1 /* 2 * Copyright (c) 1998, 2012, 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.inline.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 #endif 44 #ifdef TARGET_OS_FAMILY_solaris 45 # include "os_solaris.inline.hpp" 46 #endif 47 #ifdef TARGET_OS_FAMILY_windows 48 # include "os_windows.inline.hpp" 49 #endif 50 #ifdef TARGET_OS_FAMILY_bsd 51 # include "os_bsd.inline.hpp" 52 #endif 53 54 #if defined(__GNUC__) && !defined(IA64) 55 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 56 #define ATTR __attribute__((noinline)) 57 #else 58 #define ATTR 59 #endif 60 61 62 #ifdef DTRACE_ENABLED 63 64 // Only bother with this argument setup if dtrace is available 65 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 66 67 68 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 69 char* bytes = NULL; \ 70 int len = 0; \ 71 jlong jtid = SharedRuntime::get_java_tid(thread); \ 72 Symbol* klassname = ((oop)obj)->klass()->name(); \ 73 if (klassname != NULL) { \ 74 bytes = (char*)klassname->bytes(); \ 75 len = klassname->utf8_length(); \ 76 } 77 78 #ifndef USDT2 79 80 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 81 jlong, uintptr_t, char*, int); 82 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 83 jlong, uintptr_t, char*, int); 84 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 85 jlong, uintptr_t, char*, int); 86 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 87 jlong, uintptr_t, char*, int); 88 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 89 jlong, uintptr_t, char*, int); 90 91 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 92 { \ 93 if (DTraceMonitorProbes) { \ 94 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 95 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 96 (monitor), bytes, len, (millis)); \ 97 } \ 98 } 99 100 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 101 { \ 102 if (DTraceMonitorProbes) { \ 103 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 104 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 105 (uintptr_t)(monitor), bytes, len); \ 106 } \ 107 } 108 109 #else /* USDT2 */ 110 111 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 112 { \ 113 if (DTraceMonitorProbes) { \ 114 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 115 HOTSPOT_MONITOR_WAIT(jtid, \ 116 (monitor), bytes, len, (millis)); \ 117 } \ 118 } 119 120 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 121 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 122 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 123 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 124 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 125 126 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 127 { \ 128 if (DTraceMonitorProbes) { \ 129 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 130 HOTSPOT_MONITOR_##probe(jtid, \ 131 (uintptr_t)(monitor), bytes, len); \ 132 } \ 133 } 134 135 #endif /* USDT2 */ 136 #else // ndef DTRACE_ENABLED 137 138 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 139 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 140 141 #endif // ndef DTRACE_ENABLED 142 143 // Tunables ... 144 // The knob* variables are effectively final. Once set they should 145 // never be modified hence. Consider using __read_mostly with GCC. 146 147 int ObjectMonitor::Knob_Verbose = 0 ; 148 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 149 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 150 static int Knob_HandOff = 0 ; 151 static int Knob_ReportSettings = 0 ; 152 153 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 154 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 155 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 156 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 157 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 158 static int Knob_SpinEarly = 1 ; 159 static int Knob_SuccEnabled = 1 ; // futile wake throttling 160 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 161 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 162 static int Knob_Bonus = 100 ; // spin success bonus 163 static int Knob_BonusB = 100 ; // spin success bonus 164 static int Knob_Penalty = 200 ; // spin failure penalty 165 static int Knob_Poverty = 1000 ; 166 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 167 static int Knob_FixedSpin = 0 ; 168 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 169 static int Knob_UsePause = 1 ; 170 static int Knob_ExitPolicy = 0 ; 171 static int Knob_PreSpin = 10 ; // 20-100 likely better 172 static int Knob_ResetEvent = 0 ; 173 static int BackOffMask = 0 ; 174 175 static int Knob_FastHSSEC = 0 ; 176 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 177 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 178 static volatile int InitDone = 0 ; 179 180 #define TrySpin TrySpin_VaryDuration 181 182 // ----------------------------------------------------------------------------- 183 // Theory of operations -- Monitors lists, thread residency, etc: 184 // 185 // * A thread acquires ownership of a monitor by successfully 186 // CAS()ing the _owner field from null to non-null. 187 // 188 // * Invariant: A thread appears on at most one monitor list -- 189 // cxq, EntryList or WaitSet -- at any one time. 190 // 191 // * Contending threads "push" themselves onto the cxq with CAS 192 // and then spin/park. 193 // 194 // * After a contending thread eventually acquires the lock it must 195 // dequeue itself from either the EntryList or the cxq. 196 // 197 // * The exiting thread identifies and unparks an "heir presumptive" 198 // tentative successor thread on the EntryList. Critically, the 199 // exiting thread doesn't unlink the successor thread from the EntryList. 200 // After having been unparked, the wakee will recontend for ownership of 201 // the monitor. The successor (wakee) will either acquire the lock or 202 // re-park itself. 203 // 204 // Succession is provided for by a policy of competitive handoff. 205 // The exiting thread does _not_ grant or pass ownership to the 206 // successor thread. (This is also referred to as "handoff" succession"). 207 // Instead the exiting thread releases ownership and possibly wakes 208 // a successor, so the successor can (re)compete for ownership of the lock. 209 // If the EntryList is empty but the cxq is populated the exiting 210 // thread will drain the cxq into the EntryList. It does so by 211 // by detaching the cxq (installing null with CAS) and folding 212 // the threads from the cxq into the EntryList. The EntryList is 213 // doubly linked, while the cxq is singly linked because of the 214 // CAS-based "push" used to enqueue recently arrived threads (RATs). 215 // 216 // * Concurrency invariants: 217 // 218 // -- only the monitor owner may access or mutate the EntryList. 219 // The mutex property of the monitor itself protects the EntryList 220 // from concurrent interference. 221 // -- Only the monitor owner may detach the cxq. 222 // 223 // * The monitor entry list operations avoid locks, but strictly speaking 224 // they're not lock-free. Enter is lock-free, exit is not. 225 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 226 // 227 // * The cxq can have multiple concurrent "pushers" but only one concurrent 228 // detaching thread. This mechanism is immune from the ABA corruption. 229 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 230 // 231 // * Taken together, the cxq and the EntryList constitute or form a 232 // single logical queue of threads stalled trying to acquire the lock. 233 // We use two distinct lists to improve the odds of a constant-time 234 // dequeue operation after acquisition (in the ::enter() epilog) and 235 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 236 // A key desideratum is to minimize queue & monitor metadata manipulation 237 // that occurs while holding the monitor lock -- that is, we want to 238 // minimize monitor lock holds times. Note that even a small amount of 239 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 240 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 241 // locks and monitor metadata. 242 // 243 // Cxq points to the the set of Recently Arrived Threads attempting entry. 244 // Because we push threads onto _cxq with CAS, the RATs must take the form of 245 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 246 // the unlocking thread notices that EntryList is null but _cxq is != null. 247 // 248 // The EntryList is ordered by the prevailing queue discipline and 249 // can be organized in any convenient fashion, such as a doubly-linked list or 250 // a circular doubly-linked list. Critically, we want insert and delete operations 251 // to operate in constant-time. If we need a priority queue then something akin 252 // to Solaris' sleepq would work nicely. Viz., 253 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 254 // Queue discipline is enforced at ::exit() time, when the unlocking thread 255 // drains the cxq into the EntryList, and orders or reorders the threads on the 256 // EntryList accordingly. 257 // 258 // Barring "lock barging", this mechanism provides fair cyclic ordering, 259 // somewhat similar to an elevator-scan. 260 // 261 // * The monitor synchronization subsystem avoids the use of native 262 // synchronization primitives except for the narrow platform-specific 263 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 264 // the semantics of park-unpark. Put another way, this monitor implementation 265 // depends only on atomic operations and park-unpark. The monitor subsystem 266 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 267 // underlying OS manages the READY<->RUN transitions. 268 // 269 // * Waiting threads reside on the WaitSet list -- wait() puts 270 // the caller onto the WaitSet. 271 // 272 // * notify() or notifyAll() simply transfers threads from the WaitSet to 273 // either the EntryList or cxq. Subsequent exit() operations will 274 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 275 // it's likely the notifyee would simply impale itself on the lock held 276 // by the notifier. 277 // 278 // * An interesting alternative is to encode cxq as (List,LockByte) where 279 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 280 // variable, like _recursions, in the scheme. The threads or Events that form 281 // the list would have to be aligned in 256-byte addresses. A thread would 282 // try to acquire the lock or enqueue itself with CAS, but exiting threads 283 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 284 // Note that is is *not* word-tearing, but it does presume that full-word 285 // CAS operations are coherent with intermix with STB operations. That's true 286 // on most common processors. 287 // 288 // * See also http://blogs.sun.com/dave 289 290 291 // ----------------------------------------------------------------------------- 292 // Enter support 293 294 bool ObjectMonitor::try_enter(Thread* THREAD) { 295 if (THREAD != _owner) { 296 if (THREAD->is_lock_owned ((address)_owner)) { 297 assert(_recursions == 0, "internal state error"); 298 _owner = THREAD ; 299 _recursions = 1 ; 300 OwnerIsThread = 1 ; 301 return true; 302 } 303 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 304 return false; 305 } 306 return true; 307 } else { 308 _recursions++; 309 return true; 310 } 311 } 312 313 void ATTR ObjectMonitor::enter(TRAPS) { 314 // The following code is ordered to check the most common cases first 315 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 316 Thread * const Self = THREAD ; 317 void * cur ; 318 319 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 320 if (cur == NULL) { 321 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 322 assert (_recursions == 0 , "invariant") ; 323 assert (_owner == Self, "invariant") ; 324 // CONSIDER: set or assert OwnerIsThread == 1 325 return ; 326 } 327 328 if (cur == Self) { 329 // TODO-FIXME: check for integer overflow! BUGID 6557169. 