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.hpp" 38 #include "services/threadService.hpp" 39 #include "trace/tracing.hpp" 40 #include "trace/traceMacros.hpp" 41 #include "utilities/dtrace.hpp" 42 #include "utilities/preserveException.hpp" 43 #ifdef TARGET_OS_FAMILY_linux 44 # include "os_linux.inline.hpp" 45 # include "thread_linux.inline.hpp" 46 #endif 47 #ifdef TARGET_OS_FAMILY_solaris 48 # include "os_solaris.inline.hpp" 49 # include "thread_solaris.inline.hpp" 50 #endif 51 #ifdef TARGET_OS_FAMILY_windows 52 # include "os_windows.inline.hpp" 53 # include "thread_windows.inline.hpp" 54 #endif 55 #ifdef TARGET_OS_FAMILY_bsd 56 # include "os_bsd.inline.hpp" 57 # include "thread_bsd.inline.hpp" 58 #endif 59 60 #if defined(__GNUC__) && !defined(IA64) 61 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 62 #define ATTR __attribute__((noinline)) 63 #else 64 #define ATTR 65 #endif 66 67 68 #ifdef DTRACE_ENABLED 69 70 // Only bother with this argument setup if dtrace is available 71 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 72 73 74 #define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \ 75 char* bytes = NULL; \ 76 int len = 0; \ 77 jlong jtid = SharedRuntime::get_java_tid(thread); \ 78 Symbol* klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \ 79 if (klassname != NULL) { \ 80 bytes = (char*)klassname->bytes(); \ 81 len = klassname->utf8_length(); \ 82 } 83 84 #ifndef USDT2 85 86 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 87 jlong, uintptr_t, char*, int); 88 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 89 jlong, uintptr_t, char*, int); 90 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 91 jlong, uintptr_t, char*, int); 92 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 93 jlong, uintptr_t, char*, int); 94 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 95 jlong, uintptr_t, char*, int); 96 97 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ 98 { \ 99 if (DTraceMonitorProbes) { \ 100 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 101 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 102 (monitor), bytes, len, (millis)); \ 103 } \ 104 } 105 106 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ 107 { \ 108 if (DTraceMonitorProbes) { \ 109 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 110 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 111 (uintptr_t)(monitor), bytes, len); \ 112 } \ 113 } 114 115 #else /* USDT2 */ 116 117 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ 118 { \ 119 if (DTraceMonitorProbes) { \ 120 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 121 HOTSPOT_MONITOR_WAIT(jtid, \ 122 (monitor), bytes, len, (millis)); \ 123 } \ 124 } 125 126 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 127 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 128 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 129 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 130 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 131 132 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ 133 { \ 134 if (DTraceMonitorProbes) { \ 135 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 136 HOTSPOT_MONITOR_##probe(jtid, \ 137 (uintptr_t)(monitor), bytes, len); \ 138 } \ 139 } 140 141 #endif /* USDT2 */ 142 #else // ndef DTRACE_ENABLED 143 144 #define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;} 145 #define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;} 146 147 #endif // ndef DTRACE_ENABLED 148 149 // Tunables ... 150 // The knob* variables are effectively final. Once set they should 151 // never be modified hence. Consider using __read_mostly with GCC. 152 153 int ObjectMonitor::Knob_Verbose = 0 ; 154 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 155 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 156 static int Knob_HandOff = 0 ; 157 static int Knob_ReportSettings = 0 ; 158 159 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 160 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 161 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 162 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 163 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 164 static int Knob_SpinEarly = 1 ; 165 static int Knob_SuccEnabled = 1 ; // futile wake throttling 166 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 167 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 168 static int Knob_Bonus = 100 ; // spin success bonus 169 static int Knob_BonusB = 100 ; // spin success bonus 170 static int Knob_Penalty = 200 ; // spin failure penalty 171 static int Knob_Poverty = 1000 ; 172 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 173 static int Knob_FixedSpin = 0 ; 174 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 175 static int Knob_UsePause = 1 ; 176 static int Knob_ExitPolicy = 0 ; 177 static int Knob_PreSpin = 10 ; // 20-100 likely better 178 static int Knob_ResetEvent = 0 ; 179 static int BackOffMask = 0 ; 180 181 static int Knob_FastHSSEC = 0 ; 182 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 183 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 184 static volatile int InitDone = 0 ; 185 186 #define TrySpin TrySpin_VaryDuration 187 188 // ----------------------------------------------------------------------------- 189 // Theory of operations -- Monitors lists, thread residency, etc: 190 // 191 // * A thread acquires ownership of a monitor by successfully 192 // CAS()ing the _owner field from null to non-null. 193 // 194 // * Invariant: A thread appears on at most one monitor list -- 195 // cxq, EntryList or WaitSet -- at any one time. 196 // 197 // * Contending threads "push" themselves onto the cxq with CAS 198 // and then spin/park. 199 // 200 // * After a contending thread eventually acquires the lock it must 201 // dequeue itself from either the EntryList or the cxq. 202 // 203 // * The exiting thread identifies and unparks an "heir presumptive" 204 // tentative successor thread on the EntryList. Critically, the 205 // exiting thread doesn't unlink the successor thread from the EntryList. 