1 /* 2 * Copyright (c) 1998, 2013, 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 "trace/tracing.hpp" 40 #include "trace/traceMacros.hpp" 41 #include "utilities/dtrace.hpp" 42 #include "utilities/macros.hpp" 43 #include "utilities/preserveException.hpp" 44 #ifdef TARGET_OS_FAMILY_linux 45 # include "os_linux.inline.hpp" 46 #endif 47 #ifdef TARGET_OS_FAMILY_solaris 48 # include "os_solaris.inline.hpp" 49 #endif 50 #ifdef TARGET_OS_FAMILY_windows 51 # include "os_windows.inline.hpp" 52 #endif 53 #ifdef TARGET_OS_FAMILY_bsd 54 # include "os_bsd.inline.hpp" 55 #endif 56 57 #if defined(__GNUC__) && !defined(IA64) 58 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 59 #define ATTR __attribute__((noinline)) 60 #else 61 #define ATTR 62 #endif 63 64 65 #ifdef DTRACE_ENABLED 66 67 // Only bother with this argument setup if dtrace is available 68 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 69 70 71 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 72 char* bytes = NULL; \ 73 int len = 0; \ 74 jlong jtid = SharedRuntime::get_java_tid(thread); \ 75 Symbol* klassname = ((oop)obj)->klass()->name(); \ 76 if (klassname != NULL) { \ 77 bytes = (char*)klassname->bytes(); \ 78 len = klassname->utf8_length(); \ 79 } 80 81 #ifndef USDT2 82 83 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 84 jlong, uintptr_t, char*, int); 85 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 86 jlong, uintptr_t, char*, int); 87 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 88 jlong, uintptr_t, char*, int); 89 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 90 jlong, uintptr_t, char*, int); 91 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 92 jlong, uintptr_t, char*, int); 93 94 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 95 { \ 96 if (DTraceMonitorProbes) { \ 97 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 98 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 99 (monitor), bytes, len, (millis)); \ 100 } \ 101 } 102 103 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 104 { \ 105 if (DTraceMonitorProbes) { \ 106 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 107 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 108 (uintptr_t)(monitor), bytes, len); \ 109 } \ 110 } 111 112 #else /* USDT2 */ 113 114 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 115 { \ 116 if (DTraceMonitorProbes) { \ 117 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 118 HOTSPOT_MONITOR_WAIT(jtid, \ 119 (monitor), bytes, len, (millis)); \ 120 } \ 121 } 122 123 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 124 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 125 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 126 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 127 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 128 129 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 130 { \ 131 if (DTraceMonitorProbes) { \ 132 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 133 HOTSPOT_MONITOR_##probe(jtid, \ 134 (uintptr_t)(monitor), bytes, len); \ 135 } \ 136 } 137 138 #endif /* USDT2 */ 139 #else // ndef DTRACE_ENABLED 140 141 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 142 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 143 144 #endif // ndef DTRACE_ENABLED 145 146 // Tunables ... 147 // The knob* variables are effectively final. Once set they should 148 // never be modified hence. Consider using __read_mostly with GCC. 149 150 int ObjectMonitor::Knob_Verbose = 0 ; 151 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 152 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 153 static int Knob_HandOff = 0 ; 154 static int Knob_ReportSettings = 0 ; 155 156 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 157 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 158 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 159 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 160 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 161 static int Knob_SpinEarly = 1 ; 162 static int Knob_SuccEnabled = 1 ; // futile wake throttling 163 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 164 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 165 static int Knob_Bonus = 100 ; // spin success bonus 166 static int Knob_BonusB = 100 ; // spin success bonus 167 static int Knob_Penalty = 200 ; // spin failure penalty 168 static int Knob_Poverty = 1000 ; 169 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 170 static int Knob_FixedSpin = 0 ; 171 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 172 static int Knob_UsePause = 1 ; 173 static int Knob_ExitPolicy = 0 ; 174 static int Knob_PreSpin = 10 ; // 20-100 likely better 175 static int Knob_ResetEvent = 0 ; 176 static int BackOffMask = 0 ; 177 178 static int Knob_FastHSSEC = 0 ; 179 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 180 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 181 static volatile int InitDone = 0 ; 182 183 #define TrySpin TrySpin_VaryDuration 184 185 // ----------------------------------------------------------------------------- 186 // Theory of operations -- Monitors lists, thread residency, etc: 187 // 188 // * A thread acquires ownership of a monitor by successfully 189 // CAS()ing the _owner field from null to non-null. 190 // 191 // * Invariant: A thread appears on at most one monitor list -- 192 // cxq, EntryList or WaitSet -- at any one time. 193 // 194 // * Contending threads "push" themselves onto the cxq with CAS 195 // and then spin/park. 196 // 197 // * After a contending thread eventually acquires the lock it must 198 // dequeue itself from either the EntryList or the cxq. 199 // 200 // * The exiting thread identifies and unparks an "heir presumptive" 201 // tentative successor thread on the EntryList. Critically, the 202 // exiting thread doesn't unlink the successor thread from the EntryList. 203 // After having been unparked, the wakee will recontend for ownership of 204 // the monitor. The successor (wakee) will either acquire the lock or 205 // re-park itself. 206 // 207 // Succession is provided for by a policy of competitive handoff. 208 // The exiting thread does _not_ grant or pass ownership to the 209 // successor thread. (This is also referred to as "handoff" succession"). 210 // Instead the exiting thread releases ownership and possibly wakes 211 // a successor, so the successor can (re)compete for ownership of the lock. 212 // If the EntryList is empty but the cxq is populated the exiting 213 // thread will drain the cxq into the EntryList. It does so by 214 // by detaching the cxq (installing null with CAS) and folding 215 // the threads from the cxq into the EntryList. The EntryList is 216 // doubly linked, while the cxq is singly linked because of the 217 // CAS-based "push" used to enqueue recently arrived threads (RATs). 218 // 219 // * Concurrency invariants: 220 // 221 // -- only the monitor owner may access or mutate the EntryList. 222 // The mutex property of the monitor itself protects the EntryList 223 // from concurrent interference. 224 // -- Only the monitor owner may detach the cxq. 225 // 226 // * The monitor entry list operations avoid locks, but strictly speaking 227 // they're not lock-free. Enter is lock-free, exit is not. 228 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 229 // 230 // * The cxq can have multiple concurrent "pushers" but only one concurrent 231 // detaching thread. This mechanism is immune from the ABA corruption. 232 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 233 // 234 // * Taken together, the cxq and the EntryList constitute or form a 235 // single logical queue of threads stalled trying to acquire the lock. 236 // We use two distinct lists to improve the odds of a constant-time 237 // dequeue operation after acquisition (in the ::enter() epilog) and 238 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 239 // A key desideratum is to minimize queue & monitor metadata manipulation 240 // that occurs while holding the monitor lock -- that is, we want to 241 // minimize monitor lock holds times. Note that even a small amount of 242 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 243 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 244 // locks and monitor metadata. 245 // 246 // Cxq points to the the set of Recently Arrived Threads attempting entry. 247 // Because we push threads onto _cxq with CAS, the RATs must take the form of 248 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 249 // the unlocking thread notices that EntryList is null but _cxq is != null. 250 // 251 // The EntryList is ordered by the prevailing queue discipline and 252 // can be organized in any convenient fashion, such as a doubly-linked list or 253 // a circular doubly-linked list. Critically, we want insert and delete operations 254 // to operate in constant-time. If we need a priority queue then something akin 255 // to Solaris' sleepq would work nicely. Viz., 256 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 257 // Queue discipline is enforced at ::exit() time, when the unlocking thread 258 // drains the cxq into the EntryList, and orders or reorders the threads on the 259 // EntryList accordingly. 260 // 261 // Barring "lock barging", this mechanism provides fair cyclic ordering, 262 // somewhat similar to an elevator-scan. 263 // 264 // * The monitor synchronization subsystem avoids the use of native 265 // synchronization primitives except for the narrow platform-specific 266 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 267 // the semantics of park-unpark. Put another way, this monitor implementation 268 // depends only on atomic operations and park-unpark. The monitor subsystem 269 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 270 // underlying OS manages the READY<->RUN transitions. 271 // 272 // * Waiting threads reside on the WaitSet list -- wait() puts 273 // the caller onto the WaitSet. 274 // 275 // * notify() or notifyAll() simply transfers threads from the WaitSet to 276 // either the EntryList or cxq. Subsequent exit() operations will 277 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 278 // it's likely the notifyee would simply impale itself on the lock held 279 // by the notifier. 280 // 281 // * An interesting alternative is to encode cxq as (List,LockByte) where 282 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 283 // variable, like _recursions, in the scheme. The threads or Events that form 284 // the list would have to be aligned in 256-byte addresses. A thread would 285 // try to acquire the lock or enqueue itself with CAS, but exiting threads 286 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 287 // Note that is is *not* word-tearing, but it does presume that full-word 288 // CAS operations are coherent with intermix with STB operations. That's true 289 // on most common processors. 290 // 291 // * See also http://blogs.sun.com/dave 292 293 294 // ----------------------------------------------------------------------------- 295 // Enter support 296 297 bool ObjectMonitor::try_enter(Thread* THREAD) { 298 if (THREAD != _owner) { 299 if (THREAD->is_lock_owned ((address)_owner)) { 300 assert(_recursions == 0, "internal state error"); 301 _owner = THREAD ; 302 _recursions = 1 ; 303 OwnerIsThread = 1 ; 304 return true; 305 } 306 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 307 return false; 308 } 309 return true; 310 } else { 311 _recursions++; 312 return true; 313 } 314 } 315 316 void ATTR ObjectMonitor::enter(TRAPS) { 317 // The following code is ordered to check the most common cases first 318 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 319 Thread * const Self = THREAD ; 320 void * cur ; 321 322 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 323 if (cur == NULL) { 324 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 325 assert (_recursions == 0 , "invariant") ; 326 assert (_owner == Self, "invariant") ; 327 // CONSIDER: set or assert OwnerIsThread == 1 328 return ; 329 } 330 331 if (cur == Self) { 332 // TODO-FIXME: check for integer overflow! BUGID 6557169. 