330 _recursions ++ ; 331 return ; 332 } 333 334 if (Self->is_lock_owned ((address)cur)) { 335 assert (_recursions == 0, "internal state error"); 336 _recursions = 1 ; 337 // Commute owner from a thread-specific on-stack BasicLockObject address to 338 // a full-fledged "Thread *". 339 _owner = Self ; 340 OwnerIsThread = 1 ; 341 return ; 342 } 343 344 // We've encountered genuine contention. 345 assert (Self->_Stalled == 0, "invariant") ; 346 Self->_Stalled = intptr_t(this) ; 347 348 // Try one round of spinning *before* enqueueing Self 349 // and before going through the awkward and expensive state 350 // transitions. The following spin is strictly optional ... 351 // Note that if we acquire the monitor from an initial spin 352 // we forgo posting JVMTI events and firing DTRACE probes. 353 if (Knob_SpinEarly && TrySpin (Self) > 0) { 354 assert (_owner == Self , "invariant") ; 355 assert (_recursions == 0 , "invariant") ; 356 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 357 Self->_Stalled = 0 ; 358 return ; 359 } 360 361 assert (_owner != Self , "invariant") ; 362 assert (_succ != Self , "invariant") ; 363 assert (Self->is_Java_thread() , "invariant") ; 364 JavaThread * jt = (JavaThread *) Self ; 365 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 366 assert (jt->thread_state() != _thread_blocked , "invariant") ; 367 assert (this->object() != NULL , "invariant") ; 368 assert (_count >= 0, "invariant") ; 369 370 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 371 // Ensure the object-monitor relationship remains stable while there's contention. 372 Atomic::inc_ptr(&_count); 373 374 { // Change java thread status to indicate blocked on monitor enter. 375 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 376 377 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 378 if (JvmtiExport::should_post_monitor_contended_enter()) { 379 JvmtiExport::post_monitor_contended_enter(jt, this); 380 } 381 382 OSThreadContendState osts(Self->osthread()); 383 ThreadBlockInVM tbivm(jt); 384 385 Self->set_current_pending_monitor(this); 386 387 // TODO-FIXME: change the following for(;;) loop to straight-line code. 388 for (;;) { 389 jt->set_suspend_equivalent(); 390 // cleared by handle_special_suspend_equivalent_condition() 391 // or java_suspend_self() 392 393 EnterI (THREAD) ; 394 395 if (!ExitSuspendEquivalent(jt)) break ; 396 397 // 398 // We have acquired the contended monitor, but while we were 399 // waiting another thread suspended us. We don't want to enter 400 // the monitor while suspended because that would surprise the 401 // thread that suspended us. 402 // 403 _recursions = 0 ; 404 _succ = NULL ; 405 exit (Self) ; 406 407 jt->java_suspend_self(); 408 } 409 Self->set_current_pending_monitor(NULL); 410 } 411 412 Atomic::dec_ptr(&_count); 413 assert (_count >= 0, "invariant") ; 414 Self->_Stalled = 0 ; 415 416 // Must either set _recursions = 0 or ASSERT _recursions == 0. 417 assert (_recursions == 0 , "invariant") ; 418 assert (_owner == Self , "invariant") ; 419 assert (_succ != Self , "invariant") ; 420 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 421 422 // The thread -- now the owner -- is back in vm mode. 423 // Report the glorious news via TI,DTrace and jvmstat. 424 // The probe effect is non-trivial. All the reportage occurs 425 // while we hold the monitor, increasing the length of the critical 426 // section. Amdahl's parallel speedup law comes vividly into play. 427 // 428 // Another option might be to aggregate the events (thread local or 429 // per-monitor aggregation) and defer reporting until a more opportune 430 // time -- such as next time some thread encounters contention but has 431 // yet to acquire the lock. While spinning that thread could 432 // spinning we could increment JVMStat counters, etc. 433 434 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 435 if (JvmtiExport::should_post_monitor_contended_entered()) { 436 JvmtiExport::post_monitor_contended_entered(jt, this); 437 } 438 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 439 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 440 } 441 } 442 443 444 // Caveat: TryLock() is not necessarily serializing if it returns failure. 445 // Callers must compensate as needed. 446 447 int ObjectMonitor::TryLock (Thread * Self) { 448 for (;;) { 449 void * own = _owner ; 450 if (own != NULL) return 0 ; 451 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 452 // Either guarantee _recursions == 0 or set _recursions = 0. 453 assert (_recursions == 0, "invariant") ; 454 assert (_owner == Self, "invariant") ; 455 // CONSIDER: set or assert that OwnerIsThread == 1 456 return 1 ; 457 } 458 // The lock had been free momentarily, but we lost the race to the lock. 459 // Interference -- the CAS failed. 460 // We can either return -1 or retry. 461 // Retry doesn't make as much sense because the lock was just acquired. 462 if (true) return -1 ; 463 } 464 } 465 466 void ATTR ObjectMonitor::EnterI (TRAPS) { 467 Thread * Self = THREAD ; 468 assert (Self->is_Java_thread(), "invariant") ; 469 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 470 471 // Try the lock - TATAS 472 if (TryLock (Self) > 0) { 473 assert (_succ != Self , "invariant") ; 474 assert (_owner == Self , "invariant") ; 475 assert (_Responsible != Self , "invariant") ; 476 return ; 477 } 478 479 DeferredInitialize () ; 480 481 // We try one round of spinning *before* enqueueing Self. 482 // 483 // If the _owner is ready but OFFPROC we could use a YieldTo() 484 // operation to donate the remainder of this thread's quantum 485 // to the owner. This has subtle but beneficial affinity 486 // effects. 487 488 if (TrySpin (Self) > 0) { 489 assert (_owner == Self , "invariant") ; 490 assert (_succ != Self , "invariant") ; 491 assert (_Responsible != Self , "invariant") ; 492 return ; 493 } 494 495 // The Spin failed -- Enqueue and park the thread ... 496 assert (_succ != Self , "invariant") ; 497 assert (_owner != Self , "invariant") ; 498 assert (_Responsible != Self , "invariant") ; 499 500 // Enqueue "Self" on ObjectMonitor's _cxq. 501 // 502 // Node acts as a proxy for Self. 503 // As an aside, if were to ever rewrite the synchronization code mostly 504 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 505 // Java objects. This would avoid awkward lifecycle and liveness issues, 506 // as well as eliminate a subset of ABA issues. 507 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 508 // 509 510 ObjectWaiter node(Self) ; 511 Self->_ParkEvent->reset() ; 512 node._prev = (ObjectWaiter *) 0xBAD ; 513 node.TState = ObjectWaiter::TS_CXQ ; 514 515 // Push "Self" onto the front of the _cxq. 516 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 517 // Note that spinning tends to reduce the rate at which threads 518 // enqueue and dequeue on EntryList|cxq. 519 ObjectWaiter * nxt ; 520 for (;;) { 521 node._next = nxt = _cxq ; 522 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 523 524 // Interference - the CAS failed because _cxq changed. Just retry. 525 // As an optional optimization we retry the lock. 526 if (TryLock (Self) > 0) { 527 assert (_succ != Self , "invariant") ; 528 assert (_owner == Self , "invariant") ; 529 assert (_Responsible != Self , "invariant") ; 530 return ; 531 } 532 } 533 534 // Check for cxq|EntryList edge transition to non-null. This indicates 535 // the onset of contention. While contention persists exiting threads 536 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 537 // operations revert to the faster 1-0 mode. This enter operation may interleave 538 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 539 // arrange for one of the contending thread to use a timed park() operations 540 // to detect and recover from the race. (Stranding is form of progress failure 541 // where the monitor is unlocked but all the contending threads remain parked). 542 // That is, at least one of the contended threads will periodically poll _owner. 543 // One of the contending threads will become the designated "Responsible" thread. 544 // The Responsible thread uses a timed park instead of a normal indefinite park 545 // operation -- it periodically wakes and checks for and recovers from potential 546 // strandings admitted by 1-0 exit operations. We need at most one Responsible 547 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 548 // be responsible for a monitor. 549 // 550 // Currently, one of the contended threads takes on the added role of "Responsible". 551 // A viable alternative would be to use a dedicated "stranding checker" thread 552 // that periodically iterated over all the threads (or active monitors) and unparked 553 // successors where there was risk of stranding. This would help eliminate the 554 // timer scalability issues we see on some platforms as we'd only have one thread 555 // -- the checker -- parked on a timer. 556 557 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 558 // Try to assume the role of responsible thread for the monitor. 559 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 560 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 561 } 562 563 // The lock have been released while this thread was occupied queueing 564 // itself onto _cxq. To close the race and avoid "stranding" and 565 // progress-liveness failure we must resample-retry _owner before parking. 566 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 567 // In this case the ST-MEMBAR is accomplished with CAS(). 568 // 569 // TODO: Defer all thread state transitions until park-time. 570 // Since state transitions are heavy and inefficient we'd like 571 // to defer the state transitions until absolutely necessary, 572 // and in doing so avoid some transitions ... 573 574 TEVENT (Inflated enter - Contention) ; 575 int nWakeups = 0 ; 576 int RecheckInterval = 1 ; 577 578 for (;;) { 579 580 if (TryLock (Self) > 0) break ; 581 assert (_owner != Self, "invariant") ; 582 583 if ((SyncFlags & 2) && _Responsible == NULL) { 584 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 585 } 586 587 // park self 588 if (_Responsible == Self || (SyncFlags & 1)) { 589 TEVENT (Inflated enter - park TIMED) ; 590 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 591 // Increase the RecheckInterval, but clamp the value. 592 RecheckInterval *= 8 ; 593 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 594 } else { 595 TEVENT (Inflated enter - park UNTIMED) ; 596 Self->_ParkEvent->park() ; 597 } 598 599 if (TryLock(Self) > 0) break ; 600 601 // The lock is still contested. 602 // Keep a tally of the # of futile wakeups. 603 // Note that the counter is not protected by a lock or updated by atomics. 604 // That is by design - we trade "lossy" counters which are exposed to 605 // races during updates for a lower probe effect. 606 TEVENT (Inflated enter - Futile wakeup) ; 607 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 608 ObjectMonitor::_sync_FutileWakeups->inc() ; 609 } 610 ++ nWakeups ; 611 612 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 613 // We can defer clearing _succ until after the spin completes 614 // TrySpin() must tolerate being called with _succ == Self. 615 // Try yet another round of adaptive spinning. 616 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 617 618 // We can find that we were unpark()ed and redesignated _succ while 619 // we were spinning. That's harmless. If we iterate and call park(), 620 // park() will consume the event and return immediately and we'll 621 // just spin again. This pattern can repeat, leaving _succ to simply 622 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 623 // Alternately, we can sample fired() here, and if set, forgo spinning 624 // in the next iteration. 