1 /* 2 * Copyright (c) 1998, 2011, 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/osThread.hpp" 36 #include "runtime/stubRoutines.hpp" 37 #include "runtime/thread.hpp" 38 #include "services/threadService.hpp" 39 #include "utilities/dtrace.hpp" 40 #include "utilities/preserveException.hpp" 41 #ifdef TARGET_OS_FAMILY_linux 42 # include "os_linux.inline.hpp" 43 # include "thread_linux.inline.hpp" 44 #endif 45 #ifdef TARGET_OS_FAMILY_solaris 46 # include "os_solaris.inline.hpp" 47 # include "thread_solaris.inline.hpp" 48 #endif 49 #ifdef TARGET_OS_FAMILY_windows 50 # include "os_windows.inline.hpp" 51 # include "thread_windows.inline.hpp" 52 #endif 53 #ifdef TARGET_OS_FAMILY_bsd 54 # include "os_bsd.inline.hpp" 55 # include "thread_bsd.inline.hpp" 56 #endif 57 58 #if defined(__GNUC__) && !defined(IA64) 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 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 72 jlong, uintptr_t, char*, int); 73 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 74 jlong, uintptr_t, char*, int); 75 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 76 jlong, uintptr_t, char*, int); 77 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 78 jlong, uintptr_t, char*, int); 79 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 80 jlong, uintptr_t, char*, int); 81 82 #define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \ 83 char* bytes = NULL; \ 84 int len = 0; \ 85 jlong jtid = SharedRuntime::get_java_tid(thread); \ 86 Symbol* klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \ 87 if (klassname != NULL) { \ 88 bytes = (char*)klassname->bytes(); \ 89 len = klassname->utf8_length(); \ 90 } 91 92 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ 93 { \ 94 if (DTraceMonitorProbes) { \ 95 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 96 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 97 (monitor), bytes, len, (millis)); \ 98 } \ 99 } 100 101 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ 102 { \ 103 if (DTraceMonitorProbes) { \ 104 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 105 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 106 (uintptr_t)(monitor), bytes, len); \ 107 } \ 108 } 109 110 #else // ndef DTRACE_ENABLED 111 112 #define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;} 113 #define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;} 114 115 #endif // ndef DTRACE_ENABLED 116 117 // Tunables ... 118 // The knob* variables are effectively final. Once set they should 119 // never be modified hence. Consider using __read_mostly with GCC. 120 121 int ObjectMonitor::Knob_Verbose = 0 ; 122 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 123 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 124 static int Knob_HandOff = 0 ; 125 static int Knob_ReportSettings = 0 ; 126 127 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 128 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 129 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 130 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 131 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 132 static int Knob_SpinEarly = 1 ; 133 static int Knob_SuccEnabled = 1 ; // futile wake throttling 134 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 135 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 136 static int Knob_Bonus = 100 ; // spin success bonus 137 static int Knob_BonusB = 100 ; // spin success bonus 138 static int Knob_Penalty = 200 ; // spin failure penalty 139 static int Knob_Poverty = 1000 ; 140 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 141 static int Knob_FixedSpin = 0 ; 142 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 143 static int Knob_UsePause = 1 ; 144 static int Knob_ExitPolicy = 0 ; 145 static int Knob_PreSpin = 10 ; // 20-100 likely better 146 static int Knob_ResetEvent = 0 ; 147 static int BackOffMask = 0 ; 148 149 static int Knob_FastHSSEC = 0 ; 150 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 151 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 152 static volatile int InitDone = 0 ; 153 154 #define TrySpin TrySpin_VaryDuration 155 156 // ----------------------------------------------------------------------------- 157 // Theory of operations -- Monitors lists, thread residency, etc: 158 // 159 // * A thread acquires ownership of a monitor by successfully 160 // CAS()ing the _owner field from null to non-null. 161 // 162 // * Invariant: A thread appears on at most one monitor list -- 163 // cxq, EntryList or WaitSet -- at any one time. 164 // 165 // * Contending threads "push" themselves onto the cxq with CAS 166 // and then spin/park. 167 // 168 // * After a contending thread eventually acquires the lock it must 169 // dequeue itself from either the EntryList or the cxq. 170 // 171 // * The exiting thread identifies and unparks an "heir presumptive" 172 // tentative successor thread on the EntryList. Critically, the 173 // exiting thread doesn't unlink the successor thread from the EntryList. 174 // After having been unparked, the wakee will recontend for ownership of 175 // the monitor. The successor (wakee) will either acquire the lock or 176 // re-park itself. 177 // 178 // Succession is provided for by a policy of competitive handoff. 179 // The exiting thread does _not_ grant or pass ownership to the 180 // successor thread. (This is also referred to as "handoff" succession"). 181 // Instead the exiting thread releases ownership and possibly wakes 182 // a successor, so the successor can (re)compete for ownership of the lock. 183 // If the EntryList is empty but the cxq is populated the exiting 184 // thread will drain the cxq into the EntryList. It does so by 185 // by detaching the cxq (installing null with CAS) and folding 186 // the threads from the cxq into the EntryList. The EntryList is 187 // doubly linked, while the cxq is singly linked because of the 188 // CAS-based "push" used to enqueue recently arrived threads (RATs). 189 // 190 // * Concurrency invariants: 191 // 192 // -- only the monitor owner may access or mutate the EntryList. 193 // The mutex property of the monitor itself protects the EntryList 194 // from concurrent interference. 195 // -- Only the monitor owner may detach the cxq. 196 // 197 // * The monitor entry list operations avoid locks, but strictly speaking 198 // they're not lock-free. Enter is lock-free, exit is not. 199 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 200 // 201 // * The cxq can have multiple concurrent "pushers" but only one concurrent 202 // detaching thread. This mechanism is immune from the ABA corruption. 203 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 204 // 205 // * Taken together, the cxq and the EntryList constitute or form a 206 // single logical queue of threads stalled trying to acquire the lock. 207 // We use two distinct lists to improve the odds of a constant-time 208 // dequeue operation after acquisition (in the ::enter() epilog) and 209 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 210 // A key desideratum is to minimize queue & monitor metadata manipulation 211 // that occurs while holding the monitor lock -- that is, we want to 212 // minimize monitor lock holds times. Note that even a small amount of 213 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 214 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 215 // locks and monitor metadata. 216 // 217 // Cxq points to the the set of Recently Arrived Threads attempting entry. 218 // Because we push threads onto _cxq with CAS, the RATs must take the form of 219 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 220 // the unlocking thread notices that EntryList is null but _cxq is != null. 221 // 222 // The EntryList is ordered by the prevailing queue discipline and 223 // can be organized in any convenient fashion, such as a doubly-linked list or 224 // a circular doubly-linked list. Critically, we want insert and delete operations 225 // to operate in constant-time. If we need a priority queue then something akin 226 // to Solaris' sleepq would work nicely. Viz., 227 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 228 // Queue discipline is enforced at ::exit() time, when the unlocking thread 229 // drains the cxq into the EntryList, and orders or reorders the threads on the 230 // EntryList accordingly. 231 // 232 // Barring "lock barging", this mechanism provides fair cyclic ordering, 233 // somewhat similar to an elevator-scan. 234 // 235 // * The monitor synchronization subsystem avoids the use of native 236 // synchronization primitives except for the narrow platform-specific 237 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 238 // the semantics of park-unpark. Put another way, this monitor implementation 239 // depends only on atomic operations and park-unpark. The monitor subsystem 240 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 241 // underlying OS manages the READY<->RUN transitions. 242 // 243 // * Waiting threads reside on the WaitSet list -- wait() puts 244 // the caller onto the WaitSet. 245 // 246 // * notify() or notifyAll() simply transfers threads from the WaitSet to 247 // either the EntryList or cxq. Subsequent exit() operations will 248 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 249 // it's likely the notifyee would simply impale itself on the lock held 250 // by the notifier. 251 // 252 // * An interesting alternative is to encode cxq as (List,LockByte) where 253 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 254 // variable, like _recursions, in the scheme. The threads or Events that form 255 // the list would have to be aligned in 256-byte addresses. A thread would 256 // try to acquire the lock or enqueue itself with CAS, but exiting threads 257 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 258 // Note that is is *not* word-tearing, but it does presume that full-word 259 // CAS operations are coherent with intermix with STB operations. That's true 260 // on most common processors. 261 // 262 // * See also http://blogs.sun.com/dave 263 264 265 // ----------------------------------------------------------------------------- 266 // Enter support 267 268 bool ObjectMonitor::try_enter(Thread* THREAD) { 269 if (THREAD != _owner) { 270 if (THREAD->is_lock_owned ((address)_owner)) { 271 assert(_recursions == 0, "internal state error"); 272 _owner = THREAD ; 273 _recursions = 1 ; 274 OwnerIsThread = 1 ; 275 return true; 276 } 277 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 278 return false; 279 } 280 return true; 281 } else { 282 _recursions++; 283 return true; 284 } 285 } 286 287 void ATTR ObjectMonitor::enter(TRAPS) { 288 // The following code is ordered to check the most common cases first 289 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 290 Thread * const Self = THREAD ; 291 void * cur ; 292 293 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 294 if (cur == NULL) { 295 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 296 assert (_recursions == 0 , "invariant") ; 297 assert (_owner == Self, "invariant") ; 298 // CONSIDER: set or assert OwnerIsThread == 1 299 return ; 300 } 301 302 if (cur == Self) { 303 // TODO-FIXME: check for integer overflow! BUGID 6557169. 304 _recursions ++ ; 305 return ; 306 } 307 308 if (Self->is_lock_owned ((address)cur)) { 309 assert (_recursions == 0, "internal state error"); 310 _recursions = 1 ; 311 // Commute owner from a thread-specific on-stack BasicLockObject address to 312 // a full-fledged "Thread *". 313 _owner = Self ; 314 OwnerIsThread = 1 ; 315 return ; 316 } 317 318 // We've encountered genuine contention. 