333 _recursions ++ ; 334 return ; 335 } 336 337 if (Self->is_lock_owned ((address)cur)) { 338 assert (_recursions == 0, "internal state error"); 339 _recursions = 1 ; 340 // Commute owner from a thread-specific on-stack BasicLockObject address to 341 // a full-fledged "Thread *". 342 _owner = Self ; 343 OwnerIsThread = 1 ; 344 return ; 345 } 346 347 // We've encountered genuine contention. 348 assert (Self->_Stalled == 0, "invariant") ; 349 Self->_Stalled = intptr_t(this) ; 350 351 // Try one round of spinning *before* enqueueing Self 352 // and before going through the awkward and expensive state 353 // transitions. The following spin is strictly optional ... 354 // Note that if we acquire the monitor from an initial spin 355 // we forgo posting JVMTI events and firing DTRACE probes. 356 if (Knob_SpinEarly && TrySpin (Self) > 0) { 357 assert (_owner == Self , "invariant") ; 358 assert (_recursions == 0 , "invariant") ; 359 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 360 Self->_Stalled = 0 ; 361 return ; 362 } 363 364 assert (_owner != Self , "invariant") ; 365 assert (_succ != Self , "invariant") ; 366 assert (Self->is_Java_thread() , "invariant") ; 367 JavaThread * jt = (JavaThread *) Self ; 368 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 369 assert (jt->thread_state() != _thread_blocked , "invariant") ; 370 assert (this->object() != NULL , "invariant") ; 371 assert (_count >= 0, "invariant") ; 372 373 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 374 // Ensure the object-monitor relationship remains stable while there's contention. 375 Atomic::inc_ptr(&_count); 376 377 EventJavaMonitorEnter event; 378 379 { // Change java thread status to indicate blocked on monitor enter. 380 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 381 382 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 383 if (JvmtiExport::should_post_monitor_contended_enter()) { 384 JvmtiExport::post_monitor_contended_enter(jt, this); 385 386 // The current thread does not yet own the monitor and does not 387 // yet appear on any queues that would get it made the successor. 388 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 389 // handler cannot accidentally consume an unpark() meant for the 390 // ParkEvent associated with this ObjectMonitor. 391 } 392 393 OSThreadContendState osts(Self->osthread()); 394 ThreadBlockInVM tbivm(jt); 395 396 Self->set_current_pending_monitor(this); 397 398 // TODO-FIXME: change the following for(;;) loop to straight-line code. 399 for (;;) { 400 jt->set_suspend_equivalent(); 401 // cleared by handle_special_suspend_equivalent_condition() 402 // or java_suspend_self() 403 404 EnterI (THREAD) ; 405 406 if (!ExitSuspendEquivalent(jt)) break ; 407 408 // 409 // We have acquired the contended monitor, but while we were 410 // waiting another thread suspended us. We don't want to enter 411 // the monitor while suspended because that would surprise the 412 // thread that suspended us. 413 // 414 _recursions = 0 ; 415 _succ = NULL ; 416 exit (false, Self) ; 417 418 jt->java_suspend_self(); 419 } 420 Self->set_current_pending_monitor(NULL); 421 } 422 423 Atomic::dec_ptr(&_count); 424 assert (_count >= 0, "invariant") ; 425 Self->_Stalled = 0 ; 426 427 // Must either set _recursions = 0 or ASSERT _recursions == 0. 428 assert (_recursions == 0 , "invariant") ; 429 assert (_owner == Self , "invariant") ; 430 assert (_succ != Self , "invariant") ; 431 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 432 433 // The thread -- now the owner -- is back in vm mode. 434 // Report the glorious news via TI,DTrace and jvmstat. 435 // The probe effect is non-trivial. All the reportage occurs 436 // while we hold the monitor, increasing the length of the critical 437 // section. Amdahl's parallel speedup law comes vividly into play. 438 // 439 // Another option might be to aggregate the events (thread local or 440 // per-monitor aggregation) and defer reporting until a more opportune 441 // time -- such as next time some thread encounters contention but has 442 // yet to acquire the lock. While spinning that thread could 443 // spinning we could increment JVMStat counters, etc. 444 445 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 446 if (JvmtiExport::should_post_monitor_contended_entered()) { 447 JvmtiExport::post_monitor_contended_entered(jt, this); 448 449 // The current thread already owns the monitor and is not going to 450 // call park() for the remainder of the monitor enter protocol. So 451 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 452 // event handler consumed an unpark() issued by the thread that 453 // just exited the monitor. 454 } 455 456 if (event.should_commit()) { 457 event.set_klass(((oop)this->object())->klass()); 458 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 459 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 460 event.commit(); 461 } 462 463 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 464 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 465 } 466 } 467 468 469 // Caveat: TryLock() is not necessarily serializing if it returns failure. 470 // Callers must compensate as needed. 471 472 int ObjectMonitor::TryLock (Thread * Self) { 473 for (;;) { 474 void * own = _owner ; 475 if (own != NULL) return 0 ; 476 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 477 // Either guarantee _recursions == 0 or set _recursions = 0. 478 assert (_recursions == 0, "invariant") ; 479 assert (_owner == Self, "invariant") ; 480 // CONSIDER: set or assert that OwnerIsThread == 1 481 return 1 ; 482 } 483 // The lock had been free momentarily, but we lost the race to the lock. 484 // Interference -- the CAS failed. 485 // We can either return -1 or retry. 486 // Retry doesn't make as much sense because the lock was just acquired. 487 if (true) return -1 ; 488 } 489 } 490 491 void ATTR ObjectMonitor::EnterI (TRAPS) { 492 Thread * Self = THREAD ; 493 assert (Self->is_Java_thread(), "invariant") ; 494 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 495 496 // Try the lock - TATAS 497 if (TryLock (Self) > 0) { 498 assert (_succ != Self , "invariant") ; 499 assert (_owner == Self , "invariant") ; 500 assert (_Responsible != Self , "invariant") ; 501 return ; 502 } 503 504 DeferredInitialize () ; 505 506 // We try one round of spinning *before* enqueueing Self. 507 // 508 // If the _owner is ready but OFFPROC we could use a YieldTo() 509 // operation to donate the remainder of this thread's quantum 510 // to the owner. This has subtle but beneficial affinity 511 // effects. 512 513 if (TrySpin (Self) > 0) { 514 assert (_owner == Self , "invariant") ; 515 assert (_succ != Self , "invariant") ; 516 assert (_Responsible != Self , "invariant") ; 517 return ; 518 } 519 520 // The Spin failed -- Enqueue and park the thread ... 521 assert (_succ != Self , "invariant") ; 522 assert (_owner != Self , "invariant") ; 523 assert (_Responsible != Self , "invariant") ; 524 525 // Enqueue "Self" on ObjectMonitor's _cxq. 526 // 527 // Node acts as a proxy for Self. 528 // As an aside, if were to ever rewrite the synchronization code mostly 529 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 530 // Java objects. This would avoid awkward lifecycle and liveness issues, 531 // as well as eliminate a subset of ABA issues. 532 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 533 // 534 535 ObjectWaiter node(Self) ; 536 Self->_ParkEvent->reset() ; 537 node._prev = (ObjectWaiter *) 0xBAD ; 538 node.TState = ObjectWaiter::TS_CXQ ; 539 540 // Push "Self" onto the front of the _cxq. 541 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 542 // Note that spinning tends to reduce the rate at which threads 543 // enqueue and dequeue on EntryList|cxq. 544 ObjectWaiter * nxt ; 545 for (;;) { 546 node._next = nxt = _cxq ; 547 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 548 549 // Interference - the CAS failed because _cxq changed. Just retry. 550 // As an optional optimization we retry the lock. 551 if (TryLock (Self) > 0) { 552 assert (_succ != Self , "invariant") ; 553 assert (_owner == Self , "invariant") ; 554 assert (_Responsible != Self , "invariant") ; 555 return ; 556 } 557 } 558 559 // Check for cxq|EntryList edge transition to non-null. This indicates 560 // the onset of contention. While contention persists exiting threads 561 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 562 // operations revert to the faster 1-0 mode. This enter operation may interleave 563 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 564 // arrange for one of the contending thread to use a timed park() operations 565 // to detect and recover from the race. (Stranding is form of progress failure 566 // where the monitor is unlocked but all the contending threads remain parked). 567 // That is, at least one of the contended threads will periodically poll _owner. 568 // One of the contending threads will become the designated "Responsible" thread. 569 // The Responsible thread uses a timed park instead of a normal indefinite park 570 // operation -- it periodically wakes and checks for and recovers from potential 571 // strandings admitted by 1-0 exit operations. We need at most one Responsible 572 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 573 // be responsible for a monitor. 574 // 575 // Currently, one of the contended threads takes on the added role of "Responsible". 576 // A viable alternative would be to use a dedicated "stranding checker" thread 577 // that periodically iterated over all the threads (or active monitors) and unparked 578 // successors where there was risk of stranding. This would help eliminate the 579 // timer scalability issues we see on some platforms as we'd only have one thread 580 // -- the checker -- parked on a timer. 581 582 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 583 // Try to assume the role of responsible thread for the monitor. 584 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 585 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 586 } 587 588 // The lock have been released while this thread was occupied queueing 589 // itself onto _cxq. To close the race and avoid "stranding" and 590 // progress-liveness failure we must resample-retry _owner before parking. 591 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 592 // In this case the ST-MEMBAR is accomplished with CAS(). 593 // 594 // TODO: Defer all thread state transitions until park-time. 595 // Since state transitions are heavy and inefficient we'd like 596 // to defer the state transitions until absolutely necessary, 597 // and in doing so avoid some transitions ... 598 599 TEVENT (Inflated enter - Contention) ; 600 int nWakeups = 0 ; 601 int RecheckInterval = 1 ; 602 603 for (;;) { 604 605 if (TryLock (Self) > 0) break ; 606 assert (_owner != Self, "invariant") ; 607 608 if ((SyncFlags & 2) && _Responsible == NULL) { 609 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 610 } 611 612 // park self 613 if (_Responsible == Self || (SyncFlags & 1)) { 614 TEVENT (Inflated enter - park TIMED) ; 615 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 616 // Increase the RecheckInterval, but clamp the value. 617 RecheckInterval *= 8 ; 618 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 619 } else { 620 TEVENT (Inflated enter - park UNTIMED) ; 621 Self->_ParkEvent->park() ; 622 } 623 624 if (TryLock(Self) > 0) break ; 625 626 // The lock is still contested. 627 // Keep a tally of the # of futile wakeups. 628 // Note that the counter is not protected by a lock or updated by atomics. 629 // That is by design - we trade "lossy" counters which are exposed to 630 // races during updates for a lower probe effect. 631 TEVENT (Inflated enter - Futile wakeup) ; 632 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 633 ObjectMonitor::_sync_FutileWakeups->inc() ; 634 } 635 ++ nWakeups ; 636 637 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 638 // We can defer clearing _succ until after the spin completes 639 // TrySpin() must tolerate being called with _succ == Self. 640 // Try yet another round of adaptive spinning. 