625 626 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 627 Self->_ParkEvent->reset() ; 628 OrderAccess::fence() ; 629 } 630 if (_succ == Self) _succ = NULL ; 631 632 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 633 OrderAccess::fence() ; 634 } 635 636 // Egress : 637 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 638 // Normally we'll find Self on the EntryList . 639 // From the perspective of the lock owner (this thread), the 640 // EntryList is stable and cxq is prepend-only. 641 // The head of cxq is volatile but the interior is stable. 642 // In addition, Self.TState is stable. 643 644 assert (_owner == Self , "invariant") ; 645 assert (object() != NULL , "invariant") ; 646 // I'd like to write: 647 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 648 // but as we're at a safepoint that's not safe. 649 650 UnlinkAfterAcquire (Self, &node) ; 651 if (_succ == Self) _succ = NULL ; 652 653 assert (_succ != Self, "invariant") ; 654 if (_Responsible == Self) { 655 _Responsible = NULL ; 656 // Dekker pivot-point. 657 // Consider OrderAccess::storeload() here 658 659 // We may leave threads on cxq|EntryList without a designated 660 // "Responsible" thread. This is benign. When this thread subsequently 661 // exits the monitor it can "see" such preexisting "old" threads -- 662 // threads that arrived on the cxq|EntryList before the fence, above -- 663 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 664 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 665 // non-null and elect a new "Responsible" timer thread. 666 // 667 // This thread executes: 668 // ST Responsible=null; MEMBAR (in enter epilog - here) 669 // LD cxq|EntryList (in subsequent exit) 670 // 671 // Entering threads in the slow/contended path execute: 672 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 673 // The (ST cxq; MEMBAR) is accomplished with CAS(). 674 // 675 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 676 // exit operation from floating above the ST Responsible=null. 677 // 678 // In *practice* however, EnterI() is always followed by some atomic 679 // operation such as the decrement of _count in ::enter(). Those atomics 680 // obviate the need for the explicit MEMBAR, above. 681 } 682 683 // We've acquired ownership with CAS(). 684 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 685 // But since the CAS() this thread may have also stored into _succ, 686 // EntryList, cxq or Responsible. These meta-data updates must be 687 // visible __before this thread subsequently drops the lock. 688 // Consider what could occur if we didn't enforce this constraint -- 689 // STs to monitor meta-data and user-data could reorder with (become 690 // visible after) the ST in exit that drops ownership of the lock. 691 // Some other thread could then acquire the lock, but observe inconsistent 692 // or old monitor meta-data and heap data. That violates the JMM. 693 // To that end, the 1-0 exit() operation must have at least STST|LDST 694 // "release" barrier semantics. Specifically, there must be at least a 695 // STST|LDST barrier in exit() before the ST of null into _owner that drops 696 // the lock. The barrier ensures that changes to monitor meta-data and data 697 // protected by the lock will be visible before we release the lock, and 698 // therefore before some other thread (CPU) has a chance to acquire the lock. 699 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 700 // 701 // Critically, any prior STs to _succ or EntryList must be visible before 702 // the ST of null into _owner in the *subsequent* (following) corresponding 703 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 704 // execute a serializing instruction. 705 706 if (SyncFlags & 8) { 707 OrderAccess::fence() ; 708 } 709 return ; 710 } 711 712 // ReenterI() is a specialized inline form of the latter half of the 713 // contended slow-path from EnterI(). We use ReenterI() only for 714 // monitor reentry in wait(). 715 // 716 // In the future we should reconcile EnterI() and ReenterI(), adding 717 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 718 // loop accordingly. 719 720 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 721 assert (Self != NULL , "invariant") ; 722 assert (SelfNode != NULL , "invariant") ; 723 assert (SelfNode->_thread == Self , "invariant") ; 724 assert (_waiters > 0 , "invariant") ; 725 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 726 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 727 JavaThread * jt = (JavaThread *) Self ; 728 729 int nWakeups = 0 ; 730 for (;;) { 731 ObjectWaiter::TStates v = SelfNode->TState ; 732 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 733 assert (_owner != Self, "invariant") ; 734 735 if (TryLock (Self) > 0) break ; 736 if (TrySpin (Self) > 0) break ; 737 738 TEVENT (Wait Reentry - parking) ; 739 740 // State transition wrappers around park() ... 741 // ReenterI() wisely defers state transitions until 742 // it's clear we must park the thread. 743 { 744 OSThreadContendState osts(Self->osthread()); 745 ThreadBlockInVM tbivm(jt); 746 747 // cleared by handle_special_suspend_equivalent_condition() 748 // or java_suspend_self() 749 jt->set_suspend_equivalent(); 750 if (SyncFlags & 1) { 751 Self->_ParkEvent->park ((jlong)1000) ; 752 } else { 753 Self->_ParkEvent->park () ; 754 } 755 756 // were we externally suspended while we were waiting? 757 for (;;) { 758 if (!ExitSuspendEquivalent (jt)) break ; 759 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 760 jt->java_suspend_self(); 761 jt->set_suspend_equivalent(); 762 } 763 } 764 765 // Try again, but just so we distinguish between futile wakeups and 766 // successful wakeups. The following test isn't algorithmically 767 // necessary, but it helps us maintain sensible statistics. 768 if (TryLock(Self) > 0) break ; 769 770 // The lock is still contested. 771 // Keep a tally of the # of futile wakeups. 772 // Note that the counter is not protected by a lock or updated by atomics. 773 // That is by design - we trade "lossy" counters which are exposed to 774 // races during updates for a lower probe effect. 775 TEVENT (Wait Reentry - futile wakeup) ; 776 ++ nWakeups ; 777 778 // Assuming this is not a spurious wakeup we'll normally 779 // find that _succ == Self. 780 if (_succ == Self) _succ = NULL ; 781 782 // Invariant: after clearing _succ a contending thread 783 // *must* retry _owner before parking. 784 OrderAccess::fence() ; 785 786 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 787 ObjectMonitor::_sync_FutileWakeups->inc() ; 788 } 789 } 790 791 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 792 // Normally we'll find Self on the EntryList. 793 // Unlinking from the EntryList is constant-time and atomic-free. 794 // From the perspective of the lock owner (this thread), the 795 // EntryList is stable and cxq is prepend-only. 796 // The head of cxq is volatile but the interior is stable. 797 // In addition, Self.TState is stable. 798 799 assert (_owner == Self, "invariant") ; 800 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 801 UnlinkAfterAcquire (Self, SelfNode) ; 802 if (_succ == Self) _succ = NULL ; 803 assert (_succ != Self, "invariant") ; 804 SelfNode->TState = ObjectWaiter::TS_RUN ; 805 OrderAccess::fence() ; // see comments at the end of EnterI() 806 } 807 808 // after the thread acquires the lock in ::enter(). Equally, we could defer 809 // unlinking the thread until ::exit()-time. 810 811 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 812 { 813 assert (_owner == Self, "invariant") ; 814 assert (SelfNode->_thread == Self, "invariant") ; 815 816 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 817 // Normal case: remove Self from the DLL EntryList . 818 // This is a constant-time operation. 819 ObjectWaiter * nxt = SelfNode->_next ; 820 ObjectWaiter * prv = SelfNode->_prev ; 821 if (nxt != NULL) nxt->_prev = prv ; 822 if (prv != NULL) prv->_next = nxt ; 823 if (SelfNode == _EntryList ) _EntryList = nxt ; 824 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 825 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 826 TEVENT (Unlink from EntryList) ; 827 } else { 828 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 829 // Inopportune interleaving -- Self is still on the cxq. 830 // This usually means the enqueue of self raced an exiting thread. 831 // Normally we'll find Self near the front of the cxq, so 832 // dequeueing is typically fast. If needbe we can accelerate 833 // this with some MCS/CHL-like bidirectional list hints and advisory 834 // back-links so dequeueing from the interior will normally operate 835 // in constant-time. 836 // Dequeue Self from either the head (with CAS) or from the interior 837 // with a linear-time scan and normal non-atomic memory operations. 838 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 839 // and then unlink Self from EntryList. We have to drain eventually, 840 // so it might as well be now. 841 842 ObjectWaiter * v = _cxq ; 843 assert (v != NULL, "invariant") ; 844 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 845 // The CAS above can fail from interference IFF a "RAT" arrived. 846 // In that case Self must be in the interior and can no longer be 847 // at the head of cxq. 848 if (v == SelfNode) { 849 assert (_cxq != v, "invariant") ; 850 v = _cxq ; // CAS above failed - start scan at head of list 851 } 852 ObjectWaiter * p ; 853 ObjectWaiter * q = NULL ; 854 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 855 q = p ; 856 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 857 } 858 assert (v != SelfNode, "invariant") ; 859 assert (p == SelfNode, "Node not found on cxq") ; 860 assert (p != _cxq, "invariant") ; 861 assert (q != NULL, "invariant") ; 862 assert (q->_next == p, "invariant") ; 863 q->_next = p->_next ; 864 } 865 TEVENT (Unlink from cxq) ; 866 } 867 868 // Diagnostic hygiene ... 869 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 870 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 871 SelfNode->TState = ObjectWaiter::TS_RUN ; 872 } 873 874 // ----------------------------------------------------------------------------- 875 // Exit support 876 // 877 // exit() 878 // ~~~~~~ 879 // Note that the collector can't reclaim the objectMonitor or deflate 880 // the object out from underneath the thread calling ::exit() as the 881 // thread calling ::exit() never transitions to a stable state. 882 // This inhibits GC, which in turn inhibits asynchronous (and 883 // inopportune) reclamation of "this". 884 // 885 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 886 // There's one exception to the claim above, however. EnterI() can call 887 // exit() to drop a lock if the acquirer has been externally suspended. 888 // In that case exit() is called with _thread_state as _thread_blocked, 889 // but the monitor's _count field is > 0, which inhibits reclamation. 890 // 891 // 1-0 exit 892 // ~~~~~~~~ 893 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 894 // the fast-path operators have been optimized so the common ::exit() 895 // operation is 1-0. See i486.ad fast_unlock(), for instance. 896 // The code emitted by fast_unlock() elides the usual MEMBAR. This 897 // greatly improves latency -- MEMBAR and CAS having considerable local 898 // latency on modern processors -- but at the cost of "stranding". Absent the 899 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 900 // ::enter() path, resulting in the entering thread being stranding 901 // and a progress-liveness failure. Stranding is extremely rare. 902 // We use timers (timed park operations) & periodic polling to detect 903 // and recover from stranding. Potentially stranded threads periodically 904 // wake up and poll the lock. See the usage of the _Responsible variable. 