206 // After having been unparked, the wakee will recontend for ownership of 207 // the monitor. The successor (wakee) will either acquire the lock or 208 // re-park itself. 209 // 210 // Succession is provided for by a policy of competitive handoff. 211 // The exiting thread does _not_ grant or pass ownership to the 212 // successor thread. (This is also referred to as "handoff" succession"). 213 // Instead the exiting thread releases ownership and possibly wakes 214 // a successor, so the successor can (re)compete for ownership of the lock. 215 // If the EntryList is empty but the cxq is populated the exiting 216 // thread will drain the cxq into the EntryList. It does so by 217 // by detaching the cxq (installing null with CAS) and folding 218 // the threads from the cxq into the EntryList. The EntryList is 219 // doubly linked, while the cxq is singly linked because of the 220 // CAS-based "push" used to enqueue recently arrived threads (RATs). 221 // 222 // * Concurrency invariants: 223 // 224 // -- only the monitor owner may access or mutate the EntryList. 225 // The mutex property of the monitor itself protects the EntryList 226 // from concurrent interference. 227 // -- Only the monitor owner may detach the cxq. 228 // 229 // * The monitor entry list operations avoid locks, but strictly speaking 230 // they're not lock-free. Enter is lock-free, exit is not. 231 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 232 // 233 // * The cxq can have multiple concurrent "pushers" but only one concurrent 234 // detaching thread. This mechanism is immune from the ABA corruption. 235 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 236 // 237 // * Taken together, the cxq and the EntryList constitute or form a 238 // single logical queue of threads stalled trying to acquire the lock. 239 // We use two distinct lists to improve the odds of a constant-time 240 // dequeue operation after acquisition (in the ::enter() epilog) and 241 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 242 // A key desideratum is to minimize queue & monitor metadata manipulation 243 // that occurs while holding the monitor lock -- that is, we want to 244 // minimize monitor lock holds times. Note that even a small amount of 245 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 246 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 247 // locks and monitor metadata. 248 // 249 // Cxq points to the the set of Recently Arrived Threads attempting entry. 250 // Because we push threads onto _cxq with CAS, the RATs must take the form of 251 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 252 // the unlocking thread notices that EntryList is null but _cxq is != null. 253 // 254 // The EntryList is ordered by the prevailing queue discipline and 255 // can be organized in any convenient fashion, such as a doubly-linked list or 256 // a circular doubly-linked list. Critically, we want insert and delete operations 257 // to operate in constant-time. If we need a priority queue then something akin 258 // to Solaris' sleepq would work nicely. Viz., 259 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 260 // Queue discipline is enforced at ::exit() time, when the unlocking thread 261 // drains the cxq into the EntryList, and orders or reorders the threads on the 262 // EntryList accordingly. 263 // 264 // Barring "lock barging", this mechanism provides fair cyclic ordering, 265 // somewhat similar to an elevator-scan. 266 // 267 // * The monitor synchronization subsystem avoids the use of native 268 // synchronization primitives except for the narrow platform-specific 269 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 270 // the semantics of park-unpark. Put another way, this monitor implementation 271 // depends only on atomic operations and park-unpark. The monitor subsystem 272 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 273 // underlying OS manages the READY<->RUN transitions. 274 // 275 // * Waiting threads reside on the WaitSet list -- wait() puts 276 // the caller onto the WaitSet. 277 // 278 // * notify() or notifyAll() simply transfers threads from the WaitSet to 279 // either the EntryList or cxq. Subsequent exit() operations will 280 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 281 // it's likely the notifyee would simply impale itself on the lock held 282 // by the notifier. 283 // 284 // * An interesting alternative is to encode cxq as (List,LockByte) where 285 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 286 // variable, like _recursions, in the scheme. The threads or Events that form 287 // the list would have to be aligned in 256-byte addresses. A thread would 288 // try to acquire the lock or enqueue itself with CAS, but exiting threads 289 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 290 // Note that is is *not* word-tearing, but it does presume that full-word 291 // CAS operations are coherent with intermix with STB operations. That's true 292 // on most common processors. 