641 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 642 643 // We can find that we were unpark()ed and redesignated _succ while 644 // we were spinning. That's harmless. If we iterate and call park(), 645 // park() will consume the event and return immediately and we'll 646 // just spin again. This pattern can repeat, leaving _succ to simply 647 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 648 // Alternately, we can sample fired() here, and if set, forgo spinning 649 // in the next iteration. 650 651 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 652 Self->_ParkEvent->reset() ; 653 OrderAccess::fence() ; 654 } 655 if (_succ == Self) _succ = NULL ; 656 657 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 658 OrderAccess::fence() ; 659 } 660 661 // Egress : 662 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 663 // Normally we'll find Self on the EntryList . 664 // From the perspective of the lock owner (this thread), the 665 // EntryList is stable and cxq is prepend-only. 666 // The head of cxq is volatile but the interior is stable. 667 // In addition, Self.TState is stable. 668 669 assert (_owner == Self , "invariant") ; 670 assert (object() != NULL , "invariant") ; 671 // I'd like to write: 672 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 673 // but as we're at a safepoint that's not safe. 674 675 UnlinkAfterAcquire (Self, &node) ; 676 if (_succ == Self) _succ = NULL ; 677 678 assert (_succ != Self, "invariant") ; 679 if (_Responsible == Self) { 680 _Responsible = NULL ; 681 OrderAccess::fence(); // Dekker pivot-point 682 683 // We may leave threads on cxq|EntryList without a designated 684 // "Responsible" thread. This is benign. When this thread subsequently 685 // exits the monitor it can "see" such preexisting "old" threads -- 686 // threads that arrived on the cxq|EntryList before the fence, above -- 687 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 688 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 689 // non-null and elect a new "Responsible" timer thread. 690 // 691 // This thread executes: 692 // ST Responsible=null; MEMBAR (in enter epilog - here) 693 // LD cxq|EntryList (in subsequent exit) 694 // 695 // Entering threads in the slow/contended path execute: 696 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 697 // The (ST cxq; MEMBAR) is accomplished with CAS(). 698 // 699 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 700 // exit operation from floating above the ST Responsible=null. 701 } 702 703 // We've acquired ownership with CAS(). 704 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 705 // But since the CAS() this thread may have also stored into _succ, 706 // EntryList, cxq or Responsible. These meta-data updates must be 707 // visible __before this thread subsequently drops the lock. 708 // Consider what could occur if we didn't enforce this constraint -- 709 // STs to monitor meta-data and user-data could reorder with (become 710 // visible after) the ST in exit that drops ownership of the lock. 711 // Some other thread could then acquire the lock, but observe inconsistent 712 // or old monitor meta-data and heap data. That violates the JMM. 713 // To that end, the 1-0 exit() operation must have at least STST|LDST 714 // "release" barrier semantics. Specifically, there must be at least a 715 // STST|LDST barrier in exit() before the ST of null into _owner that drops 716 // the lock. The barrier ensures that changes to monitor meta-data and data 717 // protected by the lock will be visible before we release the lock, and 718 // therefore before some other thread (CPU) has a chance to acquire the lock. 719 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 720 // 721 // Critically, any prior STs to _succ or EntryList must be visible before 722 // the ST of null into _owner in the *subsequent* (following) corresponding 723 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 724 // execute a serializing instruction. 725 726 if (SyncFlags & 8) { 727 OrderAccess::fence() ; 728 } 729 return ; 730 } 731 732 // ReenterI() is a specialized inline form of the latter half of the 733 // contended slow-path from EnterI(). We use ReenterI() only for 734 // monitor reentry in wait(). 735 // 736 // In the future we should reconcile EnterI() and ReenterI(), adding 737 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 738 // loop accordingly. 739 740 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 741 assert (Self != NULL , "invariant") ; 742 assert (SelfNode != NULL , "invariant") ; 743 assert (SelfNode->_thread == Self , "invariant") ; 744 assert (_waiters > 0 , "invariant") ; 745 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 746 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 747 JavaThread * jt = (JavaThread *) Self ; 748 749 int nWakeups = 0 ; 750 for (;;) { 751 ObjectWaiter::TStates v = SelfNode->TState ; 752 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 753 assert (_owner != Self, "invariant") ; 754 755 if (TryLock (Self) > 0) break ; 756 if (TrySpin (Self) > 0) break ; 757 758 TEVENT (Wait Reentry - parking) ; 759 760 // State transition wrappers around park() ... 761 // ReenterI() wisely defers state transitions until 762 // it's clear we must park the thread. 763 { 764 OSThreadContendState osts(Self->osthread()); 765 ThreadBlockInVM tbivm(jt); 766 767 // cleared by handle_special_suspend_equivalent_condition() 768 // or java_suspend_self() 769 jt->set_suspend_equivalent(); 770 if (SyncFlags & 1) { 771 Self->_ParkEvent->park ((jlong)1000) ; 772 } else { 773 Self->_ParkEvent->park () ; 774 } 775 776 // were we externally suspended while we were waiting? 777 for (;;) { 778 if (!ExitSuspendEquivalent (jt)) break ; 779 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 780 jt->java_suspend_self(); 781 jt->set_suspend_equivalent(); 782 } 783 } 784 785 // Try again, but just so we distinguish between futile wakeups and 786 // successful wakeups. The following test isn't algorithmically 787 // necessary, but it helps us maintain sensible statistics. 788 if (TryLock(Self) > 0) break ; 789 790 // The lock is still contested. 791 // Keep a tally of the # of futile wakeups. 792 // Note that the counter is not protected by a lock or updated by atomics. 793 // That is by design - we trade "lossy" counters which are exposed to 794 // races during updates for a lower probe effect. 795 TEVENT (Wait Reentry - futile wakeup) ; 796 ++ nWakeups ; 797 798 // Assuming this is not a spurious wakeup we'll normally 799 // find that _succ == Self. 800 if (_succ == Self) _succ = NULL ; 801 802 // Invariant: after clearing _succ a contending thread 803 // *must* retry _owner before parking. 804 OrderAccess::fence() ; 805 806 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 807 ObjectMonitor::_sync_FutileWakeups->inc() ; 808 } 809 } 810 811 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 812 // Normally we'll find Self on the EntryList. 813 // Unlinking from the EntryList is constant-time and atomic-free. 814 // From the perspective of the lock owner (this thread), the 815 // EntryList is stable and cxq is prepend-only. 816 // The head of cxq is volatile but the interior is stable. 817 // In addition, Self.TState is stable. 818 819 assert (_owner == Self, "invariant") ; 820 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 821 UnlinkAfterAcquire (Self, SelfNode) ; 822 if (_succ == Self) _succ = NULL ; 823 assert (_succ != Self, "invariant") ; 824 SelfNode->TState = ObjectWaiter::TS_RUN ; 825 OrderAccess::fence() ; // see comments at the end of EnterI() 826 } 827 828 // after the thread acquires the lock in ::enter(). Equally, we could defer 829 // unlinking the thread until ::exit()-time. 830 831 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 832 { 833 assert (_owner == Self, "invariant") ; 834 assert (SelfNode->_thread == Self, "invariant") ; 835 836 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 837 // Normal case: remove Self from the DLL EntryList . 838 // This is a constant-time operation. 839 ObjectWaiter * nxt = SelfNode->_next ; 840 ObjectWaiter * prv = SelfNode->_prev ; 841 if (nxt != NULL) nxt->_prev = prv ; 842 if (prv != NULL) prv->_next = nxt ; 843 if (SelfNode == _EntryList ) _EntryList = nxt ; 844 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 845 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 846 TEVENT (Unlink from EntryList) ; 847 } else { 848 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 849 // Inopportune interleaving -- Self is still on the cxq. 850 // This usually means the enqueue of self raced an exiting thread. 851 // Normally we'll find Self near the front of the cxq, so 852 // dequeueing is typically fast. If needbe we can accelerate 853 // this with some MCS/CHL-like bidirectional list hints and advisory 854 // back-links so dequeueing from the interior will normally operate 855 // in constant-time. 856 // Dequeue Self from either the head (with CAS) or from the interior 857 // with a linear-time scan and normal non-atomic memory operations. 858 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 859 // and then unlink Self from EntryList. We have to drain eventually, 860 // so it might as well be now. 861 862 ObjectWaiter * v = _cxq ; 863 assert (v != NULL, "invariant") ; 864 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 865 // The CAS above can fail from interference IFF a "RAT" arrived. 866 // In that case Self must be in the interior and can no longer be 867 // at the head of cxq. 868 if (v == SelfNode) { 869 assert (_cxq != v, "invariant") ; 870 v = _cxq ; // CAS above failed - start scan at head of list 871 } 872 ObjectWaiter * p ; 873 ObjectWaiter * q = NULL ; 874 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 875 q = p ; 876 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 877 } 878 assert (v != SelfNode, "invariant") ; 879 assert (p == SelfNode, "Node not found on cxq") ; 880 assert (p != _cxq, "invariant") ; 881 assert (q != NULL, "invariant") ; 882 assert (q->_next == p, "invariant") ; 883 q->_next = p->_next ; 884 } 885 TEVENT (Unlink from cxq) ; 886 } 887 888 // Diagnostic hygiene ... 889 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 890 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 891 SelfNode->TState = ObjectWaiter::TS_RUN ; 892 } 893 894 // ----------------------------------------------------------------------------- 895 // Exit support 896 // 897 // exit() 898 // ~~~~~~ 899 // Note that the collector can't reclaim the objectMonitor or deflate 900 // the object out from underneath the thread calling ::exit() as the 901 // thread calling ::exit() never transitions to a stable state. 902 // This inhibits GC, which in turn inhibits asynchronous (and 903 // inopportune) reclamation of "this". 904 // 905 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 906 // There's one exception to the claim above, however. EnterI() can call 907 // exit() to drop a lock if the acquirer has been externally suspended. 908 // In that case exit() is called with _thread_state as _thread_blocked, 909 // but the monitor's _count field is > 0, which inhibits reclamation. 910 // 911 // 1-0 exit 912 // ~~~~~~~~ 913 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 914 // the fast-path operators have been optimized so the common ::exit() 915 // operation is 1-0. See i486.ad fast_unlock(), for instance. 916 // The code emitted by fast_unlock() elides the usual MEMBAR. This 917 // greatly improves latency -- MEMBAR and CAS having considerable local 918 // latency on modern processors -- but at the cost of "stranding". Absent the 919 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 920 // ::enter() path, resulting in the entering thread being stranding 921 // and a progress-liveness failure. Stranding is extremely rare. 922 // We use timers (timed park operations) & periodic polling to detect 923 // and recover from stranding. Potentially stranded threads periodically 924 // wake up and poll the lock. See the usage of the _Responsible variable. 925 // 926 // The CAS() in enter provides for safety and exclusion, while the CAS or 927 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 928 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 929 // We detect and recover from stranding with timers. 