319 assert (Self->_Stalled == 0, "invariant") ; 320 Self->_Stalled = intptr_t(this) ; 321 322 // Try one round of spinning *before* enqueueing Self 323 // and before going through the awkward and expensive state 324 // transitions. The following spin is strictly optional ... 325 // Note that if we acquire the monitor from an initial spin 326 // we forgo posting JVMTI events and firing DTRACE probes. 327 if (Knob_SpinEarly && TrySpin (Self) > 0) { 328 assert (_owner == Self , "invariant") ; 329 assert (_recursions == 0 , "invariant") ; 330 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 331 Self->_Stalled = 0 ; 332 return ; 333 } 334 335 assert (_owner != Self , "invariant") ; 336 assert (_succ != Self , "invariant") ; 337 assert (Self->is_Java_thread() , "invariant") ; 338 JavaThread * jt = (JavaThread *) Self ; 339 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 340 assert (jt->thread_state() != _thread_blocked , "invariant") ; 341 assert (this->object() != NULL , "invariant") ; 342 assert (_count >= 0, "invariant") ; 343 344 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 345 // Ensure the object-monitor relationship remains stable while there's contention. 346 Atomic::inc_ptr(&_count); 347 348 { // Change java thread status to indicate blocked on monitor enter. 349 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 350 351 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 352 if (JvmtiExport::should_post_monitor_contended_enter()) { 353 JvmtiExport::post_monitor_contended_enter(jt, this); 354 } 355 356 OSThreadContendState osts(Self->osthread()); 357 ThreadBlockInVM tbivm(jt); 358 359 Self->set_current_pending_monitor(this); 360 361 // TODO-FIXME: change the following for(;;) loop to straight-line code. 362 for (;;) { 363 jt->set_suspend_equivalent(); 364 // cleared by handle_special_suspend_equivalent_condition() 365 // or java_suspend_self() 366 367 EnterI (THREAD) ; 368 369 if (!ExitSuspendEquivalent(jt)) break ; 370 371 // 372 // We have acquired the contended monitor, but while we were 373 // waiting another thread suspended us. We don't want to enter 374 // the monitor while suspended because that would surprise the 375 // thread that suspended us. 376 // 377 _recursions = 0 ; 378 _succ = NULL ; 379 exit (Self) ; 380 381 jt->java_suspend_self(); 382 } 383 Self->set_current_pending_monitor(NULL); 384 } 385 386 Atomic::dec_ptr(&_count); 387 assert (_count >= 0, "invariant") ; 388 Self->_Stalled = 0 ; 389 390 // Must either set _recursions = 0 or ASSERT _recursions == 0. 391 assert (_recursions == 0 , "invariant") ; 392 assert (_owner == Self , "invariant") ; 393 assert (_succ != Self , "invariant") ; 394 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 395 396 // The thread -- now the owner -- is back in vm mode. 397 // Report the glorious news via TI,DTrace and jvmstat. 398 // The probe effect is non-trivial. All the reportage occurs 399 // while we hold the monitor, increasing the length of the critical 400 // section. Amdahl's parallel speedup law comes vividly into play. 401 // 402 // Another option might be to aggregate the events (thread local or 403 // per-monitor aggregation) and defer reporting until a more opportune 404 // time -- such as next time some thread encounters contention but has 405 // yet to acquire the lock. While spinning that thread could 406 // spinning we could increment JVMStat counters, etc. 407 408 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 409 if (JvmtiExport::should_post_monitor_contended_entered()) { 410 JvmtiExport::post_monitor_contended_entered(jt, this); 411 } 412 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 413 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 414 } 415 } 416 417 418 // Caveat: TryLock() is not necessarily serializing if it returns failure. 419 // Callers must compensate as needed. 420 421 int ObjectMonitor::TryLock (Thread * Self) { 422 for (;;) { 423 void * own = _owner ; 424 if (own != NULL) return 0 ; 425 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 426 // Either guarantee _recursions == 0 or set _recursions = 0. 427 assert (_recursions == 0, "invariant") ; 428 assert (_owner == Self, "invariant") ; 429 // CONSIDER: set or assert that OwnerIsThread == 1 430 return 1 ; 431 } 432 // The lock had been free momentarily, but we lost the race to the lock. 433 // Interference -- the CAS failed. 434 // We can either return -1 or retry. 435 // Retry doesn't make as much sense because the lock was just acquired. 436 if (true) return -1 ; 437 } 438 } 439 440 void ATTR ObjectMonitor::EnterI (TRAPS) { 441 Thread * Self = THREAD ; 442 assert (Self->is_Java_thread(), "invariant") ; 443 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 444 445 // Try the lock - TATAS 446 if (TryLock (Self) > 0) { 447 assert (_succ != Self , "invariant") ; 448 assert (_owner == Self , "invariant") ; 449 assert (_Responsible != Self , "invariant") ; 450 return ; 451 } 452 453 DeferredInitialize () ; 454 455 // We try one round of spinning *before* enqueueing Self. 456 // 457 // If the _owner is ready but OFFPROC we could use a YieldTo() 458 // operation to donate the remainder of this thread's quantum 459 // to the owner. This has subtle but beneficial affinity 460 // effects. 461 462 if (TrySpin (Self) > 0) { 463 assert (_owner == Self , "invariant") ; 464 assert (_succ != Self , "invariant") ; 465 assert (_Responsible != Self , "invariant") ; 466 return ; 467 } 468 469 // The Spin failed -- Enqueue and park the thread ... 470 assert (_succ != Self , "invariant") ; 471 assert (_owner != Self , "invariant") ; 472 assert (_Responsible != Self , "invariant") ; 473 474 // Enqueue "Self" on ObjectMonitor's _cxq. 475 // 476 // Node acts as a proxy for Self. 477 // As an aside, if were to ever rewrite the synchronization code mostly 478 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 479 // Java objects. This would avoid awkward lifecycle and liveness issues, 480 // as well as eliminate a subset of ABA issues. 481 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 482 // 483 484 ObjectWaiter node(Self) ; 485 Self->_ParkEvent->reset() ; 486 node._prev = (ObjectWaiter *) 0xBAD ; 487 node.TState = ObjectWaiter::TS_CXQ ; 488 489 // Push "Self" onto the front of the _cxq. 490 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 491 // Note that spinning tends to reduce the rate at which threads 492 // enqueue and dequeue on EntryList|cxq. 493 ObjectWaiter * nxt ; 494 for (;;) { 495 node._next = nxt = _cxq ; 496 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 497 498 // Interference - the CAS failed because _cxq changed. Just retry. 499 // As an optional optimization we retry the lock. 500 if (TryLock (Self) > 0) { 501 assert (_succ != Self , "invariant") ; 502 assert (_owner == Self , "invariant") ; 503 assert (_Responsible != Self , "invariant") ; 504 return ; 505 } 506 } 507 508 // Check for cxq|EntryList edge transition to non-null. This indicates 509 // the onset of contention. While contention persists exiting threads 510 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 511 // operations revert to the faster 1-0 mode. This enter operation may interleave 512 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 513 // arrange for one of the contending thread to use a timed park() operations 514 // to detect and recover from the race. (Stranding is form of progress failure 515 // where the monitor is unlocked but all the contending threads remain parked). 516 // That is, at least one of the contended threads will periodically poll _owner. 517 // One of the contending threads will become the designated "Responsible" thread. 518 // The Responsible thread uses a timed park instead of a normal indefinite park 519 // operation -- it periodically wakes and checks for and recovers from potential 520 // strandings admitted by 1-0 exit operations. We need at most one Responsible 521 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 522 // be responsible for a monitor. 523 // 524 // Currently, one of the contended threads takes on the added role of "Responsible". 525 // A viable alternative would be to use a dedicated "stranding checker" thread 526 // that periodically iterated over all the threads (or active monitors) and unparked 527 // successors where there was risk of stranding. This would help eliminate the 528 // timer scalability issues we see on some platforms as we'd only have one thread 529 // -- the checker -- parked on a timer. 530 531 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 532 // Try to assume the role of responsible thread for the monitor. 533 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 534 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 535 } 536 537 // The lock have been released while this thread was occupied queueing 538 // itself onto _cxq. To close the race and avoid "stranding" and 539 // progress-liveness failure we must resample-retry _owner before parking. 540 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 541 // In this case the ST-MEMBAR is accomplished with CAS(). 542 // 543 // TODO: Defer all thread state transitions until park-time. 544 // Since state transitions are heavy and inefficient we'd like 545 // to defer the state transitions until absolutely necessary, 546 // and in doing so avoid some transitions ... 547 548 TEVENT (Inflated enter - Contention) ; 549 int nWakeups = 0 ; 550 int RecheckInterval = 1 ; 551 552 for (;;) { 553 554 if (TryLock (Self) > 0) break ; 555 assert (_owner != Self, "invariant") ; 556 557 if ((SyncFlags & 2) && _Responsible == NULL) { 558 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 559 } 560 561 // park self 562 if (_Responsible == Self || (SyncFlags & 1)) { 563 TEVENT (Inflated enter - park TIMED) ; 564 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 565 // Increase the RecheckInterval, but clamp the value. 566 RecheckInterval *= 8 ; 567 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 568 } else { 569 TEVENT (Inflated enter - park UNTIMED) ; 570 Self->_ParkEvent->park() ; 571 } 572 573 if (TryLock(Self) > 0) break ; 574 575 // The lock is still contested. 576 // Keep a tally of the # of futile wakeups. 577 // Note that the counter is not protected by a lock or updated by atomics. 578 // That is by design - we trade "lossy" counters which are exposed to 579 // races during updates for a lower probe effect. 580 TEVENT (Inflated enter - Futile wakeup) ; 581 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 582 ObjectMonitor::_sync_FutileWakeups->inc() ; 583 } 584 ++ nWakeups ; 585 586 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 587 // We can defer clearing _succ until after the spin completes 588 // TrySpin() must tolerate being called with _succ == Self. 589 // Try yet another round of adaptive spinning. 590 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 591 592 // We can find that we were unpark()ed and redesignated _succ while 593 // we were spinning. That's harmless. If we iterate and call park(), 594 // park() will consume the event and return immediately and we'll 595 // just spin again. This pattern can repeat, leaving _succ to simply 596 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 597 // Alternately, we can sample fired() here, and if set, forgo spinning 598 // in the next iteration. 599 600 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 601 Self->_ParkEvent->reset() ; 602 OrderAccess::fence() ; 603 } 604 if (_succ == Self) _succ = NULL ; 605 606 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 607 OrderAccess::fence() ; 608 } 609 610 // Egress : 611 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 612 // Normally we'll find Self on the EntryList . 613 // From the perspective of the lock owner (this thread), the 614 // EntryList is stable and cxq is prepend-only. 615 // The head of cxq is volatile but the interior is stable. 