293 // 294 // * See also http://blogs.sun.com/dave 295 296 297 // ----------------------------------------------------------------------------- 298 // Enter support 299 300 bool ObjectMonitor::try_enter(Thread* THREAD) { 301 if (THREAD != _owner) { 302 if (THREAD->is_lock_owned ((address)_owner)) { 303 assert(_recursions == 0, "internal state error"); 304 _owner = THREAD ; 305 _recursions = 1 ; 306 OwnerIsThread = 1 ; 307 return true; 308 } 309 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 310 return false; 311 } 312 return true; 313 } else { 314 _recursions++; 315 return true; 316 } 317 } 318 319 void ATTR ObjectMonitor::enter(TRAPS) { 320 // The following code is ordered to check the most common cases first 321 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 322 Thread * const Self = THREAD ; 323 void * cur ; 324 325 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 326 if (cur == NULL) { 327 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 328 assert (_recursions == 0 , "invariant") ; 329 assert (_owner == Self, "invariant") ; 330 // CONSIDER: set or assert OwnerIsThread == 1 331 return ; 332 } 333 334 if (cur == Self) { 335 // TODO-FIXME: check for integer overflow! BUGID 6557169. 336 _recursions ++ ; 337 return ; 338 } 339 340 if (Self->is_lock_owned ((address)cur)) { 341 assert (_recursions == 0, "internal state error"); 342 _recursions = 1 ; 343 // Commute owner from a thread-specific on-stack BasicLockObject address to 344 // a full-fledged "Thread *". 345 _owner = Self ; 346 OwnerIsThread = 1 ; 347 return ; 348 } 349 350 // We've encountered genuine contention. 351 assert (Self->_Stalled == 0, "invariant") ; 352 Self->_Stalled = intptr_t(this) ; 353 354 // Try one round of spinning *before* enqueueing Self 355 // and before going through the awkward and expensive state 356 // transitions. The following spin is strictly optional ... 357 // Note that if we acquire the monitor from an initial spin 358 // we forgo posting JVMTI events and firing DTRACE probes. 359 if (Knob_SpinEarly && TrySpin (Self) > 0) { 360 assert (_owner == Self , "invariant") ; 361 assert (_recursions == 0 , "invariant") ; 362 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 363 Self->_Stalled = 0 ; 364 return ; 365 } 366 367 assert (_owner != Self , "invariant") ; 368 assert (_succ != Self , "invariant") ; 369 assert (Self->is_Java_thread() , "invariant") ; 370 JavaThread * jt = (JavaThread *) Self ; 371 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 372 assert (jt->thread_state() != _thread_blocked , "invariant") ; 373 assert (this->object() != NULL , "invariant") ; 374 assert (_count >= 0, "invariant") ; 375 376 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 377 // Ensure the object-monitor relationship remains stable while there's contention. 378 Atomic::inc_ptr(&_count); 379 380 EventJavaMonitorEnter event; 381 382 { // Change java thread status to indicate blocked on monitor enter. 383 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 384 385 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 386 if (JvmtiExport::should_post_monitor_contended_enter()) { 387 JvmtiExport::post_monitor_contended_enter(jt, this); 388 } 389 390 OSThreadContendState osts(Self->osthread()); 391 ThreadBlockInVM tbivm(jt); 392 393 Self->set_current_pending_monitor(this); 394 395 // TODO-FIXME: change the following for(;;) loop to straight-line code. 396 for (;;) { 397 jt->set_suspend_equivalent(); 398 // cleared by handle_special_suspend_equivalent_condition() 399 // or java_suspend_self() 400 401 EnterI (THREAD) ; 402 403 if (!ExitSuspendEquivalent(jt)) break ; 404 405 // 406 // We have acquired the contended monitor, but while we were 407 // waiting another thread suspended us. We don't want to enter 408 // the monitor while suspended because that would surprise the 409 // thread that suspended us. 410 // 411 _recursions = 0 ; 412 _succ = NULL ; 413 exit (false, Self) ; 414 415 jt->java_suspend_self(); 416 } 417 Self->set_current_pending_monitor(NULL); 418 } 419 420 Atomic::dec_ptr(&_count); 421 assert (_count >= 0, "invariant") ; 422 Self->_Stalled = 0 ; 423 424 // Must either set _recursions = 0 or ASSERT _recursions == 0. 425 assert (_recursions == 0 , "invariant") ; 426 assert (_owner == Self , "invariant") ; 427 assert (_succ != Self , "invariant") ; 428 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 429 430 // The thread -- now the owner -- is back in vm mode. 431 // Report the glorious news via TI,DTrace and jvmstat. 432 // The probe effect is non-trivial. All the reportage occurs 433 // while we hold the monitor, increasing the length of the critical 434 // section. Amdahl's parallel speedup law comes vividly into play. 435 // 436 // Another option might be to aggregate the events (thread local or 437 // per-monitor aggregation) and defer reporting until a more opportune 438 // time -- such as next time some thread encounters contention but has 439 // yet to acquire the lock. While spinning that thread could 440 // spinning we could increment JVMStat counters, etc. 441 442 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 443 if (JvmtiExport::should_post_monitor_contended_entered()) { 444 JvmtiExport::post_monitor_contended_entered(jt, this); 445 } 446 447 if (event.should_commit()) { 448 event.set_klass(((oop)this->object())->klass()); 449 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 450 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 451 event.commit(); 452 } 453 454 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 455 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 456 } 457 } 458 459 460 // Caveat: TryLock() is not necessarily serializing if it returns failure. 461 // Callers must compensate as needed. 462 463 int ObjectMonitor::TryLock (Thread * Self) { 464 for (;;) { 465 void * own = _owner ; 466 if (own != NULL) return 0 ; 467 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 468 // Either guarantee _recursions == 0 or set _recursions = 0. 469 assert (_recursions == 0, "invariant") ; 470 assert (_owner == Self, "invariant") ; 471 // CONSIDER: set or assert that OwnerIsThread == 1 472 return 1 ; 473 } 474 // The lock had been free momentarily, but we lost the race to the lock. 475 // Interference -- the CAS failed. 