930 // 931 // If a thread transiently strands it'll park until (a) another 932 // thread acquires the lock and then drops the lock, at which time the 933 // exiting thread will notice and unpark the stranded thread, or, (b) 934 // the timer expires. If the lock is high traffic then the stranding latency 935 // will be low due to (a). If the lock is low traffic then the odds of 936 // stranding are lower, although the worst-case stranding latency 937 // is longer. Critically, we don't want to put excessive load in the 938 // platform's timer subsystem. We want to minimize both the timer injection 939 // rate (timers created/sec) as well as the number of timers active at 940 // any one time. (more precisely, we want to minimize timer-seconds, which is 941 // the integral of the # of active timers at any instant over time). 942 // Both impinge on OS scalability. Given that, at most one thread parked on 943 // a monitor will use a timer. 944 945 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 946 Thread * Self = THREAD ; 947 if (THREAD != _owner) { 948 if (THREAD->is_lock_owned((address) _owner)) { 949 // Transmute _owner from a BasicLock pointer to a Thread address. 950 // We don't need to hold _mutex for this transition. 951 // Non-null to Non-null is safe as long as all readers can 952 // tolerate either flavor. 953 assert (_recursions == 0, "invariant") ; 954 _owner = THREAD ; 955 _recursions = 0 ; 956 OwnerIsThread = 1 ; 957 } else { 958 // NOTE: we need to handle unbalanced monitor enter/exit 959 // in native code by throwing an exception. 960 // TODO: Throw an IllegalMonitorStateException ? 961 TEVENT (Exit - Throw IMSX) ; 962 assert(false, "Non-balanced monitor enter/exit!"); 963 if (false) { 964 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 965 } 966 return; 967 } 968 } 969 970 if (_recursions != 0) { 971 _recursions--; // this is simple recursive enter 972 TEVENT (Inflated exit - recursive) ; 973 return ; 974 } 975 976 // Invariant: after setting Responsible=null an thread must execute 977 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 978 if ((SyncFlags & 4) == 0) { 979 _Responsible = NULL ; 980 } 981 982 #if INCLUDE_TRACE 983 // get the owner's thread id for the MonitorEnter event 984 // if it is enabled and the thread isn't suspended 985 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 986 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 987 } 988 #endif 989 990 for (;;) { 991 assert (THREAD == _owner, "invariant") ; 992 993 994 if (Knob_ExitPolicy == 0) { 995 // release semantics: prior loads and stores from within the critical section 996 // must not float (reorder) past the following store that drops the lock. 997 // On SPARC that requires MEMBAR #loadstore|#storestore. 998 // But of course in TSO #loadstore|#storestore is not required. 999 // I'd like to write one of the following: 1000 // A. OrderAccess::release() ; _owner = NULL 1001 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 1002 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 1003 // store into a _dummy variable. That store is not needed, but can result 1004 // in massive wasteful coherency traffic on classic SMP systems. 1005 // Instead, I use release_store(), which is implemented as just a simple 1006 // ST on x64, x86 and SPARC. 1007 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1008 OrderAccess::storeload() ; // See if we need to wake a successor 1009 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1010 TEVENT (Inflated exit - simple egress) ; 1011 return ; 1012 } 1013 TEVENT (Inflated exit - complex egress) ; 1014 1015 // Normally the exiting thread is responsible for ensuring succession, 1016 // but if other successors are ready or other entering threads are spinning 1017 // then this thread can simply store NULL into _owner and exit without 1018 // waking a successor. The existence of spinners or ready successors 1019 // guarantees proper succession (liveness). Responsibility passes to the 1020 // ready or running successors. The exiting thread delegates the duty. 1021 // More precisely, if a successor already exists this thread is absolved 1022 // of the responsibility of waking (unparking) one. 1023 // 1024 // The _succ variable is critical to reducing futile wakeup frequency. 1025 // _succ identifies the "heir presumptive" thread that has been made 1026 // ready (unparked) but that has not yet run. We need only one such 1027 // successor thread to guarantee progress. 1028 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1029 // section 3.3 "Futile Wakeup Throttling" for details. 1030 // 1031 // Note that spinners in Enter() also set _succ non-null. 1032 // In the current implementation spinners opportunistically set 1033 // _succ so that exiting threads might avoid waking a successor. 1034 // Another less appealing alternative would be for the exiting thread 1035 // to drop the lock and then spin briefly to see if a spinner managed 1036 // to acquire the lock. If so, the exiting thread could exit 1037 // immediately without waking a successor, otherwise the exiting 1038 // thread would need to dequeue and wake a successor. 1039 // (Note that we'd need to make the post-drop spin short, but no 1040 // shorter than the worst-case round-trip cache-line migration time. 1041 // The dropped lock needs to become visible to the spinner, and then 1042 // the acquisition of the lock by the spinner must become visible to 1043 // the exiting thread). 1044 // 1045 1046 // It appears that an heir-presumptive (successor) must be made ready. 1047 // Only the current lock owner can manipulate the EntryList or 1048 // drain _cxq, so we need to reacquire the lock. If we fail 1049 // to reacquire the lock the responsibility for ensuring succession 1050 // falls to the new owner. 1051 // 1052 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1053 return ; 1054 } 1055 TEVENT (Exit - Reacquired) ; 1056 } else { 1057 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1058 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1059 OrderAccess::storeload() ; 1060 // Ratify the previously observed values. 1061 if (_cxq == NULL || _succ != NULL) { 1062 TEVENT (Inflated exit - simple egress) ; 1063 return ; 1064 } 1065 1066 // inopportune interleaving -- the exiting thread (this thread) 1067 // in the fast-exit path raced an entering thread in the slow-enter 1068 // path. 1069 // We have two choices: 1070 // A. Try to reacquire the lock. 1071 // If the CAS() fails return immediately, otherwise 1072 // we either restart/rerun the exit operation, or simply 1073 // fall-through into the code below which wakes a successor. 1074 // B. If the elements forming the EntryList|cxq are TSM 1075 // we could simply unpark() the lead thread and return 1076 // without having set _succ. 1077 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1078 TEVENT (Inflated exit - reacquired succeeded) ; 1079 return ; 1080 } 1081 TEVENT (Inflated exit - reacquired failed) ; 1082 } else { 1083 TEVENT (Inflated exit - complex egress) ; 1084 } 1085 } 1086 1087 guarantee (_owner == THREAD, "invariant") ; 1088 1089 ObjectWaiter * w = NULL ; 1090 int QMode = Knob_QMode ; 1091 1092 if (QMode == 2 && _cxq != NULL) { 1093 // QMode == 2 : cxq has precedence over EntryList. 1094 // Try to directly wake a successor from the cxq. 1095 // If successful, the successor will need to unlink itself from cxq. 1096 w = _cxq ; 1097 assert (w != NULL, "invariant") ; 1098 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1099 ExitEpilog (Self, w) ; 1100 return ; 1101 } 1102 1103 if (QMode == 3 && _cxq != NULL) { 1104 // Aggressively drain cxq into EntryList at the first opportunity. 1105 // This policy ensure that recently-run threads live at the head of EntryList. 1106 // Drain _cxq into EntryList - bulk transfer. 1107 // First, detach _cxq. 1108 // The following loop is tantamount to: w = swap (&cxq, NULL) 1109 w = _cxq ; 1110 for (;;) { 1111 assert (w != NULL, "Invariant") ; 1112 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1113 if (u == w) break ; 1114 w = u ; 1115 } 1116 assert (w != NULL , "invariant") ; 1117 1118 ObjectWaiter * q = NULL ; 1119 ObjectWaiter * p ; 1120 for (p = w ; p != NULL ; p = p->_next) { 1121 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1122 p->TState = ObjectWaiter::TS_ENTER ; 1123 p->_prev = q ; 1124 q = p ; 1125 } 1126 1127 // Append the RATs to the EntryList 1128 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1129 ObjectWaiter * Tail ; 1130 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1131 if (Tail == NULL) { 1132 _EntryList = w ; 1133 } else { 1134 Tail->_next = w ; 1135 w->_prev = Tail ; 1136 } 1137 1138 // Fall thru into code that tries to wake a successor from EntryList 1139 } 1140 1141 if (QMode == 4 && _cxq != NULL) { 1142 // Aggressively drain cxq into EntryList at the first opportunity. 1143 // This policy ensure that recently-run threads live at the head of EntryList. 1144 1145 // Drain _cxq into EntryList - bulk transfer. 1146 // First, detach _cxq. 1147 // The following loop is tantamount to: w = swap (&cxq, NULL) 1148 w = _cxq ; 1149 for (;;) { 1150 assert (w != NULL, "Invariant") ; 1151 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1152 if (u == w) break ; 1153 w = u ; 1154 } 1155 assert (w != NULL , "invariant") ; 1156 1157 ObjectWaiter * q = NULL ; 1158 ObjectWaiter * p ; 1159 for (p = w ; p != NULL ; p = p->_next) { 1160 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1161 p->TState = ObjectWaiter::TS_ENTER ; 1162 p->_prev = q ; 1163 q = p ; 1164 } 1165 1166 // Prepend the RATs to the EntryList 1167 if (_EntryList != NULL) { 1168 q->_next = _EntryList ; 1169 _EntryList->_prev = q ; 1170 } 1171 _EntryList = w ; 1172 1173 // Fall thru into code that tries to wake a successor from EntryList 1174 } 1175 1176 w = _EntryList ; 1177 if (w != NULL) { 1178 // I'd like to write: guarantee (w->_thread != Self). 1179 // But in practice an exiting thread may find itself on the EntryList. 1180 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1181 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1182 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1183 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1184 // release the lock "O". T2 resumes immediately after the ST of null into 1185 // _owner, above. T2 notices that the EntryList is populated, so it 1186 // reacquires the lock and then finds itself on the EntryList. 1187 // Given all that, we have to tolerate the circumstance where "w" is 1188 // associated with Self. 1189 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1190 ExitEpilog (Self, w) ; 1191 return ; 1192 } 1193 1194 // If we find that both _cxq and EntryList are null then just 1195 // re-run the exit protocol from the top. 1196 w = _cxq ; 1197 if (w == NULL) continue ; 1198 1199 // Drain _cxq into EntryList - bulk transfer. 1200 // First, detach _cxq. 1201 // The following loop is tantamount to: w = swap (&cxq, NULL) 1202 for (;;) { 1203 assert (w != NULL, "Invariant") ; 1204 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1205 if (u == w) break ; 1206 w = u ; 1207 } 1208 TEVENT (Inflated exit - drain cxq into EntryList) ; 1209 1210 assert (w != NULL , "invariant") ; 1211 assert (_EntryList == NULL , "invariant") ; 1212 1213 // Convert the LIFO SLL anchored by _cxq into a DLL. 1214 // The list reorganization step operates in O(LENGTH(w)) time. 1215 // It's critical that this step operate quickly as 1216 // "Self" still holds the outer-lock, restricting parallelism 1217 // and effectively lengthening the critical section. 