905 // 906 // The CAS() in enter provides for safety and exclusion, while the CAS or 907 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 908 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 909 // We detect and recover from stranding with timers. 910 // 911 // If a thread transiently strands it'll park until (a) another 912 // thread acquires the lock and then drops the lock, at which time the 913 // exiting thread will notice and unpark the stranded thread, or, (b) 914 // the timer expires. If the lock is high traffic then the stranding latency 915 // will be low due to (a). If the lock is low traffic then the odds of 916 // stranding are lower, although the worst-case stranding latency 917 // is longer. Critically, we don't want to put excessive load in the 918 // platform's timer subsystem. We want to minimize both the timer injection 919 // rate (timers created/sec) as well as the number of timers active at 920 // any one time. (more precisely, we want to minimize timer-seconds, which is 921 // the integral of the # of active timers at any instant over time). 922 // Both impinge on OS scalability. Given that, at most one thread parked on 923 // a monitor will use a timer. 924 925 void ATTR ObjectMonitor::exit(TRAPS) { 926 Thread * Self = THREAD ; 927 if (THREAD != _owner) { 928 if (THREAD->is_lock_owned((address) _owner)) { 929 // Transmute _owner from a BasicLock pointer to a Thread address. 930 // We don't need to hold _mutex for this transition. 931 // Non-null to Non-null is safe as long as all readers can 932 // tolerate either flavor. 933 assert (_recursions == 0, "invariant") ; 934 _owner = THREAD ; 935 _recursions = 0 ; 936 OwnerIsThread = 1 ; 937 } else { 938 // NOTE: we need to handle unbalanced monitor enter/exit 939 // in native code by throwing an exception. 940 // TODO: Throw an IllegalMonitorStateException ? 941 TEVENT (Exit - Throw IMSX) ; 942 assert(false, "Non-balanced monitor enter/exit!"); 943 if (false) { 944 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 945 } 946 return; 947 } 948 } 949 950 if (_recursions != 0) { 951 _recursions--; // this is simple recursive enter 952 TEVENT (Inflated exit - recursive) ; 953 return ; 954 } 955 956 // Invariant: after setting Responsible=null an thread must execute 957 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 958 if ((SyncFlags & 4) == 0) { 959 _Responsible = NULL ; 960 } 961 962 for (;;) { 963 assert (THREAD == _owner, "invariant") ; 964 965 966 if (Knob_ExitPolicy == 0) { 967 // release semantics: prior loads and stores from within the critical section 968 // must not float (reorder) past the following store that drops the lock. 969 // On SPARC that requires MEMBAR #loadstore|#storestore. 970 // But of course in TSO #loadstore|#storestore is not required. 971 // I'd like to write one of the following: 972 // A. OrderAccess::release() ; _owner = NULL 973 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 974 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 975 // store into a _dummy variable. That store is not needed, but can result 976 // in massive wasteful coherency traffic on classic SMP systems. 977 // Instead, I use release_store(), which is implemented as just a simple 978 // ST on x64, x86 and SPARC. 979 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 980 OrderAccess::storeload() ; // See if we need to wake a successor 981 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 982 TEVENT (Inflated exit - simple egress) ; 983 return ; 984 } 985 TEVENT (Inflated exit - complex egress) ; 986 987 // Normally the exiting thread is responsible for ensuring succession, 988 // but if other successors are ready or other entering threads are spinning 989 // then this thread can simply store NULL into _owner and exit without 990 // waking a successor. The existence of spinners or ready successors 991 // guarantees proper succession (liveness). Responsibility passes to the 992 // ready or running successors. The exiting thread delegates the duty. 993 // More precisely, if a successor already exists this thread is absolved 994 // of the responsibility of waking (unparking) one. 995 // 996 // The _succ variable is critical to reducing futile wakeup frequency. 997 // _succ identifies the "heir presumptive" thread that has been made 998 // ready (unparked) but that has not yet run. We need only one such 999 // successor thread to guarantee progress. 1000 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1001 // section 3.3 "Futile Wakeup Throttling" for details. 1002 // 1003 // Note that spinners in Enter() also set _succ non-null. 1004 // In the current implementation spinners opportunistically set 1005 // _succ so that exiting threads might avoid waking a successor. 1006 // Another less appealing alternative would be for the exiting thread 1007 // to drop the lock and then spin briefly to see if a spinner managed 1008 // to acquire the lock. If so, the exiting thread could exit 1009 // immediately without waking a successor, otherwise the exiting 1010 // thread would need to dequeue and wake a successor. 1011 // (Note that we'd need to make the post-drop spin short, but no 1012 // shorter than the worst-case round-trip cache-line migration time. 1013 // The dropped lock needs to become visible to the spinner, and then 1014 // the acquisition of the lock by the spinner must become visible to 1015 // the exiting thread). 1016 // 1017 1018 // It appears that an heir-presumptive (successor) must be made ready. 1019 // Only the current lock owner can manipulate the EntryList or 1020 // drain _cxq, so we need to reacquire the lock. If we fail 1021 // to reacquire the lock the responsibility for ensuring succession 1022 // falls to the new owner. 1023 // 1024 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1025 return ; 1026 } 1027 TEVENT (Exit - Reacquired) ; 1028 } else { 1029 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1030 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1031 OrderAccess::storeload() ; 1032 // Ratify the previously observed values. 1033 if (_cxq == NULL || _succ != NULL) { 1034 TEVENT (Inflated exit - simple egress) ; 1035 return ; 1036 } 1037 1038 // inopportune interleaving -- the exiting thread (this thread) 1039 // in the fast-exit path raced an entering thread in the slow-enter 1040 // path. 1041 // We have two choices: 1042 // A. Try to reacquire the lock. 1043 // If the CAS() fails return immediately, otherwise 1044 // we either restart/rerun the exit operation, or simply 1045 // fall-through into the code below which wakes a successor. 1046 // B. If the elements forming the EntryList|cxq are TSM 1047 // we could simply unpark() the lead thread and return 1048 // without having set _succ. 1049 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1050 TEVENT (Inflated exit - reacquired succeeded) ; 1051 return ; 1052 } 1053 TEVENT (Inflated exit - reacquired failed) ; 1054 } else { 1055 TEVENT (Inflated exit - complex egress) ; 1056 } 1057 } 1058 1059 guarantee (_owner == THREAD, "invariant") ; 1060 1061 ObjectWaiter * w = NULL ; 1062 int QMode = Knob_QMode ; 1063 1064 if (QMode == 2 && _cxq != NULL) { 1065 // QMode == 2 : cxq has precedence over EntryList. 1066 // Try to directly wake a successor from the cxq. 1067 // If successful, the successor will need to unlink itself from cxq. 1068 w = _cxq ; 1069 assert (w != NULL, "invariant") ; 1070 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1071 ExitEpilog (Self, w) ; 1072 return ; 1073 } 1074 1075 if (QMode == 3 && _cxq != NULL) { 1076 // Aggressively drain cxq into EntryList at the first opportunity. 1077 // This policy ensure that recently-run threads live at the head of EntryList. 1078 // Drain _cxq into EntryList - bulk transfer. 1079 // First, detach _cxq. 1080 // The following loop is tantamount to: w = swap (&cxq, NULL) 1081 w = _cxq ; 1082 for (;;) { 1083 assert (w != NULL, "Invariant") ; 1084 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1085 if (u == w) break ; 1086 w = u ; 1087 } 1088 assert (w != NULL , "invariant") ; 1089 1090 ObjectWaiter * q = NULL ; 1091 ObjectWaiter * p ; 1092 for (p = w ; p != NULL ; p = p->_next) { 1093 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1094 p->TState = ObjectWaiter::TS_ENTER ; 1095 p->_prev = q ; 1096 q = p ; 1097 } 1098 1099 // Append the RATs to the EntryList 1100 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1101 ObjectWaiter * Tail ; 1102 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1103 if (Tail == NULL) { 1104 _EntryList = w ; 1105 } else { 1106 Tail->_next = w ; 1107 w->_prev = Tail ; 1108 } 1109 1110 // Fall thru into code that tries to wake a successor from EntryList 1111 } 1112 1113 if (QMode == 4 && _cxq != NULL) { 1114 // Aggressively drain cxq into EntryList at the first opportunity. 1115 // This policy ensure that recently-run threads live at the head of EntryList. 1116 1117 // Drain _cxq into EntryList - bulk transfer. 1118 // First, detach _cxq. 1119 // The following loop is tantamount to: w = swap (&cxq, NULL) 1120 w = _cxq ; 1121 for (;;) { 1122 assert (w != NULL, "Invariant") ; 1123 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1124 if (u == w) break ; 1125 w = u ; 1126 } 1127 assert (w != NULL , "invariant") ; 1128 1129 ObjectWaiter * q = NULL ; 1130 ObjectWaiter * p ; 1131 for (p = w ; p != NULL ; p = p->_next) { 1132 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1133 p->TState = ObjectWaiter::TS_ENTER ; 1134 p->_prev = q ; 1135 q = p ; 1136 } 1137 1138 // Prepend the RATs to the EntryList 1139 if (_EntryList != NULL) { 1140 q->_next = _EntryList ; 1141 _EntryList->_prev = q ; 1142 } 1143 _EntryList = w ; 1144 1145 // Fall thru into code that tries to wake a successor from EntryList 1146 } 1147 1148 w = _EntryList ; 1149 if (w != NULL) { 1150 // I'd like to write: guarantee (w->_thread != Self). 1151 // But in practice an exiting thread may find itself on the EntryList. 1152 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1153 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1154 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1155 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1156 // release the lock "O". T2 resumes immediately after the ST of null into 1157 // _owner, above. T2 notices that the EntryList is populated, so it 1158 // reacquires the lock and then finds itself on the EntryList. 1159 // Given all that, we have to tolerate the circumstance where "w" is 1160 // associated with Self. 1161 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1162 ExitEpilog (Self, w) ; 1163 return ; 1164 } 1165 1166 // If we find that both _cxq and EntryList are null then just 1167 // re-run the exit protocol from the top. 1168 w = _cxq ; 1169 if (w == NULL) continue ; 1170 1171 // Drain _cxq into EntryList - bulk transfer. 1172 // First, detach _cxq. 1173 // The following loop is tantamount to: w = swap (&cxq, NULL) 1174 for (;;) { 1175 assert (w != NULL, "Invariant") ; 1176 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1177 if (u == w) break ; 1178 w = u ; 1179 } 1180 TEVENT (Inflated exit - drain cxq into EntryList) ; 1181 1182 assert (w != NULL , "invariant") ; 1183 assert (_EntryList == NULL , "invariant") ; 1184 1185 // Convert the LIFO SLL anchored by _cxq into a DLL. 1186 // The list reorganization step operates in O(LENGTH(w)) time. 1187 // It's critical that this step operate quickly as 1188 // "Self" still holds the outer-lock, restricting parallelism 1189 // and effectively lengthening the critical section. 