476 // We can either return -1 or retry. 477 // Retry doesn't make as much sense because the lock was just acquired. 478 if (true) return -1 ; 479 } 480 } 481 482 void ATTR ObjectMonitor::EnterI (TRAPS) { 483 Thread * Self = THREAD ; 484 assert (Self->is_Java_thread(), "invariant") ; 485 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 486 487 // Try the lock - TATAS 488 if (TryLock (Self) > 0) { 489 assert (_succ != Self , "invariant") ; 490 assert (_owner == Self , "invariant") ; 491 assert (_Responsible != Self , "invariant") ; 492 return ; 493 } 494 495 DeferredInitialize () ; 496 497 // We try one round of spinning *before* enqueueing Self. 498 // 499 // If the _owner is ready but OFFPROC we could use a YieldTo() 500 // operation to donate the remainder of this thread's quantum 501 // to the owner. This has subtle but beneficial affinity 502 // effects. 503 504 if (TrySpin (Self) > 0) { 505 assert (_owner == Self , "invariant") ; 506 assert (_succ != Self , "invariant") ; 507 assert (_Responsible != Self , "invariant") ; 508 return ; 509 } 510 511 // The Spin failed -- Enqueue and park the thread ... 512 assert (_succ != Self , "invariant") ; 513 assert (_owner != Self , "invariant") ; 514 assert (_Responsible != Self , "invariant") ; 515 516 // Enqueue "Self" on ObjectMonitor's _cxq. 517 // 518 // Node acts as a proxy for Self. 519 // As an aside, if were to ever rewrite the synchronization code mostly 520 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 521 // Java objects. This would avoid awkward lifecycle and liveness issues, 522 // as well as eliminate a subset of ABA issues. 523 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 524 // 525 526 ObjectWaiter node(Self) ; 527 Self->_ParkEvent->reset() ; 528 node._prev = (ObjectWaiter *) 0xBAD ; 529 node.TState = ObjectWaiter::TS_CXQ ; 530 531 // Push "Self" onto the front of the _cxq. 532 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 533 // Note that spinning tends to reduce the rate at which threads 534 // enqueue and dequeue on EntryList|cxq. 535 ObjectWaiter * nxt ; 536 for (;;) { 537 node._next = nxt = _cxq ; 538 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 539 540 // Interference - the CAS failed because _cxq changed. Just retry. 541 // As an optional optimization we retry the lock. 542 if (TryLock (Self) > 0) { 543 assert (_succ != Self , "invariant") ; 544 assert (_owner == Self , "invariant") ; 545 assert (_Responsible != Self , "invariant") ; 546 return ; 547 } 548 } 549 550 // Check for cxq|EntryList edge transition to non-null. This indicates 551 // the onset of contention. While contention persists exiting threads 552 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 553 // operations revert to the faster 1-0 mode. This enter operation may interleave 554 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 555 // arrange for one of the contending thread to use a timed park() operations 556 // to detect and recover from the race. (Stranding is form of progress failure 557 // where the monitor is unlocked but all the contending threads remain parked). 558 // That is, at least one of the contended threads will periodically poll _owner. 559 // One of the contending threads will become the designated "Responsible" thread. 560 // The Responsible thread uses a timed park instead of a normal indefinite park 561 // operation -- it periodically wakes and checks for and recovers from potential 562 // strandings admitted by 1-0 exit operations. We need at most one Responsible 563 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 564 // be responsible for a monitor. 565 // 566 // Currently, one of the contended threads takes on the added role of "Responsible". 567 // A viable alternative would be to use a dedicated "stranding checker" thread 568 // that periodically iterated over all the threads (or active monitors) and unparked 569 // successors where there was risk of stranding. This would help eliminate the 570 // timer scalability issues we see on some platforms as we'd only have one thread 571 // -- the checker -- parked on a timer. 572 573 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 574 // Try to assume the role of responsible thread for the monitor. 575 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 576 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 577 } 578 579 // The lock have been released while this thread was occupied queueing 580 // itself onto _cxq. To close the race and avoid "stranding" and 581 // progress-liveness failure we must resample-retry _owner before parking. 582 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 583 // In this case the ST-MEMBAR is accomplished with CAS(). 584 // 585 // TODO: Defer all thread state transitions until park-time. 586 // Since state transitions are heavy and inefficient we'd like 587 // to defer the state transitions until absolutely necessary, 588 // and in doing so avoid some transitions ... 589 590 TEVENT (Inflated enter - Contention) ; 591 int nWakeups = 0 ; 592 int RecheckInterval = 1 ; 593 594 for (;;) { 595 596 if (TryLock (Self) > 0) break ; 597 assert (_owner != Self, "invariant") ; 598 599 if ((SyncFlags & 2) && _Responsible == NULL) { 600 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 601 } 602 603 // park self 604 if (_Responsible == Self || (SyncFlags & 1)) { 605 TEVENT (Inflated enter - park TIMED) ; 606 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 607 // Increase the RecheckInterval, but clamp the value. 