1218 // Invariant: s chases t chases u. 1219 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1220 // we have faster access to the tail. 1221 1222 if (QMode == 1) { 1223 // QMode == 1 : drain cxq to EntryList, reversing order 1224 // We also reverse the order of the list. 1225 ObjectWaiter * s = NULL ; 1226 ObjectWaiter * t = w ; 1227 ObjectWaiter * u = NULL ; 1228 while (t != NULL) { 1229 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1230 t->TState = ObjectWaiter::TS_ENTER ; 1231 u = t->_next ; 1232 t->_prev = u ; 1233 t->_next = s ; 1234 s = t; 1235 t = u ; 1236 } 1237 _EntryList = s ; 1238 assert (s != NULL, "invariant") ; 1239 } else { 1240 // QMode == 0 or QMode == 2 1241 _EntryList = w ; 1242 ObjectWaiter * q = NULL ; 1243 ObjectWaiter * p ; 1244 for (p = w ; p != NULL ; p = p->_next) { 1245 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1246 p->TState = ObjectWaiter::TS_ENTER ; 1247 p->_prev = q ; 1248 q = p ; 1249 } 1250 } 1251 1252 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1253 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1254 1255 // See if we can abdicate to a spinner instead of waking a thread. 1256 // A primary goal of the implementation is to reduce the 1257 // context-switch rate. 1258 if (_succ != NULL) continue; 1259 1260 w = _EntryList ; 1261 if (w != NULL) { 1262 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1263 ExitEpilog (Self, w) ; 1264 return ; 1265 } 1266 } 1267 } 1268 1269 // ExitSuspendEquivalent: 1270 // A faster alternate to handle_special_suspend_equivalent_condition() 1271 // 1272 // handle_special_suspend_equivalent_condition() unconditionally 1273 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1274 // operations have high latency. Note that in ::enter() we call HSSEC 1275 // while holding the monitor, so we effectively lengthen the critical sections. 1276 // 1277 // There are a number of possible solutions: 1278 // 1279 // A. To ameliorate the problem we might also defer state transitions 1280 // to as late as possible -- just prior to parking. 1281 // Given that, we'd call HSSEC after having returned from park(), 1282 // but before attempting to acquire the monitor. This is only a 1283 // partial solution. It avoids calling HSSEC while holding the 1284 // monitor (good), but it still increases successor reacquisition latency -- 1285 // the interval between unparking a successor and the time the successor 1286 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1287 // If we use this technique we can also avoid EnterI()-exit() loop 1288 // in ::enter() where we iteratively drop the lock and then attempt 1289 // to reacquire it after suspending. 1290 // 1291 // B. In the future we might fold all the suspend bits into a 1292 // composite per-thread suspend flag and then update it with CAS(). 1293 // Alternately, a Dekker-like mechanism with multiple variables 1294 // would suffice: 1295 // ST Self->_suspend_equivalent = false 1296 // MEMBAR 1297 // LD Self_>_suspend_flags 1298 // 1299 1300 1301 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1302 int Mode = Knob_FastHSSEC ; 1303 if (Mode && !jSelf->is_external_suspend()) { 1304 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1305 jSelf->clear_suspend_equivalent() ; 1306 if (2 == Mode) OrderAccess::storeload() ; 1307 if (!jSelf->is_external_suspend()) return false ; 1308 // We raced a suspension -- fall thru into the slow path 1309 TEVENT (ExitSuspendEquivalent - raced) ; 1310 jSelf->set_suspend_equivalent() ; 1311 } 1312 return jSelf->handle_special_suspend_equivalent_condition() ; 1313 } 1314 1315 1316 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1317 assert (_owner == Self, "invariant") ; 1318 1319 // Exit protocol: 1320 // 1. ST _succ = wakee 1321 // 2. membar #loadstore|#storestore; 1322 // 2. ST _owner = NULL 1323 // 3. unpark(wakee) 1324 1325 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1326 ParkEvent * Trigger = Wakee->_event ; 1327 1328 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1329 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1330 // out-of-scope (non-extant). 1331 Wakee = NULL ; 1332 1333 // Drop the lock 1334 OrderAccess::release_store_ptr (&_owner, NULL) ; 1335 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1336 1337 if (SafepointSynchronize::do_call_back()) { 1338 TEVENT (unpark before SAFEPOINT) ; 1339 } 1340 1341 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1342 Trigger->unpark() ; 1343 1344 // Maintain stats and report events to JVMTI 1345 if (ObjectMonitor::_sync_Parks != NULL) { 1346 ObjectMonitor::_sync_Parks->inc() ; 1347 } 1348 } 1349 1350 1351 // ----------------------------------------------------------------------------- 1352 // Class Loader deadlock handling. 1353 // 1354 // complete_exit exits a lock returning recursion count 1355 // complete_exit/reenter operate as a wait without waiting 1356 // complete_exit requires an inflated monitor 1357 // The _owner field is not always the Thread addr even with an 1358 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1359 // thread due to contention. 1360 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1361 Thread * const Self = THREAD; 1362 assert(Self->is_Java_thread(), "Must be Java thread!"); 1363 JavaThread *jt = (JavaThread *)THREAD; 1364 1365 DeferredInitialize(); 1366 1367 if (THREAD != _owner) { 1368 if (THREAD->is_lock_owned ((address)_owner)) { 1369 assert(_recursions == 0, "internal state error"); 1370 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1371 _recursions = 0 ; 1372 OwnerIsThread = 1 ; 1373 } 1374 } 1375 1376 guarantee(Self == _owner, "complete_exit not owner"); 1377 intptr_t save = _recursions; // record the old recursion count 1378 _recursions = 0; // set the recursion level to be 0 1379 exit (true, Self) ; // exit the monitor 1380 guarantee (_owner != Self, "invariant"); 1381 return save; 1382 } 1383 1384 // reenter() enters a lock and sets recursion count 1385 // complete_exit/reenter operate as a wait without waiting 1386 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1387 Thread * const Self = THREAD; 1388 assert(Self->is_Java_thread(), "Must be Java thread!"); 1389 JavaThread *jt = (JavaThread *)THREAD; 1390 1391 guarantee(_owner != Self, "reenter already owner"); 1392 enter (THREAD); // enter the monitor 1393 guarantee (_recursions == 0, "reenter recursion"); 1394 _recursions = recursions; 1395 return; 1396 } 1397 1398 1399 // ----------------------------------------------------------------------------- 1400 // A macro is used below because there may already be a pending 1401 // exception which should not abort the execution of the routines 1402 // which use this (which is why we don't put this into check_slow and 1403 // call it with a CHECK argument). 1404 1405 #define CHECK_OWNER() \ 1406 do { \ 1407 if (THREAD != _owner) { \ 1408 if (THREAD->is_lock_owned((address) _owner)) { \ 1409 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1410 _recursions = 0; \ 1411 OwnerIsThread = 1 ; \ 1412 } else { \ 1413 TEVENT (Throw IMSX) ; \ 1414 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1415 } \ 1416 } \ 1417 } while (false) 1418 1419 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1420 // TODO-FIXME: remove check_slow() -- it's likely dead. 1421 1422 void ObjectMonitor::check_slow(TRAPS) { 1423 TEVENT (check_slow - throw IMSX) ; 1424 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1425 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1426 } 1427 1428 static int Adjust (volatile int * adr, int dx) { 1429 int v ; 1430 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1431 return v ; 1432 } 1433 1434 // helper method for posting a monitor wait event 1435 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1436 jlong notifier_tid, 1437 jlong timeout, 1438 bool timedout) { 1439 event->set_klass(((oop)this->object())->klass()); 1440 event->set_timeout((TYPE_ULONG)timeout); 1441 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1442 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1443 event->set_timedOut((TYPE_BOOLEAN)timedout); 1444 event->commit(); 1445 } 1446 1447 // ----------------------------------------------------------------------------- 1448 // Wait/Notify/NotifyAll 1449 // 1450 // Note: a subset of changes to ObjectMonitor::wait() 1451 // will need to be replicated in complete_exit above 1452 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1453 Thread * const Self = THREAD ; 1454 assert(Self->is_Java_thread(), "Must be Java thread!"); 1455 JavaThread *jt = (JavaThread *)THREAD; 1456 1457 DeferredInitialize () ; 1458 1459 // Throw IMSX or IEX. 1460 CHECK_OWNER(); 1461 1462 EventJavaMonitorWait event; 1463 1464 // check for a pending interrupt 1465 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1466 // post monitor waited event. Note that this is past-tense, we are done waiting. 1467 if (JvmtiExport::should_post_monitor_waited()) { 1468 // Note: 'false' parameter is passed here because the 1469 // wait was not timed out due to thread interrupt. 1470 JvmtiExport::post_monitor_waited(jt, this, false); 1471 1472 // In this short circuit of the monitor wait protocol, the 1473 // current thread never drops ownership of the monitor and 1474 // never gets added to the wait queue so the current thread 1475 // cannot be made the successor. This means that the 1476 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1477 // consume an unpark() meant for the ParkEvent associated with 1478 // this ObjectMonitor. 1479 } 1480 if (event.should_commit()) { 1481 post_monitor_wait_event(&event, 0, millis, false); 1482 } 1483 TEVENT (Wait - Throw IEX) ; 1484 THROW(vmSymbols::java_lang_InterruptedException()); 1485 return ; 1486 } 1487 1488 TEVENT (Wait) ; 1489 1490 assert (Self->_Stalled == 0, "invariant") ; 1491 Self->_Stalled = intptr_t(this) ; 1492 jt->set_current_waiting_monitor(this); 1493 1494 // create a node to be put into the queue 1495 // Critically, after we reset() the event but prior to park(), we must check 1496 // for a pending interrupt. 1497 ObjectWaiter node(Self); 1498 node.TState = ObjectWaiter::TS_WAIT ; 1499 Self->_ParkEvent->reset() ; 1500 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1501 1502 // Enter the waiting queue, which is a circular doubly linked list in this case 1503 // but it could be a priority queue or any data structure. 1504 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1505 // by the the owner of the monitor *except* in the case where park() 1506 // returns because of a timeout of interrupt. Contention is exceptionally rare 1507 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1508 1509 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1510 AddWaiter (&node) ; 1511 Thread::SpinRelease (&_WaitSetLock) ; 1512 1513 if ((SyncFlags & 4) == 0) { 1514 _Responsible = NULL ; 1515 } 1516 intptr_t save = _recursions; // record the old recursion count 1517 _waiters++; // increment the number of waiters 1518 _recursions = 0; // set the recursion level to be 1 1519 exit (true, Self) ; // exit the monitor 1520 guarantee (_owner != Self, "invariant") ; 1521 1522 // The thread is on the WaitSet list - now park() it. 1523 // On MP systems it's conceivable that a brief spin before we park 1524 // could be profitable. 1525 // 1526 // TODO-FIXME: change the following logic to a loop of the form 1527 // while (!timeout && !interrupted && _notified == 0) park() 1528 1529 int ret = OS_OK ; 1530 int WasNotified = 0 ; 1531 { // State transition wrappers 1532 OSThread* osthread = Self->osthread(); 1533 OSThreadWaitState osts(osthread, true); 1534 { 1535 ThreadBlockInVM tbivm(jt); 1536 // Thread is in thread_blocked state and oop access is unsafe. 1537 jt->set_suspend_equivalent(); 1538 1539 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1540 // Intentionally empty 1541 } else 1542 if (node._notified == 0) { 1543 if (millis <= 0) { 1544 Self->_ParkEvent->park () ; 1545 } else { 1546 ret = Self->_ParkEvent->park (millis) ; 1547 } 1548 } 1549 1550 // were we externally suspended while we were waiting? 1551 if (ExitSuspendEquivalent (jt)) { 1552 // TODO-FIXME: add -- if succ == Self then succ = null. 1553 jt->java_suspend_self(); 1554 } 1555 1556 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1557 1558 1559 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1560 // from the WaitSet to the EntryList. 1561 // See if we need to remove Node from the WaitSet. 1562 // We use double-checked locking to avoid grabbing _WaitSetLock 1563 // if the thread is not on the wait queue. 