1190 // Invariant: s chases t chases u. 1191 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1192 // we have faster access to the tail. 1193 1194 if (QMode == 1) { 1195 // QMode == 1 : drain cxq to EntryList, reversing order 1196 // We also reverse the order of the list. 1197 ObjectWaiter * s = NULL ; 1198 ObjectWaiter * t = w ; 1199 ObjectWaiter * u = NULL ; 1200 while (t != NULL) { 1201 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1202 t->TState = ObjectWaiter::TS_ENTER ; 1203 u = t->_next ; 1204 t->_prev = u ; 1205 t->_next = s ; 1206 s = t; 1207 t = u ; 1208 } 1209 _EntryList = s ; 1210 assert (s != NULL, "invariant") ; 1211 } else { 1212 // QMode == 0 or QMode == 2 1213 _EntryList = w ; 1214 ObjectWaiter * q = NULL ; 1215 ObjectWaiter * p ; 1216 for (p = w ; p != NULL ; p = p->_next) { 1217 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1218 p->TState = ObjectWaiter::TS_ENTER ; 1219 p->_prev = q ; 1220 q = p ; 1221 } 1222 } 1223 1224 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1225 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1226 1227 // See if we can abdicate to a spinner instead of waking a thread. 1228 // A primary goal of the implementation is to reduce the 1229 // context-switch rate. 1230 if (_succ != NULL) continue; 1231 1232 w = _EntryList ; 1233 if (w != NULL) { 1234 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1235 ExitEpilog (Self, w) ; 1236 return ; 1237 } 1238 } 1239 } 1240 1241 // ExitSuspendEquivalent: 1242 // A faster alternate to handle_special_suspend_equivalent_condition() 1243 // 1244 // handle_special_suspend_equivalent_condition() unconditionally 1245 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1246 // operations have high latency. Note that in ::enter() we call HSSEC 1247 // while holding the monitor, so we effectively lengthen the critical sections. 1248 // 1249 // There are a number of possible solutions: 1250 // 1251 // A. To ameliorate the problem we might also defer state transitions 1252 // to as late as possible -- just prior to parking. 1253 // Given that, we'd call HSSEC after having returned from park(), 1254 // but before attempting to acquire the monitor. This is only a 1255 // partial solution. It avoids calling HSSEC while holding the 1256 // monitor (good), but it still increases successor reacquisition latency -- 1257 // the interval between unparking a successor and the time the successor 1258 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1259 // If we use this technique we can also avoid EnterI()-exit() loop 1260 // in ::enter() where we iteratively drop the lock and then attempt 1261 // to reacquire it after suspending. 1262 // 1263 // B. In the future we might fold all the suspend bits into a 1264 // composite per-thread suspend flag and then update it with CAS(). 1265 // Alternately, a Dekker-like mechanism with multiple variables 1266 // would suffice: 1267 // ST Self->_suspend_equivalent = false 1268 // MEMBAR 1269 // LD Self_>_suspend_flags 1270 // 1271 1272 1273 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1274 int Mode = Knob_FastHSSEC ; 1275 if (Mode && !jSelf->is_external_suspend()) { 1276 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1277 jSelf->clear_suspend_equivalent() ; 1278 if (2 == Mode) OrderAccess::storeload() ; 1279 if (!jSelf->is_external_suspend()) return false ; 1280 // We raced a suspension -- fall thru into the slow path 1281 TEVENT (ExitSuspendEquivalent - raced) ; 1282 jSelf->set_suspend_equivalent() ; 1283 } 1284 return jSelf->handle_special_suspend_equivalent_condition() ; 1285 } 1286 1287 1288 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1289 assert (_owner == Self, "invariant") ; 1290 1291 // Exit protocol: 1292 // 1. ST _succ = wakee 1293 // 2. membar #loadstore|#storestore; 1294 // 2. ST _owner = NULL 1295 // 3. unpark(wakee) 1296 1297 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1298 ParkEvent * Trigger = Wakee->_event ; 1299 1300 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1301 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1302 // out-of-scope (non-extant). 1303 Wakee = NULL ; 1304 1305 // Drop the lock 1306 OrderAccess::release_store_ptr (&_owner, NULL) ; 1307 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1308 1309 if (SafepointSynchronize::do_call_back()) { 1310 TEVENT (unpark before SAFEPOINT) ; 1311 } 1312 1313 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1314 Trigger->unpark() ; 1315 1316 // Maintain stats and report events to JVMTI 1317 if (ObjectMonitor::_sync_Parks != NULL) { 1318 ObjectMonitor::_sync_Parks->inc() ; 1319 } 1320 } 1321 1322 1323 // ----------------------------------------------------------------------------- 1324 // Class Loader deadlock handling. 1325 // 1326 // complete_exit exits a lock returning recursion count 1327 // complete_exit/reenter operate as a wait without waiting 1328 // complete_exit requires an inflated monitor 1329 // The _owner field is not always the Thread addr even with an 1330 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1331 // thread due to contention. 1332 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1333 Thread * const Self = THREAD; 1334 assert(Self->is_Java_thread(), "Must be Java thread!"); 1335 JavaThread *jt = (JavaThread *)THREAD; 1336 1337 DeferredInitialize(); 1338 1339 if (THREAD != _owner) { 1340 if (THREAD->is_lock_owned ((address)_owner)) { 1341 assert(_recursions == 0, "internal state error"); 1342 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1343 _recursions = 0 ; 1344 OwnerIsThread = 1 ; 1345 } 1346 } 1347 1348 guarantee(Self == _owner, "complete_exit not owner"); 1349 intptr_t save = _recursions; // record the old recursion count 1350 _recursions = 0; // set the recursion level to be 0 1351 exit (Self) ; // exit the monitor 1352 guarantee (_owner != Self, "invariant"); 1353 return save; 1354 } 1355 1356 // reenter() enters a lock and sets recursion count 1357 // complete_exit/reenter operate as a wait without waiting 1358 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1359 Thread * const Self = THREAD; 1360 assert(Self->is_Java_thread(), "Must be Java thread!"); 1361 JavaThread *jt = (JavaThread *)THREAD; 1362 1363 guarantee(_owner != Self, "reenter already owner"); 1364 enter (THREAD); // enter the monitor 1365 guarantee (_recursions == 0, "reenter recursion"); 1366 _recursions = recursions; 1367 return; 1368 } 1369 1370 1371 // ----------------------------------------------------------------------------- 1372 // A macro is used below because there may already be a pending 1373 // exception which should not abort the execution of the routines 1374 // which use this (which is why we don't put this into check_slow and 1375 // call it with a CHECK argument). 1376 1377 #define CHECK_OWNER() \ 1378 do { \ 1379 if (THREAD != _owner) { \ 1380 if (THREAD->is_lock_owned((address) _owner)) { \ 1381 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1382 _recursions = 0; \ 1383 OwnerIsThread = 1 ; \ 1384 } else { \ 1385 TEVENT (Throw IMSX) ; \ 1386 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1387 } \ 1388 } \ 1389 } while (false) 1390 1391 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1392 // TODO-FIXME: remove check_slow() -- it's likely dead. 1393 1394 void ObjectMonitor::check_slow(TRAPS) { 1395 TEVENT (check_slow - throw IMSX) ; 1396 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1397 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1398 } 1399 1400 static int Adjust (volatile int * adr, int dx) { 1401 int v ; 1402 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1403 return v ; 1404 } 1405 // ----------------------------------------------------------------------------- 1406 // Wait/Notify/NotifyAll 1407 // 1408 // Note: a subset of changes to ObjectMonitor::wait() 1409 // will need to be replicated in complete_exit above 1410 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1411 Thread * const Self = THREAD ; 1412 assert(Self->is_Java_thread(), "Must be Java thread!"); 1413 JavaThread *jt = (JavaThread *)THREAD; 1414 1415 DeferredInitialize () ; 1416 1417 // Throw IMSX or IEX. 1418 CHECK_OWNER(); 1419 1420 // check for a pending interrupt 1421 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1422 // post monitor waited event. Note that this is past-tense, we are done waiting. 1423 if (JvmtiExport::should_post_monitor_waited()) { 1424 // Note: 'false' parameter is passed here because the 1425 // wait was not timed out due to thread interrupt. 1426 JvmtiExport::post_monitor_waited(jt, this, false); 1427 } 1428 TEVENT (Wait - Throw IEX) ; 1429 THROW(vmSymbols::java_lang_InterruptedException()); 1430 return ; 1431 } 1432 TEVENT (Wait) ; 1433 1434 assert (Self->_Stalled == 0, "invariant") ; 1435 Self->_Stalled = intptr_t(this) ; 1436 jt->set_current_waiting_monitor(this); 1437 1438 // create a node to be put into the queue 1439 // Critically, after we reset() the event but prior to park(), we must check 1440 // for a pending interrupt. 1441 ObjectWaiter node(Self); 1442 node.TState = ObjectWaiter::TS_WAIT ; 1443 Self->_ParkEvent->reset() ; 1444 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1445 1446 // Enter the waiting queue, which is a circular doubly linked list in this case 1447 // but it could be a priority queue or any data structure. 1448 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1449 // by the the owner of the monitor *except* in the case where park() 1450 // returns because of a timeout of interrupt. Contention is exceptionally rare 1451 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1452 1453 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1454 AddWaiter (&node) ; 1455 Thread::SpinRelease (&_WaitSetLock) ; 1456 1457 if ((SyncFlags & 4) == 0) { 1458 _Responsible = NULL ; 1459 } 1460 intptr_t save = _recursions; // record the old recursion count 1461 _waiters++; // increment the number of waiters 1462 _recursions = 0; // set the recursion level to be 1 1463 exit (Self) ; // exit the monitor 1464 guarantee (_owner != Self, "invariant") ; 1465 1466 // As soon as the ObjectMonitor's ownership is dropped in the exit() 1467 // call above, another thread can enter() the ObjectMonitor, do the 1468 // notify(), and exit() the ObjectMonitor. If the other thread's 1469 // exit() call chooses this thread as the successor and the unpark() 1470 // call happens to occur while this thread is posting a 1471 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event 1472 // handler using RawMonitors and consuming the unpark(). 1473 // 1474 // To avoid the problem, we re-post the event. This does no harm 1475 // even if the original unpark() was not consumed because we are the 1476 // chosen successor for this monitor. 1477 if (node._notified != 0 && _succ == Self) { 1478 node._event->unpark(); 1479 } 1480 1481 // The thread is on the WaitSet list - now park() it. 1482 // On MP systems it's conceivable that a brief spin before we park 1483 // could be profitable. 1484 // 1485 // TODO-FIXME: change the following logic to a loop of the form 1486 // while (!timeout && !interrupted && _notified == 0) park() 1487 1488 int ret = OS_OK ; 1489 int WasNotified = 0 ; 1490 { // State transition wrappers 1491 OSThread* osthread = Self->osthread(); 1492 OSThreadWaitState osts(osthread, true); 1493 { 1494 ThreadBlockInVM tbivm(jt); 1495 // Thread is in thread_blocked state and oop access is unsafe. 1496 jt->set_suspend_equivalent(); 1497 1498 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1499 // Intentionally empty 1500 } else 1501 if (node._notified == 0) { 1502 if (millis <= 0) { 1503 Self->_ParkEvent->park () ; 1504 } else { 1505 ret = Self->_ParkEvent->park (millis) ; 1506 } 1507 } 1508 1509 // were we externally suspended while we were waiting? 1510 if (ExitSuspendEquivalent (jt)) { 1511 // TODO-FIXME: add -- if succ == Self then succ = null. 1512 jt->java_suspend_self(); 1513 } 1514 1515 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1516 1517 1518 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1519 // from the WaitSet to the EntryList. 