616 // In addition, Self.TState is stable. 617 618 assert (_owner == Self , "invariant") ; 619 assert (object() != NULL , "invariant") ; 620 // I'd like to write: 621 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 622 // but as we're at a safepoint that's not safe. 623 624 UnlinkAfterAcquire (Self, &node) ; 625 if (_succ == Self) _succ = NULL ; 626 627 assert (_succ != Self, "invariant") ; 628 if (_Responsible == Self) { 629 _Responsible = NULL ; 630 // Dekker pivot-point. 631 // Consider OrderAccess::storeload() here 632 633 // We may leave threads on cxq|EntryList without a designated 634 // "Responsible" thread. This is benign. When this thread subsequently 635 // exits the monitor it can "see" such preexisting "old" threads -- 636 // threads that arrived on the cxq|EntryList before the fence, above -- 637 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 638 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 639 // non-null and elect a new "Responsible" timer thread. 640 // 641 // This thread executes: 642 // ST Responsible=null; MEMBAR (in enter epilog - here) 643 // LD cxq|EntryList (in subsequent exit) 644 // 645 // Entering threads in the slow/contended path execute: 646 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 647 // The (ST cxq; MEMBAR) is accomplished with CAS(). 648 // 649 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 650 // exit operation from floating above the ST Responsible=null. 651 // 652 // In *practice* however, EnterI() is always followed by some atomic 653 // operation such as the decrement of _count in ::enter(). Those atomics 654 // obviate the need for the explicit MEMBAR, above. 655 } 656 657 // We've acquired ownership with CAS(). 658 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 659 // But since the CAS() this thread may have also stored into _succ, 660 // EntryList, cxq or Responsible. These meta-data updates must be 661 // visible __before this thread subsequently drops the lock. 662 // Consider what could occur if we didn't enforce this constraint -- 663 // STs to monitor meta-data and user-data could reorder with (become 664 // visible after) the ST in exit that drops ownership of the lock. 665 // Some other thread could then acquire the lock, but observe inconsistent 666 // or old monitor meta-data and heap data. That violates the JMM. 667 // To that end, the 1-0 exit() operation must have at least STST|LDST 668 // "release" barrier semantics. Specifically, there must be at least a 669 // STST|LDST barrier in exit() before the ST of null into _owner that drops 670 // the lock. The barrier ensures that changes to monitor meta-data and data 671 // protected by the lock will be visible before we release the lock, and 672 // therefore before some other thread (CPU) has a chance to acquire the lock. 673 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 674 // 675 // Critically, any prior STs to _succ or EntryList must be visible before 676 // the ST of null into _owner in the *subsequent* (following) corresponding 677 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 678 // execute a serializing instruction. 679 680 if (SyncFlags & 8) { 681 OrderAccess::fence() ; 682 } 683 return ; 684 } 685 686 // ReenterI() is a specialized inline form of the latter half of the 687 // contended slow-path from EnterI(). We use ReenterI() only for 688 // monitor reentry in wait(). 689 // 690 // In the future we should reconcile EnterI() and ReenterI(), adding 691 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 692 // loop accordingly. 693 694 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 695 assert (Self != NULL , "invariant") ; 696 assert (SelfNode != NULL , "invariant") ; 697 assert (SelfNode->_thread == Self , "invariant") ; 698 assert (_waiters > 0 , "invariant") ; 699 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 700 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 701 JavaThread * jt = (JavaThread *) Self ; 702 703 int nWakeups = 0 ; 704 for (;;) { 705 ObjectWaiter::TStates v = SelfNode->TState ; 706 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 707 assert (_owner != Self, "invariant") ; 708 709 if (TryLock (Self) > 0) break ; 710 if (TrySpin (Self) > 0) break ; 711 712 TEVENT (Wait Reentry - parking) ; 713 714 // State transition wrappers around park() ... 715 // ReenterI() wisely defers state transitions until 716 // it's clear we must park the thread. 717 { 718 OSThreadContendState osts(Self->osthread()); 719 ThreadBlockInVM tbivm(jt); 720 721 // cleared by handle_special_suspend_equivalent_condition() 722 // or java_suspend_self() 723 jt->set_suspend_equivalent(); 724 if (SyncFlags & 1) { 725 Self->_ParkEvent->park ((jlong)1000) ; 726 } else { 727 Self->_ParkEvent->park () ; 728 } 729 730 // were we externally suspended while we were waiting? 731 for (;;) { 732 if (!ExitSuspendEquivalent (jt)) break ; 733 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 734 jt->java_suspend_self(); 735 jt->set_suspend_equivalent(); 736 } 737 } 738 739 // Try again, but just so we distinguish between futile wakeups and 740 // successful wakeups. The following test isn't algorithmically 741 // necessary, but it helps us maintain sensible statistics. 742 if (TryLock(Self) > 0) break ; 743 744 // The lock is still contested. 745 // Keep a tally of the # of futile wakeups. 746 // Note that the counter is not protected by a lock or updated by atomics. 747 // That is by design - we trade "lossy" counters which are exposed to 748 // races during updates for a lower probe effect. 749 TEVENT (Wait Reentry - futile wakeup) ; 750 ++ nWakeups ; 751 752 // Assuming this is not a spurious wakeup we'll normally 753 // find that _succ == Self. 754 if (_succ == Self) _succ = NULL ; 755 756 // Invariant: after clearing _succ a contending thread 757 // *must* retry _owner before parking. 758 OrderAccess::fence() ; 759 760 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 761 ObjectMonitor::_sync_FutileWakeups->inc() ; 762 } 763 } 764 765 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 766 // Normally we'll find Self on the EntryList. 767 // Unlinking from the EntryList is constant-time and atomic-free. 768 // From the perspective of the lock owner (this thread), the 769 // EntryList is stable and cxq is prepend-only. 770 // The head of cxq is volatile but the interior is stable. 771 // In addition, Self.TState is stable. 772 773 assert (_owner == Self, "invariant") ; 774 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 775 UnlinkAfterAcquire (Self, SelfNode) ; 776 if (_succ == Self) _succ = NULL ; 777 assert (_succ != Self, "invariant") ; 778 SelfNode->TState = ObjectWaiter::TS_RUN ; 779 OrderAccess::fence() ; // see comments at the end of EnterI() 780 } 781 782 // after the thread acquires the lock in ::enter(). Equally, we could defer 783 // unlinking the thread until ::exit()-time. 784 785 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 786 { 787 assert (_owner == Self, "invariant") ; 788 assert (SelfNode->_thread == Self, "invariant") ; 789 790 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 791 // Normal case: remove Self from the DLL EntryList . 792 // This is a constant-time operation. 793 ObjectWaiter * nxt = SelfNode->_next ; 794 ObjectWaiter * prv = SelfNode->_prev ; 795 if (nxt != NULL) nxt->_prev = prv ; 796 if (prv != NULL) prv->_next = nxt ; 797 if (SelfNode == _EntryList ) _EntryList = nxt ; 798 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 799 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 800 TEVENT (Unlink from EntryList) ; 801 } else { 802 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 803 // Inopportune interleaving -- Self is still on the cxq. 804 // This usually means the enqueue of self raced an exiting thread. 805 // Normally we'll find Self near the front of the cxq, so 806 // dequeueing is typically fast. If needbe we can accelerate 807 // this with some MCS/CHL-like bidirectional list hints and advisory 808 // back-links so dequeueing from the interior will normally operate 809 // in constant-time. 810 // Dequeue Self from either the head (with CAS) or from the interior 811 // with a linear-time scan and normal non-atomic memory operations. 812 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 813 // and then unlink Self from EntryList. We have to drain eventually, 814 // so it might as well be now. 815 816 ObjectWaiter * v = _cxq ; 817 assert (v != NULL, "invariant") ; 818 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 819 // The CAS above can fail from interference IFF a "RAT" arrived. 820 // In that case Self must be in the interior and can no longer be 821 // at the head of cxq. 822 if (v == SelfNode) { 823 assert (_cxq != v, "invariant") ; 824 v = _cxq ; // CAS above failed - start scan at head of list 825 } 826 ObjectWaiter * p ; 827 ObjectWaiter * q = NULL ; 828 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 829 q = p ; 830 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 831 } 832 assert (v != SelfNode, "invariant") ; 833 assert (p == SelfNode, "Node not found on cxq") ; 834 assert (p != _cxq, "invariant") ; 835 assert (q != NULL, "invariant") ; 836 assert (q->_next == p, "invariant") ; 837 q->_next = p->_next ; 838 } 839 TEVENT (Unlink from cxq) ; 840 } 841 842 // Diagnostic hygiene ... 843 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 844 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 845 SelfNode->TState = ObjectWaiter::TS_RUN ; 846 } 847 848 // ----------------------------------------------------------------------------- 849 // Exit support 850 // 851 // exit() 852 // ~~~~~~ 853 // Note that the collector can't reclaim the objectMonitor or deflate 854 // the object out from underneath the thread calling ::exit() as the 855 // thread calling ::exit() never transitions to a stable state. 856 // This inhibits GC, which in turn inhibits asynchronous (and 857 // inopportune) reclamation of "this". 858 // 859 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 860 // There's one exception to the claim above, however. EnterI() can call 861 // exit() to drop a lock if the acquirer has been externally suspended. 862 // In that case exit() is called with _thread_state as _thread_blocked, 863 // but the monitor's _count field is > 0, which inhibits reclamation. 864 // 865 // 1-0 exit 866 // ~~~~~~~~ 867 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 868 // the fast-path operators have been optimized so the common ::exit() 869 // operation is 1-0. See i486.ad fast_unlock(), for instance. 870 // The code emitted by fast_unlock() elides the usual MEMBAR. This 871 // greatly improves latency -- MEMBAR and CAS having considerable local 872 // latency on modern processors -- but at the cost of "stranding". Absent the 873 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 874 // ::enter() path, resulting in the entering thread being stranding 875 // and a progress-liveness failure. Stranding is extremely rare. 876 // We use timers (timed park operations) & periodic polling to detect 877 // and recover from stranding. Potentially stranded threads periodically 878 // wake up and poll the lock. See the usage of the _Responsible variable. 879 // 880 // The CAS() in enter provides for safety and exclusion, while the CAS or 881 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 882 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 883 // We detect and recover from stranding with timers. 