1564 // 1565 // Note that we don't need a fence before the fetch of TState. 1566 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1567 // written by the is thread. (perhaps the fetch might even be satisfied 1568 // by a look-aside into the processor's own store buffer, although given 1569 // the length of the code path between the prior ST and this load that's 1570 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1571 // then we'll acquire the lock and then re-fetch a fresh TState value. 1572 // That is, we fail toward safety. 1573 1574 if (node.TState == ObjectWaiter::TS_WAIT) { 1575 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1576 if (node.TState == ObjectWaiter::TS_WAIT) { 1577 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1578 assert(node._notified == 0, "invariant"); 1579 node.TState = ObjectWaiter::TS_RUN ; 1580 } 1581 Thread::SpinRelease (&_WaitSetLock) ; 1582 } 1583 1584 // The thread is now either on off-list (TS_RUN), 1585 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1586 // The Node's TState variable is stable from the perspective of this thread. 1587 // No other threads will asynchronously modify TState. 1588 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1589 OrderAccess::loadload() ; 1590 if (_succ == Self) _succ = NULL ; 1591 WasNotified = node._notified ; 1592 1593 // Reentry phase -- reacquire the monitor. 1594 // re-enter contended monitor after object.wait(). 1595 // retain OBJECT_WAIT state until re-enter successfully completes 1596 // Thread state is thread_in_vm and oop access is again safe, 1597 // although the raw address of the object may have changed. 1598 // (Don't cache naked oops over safepoints, of course). 1599 1600 // post monitor waited event. Note that this is past-tense, we are done waiting. 1601 if (JvmtiExport::should_post_monitor_waited()) { 1602 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1603 1604 if (node._notified != 0 && _succ == Self) { 1605 // In this part of the monitor wait-notify-reenter protocol it 1606 // is possible (and normal) for another thread to do a fastpath 1607 // monitor enter-exit while this thread is still trying to get 1608 // to the reenter portion of the protocol. 1609 // 1610 // The ObjectMonitor was notified and the current thread is 1611 // the successor which also means that an unpark() has already 1612 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1613 // consume the unpark() that was done when the successor was 1614 // set because the same ParkEvent is shared between Java 1615 // monitors and JVM/TI RawMonitors (for now). 1616 // 1617 // We redo the unpark() to ensure forward progress, i.e., we 1618 // don't want all pending threads hanging (parked) with none 1619 // entering the unlocked monitor. 1620 node._event->unpark(); 1621 } 1622 } 1623 1624 if (event.should_commit()) { 1625 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1626 } 1627 1628 OrderAccess::fence() ; 1629 1630 assert (Self->_Stalled != 0, "invariant") ; 1631 Self->_Stalled = 0 ; 1632 1633 assert (_owner != Self, "invariant") ; 1634 ObjectWaiter::TStates v = node.TState ; 1635 if (v == ObjectWaiter::TS_RUN) { 1636 enter (Self) ; 1637 } else { 1638 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1639 ReenterI (Self, &node) ; 1640 node.wait_reenter_end(this); 1641 } 1642 1643 // Self has reacquired the lock. 1644 // Lifecycle - the node representing Self must not appear on any queues. 1645 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1646 // want residual elements associated with this thread left on any lists. 1647 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1648 assert (_owner == Self, "invariant") ; 1649 assert (_succ != Self , "invariant") ; 1650 } // OSThreadWaitState() 1651 1652 jt->set_current_waiting_monitor(NULL); 1653 1654 guarantee (_recursions == 0, "invariant") ; 1655 _recursions = save; // restore the old recursion count 1656 _waiters--; // decrement the number of waiters 1657 1658 // Verify a few postconditions 1659 assert (_owner == Self , "invariant") ; 1660 assert (_succ != Self , "invariant") ; 1661 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1662 1663 if (SyncFlags & 32) { 1664 OrderAccess::fence() ; 1665 } 1666 1667 // check if the notification happened 1668 if (!WasNotified) { 1669 // no, it could be timeout or Thread.interrupt() or both 1670 // check for interrupt event, otherwise it is timeout 1671 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1672 TEVENT (Wait - throw IEX from epilog) ; 1673 THROW(vmSymbols::java_lang_InterruptedException()); 1674 } 1675 } 1676 1677 // NOTE: Spurious wake up will be consider as timeout. 1678 // Monitor notify has precedence over thread interrupt. 1679 } 1680 1681 1682 // Consider: 1683 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1684 // then instead of transferring a thread from the WaitSet to the EntryList 1685 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1686 1687 void ObjectMonitor::notify(TRAPS) { 1688 CHECK_OWNER(); 1689 if (_WaitSet == NULL) { 1690 TEVENT (Empty-Notify) ; 1691 return ; 1692 } 1693 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1694 1695 int Policy = Knob_MoveNotifyee ; 1696 1697 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1698 ObjectWaiter * iterator = DequeueWaiter() ; 1699 if (iterator != NULL) { 1700 TEVENT (Notify1 - Transfer) ; 1701 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1702 guarantee (iterator->_notified == 0, "invariant") ; 1703 if (Policy != 4) { 1704 iterator->TState = ObjectWaiter::TS_ENTER ; 1705 } 1706 iterator->_notified = 1 ; 1707 Thread * Self = THREAD; 1708 iterator->_notifier_tid = Self->osthread()->thread_id(); 1709 1710 ObjectWaiter * List = _EntryList ; 1711 if (List != NULL) { 1712 assert (List->_prev == NULL, "invariant") ; 1713 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1714 assert (List != iterator, "invariant") ; 1715 } 1716 1717 if (Policy == 0) { // prepend to EntryList 1718 if (List == NULL) { 1719 iterator->_next = iterator->_prev = NULL ; 1720 _EntryList = iterator ; 1721 } else { 1722 List->_prev = iterator ; 1723 iterator->_next = List ; 1724 iterator->_prev = NULL ; 1725 _EntryList = iterator ; 1726 } 1727 } else 1728 if (Policy == 1) { // append to EntryList 1729 if (List == NULL) { 1730 iterator->_next = iterator->_prev = NULL ; 1731 _EntryList = iterator ; 1732 } else { 1733 // CONSIDER: finding the tail currently requires a linear-time walk of 1734 // the EntryList. We can make tail access constant-time by converting to 1735 // a CDLL instead of using our current DLL. 1736 ObjectWaiter * Tail ; 1737 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1738 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1739 Tail->_next = iterator ; 1740 iterator->_prev = Tail ; 1741 iterator->_next = NULL ; 1742 } 1743 } else 1744 if (Policy == 2) { // prepend to cxq 1745 // prepend to cxq 1746 if (List == NULL) { 1747 iterator->_next = iterator->_prev = NULL ; 1748 _EntryList = iterator ; 1749 } else { 1750 iterator->TState = ObjectWaiter::TS_CXQ ; 1751 for (;;) { 1752 ObjectWaiter * Front = _cxq ; 1753 iterator->_next = Front ; 1754 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1755 break ; 1756 } 1757 } 1758 } 1759 } else 1760 if (Policy == 3) { // append to cxq 1761 iterator->TState = ObjectWaiter::TS_CXQ ; 1762 for (;;) { 1763 ObjectWaiter * Tail ; 1764 Tail = _cxq ; 1765 if (Tail == NULL) { 1766 iterator->_next = NULL ; 1767 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1768 break ; 1769 } 1770 } else { 1771 while (Tail->_next != NULL) Tail = Tail->_next ; 1772 Tail->_next = iterator ; 1773 iterator->_prev = Tail ; 1774 iterator->_next = NULL ; 1775 break ; 1776 } 1777 } 1778 } else { 1779 ParkEvent * ev = iterator->_event ; 1780 iterator->TState = ObjectWaiter::TS_RUN ; 1781 OrderAccess::fence() ; 1782 ev->unpark() ; 1783 } 1784 1785 if (Policy < 4) { 1786 iterator->wait_reenter_begin(this); 1787 } 1788 1789 // _WaitSetLock protects the wait queue, not the EntryList. We could 1790 // move the add-to-EntryList operation, above, outside the critical section 1791 // protected by _WaitSetLock. In practice that's not useful. With the 1792 // exception of wait() timeouts and interrupts the monitor owner 1793 // is the only thread that grabs _WaitSetLock. There's almost no contention 1794 // on _WaitSetLock so it's not profitable to reduce the length of the 1795 // critical section. 1796 } 1797 1798 Thread::SpinRelease (&_WaitSetLock) ; 1799 1800 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1801 ObjectMonitor::_sync_Notifications->inc() ; 1802 } 1803 } 1804 1805 1806 void ObjectMonitor::notifyAll(TRAPS) { 1807 CHECK_OWNER(); 1808 ObjectWaiter* iterator; 1809 if (_WaitSet == NULL) { 1810 TEVENT (Empty-NotifyAll) ; 1811 return ; 1812 } 1813 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1814 1815 int Policy = Knob_MoveNotifyee ; 1816 int Tally = 0 ; 1817 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1818 1819 for (;;) { 1820 iterator = DequeueWaiter () ; 1821 if (iterator == NULL) break ; 1822 TEVENT (NotifyAll - Transfer1) ; 1823 ++Tally ; 1824 1825 // Disposition - what might we do with iterator ? 1826 // a. add it directly to the EntryList - either tail or head. 1827 // b. push it onto the front of the _cxq. 1828 // For now we use (a). 1829 1830 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1831 guarantee (iterator->_notified == 0, "invariant") ; 1832 iterator->_notified = 1 ; 1833 Thread * Self = THREAD; 1834 iterator->_notifier_tid = Self->osthread()->thread_id(); 1835 if (Policy != 4) { 1836 iterator->TState = ObjectWaiter::TS_ENTER ; 1837 } 1838 1839 ObjectWaiter * List = _EntryList ; 1840 if (List != NULL) { 1841 assert (List->_prev == NULL, "invariant") ; 1842 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1843 assert (List != iterator, "invariant") ; 1844 } 1845 1846 if (Policy == 0) { // prepend to EntryList 1847 if (List == NULL) { 1848 iterator->_next = iterator->_prev = NULL ; 1849 _EntryList = iterator ; 1850 } else { 1851 List->_prev = iterator ; 1852 iterator->_next = List ; 1853 iterator->_prev = NULL ; 1854 _EntryList = iterator ; 1855 } 1856 } else 1857 if (Policy == 1) { // append to EntryList 1858 if (List == NULL) { 1859 iterator->_next = iterator->_prev = NULL ; 1860 _EntryList = iterator ; 1861 } else { 1862 // CONSIDER: finding the tail currently requires a linear-time walk of 1863 // the EntryList. We can make tail access constant-time by converting to 1864 // a CDLL instead of using our current DLL. 1865 ObjectWaiter * Tail ; 1866 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1867 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1868 Tail->_next = iterator ; 1869 iterator->_prev = Tail ; 1870 iterator->_next = NULL ; 1871 } 1872 } else 1873 if (Policy == 2) { // prepend to cxq 1874 // prepend to cxq 1875 iterator->TState = ObjectWaiter::TS_CXQ ; 1876 for (;;) { 1877 ObjectWaiter * Front = _cxq ; 1878 iterator->_next = Front ; 1879 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1880 break ; 1881 } 1882 } 1883 } else 1884 if (Policy == 3) { // append to cxq 1885 iterator->TState = ObjectWaiter::TS_CXQ ; 1886 for (;;) { 1887 ObjectWaiter * Tail ; 1888 Tail = _cxq ; 1889 if (Tail == NULL) { 1890 iterator->_next = NULL ; 1891 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1892 break ; 1893 } 1894 } else { 1895 while (Tail->_next != NULL) Tail = Tail->_next ; 1896 Tail->_next = iterator ; 1897 iterator->_prev = Tail ; 1898 iterator->_next = NULL ; 1899 break ; 1900 } 1901 } 1902 } else { 1903 ParkEvent * ev = iterator->_event ; 1904 iterator->TState = ObjectWaiter::TS_RUN ; 1905 OrderAccess::fence() ; 1906 ev->unpark() ; 1907 } 1908 1909 if (Policy < 4) { 1910 iterator->wait_reenter_begin(this); 1911 } 1912 1913 // _WaitSetLock protects the wait queue, not the EntryList. We could 1914 // move the add-to-EntryList operation, above, outside the critical section 1915 // protected by _WaitSetLock. In practice that's not useful. With the 1916 // exception of wait() timeouts and interrupts the monitor owner 1917 // is the only thread that grabs _WaitSetLock. There's almost no contention 1918 // on _WaitSetLock so it's not profitable to reduce the length of the 1919 // critical section. 