608 RecheckInterval *= 8 ; 609 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 610 } else { 611 TEVENT (Inflated enter - park UNTIMED) ; 612 Self->_ParkEvent->park() ; 613 } 614 615 if (TryLock(Self) > 0) break ; 616 617 // The lock is still contested. 618 // Keep a tally of the # of futile wakeups. 619 // Note that the counter is not protected by a lock or updated by atomics. 620 // That is by design - we trade "lossy" counters which are exposed to 621 // races during updates for a lower probe effect. 622 TEVENT (Inflated enter - Futile wakeup) ; 623 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 624 ObjectMonitor::_sync_FutileWakeups->inc() ; 625 } 626 ++ nWakeups ; 627 628 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 629 // We can defer clearing _succ until after the spin completes 630 // TrySpin() must tolerate being called with _succ == Self. 631 // Try yet another round of adaptive spinning. 632 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 633 634 // We can find that we were unpark()ed and redesignated _succ while 635 // we were spinning. That's harmless. If we iterate and call park(), 636 // park() will consume the event and return immediately and we'll 637 // just spin again. This pattern can repeat, leaving _succ to simply 638 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 639 // Alternately, we can sample fired() here, and if set, forgo spinning 640 // in the next iteration. 641 642 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 643 Self->_ParkEvent->reset() ; 644 OrderAccess::fence() ; 645 } 646 if (_succ == Self) _succ = NULL ; 647 648 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 649 OrderAccess::fence() ; 650 } 651 652 // Egress : 653 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 654 // Normally we'll find Self on the EntryList . 655 // From the perspective of the lock owner (this thread), the 656 // EntryList is stable and cxq is prepend-only. 657 // The head of cxq is volatile but the interior is stable. 658 // In addition, Self.TState is stable. 659 660 assert (_owner == Self , "invariant") ; 661 assert (object() != NULL , "invariant") ; 662 // I'd like to write: 663 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 664 // but as we're at a safepoint that's not safe. 665 666 UnlinkAfterAcquire (Self, &node) ; 667 if (_succ == Self) _succ = NULL ; 668 669 assert (_succ != Self, "invariant") ; 670 if (_Responsible == Self) { 671 _Responsible = NULL ; 672 OrderAccess::fence(); // Dekker pivot-point 673 674 // We may leave threads on cxq|EntryList without a designated 675 // "Responsible" thread. This is benign. When this thread subsequently 676 // exits the monitor it can "see" such preexisting "old" threads -- 677 // threads that arrived on the cxq|EntryList before the fence, above -- 678 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 679 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 680 // non-null and elect a new "Responsible" timer thread. 681 // 682 // This thread executes: 683 // ST Responsible=null; MEMBAR (in enter epilog - here) 684 // LD cxq|EntryList (in subsequent exit) 685 // 686 // Entering threads in the slow/contended path execute: 687 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 688 // The (ST cxq; MEMBAR) is accomplished with CAS(). 689 // 690 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 691 // exit operation from floating above the ST Responsible=null. 692 } 693 694 // We've acquired ownership with CAS(). 695 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 696 // But since the CAS() this thread may have also stored into _succ, 697 // EntryList, cxq or Responsible. These meta-data updates must be 698 // visible __before this thread subsequently drops the lock. 699 // Consider what could occur if we didn't enforce this constraint -- 700 // STs to monitor meta-data and user-data could reorder with (become 701 // visible after) the ST in exit that drops ownership of the lock. 702 // Some other thread could then acquire the lock, but observe inconsistent 703 // or old monitor meta-data and heap data. That violates the JMM. 704 // To that end, the 1-0 exit() operation must have at least STST|LDST 705 // "release" barrier semantics. Specifically, there must be at least a 706 // STST|LDST barrier in exit() before the ST of null into _owner that drops 707 // the lock. The barrier ensures that changes to monitor meta-data and data 708 // protected by the lock will be visible before we release the lock, and 709 // therefore before some other thread (CPU) has a chance to acquire the lock. 710 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 711 // 712 // Critically, any prior STs to _succ or EntryList must be visible before 713 // the ST of null into _owner in the *subsequent* (following) corresponding 714 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 715 // execute a serializing instruction. 716 717 if (SyncFlags & 8) { 718 OrderAccess::fence() ; 719 } 720 return ; 721 } 722 723 // ReenterI() is a specialized inline form of the latter half of the 724 // contended slow-path from EnterI(). We use ReenterI() only for 725 // monitor reentry in wait(). 726 // 727 // In the future we should reconcile EnterI() and ReenterI(), adding 728 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 729 // loop accordingly. 730 731 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 732 assert (Self != NULL , "invariant") ; 733 assert (SelfNode != NULL , "invariant") ; 734 assert (SelfNode->_thread == Self , "invariant") ; 735 assert (_waiters > 0 , "invariant") ; 736 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 737 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 738 JavaThread * jt = (JavaThread *) Self ; 739 740 int nWakeups = 0 ; 741 for (;;) { 742 ObjectWaiter::TStates v = SelfNode->TState ; 743 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 744 assert (_owner != Self, "invariant") ; 745 746 if (TryLock (Self) > 0) break ; 747 if (TrySpin (Self) > 0) break ; 748 749 TEVENT (Wait Reentry - parking) ; 750 751 // State transition wrappers around park() ... 