884 // 885 // If a thread transiently strands it'll park until (a) another 886 // thread acquires the lock and then drops the lock, at which time the 887 // exiting thread will notice and unpark the stranded thread, or, (b) 888 // the timer expires. If the lock is high traffic then the stranding latency 889 // will be low due to (a). If the lock is low traffic then the odds of 890 // stranding are lower, although the worst-case stranding latency 891 // is longer. Critically, we don't want to put excessive load in the 892 // platform's timer subsystem. We want to minimize both the timer injection 893 // rate (timers created/sec) as well as the number of timers active at 894 // any one time. (more precisely, we want to minimize timer-seconds, which is 895 // the integral of the # of active timers at any instant over time). 896 // Both impinge on OS scalability. Given that, at most one thread parked on 897 // a monitor will use a timer. 898 899 void ATTR ObjectMonitor::exit(TRAPS) { 900 Thread * Self = THREAD ; 901 if (THREAD != _owner) { 902 if (THREAD->is_lock_owned((address) _owner)) { 903 // Transmute _owner from a BasicLock pointer to a Thread address. 904 // We don't need to hold _mutex for this transition. 905 // Non-null to Non-null is safe as long as all readers can 906 // tolerate either flavor. 907 assert (_recursions == 0, "invariant") ; 908 _owner = THREAD ; 909 _recursions = 0 ; 910 OwnerIsThread = 1 ; 911 } else { 912 // NOTE: we need to handle unbalanced monitor enter/exit 913 // in native code by throwing an exception. 914 // TODO: Throw an IllegalMonitorStateException ? 915 TEVENT (Exit - Throw IMSX) ; 916 assert(false, "Non-balanced monitor enter/exit!"); 917 if (false) { 918 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 919 } 920 return; 921 } 922 } 923 924 if (_recursions != 0) { 925 _recursions--; // this is simple recursive enter 926 TEVENT (Inflated exit - recursive) ; 927 return ; 928 } 929 930 // Invariant: after setting Responsible=null an thread must execute 931 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 932 if ((SyncFlags & 4) == 0) { 933 _Responsible = NULL ; 934 } 935 936 for (;;) { 937 assert (THREAD == _owner, "invariant") ; 938 939 940 if (Knob_ExitPolicy == 0) { 941 // release semantics: prior loads and stores from within the critical section 942 // must not float (reorder) past the following store that drops the lock. 943 // On SPARC that requires MEMBAR #loadstore|#storestore. 944 // But of course in TSO #loadstore|#storestore is not required. 945 // I'd like to write one of the following: 946 // A. OrderAccess::release() ; _owner = NULL 947 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 948 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 949 // store into a _dummy variable. That store is not needed, but can result 950 // in massive wasteful coherency traffic on classic SMP systems. 951 // Instead, I use release_store(), which is implemented as just a simple 952 // ST on x64, x86 and SPARC. 953 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 954 OrderAccess::storeload() ; // See if we need to wake a successor 955 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 956 TEVENT (Inflated exit - simple egress) ; 957 return ; 958 } 959 TEVENT (Inflated exit - complex egress) ; 960 961 // Normally the exiting thread is responsible for ensuring succession, 962 // but if other successors are ready or other entering threads are spinning 963 // then this thread can simply store NULL into _owner and exit without 964 // waking a successor. The existence of spinners or ready successors 965 // guarantees proper succession (liveness). Responsibility passes to the 966 // ready or running successors. The exiting thread delegates the duty. 967 // More precisely, if a successor already exists this thread is absolved 968 // of the responsibility of waking (unparking) one. 969 // 970 // The _succ variable is critical to reducing futile wakeup frequency. 971 // _succ identifies the "heir presumptive" thread that has been made 972 // ready (unparked) but that has not yet run. We need only one such 973 // successor thread to guarantee progress. 974 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 975 // section 3.3 "Futile Wakeup Throttling" for details. 976 // 977 // Note that spinners in Enter() also set _succ non-null. 978 // In the current implementation spinners opportunistically set 979 // _succ so that exiting threads might avoid waking a successor. 980 // Another less appealing alternative would be for the exiting thread 981 // to drop the lock and then spin briefly to see if a spinner managed 982 // to acquire the lock. If so, the exiting thread could exit 983 // immediately without waking a successor, otherwise the exiting 984 // thread would need to dequeue and wake a successor. 985 // (Note that we'd need to make the post-drop spin short, but no 986 // shorter than the worst-case round-trip cache-line migration time. 987 // The dropped lock needs to become visible to the spinner, and then 988 // the acquisition of the lock by the spinner must become visible to 989 // the exiting thread). 990 // 991 992 // It appears that an heir-presumptive (successor) must be made ready. 993 // Only the current lock owner can manipulate the EntryList or 994 // drain _cxq, so we need to reacquire the lock. If we fail 995 // to reacquire the lock the responsibility for ensuring succession 996 // falls to the new owner. 997 // 998 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 999 return ; 1000 } 1001 TEVENT (Exit - Reacquired) ; 1002 } else { 1003 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1004 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1005 OrderAccess::storeload() ; 1006 // Ratify the previously observed values. 1007 if (_cxq == NULL || _succ != NULL) { 1008 TEVENT (Inflated exit - simple egress) ; 1009 return ; 1010 } 1011 1012 // inopportune interleaving -- the exiting thread (this thread) 1013 // in the fast-exit path raced an entering thread in the slow-enter 1014 // path. 1015 // We have two choices: 1016 // A. Try to reacquire the lock. 1017 // If the CAS() fails return immediately, otherwise 1018 // we either restart/rerun the exit operation, or simply 1019 // fall-through into the code below which wakes a successor. 1020 // B. If the elements forming the EntryList|cxq are TSM 1021 // we could simply unpark() the lead thread and return 1022 // without having set _succ. 1023 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1024 TEVENT (Inflated exit - reacquired succeeded) ; 1025 return ; 1026 } 1027 TEVENT (Inflated exit - reacquired failed) ; 1028 } else { 1029 TEVENT (Inflated exit - complex egress) ; 1030 } 1031 } 1032 1033 guarantee (_owner == THREAD, "invariant") ; 1034 1035 ObjectWaiter * w = NULL ; 1036 int QMode = Knob_QMode ; 1037 1038 if (QMode == 2 && _cxq != NULL) { 1039 // QMode == 2 : cxq has precedence over EntryList. 1040 // Try to directly wake a successor from the cxq. 1041 // If successful, the successor will need to unlink itself from cxq. 1042 w = _cxq ; 1043 assert (w != NULL, "invariant") ; 1044 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1045 ExitEpilog (Self, w) ; 1046 return ; 1047 } 1048 1049 if (QMode == 3 && _cxq != NULL) { 1050 // Aggressively drain cxq into EntryList at the first opportunity. 1051 // This policy ensure that recently-run threads live at the head of EntryList. 1052 // Drain _cxq into EntryList - bulk transfer. 1053 // First, detach _cxq. 1054 // The following loop is tantamount to: w = swap (&cxq, NULL) 1055 w = _cxq ; 1056 for (;;) { 1057 assert (w != NULL, "Invariant") ; 1058 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1059 if (u == w) break ; 1060 w = u ; 1061 } 1062 assert (w != NULL , "invariant") ; 1063 1064 ObjectWaiter * q = NULL ; 1065 ObjectWaiter * p ; 1066 for (p = w ; p != NULL ; p = p->_next) { 1067 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1068 p->TState = ObjectWaiter::TS_ENTER ; 1069 p->_prev = q ; 1070 q = p ; 1071 } 1072 1073 // Append the RATs to the EntryList 1074 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1075 ObjectWaiter * Tail ; 1076 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1077 if (Tail == NULL) { 1078 _EntryList = w ; 1079 } else { 1080 Tail->_next = w ; 1081 w->_prev = Tail ; 1082 } 1083 1084 // Fall thru into code that tries to wake a successor from EntryList 1085 } 1086 1087 if (QMode == 4 && _cxq != NULL) { 1088 // Aggressively drain cxq into EntryList at the first opportunity. 1089 // This policy ensure that recently-run threads live at the head of EntryList. 1090 1091 // Drain _cxq into EntryList - bulk transfer. 1092 // First, detach _cxq. 1093 // The following loop is tantamount to: w = swap (&cxq, NULL) 1094 w = _cxq ; 1095 for (;;) { 1096 assert (w != NULL, "Invariant") ; 1097 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1098 if (u == w) break ; 1099 w = u ; 1100 } 1101 assert (w != NULL , "invariant") ; 1102 1103 ObjectWaiter * q = NULL ; 1104 ObjectWaiter * p ; 1105 for (p = w ; p != NULL ; p = p->_next) { 1106 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1107 p->TState = ObjectWaiter::TS_ENTER ; 1108 p->_prev = q ; 1109 q = p ; 1110 } 1111 1112 // Prepend the RATs to the EntryList 1113 if (_EntryList != NULL) { 1114 q->_next = _EntryList ; 1115 _EntryList->_prev = q ; 1116 } 1117 _EntryList = w ; 1118 1119 // Fall thru into code that tries to wake a successor from EntryList 1120 } 1121 1122 w = _EntryList ; 1123 if (w != NULL) { 1124 // I'd like to write: guarantee (w->_thread != Self). 1125 // But in practice an exiting thread may find itself on the EntryList. 1126 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1127 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1128 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1129 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1130 // release the lock "O". T2 resumes immediately after the ST of null into 1131 // _owner, above. T2 notices that the EntryList is populated, so it 1132 // reacquires the lock and then finds itself on the EntryList. 1133 // Given all that, we have to tolerate the circumstance where "w" is 1134 // associated with Self. 1135 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1136 ExitEpilog (Self, w) ; 1137 return ; 1138 } 1139 1140 // If we find that both _cxq and EntryList are null then just 1141 // re-run the exit protocol from the top. 1142 w = _cxq ; 1143 if (w == NULL) continue ; 1144 1145 // Drain _cxq into EntryList - bulk transfer. 1146 // First, detach _cxq. 1147 // The following loop is tantamount to: w = swap (&cxq, NULL) 1148 for (;;) { 1149 assert (w != NULL, "Invariant") ; 1150 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1151 if (u == w) break ; 1152 w = u ; 1153 } 1154 TEVENT (Inflated exit - drain cxq into EntryList) ; 1155 1156 assert (w != NULL , "invariant") ; 1157 assert (_EntryList == NULL , "invariant") ; 1158 1159 // Convert the LIFO SLL anchored by _cxq into a DLL. 1160 // The list reorganization step operates in O(LENGTH(w)) time. 1161 // It's critical that this step operate quickly as 1162 // "Self" still holds the outer-lock, restricting parallelism 1163 // and effectively lengthening the critical section. 1164 // Invariant: s chases t chases u. 1165 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1166 // we have faster access to the tail. 1167 1168 if (QMode == 1) { 1169 // QMode == 1 : drain cxq to EntryList, reversing order 1170 // We also reverse the order of the list. 1171 ObjectWaiter * s = NULL ; 1172 ObjectWaiter * t = w ; 1173 ObjectWaiter * u = NULL ; 1174 while (t != NULL) { 1175 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1176 t->TState = ObjectWaiter::TS_ENTER ; 1177 u = t->_next ; 1178 t->_prev = u ; 1179 t->_next = s ; 1180 s = t; 1181 t = u ; 1182 } 1183 _EntryList = s ; 1184 assert (s != NULL, "invariant") ; 1185 } else { 1186 // QMode == 0 or QMode == 2 1187 _EntryList = w ; 1188 ObjectWaiter * q = NULL ; 1189 ObjectWaiter * p ; 1190 for (p = w ; p != NULL ; p = p->_next) { 1191 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1192 p->TState = ObjectWaiter::TS_ENTER ; 1193 p->_prev = q ; 1194 q = p ; 1195 } 1196 } 1197 1198 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1199 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1200 1201 // See if we can abdicate to a spinner instead of waking a thread. 