1920 } 1921 1922 Thread::SpinRelease (&_WaitSetLock) ; 1923 1924 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1925 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1926 } 1927 } 1928 1929 // ----------------------------------------------------------------------------- 1930 // Adaptive Spinning Support 1931 // 1932 // Adaptive spin-then-block - rational spinning 1933 // 1934 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1935 // algorithm. On high order SMP systems it would be better to start with 1936 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1937 // a contending thread could enqueue itself on the cxq and then spin locally 1938 // on a thread-specific variable such as its ParkEvent._Event flag. 1939 // That's left as an exercise for the reader. Note that global spinning is 1940 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1941 // low latency and massive bandwidth. 1942 // 1943 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1944 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1945 // (duration) or we can fix the count at approximately the duration of 1946 // a context switch and vary the frequency. Of course we could also 1947 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1948 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1949 // 1950 // This implementation varies the duration "D", where D varies with 1951 // the success rate of recent spin attempts. (D is capped at approximately 1952 // length of a round-trip context switch). The success rate for recent 1953 // spin attempts is a good predictor of the success rate of future spin 1954 // attempts. The mechanism adapts automatically to varying critical 1955 // section length (lock modality), system load and degree of parallelism. 1956 // D is maintained per-monitor in _SpinDuration and is initialized 1957 // optimistically. Spin frequency is fixed at 100%. 1958 // 1959 // Note that _SpinDuration is volatile, but we update it without locks 1960 // or atomics. The code is designed so that _SpinDuration stays within 1961 // a reasonable range even in the presence of races. The arithmetic 1962 // operations on _SpinDuration are closed over the domain of legal values, 1963 // so at worst a race will install and older but still legal value. 1964 // At the very worst this introduces some apparent non-determinism. 1965 // We might spin when we shouldn't or vice-versa, but since the spin 1966 // count are relatively short, even in the worst case, the effect is harmless. 1967 // 1968 // Care must be taken that a low "D" value does not become an 1969 // an absorbing state. Transient spinning failures -- when spinning 1970 // is overall profitable -- should not cause the system to converge 1971 // on low "D" values. We want spinning to be stable and predictable 1972 // and fairly responsive to change and at the same time we don't want 1973 // it to oscillate, become metastable, be "too" non-deterministic, 1974 // or converge on or enter undesirable stable absorbing states. 1975 // 1976 // We implement a feedback-based control system -- using past behavior 1977 // to predict future behavior. We face two issues: (a) if the 1978 // input signal is random then the spin predictor won't provide optimal 1979 // results, and (b) if the signal frequency is too high then the control 1980 // system, which has some natural response lag, will "chase" the signal. 1981 // (b) can arise from multimodal lock hold times. Transient preemption 1982 // can also result in apparent bimodal lock hold times. 1983 // Although sub-optimal, neither condition is particularly harmful, as 1984 // in the worst-case we'll spin when we shouldn't or vice-versa. 1985 // The maximum spin duration is rather short so the failure modes aren't bad. 1986 // To be conservative, I've tuned the gain in system to bias toward 1987 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1988 // "rings" or oscillates between spinning and not spinning. This happens 1989 // when spinning is just on the cusp of profitability, however, so the 1990 // situation is not dire. The state is benign -- there's no need to add 1991 // hysteresis control to damp the transition rate between spinning and 1992 // not spinning. 1993 // 1994 1995 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1996 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1997 1998 // Spinning: Fixed frequency (100%), vary duration 1999 2000 2001 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 2002 2003 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2004 int ctr = Knob_FixedSpin ; 2005 if (ctr != 0) { 2006 while (--ctr >= 0) { 2007 if (TryLock (Self) > 0) return 1 ; 2008 SpinPause () ; 2009 } 2010 return 0 ; 2011 } 2012 2013 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 2014 if (TryLock(Self) > 0) { 2015 // Increase _SpinDuration ... 2016 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2017 // Raising _SpurDuration to the poverty line is key. 2018 int x = _SpinDuration ; 2019 if (x < Knob_SpinLimit) { 2020 if (x < Knob_Poverty) x = Knob_Poverty ; 2021 _SpinDuration = x + Knob_BonusB ; 2022 } 2023 return 1 ; 2024 } 2025 SpinPause () ; 2026 } 2027 2028 // Admission control - verify preconditions for spinning 2029 // 2030 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2031 // becoming an absorbing state. Put another way, we spin briefly to 2032 // sample, just in case the system load, parallelism, contention, or lock 2033 // modality changed. 2034 // 2035 // Consider the following alternative: 2036 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2037 // spin attempt. "Periodically" might mean after a tally of 2038 // the # of failed spin attempts (or iterations) reaches some threshold. 2039 // This takes us into the realm of 1-out-of-N spinning, where we 2040 // hold the duration constant but vary the frequency. 2041 2042 ctr = _SpinDuration ; 2043 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2044 if (ctr <= 0) return 0 ; 2045 2046 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2047 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2048 TEVENT (Spin abort - notrunnable [TOP]); 2049 return 0 ; 2050 } 2051 2052 int MaxSpin = Knob_MaxSpinners ; 2053 if (MaxSpin >= 0) { 2054 if (_Spinner > MaxSpin) { 2055 TEVENT (Spin abort -- too many spinners) ; 2056 return 0 ; 2057 } 2058 // Slighty racy, but benign ... 2059 Adjust (&_Spinner, 1) ; 2060 } 2061 2062 // We're good to spin ... spin ingress. 2063 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2064 // when preparing to LD...CAS _owner, etc and the CAS is likely 2065 // to succeed. 2066 int hits = 0 ; 2067 int msk = 0 ; 2068 int caspty = Knob_CASPenalty ; 2069 int oxpty = Knob_OXPenalty ; 2070 int sss = Knob_SpinSetSucc ; 2071 if (sss && _succ == NULL ) _succ = Self ; 2072 Thread * prv = NULL ; 2073 2074 // There are three ways to exit the following loop: 2075 // 1. A successful spin where this thread has acquired the lock. 2076 // 2. Spin failure with prejudice 2077 // 3. Spin failure without prejudice 2078 2079 while (--ctr >= 0) { 2080 2081 // Periodic polling -- Check for pending GC 2082 // Threads may spin while they're unsafe. 2083 // We don't want spinning threads to delay the JVM from reaching 2084 // a stop-the-world safepoint or to steal cycles from GC. 2085 // If we detect a pending safepoint we abort in order that 2086 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2087 // this thread, if safe, doesn't steal cycles from GC. 2088 // This is in keeping with the "no loitering in runtime" rule. 2089 // We periodically check to see if there's a safepoint pending. 2090 if ((ctr & 0xFF) == 0) { 2091 if (SafepointSynchronize::do_call_back()) { 2092 TEVENT (Spin: safepoint) ; 2093 goto Abort ; // abrupt spin egress 2094 } 2095 if (Knob_UsePause & 1) SpinPause () ; 2096 2097 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2098 if (hits > 50 && scb != NULL) { 2099 int abend = (*scb)(SpinCallbackArgument, 0) ; 2100 } 2101 } 2102 2103 if (Knob_UsePause & 2) SpinPause() ; 2104 2105 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2106 // This is useful on classic SMP systems, but is of less utility on 2107 // N1-style CMT platforms. 2108 // 2109 // Trade-off: lock acquisition latency vs coherency bandwidth. 2110 // Lock hold times are typically short. A histogram 2111 // of successful spin attempts shows that we usually acquire 2112 // the lock early in the spin. That suggests we want to 2113 // sample _owner frequently in the early phase of the spin, 2114 // but then back-off and sample less frequently as the spin 2115 // progresses. The back-off makes a good citizen on SMP big 2116 // SMP systems. Oversampling _owner can consume excessive 2117 // coherency bandwidth. Relatedly, if we _oversample _owner we 2118 // can inadvertently interfere with the the ST m->owner=null. 2119 // executed by the lock owner. 2120 if (ctr & msk) continue ; 2121 ++hits ; 2122 if ((hits & 0xF) == 0) { 2123 // The 0xF, above, corresponds to the exponent. 2124 // Consider: (msk+1)|msk 2125 msk = ((msk << 2)|3) & BackOffMask ; 2126 } 2127 2128 // Probe _owner with TATAS 2129 // If this thread observes the monitor transition or flicker 2130 // from locked to unlocked to locked, then the odds that this 2131 // thread will acquire the lock in this spin attempt go down 2132 // considerably. The same argument applies if the CAS fails 2133 // or if we observe _owner change from one non-null value to 2134 // another non-null value. In such cases we might abort 2135 // the spin without prejudice or apply a "penalty" to the 2136 // spin count-down variable "ctr", reducing it by 100, say. 2137 2138 Thread * ox = (Thread *) _owner ; 2139 if (ox == NULL) { 2140 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2141 if (ox == NULL) { 2142 // The CAS succeeded -- this thread acquired ownership 2143 // Take care of some bookkeeping to exit spin state. 2144 if (sss && _succ == Self) { 2145 _succ = NULL ; 2146 } 2147 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2148 2149 // Increase _SpinDuration : 2150 // The spin was successful (profitable) so we tend toward 2151 // longer spin attempts in the future. 2152 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2153 // If we acquired the lock early in the spin cycle it 2154 // makes sense to increase _SpinDuration proportionally. 2155 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2156 int x = _SpinDuration ; 2157 if (x < Knob_SpinLimit) { 2158 if (x < Knob_Poverty) x = Knob_Poverty ; 2159 _SpinDuration = x + Knob_Bonus ; 2160 } 2161 return 1 ; 2162 } 2163 2164 // The CAS failed ... we can take any of the following actions: 2165 // * penalize: ctr -= Knob_CASPenalty 2166 // * exit spin with prejudice -- goto Abort; 2167 // * exit spin without prejudice. 2168 // * Since CAS is high-latency, retry again immediately. 2169 prv = ox ; 2170 TEVENT (Spin: cas failed) ; 2171 if (caspty == -2) break ; 2172 if (caspty == -1) goto Abort ; 2173 ctr -= caspty ; 2174 continue ; 2175 } 2176 2177 // Did lock ownership change hands ? 2178 if (ox != prv && prv != NULL ) { 2179 TEVENT (spin: Owner changed) 2180 if (oxpty == -2) break ; 2181 if (oxpty == -1) goto Abort ; 2182 ctr -= oxpty ; 2183 } 2184 prv = ox ; 2185 2186 // Abort the spin if the owner is not executing. 2187 // The owner must be executing in order to drop the lock. 2188 // Spinning while the owner is OFFPROC is idiocy. 2189 // Consider: ctr -= RunnablePenalty ; 2190 if (Knob_OState && NotRunnable (Self, ox)) { 2191 TEVENT (Spin abort - notrunnable); 2192 goto Abort ; 2193 } 2194 if (sss && _succ == NULL ) _succ = Self ; 2195 } 2196 2197 // Spin failed with prejudice -- reduce _SpinDuration. 2198 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2199 // AIMD is globally stable. 2200 TEVENT (Spin failure) ; 2201 { 2202 int x = _SpinDuration ; 2203 if (x > 0) { 2204 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2205 // This is globally sample and tends to damp the response. 