1520 // See if we need to remove Node from the WaitSet. 1521 // We use double-checked locking to avoid grabbing _WaitSetLock 1522 // if the thread is not on the wait queue. 1523 // 1524 // Note that we don't need a fence before the fetch of TState. 1525 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1526 // written by the is thread. (perhaps the fetch might even be satisfied 1527 // by a look-aside into the processor's own store buffer, although given 1528 // the length of the code path between the prior ST and this load that's 1529 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1530 // then we'll acquire the lock and then re-fetch a fresh TState value. 1531 // That is, we fail toward safety. 1532 1533 if (node.TState == ObjectWaiter::TS_WAIT) { 1534 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1535 if (node.TState == ObjectWaiter::TS_WAIT) { 1536 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1537 assert(node._notified == 0, "invariant"); 1538 node.TState = ObjectWaiter::TS_RUN ; 1539 } 1540 Thread::SpinRelease (&_WaitSetLock) ; 1541 } 1542 1543 // The thread is now either on off-list (TS_RUN), 1544 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1545 // The Node's TState variable is stable from the perspective of this thread. 1546 // No other threads will asynchronously modify TState. 1547 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1548 OrderAccess::loadload() ; 1549 if (_succ == Self) _succ = NULL ; 1550 WasNotified = node._notified ; 1551 1552 // Reentry phase -- reacquire the monitor. 1553 // re-enter contended monitor after object.wait(). 1554 // retain OBJECT_WAIT state until re-enter successfully completes 1555 // Thread state is thread_in_vm and oop access is again safe, 1556 // although the raw address of the object may have changed. 1557 // (Don't cache naked oops over safepoints, of course). 1558 1559 // post monitor waited event. Note that this is past-tense, we are done waiting. 1560 if (JvmtiExport::should_post_monitor_waited()) { 1561 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1562 } 1563 OrderAccess::fence() ; 1564 1565 assert (Self->_Stalled != 0, "invariant") ; 1566 Self->_Stalled = 0 ; 1567 1568 assert (_owner != Self, "invariant") ; 1569 ObjectWaiter::TStates v = node.TState ; 1570 if (v == ObjectWaiter::TS_RUN) { 1571 enter (Self) ; 1572 } else { 1573 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1574 ReenterI (Self, &node) ; 1575 node.wait_reenter_end(this); 1576 } 1577 1578 // Self has reacquired the lock. 1579 // Lifecycle - the node representing Self must not appear on any queues. 1580 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1581 // want residual elements associated with this thread left on any lists. 1582 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1583 assert (_owner == Self, "invariant") ; 1584 assert (_succ != Self , "invariant") ; 1585 } // OSThreadWaitState() 1586 1587 jt->set_current_waiting_monitor(NULL); 1588 1589 guarantee (_recursions == 0, "invariant") ; 1590 _recursions = save; // restore the old recursion count 1591 _waiters--; // decrement the number of waiters 1592 1593 // Verify a few postconditions 1594 assert (_owner == Self , "invariant") ; 1595 assert (_succ != Self , "invariant") ; 1596 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1597 1598 if (SyncFlags & 32) { 1599 OrderAccess::fence() ; 1600 } 1601 1602 // check if the notification happened 1603 if (!WasNotified) { 1604 // no, it could be timeout or Thread.interrupt() or both 1605 // check for interrupt event, otherwise it is timeout 1606 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1607 TEVENT (Wait - throw IEX from epilog) ; 1608 THROW(vmSymbols::java_lang_InterruptedException()); 1609 } 1610 } 1611 1612 // NOTE: Spurious wake up will be consider as timeout. 1613 // Monitor notify has precedence over thread interrupt. 1614 } 1615 1616 1617 // Consider: 1618 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1619 // then instead of transferring a thread from the WaitSet to the EntryList 1620 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1621 1622 void ObjectMonitor::notify(TRAPS) { 1623 CHECK_OWNER(); 1624 if (_WaitSet == NULL) { 1625 TEVENT (Empty-Notify) ; 1626 return ; 1627 } 1628 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1629 1630 int Policy = Knob_MoveNotifyee ; 1631 1632 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1633 ObjectWaiter * iterator = DequeueWaiter() ; 1634 if (iterator != NULL) { 1635 TEVENT (Notify1 - Transfer) ; 1636 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1637 guarantee (iterator->_notified == 0, "invariant") ; 1638 if (Policy != 4) { 1639 iterator->TState = ObjectWaiter::TS_ENTER ; 1640 } 1641 iterator->_notified = 1 ; 1642 1643 ObjectWaiter * List = _EntryList ; 1644 if (List != NULL) { 1645 assert (List->_prev == NULL, "invariant") ; 1646 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1647 assert (List != iterator, "invariant") ; 1648 } 1649 1650 if (Policy == 0) { // prepend to EntryList 1651 if (List == NULL) { 1652 iterator->_next = iterator->_prev = NULL ; 1653 _EntryList = iterator ; 1654 } else { 1655 List->_prev = iterator ; 1656 iterator->_next = List ; 1657 iterator->_prev = NULL ; 1658 _EntryList = iterator ; 1659 } 1660 } else 1661 if (Policy == 1) { // append to EntryList 1662 if (List == NULL) { 1663 iterator->_next = iterator->_prev = NULL ; 1664 _EntryList = iterator ; 1665 } else { 1666 // CONSIDER: finding the tail currently requires a linear-time walk of 1667 // the EntryList. We can make tail access constant-time by converting to 1668 // a CDLL instead of using our current DLL. 1669 ObjectWaiter * Tail ; 1670 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1671 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1672 Tail->_next = iterator ; 1673 iterator->_prev = Tail ; 1674 iterator->_next = NULL ; 1675 } 1676 } else 1677 if (Policy == 2) { // prepend to cxq 1678 // prepend to cxq 1679 if (List == NULL) { 1680 iterator->_next = iterator->_prev = NULL ; 1681 _EntryList = iterator ; 1682 } else { 1683 iterator->TState = ObjectWaiter::TS_CXQ ; 1684 for (;;) { 1685 ObjectWaiter * Front = _cxq ; 1686 iterator->_next = Front ; 1687 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1688 break ; 1689 } 1690 } 1691 } 1692 } else 1693 if (Policy == 3) { // append to cxq 1694 iterator->TState = ObjectWaiter::TS_CXQ ; 1695 for (;;) { 1696 ObjectWaiter * Tail ; 1697 Tail = _cxq ; 1698 if (Tail == NULL) { 1699 iterator->_next = NULL ; 1700 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1701 break ; 1702 } 1703 } else { 1704 while (Tail->_next != NULL) Tail = Tail->_next ; 1705 Tail->_next = iterator ; 1706 iterator->_prev = Tail ; 1707 iterator->_next = NULL ; 1708 break ; 1709 } 1710 } 1711 } else { 1712 ParkEvent * ev = iterator->_event ; 1713 iterator->TState = ObjectWaiter::TS_RUN ; 1714 OrderAccess::fence() ; 1715 ev->unpark() ; 1716 } 1717 1718 if (Policy < 4) { 1719 iterator->wait_reenter_begin(this); 1720 } 1721 1722 // _WaitSetLock protects the wait queue, not the EntryList. We could 1723 // move the add-to-EntryList operation, above, outside the critical section 1724 // protected by _WaitSetLock. In practice that's not useful. With the 1725 // exception of wait() timeouts and interrupts the monitor owner 1726 // is the only thread that grabs _WaitSetLock. There's almost no contention 1727 // on _WaitSetLock so it's not profitable to reduce the length of the 1728 // critical section. 1729 } 1730 1731 Thread::SpinRelease (&_WaitSetLock) ; 1732 1733 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1734 ObjectMonitor::_sync_Notifications->inc() ; 1735 } 1736 } 1737 1738 1739 void ObjectMonitor::notifyAll(TRAPS) { 1740 CHECK_OWNER(); 1741 ObjectWaiter* iterator; 1742 if (_WaitSet == NULL) { 1743 TEVENT (Empty-NotifyAll) ; 1744 return ; 1745 } 1746 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1747 1748 int Policy = Knob_MoveNotifyee ; 1749 int Tally = 0 ; 1750 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1751 1752 for (;;) { 1753 iterator = DequeueWaiter () ; 1754 if (iterator == NULL) break ; 1755 TEVENT (NotifyAll - Transfer1) ; 1756 ++Tally ; 1757 1758 // Disposition - what might we do with iterator ? 1759 // a. add it directly to the EntryList - either tail or head. 1760 // b. push it onto the front of the _cxq. 1761 // For now we use (a). 1762 1763 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1764 guarantee (iterator->_notified == 0, "invariant") ; 1765 iterator->_notified = 1 ; 1766 if (Policy != 4) { 1767 iterator->TState = ObjectWaiter::TS_ENTER ; 1768 } 1769 1770 ObjectWaiter * List = _EntryList ; 1771 if (List != NULL) { 1772 assert (List->_prev == NULL, "invariant") ; 1773 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1774 assert (List != iterator, "invariant") ; 1775 } 1776 1777 if (Policy == 0) { // prepend to EntryList 1778 if (List == NULL) { 1779 iterator->_next = iterator->_prev = NULL ; 1780 _EntryList = iterator ; 1781 } else { 1782 List->_prev = iterator ; 1783 iterator->_next = List ; 1784 iterator->_prev = NULL ; 1785 _EntryList = iterator ; 1786 } 1787 } else 1788 if (Policy == 1) { // append to EntryList 1789 if (List == NULL) { 1790 iterator->_next = iterator->_prev = NULL ; 1791 _EntryList = iterator ; 1792 } else { 1793 // CONSIDER: finding the tail currently requires a linear-time walk of 1794 // the EntryList. We can make tail access constant-time by converting to 1795 // a CDLL instead of using our current DLL. 1796 ObjectWaiter * Tail ; 1797 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1798 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1799 Tail->_next = iterator ; 1800 iterator->_prev = Tail ; 1801 iterator->_next = NULL ; 1802 } 1803 } else 1804 if (Policy == 2) { // prepend to cxq 1805 // prepend to cxq 1806 iterator->TState = ObjectWaiter::TS_CXQ ; 1807 for (;;) { 1808 ObjectWaiter * Front = _cxq ; 1809 iterator->_next = Front ; 1810 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1811 break ; 1812 } 1813 } 1814 } else 1815 if (Policy == 3) { // append to cxq 1816 iterator->TState = ObjectWaiter::TS_CXQ ; 1817 for (;;) { 1818 ObjectWaiter * Tail ; 1819 Tail = _cxq ; 1820 if (Tail == NULL) { 1821 iterator->_next = NULL ; 1822 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1823 break ; 1824 } 1825 } else { 1826 while (Tail->_next != NULL) Tail = Tail->_next ; 1827 Tail->_next = iterator ; 1828 iterator->_prev = Tail ; 1829 iterator->_next = NULL ; 1830 break ; 1831 } 1832 } 1833 } else { 1834 ParkEvent * ev = iterator->_event ; 1835 iterator->TState = ObjectWaiter::TS_RUN ; 1836 OrderAccess::fence() ; 1837 ev->unpark() ; 1838 } 1839 1840 if (Policy < 4) { 1841 iterator->wait_reenter_begin(this); 1842 } 1843 1844 // _WaitSetLock protects the wait queue, not the EntryList. We could 1845 // move the add-to-EntryList operation, above, outside the critical section 1846 // protected by _WaitSetLock. In practice that's not useful. With the 1847 // exception of wait() timeouts and interrupts the monitor owner 1848 // is the only thread that grabs _WaitSetLock. There's almost no contention 1849 // on _WaitSetLock so it's not profitable to reduce the length of the 1850 // critical section. 1851 } 1852 1853 Thread::SpinRelease (&_WaitSetLock) ; 1854 1855 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1856 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1857 } 1858 } 1859 1860 // ----------------------------------------------------------------------------- 1861 // Adaptive Spinning Support 1862 // 1863 // Adaptive spin-then-block - rational spinning 1864 // 1865 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1866 // algorithm. On high order SMP systems it would be better to start with 1867 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1868 // a contending thread could enqueue itself on the cxq and then spin locally 1869 // on a thread-specific variable such as its ParkEvent._Event flag. 1870 // That's left as an exercise for the reader. Note that global spinning is 1871 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1872 // low latency and massive bandwidth. 