752 // ReenterI() wisely defers state transitions until 753 // it's clear we must park the thread. 754 { 755 OSThreadContendState osts(Self->osthread()); 756 ThreadBlockInVM tbivm(jt); 757 758 // cleared by handle_special_suspend_equivalent_condition() 759 // or java_suspend_self() 760 jt->set_suspend_equivalent(); 761 if (SyncFlags & 1) { 762 Self->_ParkEvent->park ((jlong)1000) ; 763 } else { 764 Self->_ParkEvent->park () ; 765 } 766 767 // were we externally suspended while we were waiting? 768 for (;;) { 769 if (!ExitSuspendEquivalent (jt)) break ; 770 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 771 jt->java_suspend_self(); 772 jt->set_suspend_equivalent(); 773 } 774 } 775 776 // Try again, but just so we distinguish between futile wakeups and 777 // successful wakeups. The following test isn't algorithmically 778 // necessary, but it helps us maintain sensible statistics. 779 if (TryLock(Self) > 0) break ; 780 781 // The lock is still contested. 782 // Keep a tally of the # of futile wakeups. 783 // Note that the counter is not protected by a lock or updated by atomics. 784 // That is by design - we trade "lossy" counters which are exposed to 785 // races during updates for a lower probe effect. 786 TEVENT (Wait Reentry - futile wakeup) ; 787 ++ nWakeups ; 788 789 // Assuming this is not a spurious wakeup we'll normally 790 // find that _succ == Self. 791 if (_succ == Self) _succ = NULL ; 792 793 // Invariant: after clearing _succ a contending thread 794 // *must* retry _owner before parking. 795 OrderAccess::fence() ; 796 797 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 798 ObjectMonitor::_sync_FutileWakeups->inc() ; 799 } 800 } 801 802 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 803 // Normally we'll find Self on the EntryList. 804 // Unlinking from the EntryList is constant-time and atomic-free. 805 // From the perspective of the lock owner (this thread), the 806 // EntryList is stable and cxq is prepend-only. 807 // The head of cxq is volatile but the interior is stable. 808 // In addition, Self.TState is stable. 809 810 assert (_owner == Self, "invariant") ; 811 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 812 UnlinkAfterAcquire (Self, SelfNode) ; 813 if (_succ == Self) _succ = NULL ; 814 assert (_succ != Self, "invariant") ; 815 SelfNode->TState = ObjectWaiter::TS_RUN ; 816 OrderAccess::fence() ; // see comments at the end of EnterI() 817 } 818 819 // after the thread acquires the lock in ::enter(). Equally, we could defer 820 // unlinking the thread until ::exit()-time. 821 822 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 823 { 824 assert (_owner == Self, "invariant") ; 825 assert (SelfNode->_thread == Self, "invariant") ; 826 827 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 828 // Normal case: remove Self from the DLL EntryList . 829 // This is a constant-time operation. 830 ObjectWaiter * nxt = SelfNode->_next ; 831 ObjectWaiter * prv = SelfNode->_prev ; 832 if (nxt != NULL) nxt->_prev = prv ; 833 if (prv != NULL) prv->_next = nxt ; 834 if (SelfNode == _EntryList ) _EntryList = nxt ; 835 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 836 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 837 TEVENT (Unlink from EntryList) ; 838 } else { 839 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 840 // Inopportune interleaving -- Self is still on the cxq. 841 // This usually means the enqueue of self raced an exiting thread. 842 // Normally we'll find Self near the front of the cxq, so 843 // dequeueing is typically fast. If needbe we can accelerate 844 // this with some MCS/CHL-like bidirectional list hints and advisory 845 // back-links so dequeueing from the interior will normally operate 846 // in constant-time. 847 // Dequeue Self from either the head (with CAS) or from the interior 848 // with a linear-time scan and normal non-atomic memory operations. 849 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 850 // and then unlink Self from EntryList. We have to drain eventually, 851 // so it might as well be now. 852 853 ObjectWaiter * v = _cxq ; 854 assert (v != NULL, "invariant") ; 855 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 856 // The CAS above can fail from interference IFF a "RAT" arrived. 857 // In that case Self must be in the interior and can no longer be 858 // at the head of cxq. 859 if (v == SelfNode) { 860 assert (_cxq != v, "invariant") ; 861 v = _cxq ; // CAS above failed - start scan at head of list 862 } 863 ObjectWaiter * p ; 864 ObjectWaiter * q = NULL ; 865 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 866 q = p ; 867 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 868 } 869 assert (v != SelfNode, "invariant") ; 870 assert (p == SelfNode, "Node not found on cxq") ; 871 assert (p != _cxq, "invariant") ; 872 assert (q != NULL, "invariant") ; 873 assert (q->_next == p, "invariant") ; 874 q->_next = p->_next ; 875 } 876 TEVENT (Unlink from cxq) ; 877 } 878 879 // Diagnostic hygiene ... 880 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 881 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 882 SelfNode->TState = ObjectWaiter::TS_RUN ; 883 } 884 885 // ----------------------------------------------------------------------------- 886 // Exit support 887 // 888 // exit() 889 // ~~~~~~ 890 // Note that the collector can't reclaim the objectMonitor or deflate 891 // the object out from underneath the thread calling ::exit() as the 892 // thread calling ::exit() never transitions to a stable state. 