2206 x -= Knob_Penalty ; 2207 if (x < 0) x = 0 ; 2208 _SpinDuration = x ; 2209 } 2210 } 2211 2212 Abort: 2213 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2214 if (sss && _succ == Self) { 2215 _succ = NULL ; 2216 // Invariant: after setting succ=null a contending thread 2217 // must recheck-retry _owner before parking. This usually happens 2218 // in the normal usage of TrySpin(), but it's safest 2219 // to make TrySpin() as foolproof as possible. 2220 OrderAccess::fence() ; 2221 if (TryLock(Self) > 0) return 1 ; 2222 } 2223 return 0 ; 2224 } 2225 2226 // NotRunnable() -- informed spinning 2227 // 2228 // Don't bother spinning if the owner is not eligible to drop the lock. 2229 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2230 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2231 // The thread must be runnable in order to drop the lock in timely fashion. 2232 // If the _owner is not runnable then spinning will not likely be 2233 // successful (profitable). 2234 // 2235 // Beware -- the thread referenced by _owner could have died 2236 // so a simply fetch from _owner->_thread_state might trap. 2237 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2238 // Because of the lifecycle issues the schedctl and _thread_state values 2239 // observed by NotRunnable() might be garbage. NotRunnable must 2240 // tolerate this and consider the observed _thread_state value 2241 // as advisory. 2242 // 2243 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2244 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2245 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2246 // with the LSB of _owner. Another option would be to probablistically probe 2247 // the putative _owner->TypeTag value. 2248 // 2249 // Checking _thread_state isn't perfect. Even if the thread is 2250 // in_java it might be blocked on a page-fault or have been preempted 2251 // and sitting on a ready/dispatch queue. _thread state in conjunction 2252 // with schedctl.sc_state gives us a good picture of what the 2253 // thread is doing, however. 2254 // 2255 // TODO: check schedctl.sc_state. 2256 // We'll need to use SafeFetch32() to read from the schedctl block. 2257 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2258 // 2259 // The return value from NotRunnable() is *advisory* -- the 2260 // result is based on sampling and is not necessarily coherent. 2261 // The caller must tolerate false-negative and false-positive errors. 2262 // Spinning, in general, is probabilistic anyway. 2263 2264 2265 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2266 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2267 if (!OwnerIsThread) return 0 ; 2268 2269 if (ox == NULL) return 0 ; 2270 2271 // Avoid transitive spinning ... 2272 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2273 // Immediately after T1 acquires L it's possible that T2, also 2274 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2275 // This occurs transiently after T1 acquired L but before 2276 // T1 managed to clear T1.Stalled. T2 does not need to abort 2277 // its spin in this circumstance. 2278 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2279 2280 if (BlockedOn == 1) return 1 ; 2281 if (BlockedOn != 0) { 2282 return BlockedOn != intptr_t(this) && _owner == ox ; 2283 } 2284 2285 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2286 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2287 // consider also: jst != _thread_in_Java -- but that's overspecific. 2288 return jst == _thread_blocked || jst == _thread_in_native ; 2289 } 2290 2291 2292 // ----------------------------------------------------------------------------- 2293 // WaitSet management ... 2294 2295 ObjectWaiter::ObjectWaiter(Thread* thread) { 2296 _next = NULL; 2297 _prev = NULL; 2298 _notified = 0; 2299 TState = TS_RUN ; 2300 _thread = thread; 2301 _event = thread->_ParkEvent ; 2302 _active = false; 2303 assert (_event != NULL, "invariant") ; 2304 } 2305 2306 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2307 JavaThread *jt = (JavaThread *)this->_thread; 2308 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2309 } 2310 2311 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2312 JavaThread *jt = (JavaThread *)this->_thread; 2313 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2314 } 2315 2316 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2317 assert(node != NULL, "should not dequeue NULL node"); 2318 assert(node->_prev == NULL, "node already in list"); 2319 assert(node->_next == NULL, "node already in list"); 2320 // put node at end of queue (circular doubly linked list) 2321 if (_WaitSet == NULL) { 2322 _WaitSet = node; 2323 node->_prev = node; 2324 node->_next = node; 2325 } else { 2326 ObjectWaiter* head = _WaitSet ; 2327 ObjectWaiter* tail = head->_prev; 2328 assert(tail->_next == head, "invariant check"); 2329 tail->_next = node; 2330 head->_prev = node; 2331 node->_next = head; 2332 node->_prev = tail; 2333 } 2334 } 2335 2336 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2337 // dequeue the very first waiter 2338 ObjectWaiter* waiter = _WaitSet; 2339 if (waiter) { 2340 DequeueSpecificWaiter(waiter); 2341 } 2342 return waiter; 2343 } 2344 2345 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2346 assert(node != NULL, "should not dequeue NULL node"); 2347 assert(node->_prev != NULL, "node already removed from list"); 2348 assert(node->_next != NULL, "node already removed from list"); 2349 // when the waiter has woken up because of interrupt, 2350 // timeout or other spurious wake-up, dequeue the 2351 // waiter from waiting list 2352 ObjectWaiter* next = node->_next; 2353 if (next == node) { 2354 assert(node->_prev == node, "invariant check"); 2355 _WaitSet = NULL; 2356 } else { 2357 ObjectWaiter* prev = node->_prev; 2358 assert(prev->_next == node, "invariant check"); 2359 assert(next->_prev == node, "invariant check"); 2360 next->_prev = prev; 2361 prev->_next = next; 2362 if (_WaitSet == node) { 2363 _WaitSet = next; 2364 } 2365 } 2366 node->_next = NULL; 2367 node->_prev = NULL; 2368 } 2369 2370 // ----------------------------------------------------------------------------- 2371 // PerfData support 2372 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2373 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2374 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2375 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2376 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2377 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2378 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2379 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2380 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2381 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2382 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2383 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2384 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2385 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2386 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2387 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2388 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2389 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2390 2391 // One-shot global initialization for the sync subsystem. 2392 // We could also defer initialization and initialize on-demand 2393 // the first time we call inflate(). Initialization would 2394 // be protected - like so many things - by the MonitorCache_lock. 2395 2396 void ObjectMonitor::Initialize () { 2397 static int InitializationCompleted = 0 ; 2398 assert (InitializationCompleted == 0, "invariant") ; 2399 InitializationCompleted = 1 ; 2400 if (UsePerfData) { 2401 EXCEPTION_MARK ; 2402 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2403 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2404 NEWPERFCOUNTER(_sync_Inflations) ; 2405 NEWPERFCOUNTER(_sync_Deflations) ; 2406 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2407 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2408 NEWPERFCOUNTER(_sync_Parks) ; 2409 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2410 NEWPERFCOUNTER(_sync_Notifications) ; 2411 NEWPERFCOUNTER(_sync_SlowEnter) ; 2412 NEWPERFCOUNTER(_sync_SlowExit) ; 2413 NEWPERFCOUNTER(_sync_SlowNotify) ; 2414 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2415 NEWPERFCOUNTER(_sync_FailedSpins) ; 2416 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2417 NEWPERFCOUNTER(_sync_PrivateA) ; 2418 NEWPERFCOUNTER(_sync_PrivateB) ; 2419 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2420 NEWPERFCOUNTER(_sync_MonScavenged) ; 2421 NEWPERFVARIABLE(_sync_MonExtant) ; 2422 #undef NEWPERFCOUNTER 2423 } 2424 } 2425 2426 2427 // Compile-time asserts 2428 // When possible, it's better to catch errors deterministically at 2429 // compile-time than at runtime. The down-side to using compile-time 2430 // asserts is that error message -- often something about negative array 2431 // indices -- is opaque. 2432 2433 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2434 2435 void ObjectMonitor::ctAsserts() { 2436 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2437 } 2438 2439 2440 static char * kvGet (char * kvList, const char * Key) { 2441 if (kvList == NULL) return NULL ; 2442 size_t n = strlen (Key) ; 2443 char * Search ; 2444 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2445 if (strncmp (Search, Key, n) == 0) { 2446 if (Search[n] == '=') return Search + n + 1 ; 2447 if (Search[n] == 0) return (char *) "1" ; 2448 } 2449 } 2450 return NULL ; 2451 } 2452 2453 static int kvGetInt (char * kvList, const char * Key, int Default) { 2454 char * v = kvGet (kvList, Key) ; 2455 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2456 if (Knob_ReportSettings && v != NULL) { 2457 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2458 ::fflush (stdout) ; 2459 } 2460 return rslt ; 2461 } 2462 2463 void ObjectMonitor::DeferredInitialize () { 2464 if (InitDone > 0) return ; 2465 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2466 while (InitDone != 1) ; 2467 return ; 2468 } 2469 2470 // One-shot global initialization ... 2471 // The initialization is idempotent, so we don't need locks. 2472 // In the future consider doing this via os::init_2(). 2473 // SyncKnobs consist of <Key>=<Value> pairs in the style 2474 // of environment variables. Start by converting ':' to NUL. 2475 2476 if (SyncKnobs == NULL) SyncKnobs = "" ; 2477 2478 size_t sz = strlen (SyncKnobs) ; 2479 char * knobs = (char *) malloc (sz + 2) ; 2480 if (knobs == NULL) { 2481 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; 2482 guarantee (0, "invariant") ; 2483 } 2484 strcpy (knobs, SyncKnobs) ; 2485 knobs[sz+1] = 0 ; 2486 for (char * p = knobs ; *p ; p++) { 2487 if (*p == ':') *p = 0 ; 2488 } 2489 2490 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2491 SETKNOB(ReportSettings) ; 2492 SETKNOB(Verbose) ; 2493 SETKNOB(FixedSpin) ; 2494 SETKNOB(SpinLimit) ; 2495 SETKNOB(SpinBase) ; 2496 SETKNOB(SpinBackOff); 2497 SETKNOB(CASPenalty) ; 2498 SETKNOB(OXPenalty) ; 2499 SETKNOB(LogSpins) ; 2500 SETKNOB(SpinSetSucc) ; 2501 SETKNOB(SuccEnabled) ; 2502 SETKNOB(SuccRestrict) ; 2503 SETKNOB(Penalty) ; 2504 SETKNOB(Bonus) ; 2505 SETKNOB(BonusB) ; 2506 SETKNOB(Poverty) ; 2507 SETKNOB(SpinAfterFutile) ; 2508 SETKNOB(UsePause) ; 2509 SETKNOB(SpinEarly) ; 2510 SETKNOB(OState) ; 2511 SETKNOB(MaxSpinners) ; 2512 SETKNOB(PreSpin) ; 2513 SETKNOB(ExitPolicy) ; 2514 SETKNOB(QMode); 2515 SETKNOB(ResetEvent) ; 2516 SETKNOB(MoveNotifyee) ; 2517 SETKNOB(FastHSSEC) ; 2518 #undef SETKNOB 2519 2520 if (os::is_MP()) { 2521 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2522 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2523 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2524 } else { 2525 Knob_SpinLimit = 0 ; 2526 Knob_SpinBase = 0 ; 2527 Knob_PreSpin = 0 ; 2528 Knob_FixedSpin = -1 ; 2529 } 2530 2531 if (Knob_LogSpins == 0) { 2532 ObjectMonitor::_sync_FailedSpins = NULL ; 2533 } 2534 2535 free (knobs) ; 2536 OrderAccess::fence() ; 2537 InitDone = 1 ; 2538 } 2539 2540 #ifndef PRODUCT 2541 void ObjectMonitor::verify() { 2542 } 2543 2544 void ObjectMonitor::print() { 2545 } 2546 #endif