1202 // A primary goal of the implementation is to reduce the 1203 // context-switch rate. 1204 if (_succ != NULL) continue; 1205 1206 w = _EntryList ; 1207 if (w != NULL) { 1208 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1209 ExitEpilog (Self, w) ; 1210 return ; 1211 } 1212 } 1213 } 1214 1215 // ExitSuspendEquivalent: 1216 // A faster alternate to handle_special_suspend_equivalent_condition() 1217 // 1218 // handle_special_suspend_equivalent_condition() unconditionally 1219 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1220 // operations have high latency. Note that in ::enter() we call HSSEC 1221 // while holding the monitor, so we effectively lengthen the critical sections. 1222 // 1223 // There are a number of possible solutions: 1224 // 1225 // A. To ameliorate the problem we might also defer state transitions 1226 // to as late as possible -- just prior to parking. 1227 // Given that, we'd call HSSEC after having returned from park(), 1228 // but before attempting to acquire the monitor. This is only a 1229 // partial solution. It avoids calling HSSEC while holding the 1230 // monitor (good), but it still increases successor reacquisition latency -- 1231 // the interval between unparking a successor and the time the successor 1232 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1233 // If we use this technique we can also avoid EnterI()-exit() loop 1234 // in ::enter() where we iteratively drop the lock and then attempt 1235 // to reacquire it after suspending. 1236 // 1237 // B. In the future we might fold all the suspend bits into a 1238 // composite per-thread suspend flag and then update it with CAS(). 1239 // Alternately, a Dekker-like mechanism with multiple variables 1240 // would suffice: 1241 // ST Self->_suspend_equivalent = false 1242 // MEMBAR 1243 // LD Self_>_suspend_flags 1244 // 1245 1246 1247 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1248 int Mode = Knob_FastHSSEC ; 1249 if (Mode && !jSelf->is_external_suspend()) { 1250 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1251 jSelf->clear_suspend_equivalent() ; 1252 if (2 == Mode) OrderAccess::storeload() ; 1253 if (!jSelf->is_external_suspend()) return false ; 1254 // We raced a suspension -- fall thru into the slow path 1255 TEVENT (ExitSuspendEquivalent - raced) ; 1256 jSelf->set_suspend_equivalent() ; 1257 } 1258 return jSelf->handle_special_suspend_equivalent_condition() ; 1259 } 1260 1261 1262 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1263 assert (_owner == Self, "invariant") ; 1264 1265 // Exit protocol: 1266 // 1. ST _succ = wakee 1267 // 2. membar #loadstore|#storestore; 1268 // 2. ST _owner = NULL 1269 // 3. unpark(wakee) 1270 1271 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1272 ParkEvent * Trigger = Wakee->_event ; 1273 1274 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1275 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1276 // out-of-scope (non-extant). 1277 Wakee = NULL ; 1278 1279 // Drop the lock 1280 OrderAccess::release_store_ptr (&_owner, NULL) ; 1281 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1282 1283 if (SafepointSynchronize::do_call_back()) { 1284 TEVENT (unpark before SAFEPOINT) ; 1285 } 1286 1287 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1288 Trigger->unpark() ; 1289 1290 // Maintain stats and report events to JVMTI 1291 if (ObjectMonitor::_sync_Parks != NULL) { 1292 ObjectMonitor::_sync_Parks->inc() ; 1293 } 1294 } 1295 1296 1297 // ----------------------------------------------------------------------------- 1298 // Class Loader deadlock handling. 1299 // 1300 // complete_exit exits a lock returning recursion count 1301 // complete_exit/reenter operate as a wait without waiting 1302 // complete_exit requires an inflated monitor 1303 // The _owner field is not always the Thread addr even with an 1304 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1305 // thread due to contention. 1306 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1307 Thread * const Self = THREAD; 1308 assert(Self->is_Java_thread(), "Must be Java thread!"); 1309 JavaThread *jt = (JavaThread *)THREAD; 1310 1311 DeferredInitialize(); 1312 1313 if (THREAD != _owner) { 1314 if (THREAD->is_lock_owned ((address)_owner)) { 1315 assert(_recursions == 0, "internal state error"); 1316 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1317 _recursions = 0 ; 1318 OwnerIsThread = 1 ; 1319 } 1320 } 1321 1322 guarantee(Self == _owner, "complete_exit not owner"); 1323 intptr_t save = _recursions; // record the old recursion count 1324 _recursions = 0; // set the recursion level to be 0 1325 exit (Self) ; // exit the monitor 1326 guarantee (_owner != Self, "invariant"); 1327 return save; 1328 } 1329 1330 // reenter() enters a lock and sets recursion count 1331 // complete_exit/reenter operate as a wait without waiting 1332 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1333 Thread * const Self = THREAD; 1334 assert(Self->is_Java_thread(), "Must be Java thread!"); 1335 JavaThread *jt = (JavaThread *)THREAD; 1336 1337 guarantee(_owner != Self, "reenter already owner"); 1338 enter (THREAD); // enter the monitor 1339 guarantee (_recursions == 0, "reenter recursion"); 1340 _recursions = recursions; 1341 return; 1342 } 1343 1344 1345 // ----------------------------------------------------------------------------- 1346 // A macro is used below because there may already be a pending 1347 // exception which should not abort the execution of the routines 1348 // which use this (which is why we don't put this into check_slow and 1349 // call it with a CHECK argument). 1350 1351 #define CHECK_OWNER() \ 1352 do { \ 1353 if (THREAD != _owner) { \ 1354 if (THREAD->is_lock_owned((address) _owner)) { \ 1355 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1356 _recursions = 0; \ 1357 OwnerIsThread = 1 ; \ 1358 } else { \ 1359 TEVENT (Throw IMSX) ; \ 1360 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1361 } \ 1362 } \ 1363 } while (false) 1364 1365 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1366 // TODO-FIXME: remove check_slow() -- it's likely dead. 1367 1368 void ObjectMonitor::check_slow(TRAPS) { 1369 TEVENT (check_slow - throw IMSX) ; 1370 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1371 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1372 } 1373 1374 static int Adjust (volatile int * adr, int dx) { 1375 int v ; 1376 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1377 return v ; 1378 } 1379 // ----------------------------------------------------------------------------- 1380 // Wait/Notify/NotifyAll 1381 // 1382 // Note: a subset of changes to ObjectMonitor::wait() 1383 // will need to be replicated in complete_exit above 1384 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1385 Thread * const Self = THREAD ; 1386 assert(Self->is_Java_thread(), "Must be Java thread!"); 1387 JavaThread *jt = (JavaThread *)THREAD; 1388 1389 DeferredInitialize () ; 1390 1391 // Throw IMSX or IEX. 1392 CHECK_OWNER(); 1393 1394 // check for a pending interrupt 1395 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1396 // post monitor waited event. Note that this is past-tense, we are done waiting. 1397 if (JvmtiExport::should_post_monitor_waited()) { 1398 // Note: 'false' parameter is passed here because the 1399 // wait was not timed out due to thread interrupt. 1400 JvmtiExport::post_monitor_waited(jt, this, false); 1401 } 1402 TEVENT (Wait - Throw IEX) ; 1403 THROW(vmSymbols::java_lang_InterruptedException()); 1404 return ; 1405 } 1406 TEVENT (Wait) ; 1407 1408 assert (Self->_Stalled == 0, "invariant") ; 1409 Self->_Stalled = intptr_t(this) ; 1410 jt->set_current_waiting_monitor(this); 1411 1412 // create a node to be put into the queue 1413 // Critically, after we reset() the event but prior to park(), we must check 1414 // for a pending interrupt. 1415 ObjectWaiter node(Self); 1416 node.TState = ObjectWaiter::TS_WAIT ; 1417 Self->_ParkEvent->reset() ; 1418 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1419 1420 // Enter the waiting queue, which is a circular doubly linked list in this case 1421 // but it could be a priority queue or any data structure. 1422 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1423 // by the the owner of the monitor *except* in the case where park() 1424 // returns because of a timeout of interrupt. Contention is exceptionally rare 1425 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1426 1427 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1428 AddWaiter (&node) ; 1429 Thread::SpinRelease (&_WaitSetLock) ; 1430 1431 if ((SyncFlags & 4) == 0) { 1432 _Responsible = NULL ; 1433 } 1434 intptr_t save = _recursions; // record the old recursion count 1435 _waiters++; // increment the number of waiters 1436 _recursions = 0; // set the recursion level to be 1 1437 exit (Self) ; // exit the monitor 1438 guarantee (_owner != Self, "invariant") ; 1439 1440 // As soon as the ObjectMonitor's ownership is dropped in the exit() 1441 // call above, another thread can enter() the ObjectMonitor, do the 1442 // notify(), and exit() the ObjectMonitor. If the other thread's 1443 // exit() call chooses this thread as the successor and the unpark() 1444 // call happens to occur while this thread is posting a 1445 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event 1446 // handler using RawMonitors and consuming the unpark(). 1447 // 1448 // To avoid the problem, we re-post the event. This does no harm 1449 // even if the original unpark() was not consumed because we are the 1450 // chosen successor for this monitor. 1451 if (node._notified != 0 && _succ == Self) { 1452 node._event->unpark(); 1453 } 1454 1455 // The thread is on the WaitSet list - now park() it. 1456 // On MP systems it's conceivable that a brief spin before we park 1457 // could be profitable. 1458 // 1459 // TODO-FIXME: change the following logic to a loop of the form 1460 // while (!timeout && !interrupted && _notified == 0) park() 1461 1462 int ret = OS_OK ; 1463 int WasNotified = 0 ; 1464 { // State transition wrappers 1465 OSThread* osthread = Self->osthread(); 1466 OSThreadWaitState osts(osthread, true); 1467 { 1468 ThreadBlockInVM tbivm(jt); 1469 // Thread is in thread_blocked state and oop access is unsafe. 1470 jt->set_suspend_equivalent(); 1471 1472 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1473 // Intentionally empty 1474 } else 1475 if (node._notified == 0) { 1476 if (millis <= 0) { 1477 Self->_ParkEvent->park () ; 1478 } else { 1479 ret = Self->_ParkEvent->park (millis) ; 1480 } 1481 } 1482 1483 // were we externally suspended while we were waiting? 1484 if (ExitSuspendEquivalent (jt)) { 1485 // TODO-FIXME: add -- if succ == Self then succ = null. 1486 jt->java_suspend_self(); 1487 } 1488 1489 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1490 1491 1492 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1493 // from the WaitSet to the EntryList. 1494 // See if we need to remove Node from the WaitSet. 1495 // We use double-checked locking to avoid grabbing _WaitSetLock 1496 // if the thread is not on the wait queue. 1497 // 1498 // Note that we don't need a fence before the fetch of TState. 