893 // This inhibits GC, which in turn inhibits asynchronous (and 894 // inopportune) reclamation of "this". 895 // 896 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 897 // There's one exception to the claim above, however. EnterI() can call 898 // exit() to drop a lock if the acquirer has been externally suspended. 899 // In that case exit() is called with _thread_state as _thread_blocked, 900 // but the monitor's _count field is > 0, which inhibits reclamation. 901 // 902 // 1-0 exit 903 // ~~~~~~~~ 904 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 905 // the fast-path operators have been optimized so the common ::exit() 906 // operation is 1-0. See i486.ad fast_unlock(), for instance. 907 // The code emitted by fast_unlock() elides the usual MEMBAR. This 908 // greatly improves latency -- MEMBAR and CAS having considerable local 909 // latency on modern processors -- but at the cost of "stranding". Absent the 910 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 911 // ::enter() path, resulting in the entering thread being stranding 912 // and a progress-liveness failure. Stranding is extremely rare. 913 // We use timers (timed park operations) & periodic polling to detect 914 // and recover from stranding. Potentially stranded threads periodically 915 // wake up and poll the lock. See the usage of the _Responsible variable. 916 // 917 // The CAS() in enter provides for safety and exclusion, while the CAS or 918 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 919 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 920 // We detect and recover from stranding with timers. 921 // 922 // If a thread transiently strands it'll park until (a) another 923 // thread acquires the lock and then drops the lock, at which time the 924 // exiting thread will notice and unpark the stranded thread, or, (b) 925 // the timer expires. If the lock is high traffic then the stranding latency 926 // will be low due to (a). If the lock is low traffic then the odds of 927 // stranding are lower, although the worst-case stranding latency 928 // is longer. Critically, we don't want to put excessive load in the 929 // platform's timer subsystem. We want to minimize both the timer injection 930 // rate (timers created/sec) as well as the number of timers active at 931 // any one time. (more precisely, we want to minimize timer-seconds, which is 932 // the integral of the # of active timers at any instant over time). 933 // Both impinge on OS scalability. Given that, at most one thread parked on 934 // a monitor will use a timer. 935 936 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 937 Thread * Self = THREAD ; 938 if (THREAD != _owner) { 939 if (THREAD->is_lock_owned((address) _owner)) { 940 // Transmute _owner from a BasicLock pointer to a Thread address. 941 // We don't need to hold _mutex for this transition. 942 // Non-null to Non-null is safe as long as all readers can 943 // tolerate either flavor. 944 assert (_recursions == 0, "invariant") ; 945 _owner = THREAD ; 946 _recursions = 0 ; 947 OwnerIsThread = 1 ; 948 } else { 949 // NOTE: we need to handle unbalanced monitor enter/exit 950 // in native code by throwing an exception. 951 // TODO: Throw an IllegalMonitorStateException ? 952 TEVENT (Exit - Throw IMSX) ; 953 assert(false, "Non-balanced monitor enter/exit!"); 954 if (false) { 955 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 956 } 957 return; 958 } 959 } 960 961 if (_recursions != 0) { 962 _recursions--; // this is simple recursive enter 963 TEVENT (Inflated exit - recursive) ; 964 return ; 965 } 966 967 // Invariant: after setting Responsible=null an thread must execute 968 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 969 if ((SyncFlags & 4) == 0) { 970 _Responsible = NULL ; 971 } 972 973 #if INCLUDE_TRACE 974 // get the owner's thread id for the MonitorEnter event 975 // if it is enabled and the thread isn't suspended 976 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 977 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 978 } 979 #endif 980 981 for (;;) { 982 assert (THREAD == _owner, "invariant") ; 983 984 985 if (Knob_ExitPolicy == 0) { 986 // release semantics: prior loads and stores from within the critical section 987 // must not float (reorder) past the following store that drops the lock. 988 // On SPARC that requires MEMBAR #loadstore|#storestore. 989 // But of course in TSO #loadstore|#storestore is not required. 990 // I'd like to write one of the following: 991 // A. OrderAccess::release() ; _owner = NULL 992 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 993 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 994 // store into a _dummy variable. That store is not needed, but can result 995 // in massive wasteful coherency traffic on classic SMP systems. 996 // Instead, I use release_store(), which is implemented as just a simple 997 // ST on x64, x86 and SPARC. 998 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 999 OrderAccess::storeload() ; // See if we need to wake a successor 1000 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1001 TEVENT (Inflated exit - simple egress) ; 1002 return ; 1003 } 1004 TEVENT (Inflated exit - complex egress) ; 1005 1006 // Normally the exiting thread is responsible for ensuring succession, 1007 // but if other successors are ready or other entering threads are spinning 1008 // then this thread can simply store NULL into _owner and exit without 1009 // waking a successor. The existence of spinners or ready successors 1010 // guarantees proper succession (liveness). Responsibility passes to the 1011 // ready or running successors. The exiting thread delegates the duty. 1012 // More precisely, if a successor already exists this thread is absolved 1013 // of the responsibility of waking (unparking) one. 1014 // 1015 // The _succ variable is critical to reducing futile wakeup frequency. 