1873 // 1874 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1875 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1876 // (duration) or we can fix the count at approximately the duration of 1877 // a context switch and vary the frequency. Of course we could also 1878 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1879 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1880 // 1881 // This implementation varies the duration "D", where D varies with 1882 // the success rate of recent spin attempts. (D is capped at approximately 1883 // length of a round-trip context switch). The success rate for recent 1884 // spin attempts is a good predictor of the success rate of future spin 1885 // attempts. The mechanism adapts automatically to varying critical 1886 // section length (lock modality), system load and degree of parallelism. 1887 // D is maintained per-monitor in _SpinDuration and is initialized 1888 // optimistically. Spin frequency is fixed at 100%. 1889 // 1890 // Note that _SpinDuration is volatile, but we update it without locks 1891 // or atomics. The code is designed so that _SpinDuration stays within 1892 // a reasonable range even in the presence of races. The arithmetic 1893 // operations on _SpinDuration are closed over the domain of legal values, 1894 // so at worst a race will install and older but still legal value. 1895 // At the very worst this introduces some apparent non-determinism. 1896 // We might spin when we shouldn't or vice-versa, but since the spin 1897 // count are relatively short, even in the worst case, the effect is harmless. 1898 // 1899 // Care must be taken that a low "D" value does not become an 1900 // an absorbing state. Transient spinning failures -- when spinning 1901 // is overall profitable -- should not cause the system to converge 1902 // on low "D" values. We want spinning to be stable and predictable 1903 // and fairly responsive to change and at the same time we don't want 1904 // it to oscillate, become metastable, be "too" non-deterministic, 1905 // or converge on or enter undesirable stable absorbing states. 1906 // 1907 // We implement a feedback-based control system -- using past behavior 1908 // to predict future behavior. We face two issues: (a) if the 1909 // input signal is random then the spin predictor won't provide optimal 1910 // results, and (b) if the signal frequency is too high then the control 1911 // system, which has some natural response lag, will "chase" the signal. 1912 // (b) can arise from multimodal lock hold times. Transient preemption 1913 // can also result in apparent bimodal lock hold times. 1914 // Although sub-optimal, neither condition is particularly harmful, as 1915 // in the worst-case we'll spin when we shouldn't or vice-versa. 1916 // The maximum spin duration is rather short so the failure modes aren't bad. 1917 // To be conservative, I've tuned the gain in system to bias toward 1918 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1919 // "rings" or oscillates between spinning and not spinning. This happens 1920 // when spinning is just on the cusp of profitability, however, so the 1921 // situation is not dire. The state is benign -- there's no need to add 1922 // hysteresis control to damp the transition rate between spinning and 1923 // not spinning. 1924 // 1925 1926 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1927 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1928 1929 // Spinning: Fixed frequency (100%), vary duration 1930 1931 1932 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1933 1934 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1935 int ctr = Knob_FixedSpin ; 1936 if (ctr != 0) { 1937 while (--ctr >= 0) { 1938 if (TryLock (Self) > 0) return 1 ; 1939 SpinPause () ; 1940 } 1941 return 0 ; 1942 } 1943 1944 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 1945 if (TryLock(Self) > 0) { 1946 // Increase _SpinDuration ... 1947 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1948 // Raising _SpurDuration to the poverty line is key. 1949 int x = _SpinDuration ; 1950 if (x < Knob_SpinLimit) { 1951 if (x < Knob_Poverty) x = Knob_Poverty ; 1952 _SpinDuration = x + Knob_BonusB ; 1953 } 1954 return 1 ; 1955 } 1956 SpinPause () ; 1957 } 1958 1959 // Admission control - verify preconditions for spinning 1960 // 1961 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1962 // becoming an absorbing state. Put another way, we spin briefly to 1963 // sample, just in case the system load, parallelism, contention, or lock 1964 // modality changed. 1965 // 1966 // Consider the following alternative: 1967 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1968 // spin attempt. "Periodically" might mean after a tally of 1969 // the # of failed spin attempts (or iterations) reaches some threshold. 1970 // This takes us into the realm of 1-out-of-N spinning, where we 1971 // hold the duration constant but vary the frequency. 1972 1973 ctr = _SpinDuration ; 1974 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 1975 if (ctr <= 0) return 0 ; 1976 1977 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 1978 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 1979 TEVENT (Spin abort - notrunnable [TOP]); 1980 return 0 ; 1981 } 1982 1983 int MaxSpin = Knob_MaxSpinners ; 1984 if (MaxSpin >= 0) { 1985 if (_Spinner > MaxSpin) { 1986 TEVENT (Spin abort -- too many spinners) ; 1987 return 0 ; 1988 } 1989 // Slighty racy, but benign ... 1990 Adjust (&_Spinner, 1) ; 1991 } 1992 1993 // We're good to spin ... spin ingress. 1994 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1995 // when preparing to LD...CAS _owner, etc and the CAS is likely 1996 // to succeed. 1997 int hits = 0 ; 1998 int msk = 0 ; 1999 int caspty = Knob_CASPenalty ; 2000 int oxpty = Knob_OXPenalty ; 2001 int sss = Knob_SpinSetSucc ; 2002 if (sss && _succ == NULL ) _succ = Self ; 2003 Thread * prv = NULL ; 2004 2005 // There are three ways to exit the following loop: 2006 // 1. A successful spin where this thread has acquired the lock. 2007 // 2. Spin failure with prejudice 2008 // 3. Spin failure without prejudice 2009 2010 while (--ctr >= 0) { 2011 2012 // Periodic polling -- Check for pending GC 2013 // Threads may spin while they're unsafe. 2014 // We don't want spinning threads to delay the JVM from reaching 2015 // a stop-the-world safepoint or to steal cycles from GC. 2016 // If we detect a pending safepoint we abort in order that 2017 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2018 // this thread, if safe, doesn't steal cycles from GC. 2019 // This is in keeping with the "no loitering in runtime" rule. 2020 // We periodically check to see if there's a safepoint pending. 2021 if ((ctr & 0xFF) == 0) { 2022 if (SafepointSynchronize::do_call_back()) { 2023 TEVENT (Spin: safepoint) ; 2024 goto Abort ; // abrupt spin egress 2025 } 2026 if (Knob_UsePause & 1) SpinPause () ; 2027 2028 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2029 if (hits > 50 && scb != NULL) { 2030 int abend = (*scb)(SpinCallbackArgument, 0) ; 2031 } 2032 } 2033 2034 if (Knob_UsePause & 2) SpinPause() ; 2035 2036 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2037 // This is useful on classic SMP systems, but is of less utility on 2038 // N1-style CMT platforms. 2039 // 2040 // Trade-off: lock acquisition latency vs coherency bandwidth. 2041 // Lock hold times are typically short. A histogram 2042 // of successful spin attempts shows that we usually acquire 2043 // the lock early in the spin. That suggests we want to 2044 // sample _owner frequently in the early phase of the spin, 2045 // but then back-off and sample less frequently as the spin 2046 // progresses. The back-off makes a good citizen on SMP big 2047 // SMP systems. Oversampling _owner can consume excessive 2048 // coherency bandwidth. Relatedly, if we _oversample _owner we 2049 // can inadvertently interfere with the the ST m->owner=null. 2050 // executed by the lock owner. 2051 if (ctr & msk) continue ; 2052 ++hits ; 2053 if ((hits & 0xF) == 0) { 2054 // The 0xF, above, corresponds to the exponent. 2055 // Consider: (msk+1)|msk 2056 msk = ((msk << 2)|3) & BackOffMask ; 2057 } 2058 2059 // Probe _owner with TATAS 2060 // If this thread observes the monitor transition or flicker 2061 // from locked to unlocked to locked, then the odds that this 2062 // thread will acquire the lock in this spin attempt go down 2063 // considerably. The same argument applies if the CAS fails 2064 // or if we observe _owner change from one non-null value to 2065 // another non-null value. In such cases we might abort 2066 // the spin without prejudice or apply a "penalty" to the 2067 // spin count-down variable "ctr", reducing it by 100, say. 2068 2069 Thread * ox = (Thread *) _owner ; 2070 if (ox == NULL) { 2071 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2072 if (ox == NULL) { 2073 // The CAS succeeded -- this thread acquired ownership 2074 // Take care of some bookkeeping to exit spin state. 2075 if (sss && _succ == Self) { 2076 _succ = NULL ; 2077 } 2078 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2079 2080 // Increase _SpinDuration : 2081 // The spin was successful (profitable) so we tend toward 2082 // longer spin attempts in the future. 2083 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2084 // If we acquired the lock early in the spin cycle it 2085 // makes sense to increase _SpinDuration proportionally. 2086 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2087 int x = _SpinDuration ; 2088 if (x < Knob_SpinLimit) { 2089 if (x < Knob_Poverty) x = Knob_Poverty ; 2090 _SpinDuration = x + Knob_Bonus ; 2091 } 2092 return 1 ; 2093 } 2094 2095 // The CAS failed ... we can take any of the following actions: 2096 // * penalize: ctr -= Knob_CASPenalty 2097 // * exit spin with prejudice -- goto Abort; 2098 // * exit spin without prejudice. 2099 // * Since CAS is high-latency, retry again immediately. 2100 prv = ox ; 2101 TEVENT (Spin: cas failed) ; 2102 if (caspty == -2) break ; 2103 if (caspty == -1) goto Abort ; 2104 ctr -= caspty ; 2105 continue ; 2106 } 2107 2108 // Did lock ownership change hands ? 2109 if (ox != prv && prv != NULL ) { 2110 TEVENT (spin: Owner changed) 2111 if (oxpty == -2) break ; 2112 if (oxpty == -1) goto Abort ; 2113 ctr -= oxpty ; 2114 } 2115 prv = ox ; 2116 2117 // Abort the spin if the owner is not executing. 2118 // The owner must be executing in order to drop the lock. 2119 // Spinning while the owner is OFFPROC is idiocy. 2120 // Consider: ctr -= RunnablePenalty ; 2121 if (Knob_OState && NotRunnable (Self, ox)) { 2122 TEVENT (Spin abort - notrunnable); 2123 goto Abort ; 2124 } 2125 if (sss && _succ == NULL ) _succ = Self ; 2126 } 2127 2128 // Spin failed with prejudice -- reduce _SpinDuration. 2129 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2130 // AIMD is globally stable. 2131 TEVENT (Spin failure) ; 2132 { 2133 int x = _SpinDuration ; 2134 if (x > 0) { 2135 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2136 // This is globally sample and tends to damp the response. 2137 x -= Knob_Penalty ; 2138 if (x < 0) x = 0 ; 2139 _SpinDuration = x ; 2140 } 2141 } 2142 2143 Abort: 2144 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2145 if (sss && _succ == Self) { 2146 _succ = NULL ; 2147 // Invariant: after setting succ=null a contending thread 2148 // must recheck-retry _owner before parking. This usually happens 2149 // in the normal usage of TrySpin(), but it's safest 2150 // to make TrySpin() as foolproof as possible. 