1499 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1500 // written by the is thread. (perhaps the fetch might even be satisfied 1501 // by a look-aside into the processor's own store buffer, although given 1502 // the length of the code path between the prior ST and this load that's 1503 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1504 // then we'll acquire the lock and then re-fetch a fresh TState value. 1505 // That is, we fail toward safety. 1506 1507 if (node.TState == ObjectWaiter::TS_WAIT) { 1508 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1509 if (node.TState == ObjectWaiter::TS_WAIT) { 1510 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1511 assert(node._notified == 0, "invariant"); 1512 node.TState = ObjectWaiter::TS_RUN ; 1513 } 1514 Thread::SpinRelease (&_WaitSetLock) ; 1515 } 1516 1517 // The thread is now either on off-list (TS_RUN), 1518 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1519 // The Node's TState variable is stable from the perspective of this thread. 1520 // No other threads will asynchronously modify TState. 1521 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1522 OrderAccess::loadload() ; 1523 if (_succ == Self) _succ = NULL ; 1524 WasNotified = node._notified ; 1525 1526 // Reentry phase -- reacquire the monitor. 1527 // re-enter contended monitor after object.wait(). 1528 // retain OBJECT_WAIT state until re-enter successfully completes 1529 // Thread state is thread_in_vm and oop access is again safe, 1530 // although the raw address of the object may have changed. 1531 // (Don't cache naked oops over safepoints, of course). 1532 1533 // post monitor waited event. Note that this is past-tense, we are done waiting. 1534 if (JvmtiExport::should_post_monitor_waited()) { 1535 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1536 } 1537 OrderAccess::fence() ; 1538 1539 assert (Self->_Stalled != 0, "invariant") ; 1540 Self->_Stalled = 0 ; 1541 1542 assert (_owner != Self, "invariant") ; 1543 ObjectWaiter::TStates v = node.TState ; 1544 if (v == ObjectWaiter::TS_RUN) { 1545 enter (Self) ; 1546 } else { 1547 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1548 ReenterI (Self, &node) ; 1549 node.wait_reenter_end(this); 1550 } 1551 1552 // Self has reacquired the lock. 1553 // Lifecycle - the node representing Self must not appear on any queues. 1554 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1555 // want residual elements associated with this thread left on any lists. 1556 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1557 assert (_owner == Self, "invariant") ; 1558 assert (_succ != Self , "invariant") ; 1559 } // OSThreadWaitState() 1560 1561 jt->set_current_waiting_monitor(NULL); 1562 1563 guarantee (_recursions == 0, "invariant") ; 1564 _recursions = save; // restore the old recursion count 1565 _waiters--; // decrement the number of waiters 1566 1567 // Verify a few postconditions 1568 assert (_owner == Self , "invariant") ; 1569 assert (_succ != Self , "invariant") ; 1570 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1571 1572 if (SyncFlags & 32) { 1573 OrderAccess::fence() ; 1574 } 1575 1576 // check if the notification happened 1577 if (!WasNotified) { 1578 // no, it could be timeout or Thread.interrupt() or both 1579 // check for interrupt event, otherwise it is timeout 1580 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1581 TEVENT (Wait - throw IEX from epilog) ; 1582 THROW(vmSymbols::java_lang_InterruptedException()); 1583 } 1584 } 1585 1586 // NOTE: Spurious wake up will be consider as timeout. 1587 // Monitor notify has precedence over thread interrupt. 1588 } 1589 1590 1591 // Consider: 1592 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1593 // then instead of transferring a thread from the WaitSet to the EntryList 1594 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1595 1596 void ObjectMonitor::notify(TRAPS) { 1597 CHECK_OWNER(); 1598 if (_WaitSet == NULL) { 1599 TEVENT (Empty-Notify) ; 1600 return ; 1601 } 1602 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1603 1604 int Policy = Knob_MoveNotifyee ; 1605 1606 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1607 ObjectWaiter * iterator = DequeueWaiter() ; 1608 if (iterator != NULL) { 1609 TEVENT (Notify1 - Transfer) ; 1610 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1611 guarantee (iterator->_notified == 0, "invariant") ; 1612 if (Policy != 4) { 1613 iterator->TState = ObjectWaiter::TS_ENTER ; 1614 } 1615 iterator->_notified = 1 ; 1616 1617 ObjectWaiter * List = _EntryList ; 1618 if (List != NULL) { 1619 assert (List->_prev == NULL, "invariant") ; 1620 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1621 assert (List != iterator, "invariant") ; 1622 } 1623 1624 if (Policy == 0) { // prepend to EntryList 1625 if (List == NULL) { 1626 iterator->_next = iterator->_prev = NULL ; 1627 _EntryList = iterator ; 1628 } else { 1629 List->_prev = iterator ; 1630 iterator->_next = List ; 1631 iterator->_prev = NULL ; 1632 _EntryList = iterator ; 1633 } 1634 } else 1635 if (Policy == 1) { // append to EntryList 1636 if (List == NULL) { 1637 iterator->_next = iterator->_prev = NULL ; 1638 _EntryList = iterator ; 1639 } else { 1640 // CONSIDER: finding the tail currently requires a linear-time walk of 1641 // the EntryList. We can make tail access constant-time by converting to 1642 // a CDLL instead of using our current DLL. 1643 ObjectWaiter * Tail ; 1644 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1645 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1646 Tail->_next = iterator ; 1647 iterator->_prev = Tail ; 1648 iterator->_next = NULL ; 1649 } 1650 } else 1651 if (Policy == 2) { // prepend to cxq 1652 // prepend to cxq 1653 if (List == NULL) { 1654 iterator->_next = iterator->_prev = NULL ; 1655 _EntryList = iterator ; 1656 } else { 1657 iterator->TState = ObjectWaiter::TS_CXQ ; 1658 for (;;) { 1659 ObjectWaiter * Front = _cxq ; 1660 iterator->_next = Front ; 1661 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1662 break ; 1663 } 1664 } 1665 } 1666 } else 1667 if (Policy == 3) { // append to cxq 1668 iterator->TState = ObjectWaiter::TS_CXQ ; 1669 for (;;) { 1670 ObjectWaiter * Tail ; 1671 Tail = _cxq ; 1672 if (Tail == NULL) { 1673 iterator->_next = NULL ; 1674 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1675 break ; 1676 } 1677 } else { 1678 while (Tail->_next != NULL) Tail = Tail->_next ; 1679 Tail->_next = iterator ; 1680 iterator->_prev = Tail ; 1681 iterator->_next = NULL ; 1682 break ; 1683 } 1684 } 1685 } else { 1686 ParkEvent * ev = iterator->_event ; 1687 iterator->TState = ObjectWaiter::TS_RUN ; 1688 OrderAccess::fence() ; 1689 ev->unpark() ; 1690 } 1691 1692 if (Policy < 4) { 1693 iterator->wait_reenter_begin(this); 1694 } 1695 1696 // _WaitSetLock protects the wait queue, not the EntryList. We could 1697 // move the add-to-EntryList operation, above, outside the critical section 1698 // protected by _WaitSetLock. In practice that's not useful. With the 1699 // exception of wait() timeouts and interrupts the monitor owner 1700 // is the only thread that grabs _WaitSetLock. There's almost no contention 1701 // on _WaitSetLock so it's not profitable to reduce the length of the 1702 // critical section. 1703 } 1704 1705 Thread::SpinRelease (&_WaitSetLock) ; 1706 1707 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1708 ObjectMonitor::_sync_Notifications->inc() ; 1709 } 1710 } 1711 1712 1713 void ObjectMonitor::notifyAll(TRAPS) { 1714 CHECK_OWNER(); 1715 ObjectWaiter* iterator; 1716 if (_WaitSet == NULL) { 1717 TEVENT (Empty-NotifyAll) ; 1718 return ; 1719 } 1720 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1721 1722 int Policy = Knob_MoveNotifyee ; 1723 int Tally = 0 ; 1724 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1725 1726 for (;;) { 1727 iterator = DequeueWaiter () ; 1728 if (iterator == NULL) break ; 1729 TEVENT (NotifyAll - Transfer1) ; 1730 ++Tally ; 1731 1732 // Disposition - what might we do with iterator ? 1733 // a. add it directly to the EntryList - either tail or head. 1734 // b. push it onto the front of the _cxq. 1735 // For now we use (a). 1736 1737 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1738 guarantee (iterator->_notified == 0, "invariant") ; 1739 iterator->_notified = 1 ; 1740 if (Policy != 4) { 1741 iterator->TState = ObjectWaiter::TS_ENTER ; 1742 } 1743 1744 ObjectWaiter * List = _EntryList ; 1745 if (List != NULL) { 1746 assert (List->_prev == NULL, "invariant") ; 1747 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1748 assert (List != iterator, "invariant") ; 1749 } 1750 1751 if (Policy == 0) { // prepend to EntryList 1752 if (List == NULL) { 1753 iterator->_next = iterator->_prev = NULL ; 1754 _EntryList = iterator ; 1755 } else { 1756 List->_prev = iterator ; 1757 iterator->_next = List ; 1758 iterator->_prev = NULL ; 1759 _EntryList = iterator ; 1760 } 1761 } else 1762 if (Policy == 1) { // append to EntryList 1763 if (List == NULL) { 1764 iterator->_next = iterator->_prev = NULL ; 1765 _EntryList = iterator ; 1766 } else { 1767 // CONSIDER: finding the tail currently requires a linear-time walk of 1768 // the EntryList. We can make tail access constant-time by converting to 1769 // a CDLL instead of using our current DLL. 1770 ObjectWaiter * Tail ; 1771 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1772 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1773 Tail->_next = iterator ; 1774 iterator->_prev = Tail ; 1775 iterator->_next = NULL ; 1776 } 1777 } else 1778 if (Policy == 2) { // prepend to cxq 1779 // prepend to cxq 1780 iterator->TState = ObjectWaiter::TS_CXQ ; 1781 for (;;) { 1782 ObjectWaiter * Front = _cxq ; 1783 iterator->_next = Front ; 1784 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1785 break ; 1786 } 1787 } 1788 } else 1789 if (Policy == 3) { // append to cxq 1790 iterator->TState = ObjectWaiter::TS_CXQ ; 1791 for (;;) { 1792 ObjectWaiter * Tail ; 1793 Tail = _cxq ; 1794 if (Tail == NULL) { 1795 iterator->_next = NULL ; 1796 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1797 break ; 1798 } 1799 } else { 1800 while (Tail->_next != NULL) Tail = Tail->_next ; 1801 Tail->_next = iterator ; 1802 iterator->_prev = Tail ; 1803 iterator->_next = NULL ; 1804 break ; 1805 } 1806 } 1807 } else { 1808 ParkEvent * ev = iterator->_event ; 1809 iterator->TState = ObjectWaiter::TS_RUN ; 1810 OrderAccess::fence() ; 1811 ev->unpark() ; 1812 } 1813 1814 if (Policy < 4) { 1815 iterator->wait_reenter_begin(this); 1816 } 1817 1818 // _WaitSetLock protects the wait queue, not the EntryList. We could 1819 // move the add-to-EntryList operation, above, outside the critical section 1820 // protected by _WaitSetLock. In practice that's not useful. With the 1821 // exception of wait() timeouts and interrupts the monitor owner 1822 // is the only thread that grabs _WaitSetLock. There's almost no contention 1823 // on _WaitSetLock so it's not profitable to reduce the length of the 1824 // critical section. 1825 } 1826 1827 Thread::SpinRelease (&_WaitSetLock) ; 1828 1829 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1830 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1831 } 1832 } 1833 1834 // ----------------------------------------------------------------------------- 1835 // Adaptive Spinning Support 1836 // 1837 // Adaptive spin-then-block - rational spinning 1838 // 1839 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1840 // algorithm. On high order SMP systems it would be better to start with 1841 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1842 // a contending thread could enqueue itself on the cxq and then spin locally 1843 // on a thread-specific variable such as its ParkEvent._Event flag. 1844 // That's left as an exercise for the reader. Note that global spinning is 1845 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1846 // low latency and massive bandwidth. 