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