be protected - like so many things - by the MonitorCache_lock. 2362 2363 void ObjectMonitor::Initialize () { 2364 static int InitializationCompleted = 0 ; 2365 assert (InitializationCompleted == 0, "invariant") ; 2366 InitializationCompleted = 1 ; 2367 if (UsePerfData) { 2368 EXCEPTION_MARK ; 2369 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2370 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2371 NEWPERFCOUNTER(_sync_Inflations) ; 2372 NEWPERFCOUNTER(_sync_Deflations) ; 2373 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2374 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2375 NEWPERFCOUNTER(_sync_Parks) ; 2376 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2377 NEWPERFCOUNTER(_sync_Notifications) ; 2378 NEWPERFCOUNTER(_sync_SlowEnter) ; 2379 NEWPERFCOUNTER(_sync_SlowExit) ; 2380 NEWPERFCOUNTER(_sync_SlowNotify) ; 2381 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2382 NEWPERFCOUNTER(_sync_FailedSpins) ; 2383 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2384 NEWPERFCOUNTER(_sync_PrivateA) ; 2385 NEWPERFCOUNTER(_sync_PrivateB) ; 2386 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2387 NEWPERFCOUNTER(_sync_MonScavenged) ; 2388 NEWPERFVARIABLE(_sync_MonExtant) ; 2389 #undef NEWPERFCOUNTER 2390 } 2391 } 2392 2393 2394 // Compile-time asserts 2395 // When possible, it's better to catch errors deterministically at 2396 // compile-time than at runtime. The down-side to using compile-time 2397 // asserts is that error message -- often something about negative array 2398 // indices -- is opaque. 2399 2400 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2401 2402 void ObjectMonitor::ctAsserts() { 2403 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2404 } 2405 2406 2407 static char * kvGet (char * kvList, const char * Key) { 2408 if (kvList == NULL) return NULL ; 2409 size_t n = strlen (Key) ; 2410 char * Search ; 2411 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2412 if (strncmp (Search, Key, n) == 0) { 2413 if (Search[n] == '=') return Search + n + 1 ; 2414 if (Search[n] == 0) return (char *) "1" ; 2415 } 2416 } 2417 return NULL ; 2418 } 2419 2420 static int kvGetInt (char * kvList, const char * Key, int Default) { 2421 char * v = kvGet (kvList, Key) ; 2422 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2423 if (Knob_ReportSettings && v != NULL) { 2424 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2425 ::fflush (stdout) ; 2426 } 2427 return rslt ; 2428 } 2429 2430 void ObjectMonitor::DeferredInitialize () { 2431 if (InitDone > 0) return ; 2432 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2433 while (InitDone != 1) ; 2434 return ; 2435 } 2436 2437 // One-shot global initialization ... 2438 // The initialization is idempotent, so we don't need locks. 2439 // In the future consider doing this via os::init_2(). 2440 // SyncKnobs consist of <Key>=<Value> pairs in the style 2441 // of environment variables. Start by converting ':' to NUL. 2442 2443 if (SyncKnobs == NULL) SyncKnobs = "" ; 2444 2445 size_t sz = strlen (SyncKnobs) ; 2446 char * knobs = (char *) malloc (sz + 2) ; 2447 if (knobs == NULL) { 2448 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; 2449 guarantee (0, "invariant") ; 2450 } 2451 strcpy (knobs, SyncKnobs) ; 2452 knobs[sz+1] = 0 ; 2453 for (char * p = knobs ; *p ; p++) { 2454 if (*p == ':') *p = 0 ; 2455 } 2456 2457 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2458 SETKNOB(ReportSettings) ; 2459 SETKNOB(Verbose) ; 2460 SETKNOB(FixedSpin) ; 2461 SETKNOB(SpinLimit) ; 2462 SETKNOB(SpinBase) ; 2463 SETKNOB(SpinBackOff); 2464 SETKNOB(CASPenalty) ; 2465 SETKNOB(OXPenalty) ; 2466 SETKNOB(LogSpins) ; 2467 SETKNOB(SpinSetSucc) ; 2468 SETKNOB(SuccEnabled) ; 2469 SETKNOB(SuccRestrict) ; 2470 SETKNOB(Penalty) ; 2471 SETKNOB(Bonus) ; 2472 SETKNOB(BonusB) ; 2473 SETKNOB(Poverty) ; 2474 SETKNOB(SpinAfterFutile) ; 2475 SETKNOB(UsePause) ; 2476 SETKNOB(SpinEarly) ; 2477 SETKNOB(OState) ; 2478 SETKNOB(MaxSpinners) ; 2479 SETKNOB(PreSpin) ; 2480 SETKNOB(ExitPolicy) ; 2481 SETKNOB(QMode); 2482 SETKNOB(ResetEvent) ; 2483 SETKNOB(MoveNotifyee) ; 2484 SETKNOB(FastHSSEC) ; 2485 #undef SETKNOB 2486 2487 if (os::is_MP()) { 2488 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2489 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2490 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2491 } else { 2492 Knob_SpinLimit = 0 ; 2493 Knob_SpinBase = 0 ; 2494 Knob_PreSpin = 0 ; 2495 Knob_FixedSpin = -1 ; 2496 } 2497 2498 if (Knob_LogSpins == 0) { 2499 ObjectMonitor::_sync_FailedSpins = NULL ; 2500 } 2501 2502 free (knobs) ; 2503 OrderAccess::fence() ; 2504 InitDone = 1 ; 2505 } 2506 2507 #ifndef PRODUCT 2508 void ObjectMonitor::verify() { 2509 } 2510 2511 void ObjectMonitor::print() { 2512 } 2513 #endif