1847 // 1848 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1849 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1850 // (duration) or we can fix the count at approximately the duration of 1851 // a context switch and vary the frequency. Of course we could also 1852 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1853 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1854 // 1855 // This implementation varies the duration "D", where D varies with 1856 // the success rate of recent spin attempts. (D is capped at approximately 1857 // length of a round-trip context switch). The success rate for recent 1858 // spin attempts is a good predictor of the success rate of future spin 1859 // attempts. The mechanism adapts automatically to varying critical 1860 // section length (lock modality), system load and degree of parallelism. 1861 // D is maintained per-monitor in _SpinDuration and is initialized 1862 // optimistically. Spin frequency is fixed at 100%. 1863 // 1864 // Note that _SpinDuration is volatile, but we update it without locks 1865 // or atomics. The code is designed so that _SpinDuration stays within 1866 // a reasonable range even in the presence of races. The arithmetic 1867 // operations on _SpinDuration are closed over the domain of legal values, 1868 // so at worst a race will install and older but still legal value. 1869 // At the very worst this introduces some apparent non-determinism. 1870 // We might spin when we shouldn't or vice-versa, but since the spin 1871 // count are relatively short, even in the worst case, the effect is harmless. 1872 // 1873 // Care must be taken that a low "D" value does not become an 1874 // an absorbing state. Transient spinning failures -- when spinning 1875 // is overall profitable -- should not cause the system to converge 1876 // on low "D" values. We want spinning to be stable and predictable 1877 // and fairly responsive to change and at the same time we don't want 1878 // it to oscillate, become metastable, be "too" non-deterministic, 1879 // or converge on or enter undesirable stable absorbing states. 1880 // 1881 // We implement a feedback-based control system -- using past behavior 1882 // to predict future behavior. We face two issues: (a) if the 1883 // input signal is random then the spin predictor won't provide optimal 1884 // results, and (b) if the signal frequency is too high then the control 1885 // system, which has some natural response lag, will "chase" the signal. 1886 // (b) can arise from multimodal lock hold times. Transient preemption 1887 // can also result in apparent bimodal lock hold times. 1888 // Although sub-optimal, neither condition is particularly harmful, as 1889 // in the worst-case we'll spin when we shouldn't or vice-versa. 1890 // The maximum spin duration is rather short so the failure modes aren't bad. 1891 // To be conservative, I've tuned the gain in system to bias toward 1892 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1893 // "rings" or oscillates between spinning and not spinning. This happens 1894 // when spinning is just on the cusp of profitability, however, so the 1895 // situation is not dire. The state is benign -- there's no need to add 1896 // hysteresis control to damp the transition rate between spinning and 1897 // not spinning. 1898 // 1899 1900 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1901 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1902 1903 // Spinning: Fixed frequency (100%), vary duration 1904 1905 1906 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1907 1908 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1909 int ctr = Knob_FixedSpin ; 1910 if (ctr != 0) { 1911 while (--ctr >= 0) { 1912 if (TryLock (Self) > 0) return 1 ; 1913 SpinPause () ; 1914 } 1915 return 0 ; 1916 } 1917 1918 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 1919 if (TryLock(Self) > 0) { 1920 // Increase _SpinDuration ... 1921 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1922 // Raising _SpurDuration to the poverty line is key. 1923 int x = _SpinDuration ; 1924 if (x < Knob_SpinLimit) { 1925 if (x < Knob_Poverty) x = Knob_Poverty ; 1926 _SpinDuration = x + Knob_BonusB ; 1927 } 1928 return 1 ; 1929 } 1930 SpinPause () ; 1931 } 1932 1933 // Admission control - verify preconditions for spinning 1934 // 1935 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1936 // becoming an absorbing state. Put another way, we spin briefly to 1937 // sample, just in case the system load, parallelism, contention, or lock 1938 // modality changed. 1939 // 1940 // Consider the following alternative: 1941 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1942 // spin attempt. "Periodically" might mean after a tally of 1943 // the # of failed spin attempts (or iterations) reaches some threshold. 1944 // This takes us into the realm of 1-out-of-N spinning, where we 1945 // hold the duration constant but vary the frequency. 1946 1947 ctr = _SpinDuration ; 1948 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 1949 if (ctr <= 0) return 0 ; 1950 1951 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 1952 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1953 TEVENT (Spin abort - notrunnable [TOP]); 1954 return 0 ; 1955 } 1956 1957 int MaxSpin = Knob_MaxSpinners ; 1958 if (MaxSpin >= 0) { 1959 if (_Spinner > MaxSpin) { 1960 TEVENT (Spin abort -- too many spinners) ; 1961 return 0 ; 1962 } 1963 // Slighty racy, but benign ... 1964 Adjust (&_Spinner, 1) ; 1965 } 1966 1967 // We're good to spin ... spin ingress. 1968 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1969 // when preparing to LD...CAS _owner, etc and the CAS is likely 1970 // to succeed. 1971 int hits = 0 ; 1972 int msk = 0 ; 1973 int caspty = Knob_CASPenalty ; 1974 int oxpty = Knob_OXPenalty ; 1975 int sss = Knob_SpinSetSucc ; 1976 if (sss && _succ == NULL ) _succ = Self ; 1977 Thread * prv = NULL ; 1978 1979 // There are three ways to exit the following loop: 1980 // 1. A successful spin where this thread has acquired the lock. 1981 // 2. Spin failure with prejudice 1982 // 3. Spin failure without prejudice 1983 1984 while (--ctr >= 0) { 1985 1986 // Periodic polling -- Check for pending GC 1987 // Threads may spin while they're unsafe. 1988 // We don't want spinning threads to delay the JVM from reaching 1989 // a stop-the-world safepoint or to steal cycles from GC. 1990 // If we detect a pending safepoint we abort in order that 1991 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1992 // this thread, if safe, doesn't steal cycles from GC. 1993 // This is in keeping with the "no loitering in runtime" rule. 1994 // We periodically check to see if there's a safepoint pending. 1995 if ((ctr & 0xFF) == 0) { 1996 if (SafepointSynchronize::do_call_back()) { 1997 TEVENT (Spin: safepoint) ; 1998 goto Abort ; // abrupt spin egress 1999 } 2000 if (Knob_UsePause & 1) SpinPause () ; 2001 2002 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2003 if (hits > 50 && scb != NULL) { 2004 int abend = (*scb)(SpinCallbackArgument, 0) ; 2005 } 2006 } 2007 2008 if (Knob_UsePause & 2) SpinPause() ; 2009 2010 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2011 // This is useful on classic SMP systems, but is of less utility on 2012 // N1-style CMT platforms. 2013 // 2014 // Trade-off: lock acquisition latency vs coherency bandwidth. 2015 // Lock hold times are typically short. A histogram 2016 // of successful spin attempts shows that we usually acquire 2017 // the lock early in the spin. That suggests we want to 2018 // sample _owner frequently in the early phase of the spin, 2019 // but then back-off and sample less frequently as the spin 2020 // progresses. The back-off makes a good citizen on SMP big 2021 // SMP systems. Oversampling _owner can consume excessive 2022 // coherency bandwidth. Relatedly, if we _oversample _owner we 2023 // can inadvertently interfere with the the ST m->owner=null. 2024 // executed by the lock owner. 2025 if (ctr & msk) continue ; 2026 ++hits ; 2027 if ((hits & 0xF) == 0) { 2028 // The 0xF, above, corresponds to the exponent. 2029 // Consider: (msk+1)|msk 2030 msk = ((msk << 2)|3) & BackOffMask ; 2031 } 2032 2033 // Probe _owner with TATAS 2034 // If this thread observes the monitor transition or flicker 2035 // from locked to unlocked to locked, then the odds that this 2036 // thread will acquire the lock in this spin attempt go down 2037 // considerably. The same argument applies if the CAS fails 2038 // or if we observe _owner change from one non-null value to 2039 // another non-null value. In such cases we might abort 2040 // the spin without prejudice or apply a "penalty" to the 2041 // spin count-down variable "ctr", reducing it by 100, say. 2042 2043 Thread * ox = (Thread *) _owner ; 2044 if (ox == NULL) { 2045 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2046 if (ox == NULL) { 2047 // The CAS succeeded -- this thread acquired ownership 2048 // Take care of some bookkeeping to exit spin state. 2049 if (sss && _succ == Self) { 2050 _succ = NULL ; 2051 } 2052 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2053 2054 // Increase _SpinDuration : 2055 // The spin was successful (profitable) so we tend toward 2056 // longer spin attempts in the future. 2057 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2058 // If we acquired the lock early in the spin cycle it 2059 // makes sense to increase _SpinDuration proportionally. 2060 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2061 int x = _SpinDuration ; 2062 if (x < Knob_SpinLimit) { 2063 if (x < Knob_Poverty) x = Knob_Poverty ; 2064 _SpinDuration = x + Knob_Bonus ; 2065 } 2066 return 1 ; 2067 } 2068 2069 // The CAS failed ... we can take any of the following actions: 2070 // * penalize: ctr -= Knob_CASPenalty 2071 // * exit spin with prejudice -- goto Abort; 2072 // * exit spin without prejudice. 2073 // * Since CAS is high-latency, retry again immediately. 2074 prv = ox ; 2075 TEVENT (Spin: cas failed) ; 2076 if (caspty == -2) break ; 2077 if (caspty == -1) goto Abort ; 2078 ctr -= caspty ; 2079 continue ; 2080 } 2081 2082 // Did lock ownership change hands ? 2083 if (ox != prv && prv != NULL ) { 2084 TEVENT (spin: Owner changed) 2085 if (oxpty == -2) break ; 2086 if (oxpty == -1) goto Abort ; 2087 ctr -= oxpty ; 2088 } 2089 prv = ox ; 2090 2091 // Abort the spin if the owner is not executing. 2092 // The owner must be executing in order to drop the lock. 2093 // Spinning while the owner is OFFPROC is idiocy. 2094 // Consider: ctr -= RunnablePenalty ; 2095 if (Knob_OState && NotRunnable (Self, ox)) { 2096 TEVENT (Spin abort - notrunnable); 2097 goto Abort ; 2098 } 2099 if (sss && _succ == NULL ) _succ = Self ; 2100 } 2101 2102 // Spin failed with prejudice -- reduce _SpinDuration. 2103 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2104 // AIMD is globally stable. 2105 TEVENT (Spin failure) ; 2106 { 2107 int x = _SpinDuration ; 2108 if (x > 0) { 2109 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2110 // This is globally sample and tends to damp the response. 2111 x -= Knob_Penalty ; 2112 if (x < 0) x = 0 ; 2113 _SpinDuration = x ; 2114 } 2115 } 2116 2117 Abort: 2118 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2119 if (sss && _succ == Self) { 2120 _succ = NULL ; 2121 // Invariant: after setting succ=null a contending thread 2122 // must recheck-retry _owner before parking. This usually happens 2123 // in the normal usage of TrySpin(), but it's safest 2124 // to make TrySpin() as foolproof as possible. 