2151 OrderAccess::fence() ; 2152 if (TryLock(Self) > 0) return 1 ; 2153 } 2154 return 0 ; 2155 } 2156 2157 // NotRunnable() -- informed spinning 2158 // 2159 // Don't bother spinning if the owner is not eligible to drop the lock. 2160 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2161 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2162 // The thread must be runnable in order to drop the lock in timely fashion. 2163 // If the _owner is not runnable then spinning will not likely be 2164 // successful (profitable). 2165 // 2166 // Beware -- the thread referenced by _owner could have died 2167 // so a simply fetch from _owner->_thread_state might trap. 2168 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2169 // Because of the lifecycle issues the schedctl and _thread_state values 2170 // observed by NotRunnable() might be garbage. NotRunnable must 2171 // tolerate this and consider the observed _thread_state value 2172 // as advisory. 2173 // 2174 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2175 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2176 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2177 // with the LSB of _owner. Another option would be to probablistically probe 2178 // the putative _owner->TypeTag value. 2179 // 2180 // Checking _thread_state isn't perfect. Even if the thread is 2181 // in_java it might be blocked on a page-fault or have been preempted 2182 // and sitting on a ready/dispatch queue. _thread state in conjunction 2183 // with schedctl.sc_state gives us a good picture of what the 2184 // thread is doing, however. 2185 // 2186 // TODO: check schedctl.sc_state. 2187 // We'll need to use SafeFetch32() to read from the schedctl block. 2188 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2189 // 2190 // The return value from NotRunnable() is *advisory* -- the 2191 // result is based on sampling and is not necessarily coherent. 2192 // The caller must tolerate false-negative and false-positive errors. 2193 // Spinning, in general, is probabilistic anyway. 2194 2195 2196 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2197 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2198 if (!OwnerIsThread) return 0 ; 2199 2200 if (ox == NULL) return 0 ; 2201 2202 // Avoid transitive spinning ... 2203 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2204 // Immediately after T1 acquires L it's possible that T2, also 2205 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2206 // This occurs transiently after T1 acquired L but before 2207 // T1 managed to clear T1.Stalled. T2 does not need to abort 2208 // its spin in this circumstance. 2209 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2210 2211 if (BlockedOn == 1) return 1 ; 2212 if (BlockedOn != 0) { 2213 return BlockedOn != intptr_t(this) && _owner == ox ; 2214 } 2215 2216 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2217 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2218 // consider also: jst != _thread_in_Java -- but that's overspecific. 2219 return jst == _thread_blocked || jst == _thread_in_native ; 2220 } 2221 2222 2223 // ----------------------------------------------------------------------------- 2224 // WaitSet management ... 2225 2226 ObjectWaiter::ObjectWaiter(Thread* thread) { 2227 _next = NULL; 2228 _prev = NULL; 2229 _notified = 0; 2230 TState = TS_RUN ; 2231 _thread = thread; 2232 _event = thread->_ParkEvent ; 2233 _active = false; 2234 assert (_event != NULL, "invariant") ; 2235 } 2236 2237 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2238 JavaThread *jt = (JavaThread *)this->_thread; 2239 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2240 } 2241 2242 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2243 JavaThread *jt = (JavaThread *)this->_thread; 2244 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2245 } 2246 2247 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2248 assert(node != NULL, "should not dequeue NULL node"); 2249 assert(node->_prev == NULL, "node already in list"); 2250 assert(node->_next == NULL, "node already in list"); 2251 // put node at end of queue (circular doubly linked list) 2252 if (_WaitSet == NULL) { 2253 _WaitSet = node; 2254 node->_prev = node; 2255 node->_next = node; 2256 } else { 2257 ObjectWaiter* head = _WaitSet ; 2258 ObjectWaiter* tail = head->_prev; 2259 assert(tail->_next == head, "invariant check"); 2260 tail->_next = node; 2261 head->_prev = node; 2262 node->_next = head; 2263 node->_prev = tail; 2264 } 2265 } 2266 2267 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2268 // dequeue the very first waiter 2269 ObjectWaiter* waiter = _WaitSet; 2270 if (waiter) { 2271 DequeueSpecificWaiter(waiter); 2272 } 2273 return waiter; 2274 } 2275 2276 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2277 assert(node != NULL, "should not dequeue NULL node"); 2278 assert(node->_prev != NULL, "node already removed from list"); 2279 assert(node->_next != NULL, "node already removed from list"); 2280 // when the waiter has woken up because of interrupt, 2281 // timeout or other spurious wake-up, dequeue the 2282 // waiter from waiting list 2283 ObjectWaiter* next = node->_next; 2284 if (next == node) { 2285 assert(node->_prev == node, "invariant check"); 2286 _WaitSet = NULL; 2287 } else { 2288 ObjectWaiter* prev = node->_prev; 2289 assert(prev->_next == node, "invariant check"); 2290 assert(next->_prev == node, "invariant check"); 2291 next->_prev = prev; 2292 prev->_next = next; 2293 if (_WaitSet == node) { 2294 _WaitSet = next; 2295 } 2296 } 2297 node->_next = NULL; 2298 node->_prev = NULL; 2299 } 2300 2301 // ----------------------------------------------------------------------------- 2302 // PerfData support 2303 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2304 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2305 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2306 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2307 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2308 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2309 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2310 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2311 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2312 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2313 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2314 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2315 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2316 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2317 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2318 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2319 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2320 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2321 2322 // One-shot global initialization for the sync subsystem. 2323 // We could also defer initialization and initialize on-demand 2324 // the first time we call inflate(). Initialization would 2325 // be protected - like so many things - by the MonitorCache_lock. 2326 2327 void ObjectMonitor::Initialize () { 2328 static int InitializationCompleted = 0 ; 2329 assert (InitializationCompleted == 0, "invariant") ; 2330 InitializationCompleted = 1 ; 2331 if (UsePerfData) { 2332 EXCEPTION_MARK ; 2333 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2334 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2335 NEWPERFCOUNTER(_sync_Inflations) ; 2336 NEWPERFCOUNTER(_sync_Deflations) ; 2337 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2338 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2339 NEWPERFCOUNTER(_sync_Parks) ; 2340 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2341 NEWPERFCOUNTER(_sync_Notifications) ; 2342 NEWPERFCOUNTER(_sync_SlowEnter) ; 2343 NEWPERFCOUNTER(_sync_SlowExit) ; 2344 NEWPERFCOUNTER(_sync_SlowNotify) ; 2345 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2346 NEWPERFCOUNTER(_sync_FailedSpins) ; 2347 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2348 NEWPERFCOUNTER(_sync_PrivateA) ; 2349 NEWPERFCOUNTER(_sync_PrivateB) ; 2350 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2351 NEWPERFCOUNTER(_sync_MonScavenged) ; 2352 NEWPERFVARIABLE(_sync_MonExtant) ; 2353 #undef NEWPERFCOUNTER 2354 } 2355 } 2356 2357 2358 // Compile-time asserts 2359 // When possible, it's better to catch errors deterministically at 2360 // compile-time than at runtime. The down-side to using compile-time 2361 // asserts is that error message -- often something about negative array 2362 // indices -- is opaque. 2363 2364 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2365 2366 void ObjectMonitor::ctAsserts() { 2367 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2368 } 2369 2370 2371 static char * kvGet (char * kvList, const char * Key) { 2372 if (kvList == NULL) return NULL ; 2373 size_t n = strlen (Key) ; 2374 char * Search ; 2375 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2376 if (strncmp (Search, Key, n) == 0) { 2377 if (Search[n] == '=') return Search + n + 1 ; 2378 if (Search[n] == 0) return (char *) "1" ; 2379 } 2380 } 2381 return NULL ; 2382 } 2383 2384 static int kvGetInt (char * kvList, const char * Key, int Default) { 2385 char * v = kvGet (kvList, Key) ; 2386 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2387 if (Knob_ReportSettings && v != NULL) { 2388 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2389 ::fflush (stdout) ; 2390 } 2391 return rslt ; 2392 } 2393 2394 void ObjectMonitor::DeferredInitialize () { 2395 if (InitDone > 0) return ; 2396 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2397 while (InitDone != 1) ; 2398 return ; 2399 } 2400 2401 // One-shot global initialization ... 2402 // The initialization is idempotent, so we don't need locks. 2403 // In the future consider doing this via os::init_2(). 2404 // SyncKnobs consist of <Key>=<Value> pairs in the style 2405 // of environment variables. Start by converting ':' to NUL. 2406 2407 if (SyncKnobs == NULL) SyncKnobs = "" ; 2408 2409 size_t sz = strlen (SyncKnobs) ; 2410 char * knobs = (char *) malloc (sz + 2) ; 2411 if (knobs == NULL) { 2412 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; 2413 guarantee (0, "invariant") ; 2414 } 2415 strcpy (knobs, SyncKnobs) ; 2416 knobs[sz+1] = 0 ; 2417 for (char * p = knobs ; *p ; p++) { 2418 if (*p == ':') *p = 0 ; 2419 } 2420 2421 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2422 SETKNOB(ReportSettings) ; 2423 SETKNOB(Verbose) ; 2424 SETKNOB(FixedSpin) ; 2425 SETKNOB(SpinLimit) ; 2426 SETKNOB(SpinBase) ; 2427 SETKNOB(SpinBackOff); 2428 SETKNOB(CASPenalty) ; 2429 SETKNOB(OXPenalty) ; 2430 SETKNOB(LogSpins) ; 2431 SETKNOB(SpinSetSucc) ; 2432 SETKNOB(SuccEnabled) ; 2433 SETKNOB(SuccRestrict) ; 2434 SETKNOB(Penalty) ; 2435 SETKNOB(Bonus) ; 2436 SETKNOB(BonusB) ; 2437 SETKNOB(Poverty) ; 2438 SETKNOB(SpinAfterFutile) ; 2439 SETKNOB(UsePause) ; 2440 SETKNOB(SpinEarly) ; 2441 SETKNOB(OState) ; 2442 SETKNOB(MaxSpinners) ; 2443 SETKNOB(PreSpin) ; 2444 SETKNOB(ExitPolicy) ; 2445 SETKNOB(QMode); 2446 SETKNOB(ResetEvent) ; 2447 SETKNOB(MoveNotifyee) ; 2448 SETKNOB(FastHSSEC) ; 2449 #undef SETKNOB 2450 2451 if (os::is_MP()) { 2452 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2453 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2454 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2455 } else { 2456 Knob_SpinLimit = 0 ; 2457 Knob_SpinBase = 0 ; 2458 Knob_PreSpin = 0 ; 2459 Knob_FixedSpin = -1 ; 2460 } 2461 2462 if (Knob_LogSpins == 0) { 2463 ObjectMonitor::_sync_FailedSpins = NULL ; 2464 } 2465 2466 free (knobs) ; 2467 OrderAccess::fence() ; 2468 InitDone = 1 ; 2469 } 2470 2471 #ifndef PRODUCT 2472 void ObjectMonitor::verify() { 2473 } 2474 2475 void ObjectMonitor::print() { 2476 } 2477 #endif