2125 OrderAccess::fence() ; 2126 if (TryLock(Self) > 0) return 1 ; 2127 } 2128 return 0 ; 2129 } 2130 2131 // NotRunnable() -- informed spinning 2132 // 2133 // Don't bother spinning if the owner is not eligible to drop the lock. 2134 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2135 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2136 // The thread must be runnable in order to drop the lock in timely fashion. 2137 // If the _owner is not runnable then spinning will not likely be 2138 // successful (profitable). 2139 // 2140 // Beware -- the thread referenced by _owner could have died 2141 // so a simply fetch from _owner->_thread_state might trap. 2142 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2143 // Because of the lifecycle issues the schedctl and _thread_state values 2144 // observed by NotRunnable() might be garbage. NotRunnable must 2145 // tolerate this and consider the observed _thread_state value 2146 // as advisory. 2147 // 2148 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2149 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2150 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2151 // with the LSB of _owner. Another option would be to probablistically probe 2152 // the putative _owner->TypeTag value. 2153 // 2154 // Checking _thread_state isn't perfect. Even if the thread is 2155 // in_java it might be blocked on a page-fault or have been preempted 2156 // and sitting on a ready/dispatch queue. _thread state in conjunction 2157 // with schedctl.sc_state gives us a good picture of what the 2158 // thread is doing, however. 2159 // 2160 // TODO: check schedctl.sc_state. 2161 // We'll need to use SafeFetch32() to read from the schedctl block. 2162 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2163 // 2164 // The return value from NotRunnable() is *advisory* -- the 2165 // result is based on sampling and is not necessarily coherent. 2166 // The caller must tolerate false-negative and false-positive errors. 2167 // Spinning, in general, is probabilistic anyway. 2168 2169 2170 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2171 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2172 if (!OwnerIsThread) return 0 ; 2173 2174 if (ox == NULL) return 0 ; 2175 2176 // Avoid transitive spinning ... 2177 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2178 // Immediately after T1 acquires L it's possible that T2, also 2179 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2180 // This occurs transiently after T1 acquired L but before 2181 // T1 managed to clear T1.Stalled. T2 does not need to abort 2182 // its spin in this circumstance. 2183 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2184 2185 if (BlockedOn == 1) return 1 ; 2186 if (BlockedOn != 0) { 2187 return BlockedOn != intptr_t(this) && _owner == ox ; 2188 } 2189 2190 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2191 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2192 // consider also: jst != _thread_in_Java -- but that's overspecific. 2193 return jst == _thread_blocked || jst == _thread_in_native ; 2194 } 2195 2196 2197 // ----------------------------------------------------------------------------- 2198 // WaitSet management ... 2199 2200 ObjectWaiter::ObjectWaiter(Thread* thread) { 2201 _next = NULL; 2202 _prev = NULL; 2203 _notified = 0; 2204 TState = TS_RUN ; 2205 _thread = thread; 2206 _event = thread->_ParkEvent ; 2207 _active = false; 2208 assert (_event != NULL, "invariant") ; 2209 } 2210 2211 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2212 JavaThread *jt = (JavaThread *)this->_thread; 2213 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2214 } 2215 2216 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2217 JavaThread *jt = (JavaThread *)this->_thread; 2218 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2219 } 2220 2221 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2222 assert(node != NULL, "should not dequeue NULL node"); 2223 assert(node->_prev == NULL, "node already in list"); 2224 assert(node->_next == NULL, "node already in list"); 2225 // put node at end of queue (circular doubly linked list) 2226 if (_WaitSet == NULL) { 2227 _WaitSet = node; 2228 node->_prev = node; 2229 node->_next = node; 2230 } else { 2231 ObjectWaiter* head = _WaitSet ; 2232 ObjectWaiter* tail = head->_prev; 2233 assert(tail->_next == head, "invariant check"); 2234 tail->_next = node; 2235 head->_prev = node; 2236 node->_next = head; 2237 node->_prev = tail; 2238 } 2239 } 2240 2241 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2242 // dequeue the very first waiter 2243 ObjectWaiter* waiter = _WaitSet; 2244 if (waiter) { 2245 DequeueSpecificWaiter(waiter); 2246 } 2247 return waiter; 2248 } 2249 2250 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2251 assert(node != NULL, "should not dequeue NULL node"); 2252 assert(node->_prev != NULL, "node already removed from list"); 2253 assert(node->_next != NULL, "node already removed from list"); 2254 // when the waiter has woken up because of interrupt, 2255 // timeout or other spurious wake-up, dequeue the 2256 // waiter from waiting list 2257 ObjectWaiter* next = node->_next; 2258 if (next == node) { 2259 assert(node->_prev == node, "invariant check"); 2260 _WaitSet = NULL; 2261 } else { 2262 ObjectWaiter* prev = node->_prev; 2263 assert(prev->_next == node, "invariant check"); 2264 assert(next->_prev == node, "invariant check"); 2265 next->_prev = prev; 2266 prev->_next = next; 2267 if (_WaitSet == node) { 2268 _WaitSet = next; 2269 } 2270 } 2271 node->_next = NULL; 2272 node->_prev = NULL; 2273 } 2274 2275 // ----------------------------------------------------------------------------- 2276 // PerfData support 2277 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2278 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2279 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2280 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2281 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2282 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2283 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2284 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2285 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2286 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2287 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2288 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2289 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2290 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2291 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2292 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2293 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2294 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2295 2296 // One-shot global initialization for the sync subsystem. 2297 // We could also defer initialization and initialize on-demand 2298 // the first time we call inflate(). Initialization would 2299 // be protected - like so many things - by the MonitorCache_lock. 2300 2301 void ObjectMonitor::Initialize () { 2302 static int InitializationCompleted = 0 ; 2303 assert (InitializationCompleted == 0, "invariant") ; 2304 InitializationCompleted = 1 ; 2305 if (UsePerfData) { 2306 EXCEPTION_MARK ; 2307 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2308 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2309 NEWPERFCOUNTER(_sync_Inflations) ; 2310 NEWPERFCOUNTER(_sync_Deflations) ; 2311 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2312 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2313 NEWPERFCOUNTER(_sync_Parks) ; 2314 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2315 NEWPERFCOUNTER(_sync_Notifications) ; 2316 NEWPERFCOUNTER(_sync_SlowEnter) ; 2317 NEWPERFCOUNTER(_sync_SlowExit) ; 2318 NEWPERFCOUNTER(_sync_SlowNotify) ; 2319 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2320 NEWPERFCOUNTER(_sync_FailedSpins) ; 2321 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2322 NEWPERFCOUNTER(_sync_PrivateA) ; 2323 NEWPERFCOUNTER(_sync_PrivateB) ; 2324 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2325 NEWPERFCOUNTER(_sync_MonScavenged) ; 2326 NEWPERFVARIABLE(_sync_MonExtant) ; 2327 #undef NEWPERFCOUNTER 2328 } 2329 } 2330 2331 2332 // Compile-time asserts 2333 // When possible, it's better to catch errors deterministically at 2334 // compile-time than at runtime. The down-side to using compile-time 2335 // asserts is that error message -- often something about negative array 2336 // indices -- is opaque. 2337 2338 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2339 2340 void ObjectMonitor::ctAsserts() { 2341 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2342 } 2343 2344 2345 static char * kvGet (char * kvList, const char * Key) { 2346 if (kvList == NULL) return NULL ; 2347 size_t n = strlen (Key) ; 2348 char * Search ; 2349 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2350 if (strncmp (Search, Key, n) == 0) { 2351 if (Search[n] == '=') return Search + n + 1 ; 2352 if (Search[n] == 0) return (char *) "1" ; 2353 } 2354 } 2355 return NULL ; 2356 } 2357 2358 static int kvGetInt (char * kvList, const char * Key, int Default) { 2359 char * v = kvGet (kvList, Key) ; 2360 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2361 if (Knob_ReportSettings && v != NULL) { 2362 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2363 ::fflush (stdout) ; 2364 } 2365 return rslt ; 2366 } 2367 2368 void ObjectMonitor::DeferredInitialize () { 2369 if (InitDone > 0) return ; 2370 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2371 while (InitDone != 1) ; 2372 return ; 2373 } 2374 2375 // One-shot global initialization ... 2376 // The initialization is idempotent, so we don't need locks. 2377 // In the future consider doing this via os::init_2(). 2378 // SyncKnobs consist of <Key>=<Value> pairs in the style 2379 // of environment variables. Start by converting ':' to NUL. 2380 2381 if (SyncKnobs == NULL) SyncKnobs = "" ; 2382 2383 size_t sz = strlen (SyncKnobs) ; 2384 char * knobs = (char *) malloc (sz + 2) ; 2385 if (knobs == NULL) { 2386 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; 2387 guarantee (0, "invariant") ; 2388 } 2389 strcpy (knobs, SyncKnobs) ; 2390 knobs[sz+1] = 0 ; 2391 for (char * p = knobs ; *p ; p++) { 2392 if (*p == ':') *p = 0 ; 2393 } 2394 2395 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2396 SETKNOB(ReportSettings) ; 2397 SETKNOB(Verbose) ; 2398 SETKNOB(FixedSpin) ; 2399 SETKNOB(SpinLimit) ; 2400 SETKNOB(SpinBase) ; 2401 SETKNOB(SpinBackOff); 2402 SETKNOB(CASPenalty) ; 2403 SETKNOB(OXPenalty) ; 2404 SETKNOB(LogSpins) ; 2405 SETKNOB(SpinSetSucc) ; 2406 SETKNOB(SuccEnabled) ; 2407 SETKNOB(SuccRestrict) ; 2408 SETKNOB(Penalty) ; 2409 SETKNOB(Bonus) ; 2410 SETKNOB(BonusB) ; 2411 SETKNOB(Poverty) ; 2412 SETKNOB(SpinAfterFutile) ; 2413 SETKNOB(UsePause) ; 2414 SETKNOB(SpinEarly) ; 2415 SETKNOB(OState) ; 2416 SETKNOB(MaxSpinners) ; 2417 SETKNOB(PreSpin) ; 2418 SETKNOB(ExitPolicy) ; 2419 SETKNOB(QMode); 2420 SETKNOB(ResetEvent) ; 2421 SETKNOB(MoveNotifyee) ; 2422 SETKNOB(FastHSSEC) ; 2423 #undef SETKNOB 2424 2425 if (os::is_MP()) { 2426 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2427 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2428 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2429 } else { 2430 Knob_SpinLimit = 0 ; 2431 Knob_SpinBase = 0 ; 2432 Knob_PreSpin = 0 ; 2433 Knob_FixedSpin = -1 ; 2434 } 2435 2436 if (Knob_LogSpins == 0) { 2437 ObjectMonitor::_sync_FailedSpins = NULL ; 2438 } 2439 2440 free (knobs) ; 2441 OrderAccess::fence() ; 2442 InitDone = 1 ; 2443 } 2444 2445 #ifndef PRODUCT 2446 void ObjectMonitor::verify() { 2447 } 2448 2449 void ObjectMonitor::print() { 2450 } 2451 #endif