1016 // _succ identifies the "heir presumptive" thread that has been made 1017 // ready (unparked) but that has not yet run. We need only one such 1018 // successor thread to guarantee progress. 1019 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1020 // section 3.3 "Futile Wakeup Throttling" for details. 1021 // 1022 // Note that spinners in Enter() also set _succ non-null. 1023 // In the current implementation spinners opportunistically set 1024 // _succ so that exiting threads might avoid waking a successor. 1025 // Another less appealing alternative would be for the exiting thread 1026 // to drop the lock and then spin briefly to see if a spinner managed 1027 // to acquire the lock. If so, the exiting thread could exit 1028 // immediately without waking a successor, otherwise the exiting 1029 // thread would need to dequeue and wake a successor. 1030 // (Note that we'd need to make the post-drop spin short, but no 1031 // shorter than the worst-case round-trip cache-line migration time. 1032 // The dropped lock needs to become visible to the spinner, and then 1033 // the acquisition of the lock by the spinner must become visible to 1034 // the exiting thread). 1035 // 1036 1037 // It appears that an heir-presumptive (successor) must be made ready. 1038 // Only the current lock owner can manipulate the EntryList or 1039 // drain _cxq, so we need to reacquire the lock. If we fail 1040 // to reacquire the lock the responsibility for ensuring succession 1041 // falls to the new owner. 1042 // 1043 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1044 return ; 1045 } 1046 TEVENT (Exit - Reacquired) ; 1047 } else { 1048 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1049 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1050 OrderAccess::storeload() ; 1051 // Ratify the previously observed values. 1052 if (_cxq == NULL || _succ != NULL) { 1053 TEVENT (Inflated exit - simple egress) ; 1054 return ; 1055 } 1056 1057 // inopportune interleaving -- the exiting thread (this thread) 1058 // in the fast-exit path raced an entering thread in the slow-enter 1059 // path. 1060 // We have two choices: 1061 // A. Try to reacquire the lock. 1062 // If the CAS() fails return immediately, otherwise 1063 // we either restart/rerun the exit operation, or simply 1064 // fall-through into the code below which wakes a successor. 1065 // B. If the elements forming the EntryList|cxq are TSM 1066 // we could simply unpark() the lead thread and return 1067 // without having set _succ. 1068 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1069 TEVENT (Inflated exit - reacquired succeeded) ; 1070 return ; 1071 } 1072 TEVENT (Inflated exit - reacquired failed) ; 1073 } else { 1074 TEVENT (Inflated exit - complex egress) ; 1075 } 1076 } 1077 1078 guarantee (_owner == THREAD, "invariant") ; 1079 1080 ObjectWaiter * w = NULL ; 1081 int QMode = Knob_QMode ; 1082 1083 if (QMode == 2 && _cxq != NULL) { 1084 // QMode == 2 : cxq has precedence over EntryList. 1085 // Try to directly wake a successor from the cxq. 1086 // If successful, the successor will need to unlink itself from cxq. 1087 w = _cxq ; 1088 assert (w != NULL, "invariant") ; 1089 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1090 ExitEpilog (Self, w) ; 1091 return ; 1092 } 1093 1094 if (QMode == 3 && _cxq != NULL) { 1095 // Aggressively drain cxq into EntryList at the first opportunity. 1096 // This policy ensure that recently-run threads live at the head of EntryList. 1097 // Drain _cxq into EntryList - bulk transfer. 1098 // First, detach _cxq. 1099 // The following loop is tantamount to: w = swap (&cxq, NULL) 1100 w = _cxq ; 1101 for (;;) { 1102 assert (w != NULL, "Invariant") ; 1103 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1104 if (u == w) break ; 1105 w = u ; 1106 } 1107 assert (w != NULL , "invariant") ; 1108 1109 ObjectWaiter * q = NULL ; 1110 ObjectWaiter * p ; 1111 for (p = w ; p != NULL ; p = p->_next) { 1112 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1113 p->TState = ObjectWaiter::TS_ENTER ; 1114 p->_prev = q ; 1115 q = p ; 1116 } 1117 1118 // Append the RATs to the EntryList 1119 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1120 ObjectWaiter * Tail ; 1121 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1122 if (Tail == NULL) { 1123 _EntryList = w ; 1124 } else { 1125 Tail->_next = w ; 1126 w->_prev = Tail ; 1127 } 1128 1129 // Fall thru into code that tries to wake a successor from EntryList 1130 } 1131 1132 if (QMode == 4 && _cxq != NULL) { 1133 // Aggressively drain cxq into EntryList at the first opportunity. 1134 // This policy ensure that recently-run threads live at the head of EntryList. 1135 1136 // Drain _cxq into EntryList - bulk transfer. 1137 // First, detach _cxq. 1138 // The following loop is tantamount to: w = swap (&cxq, NULL) 1139 w = _cxq ; 1140 for (;;) { 1141 assert (w != NULL, "Invariant") ; 1142 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1143 if (u == w) break ; 1144 w = u ; 1145 } 1146 assert (w != NULL , "invariant") ; 1147 1148 ObjectWaiter * q = NULL ; 1149 ObjectWaiter * p ; 1150 for (p = w ; p != NULL ; p = p->_next) { 1151 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1152 p->TState = ObjectWaiter::TS_ENTER ; 1153 p->_prev = q ; 1154 q = p ; 1155 } 1156 1157 // Prepend the RATs to the EntryList 1158 if (_EntryList != NULL) { 1159 q->_next = _EntryList ; 1160 _EntryList->_prev = q ; 1161 } 1162 _EntryList = w ; 1163 1164 // Fall thru into code that tries to wake a successor from EntryList 1165 } 1166 1167 w = _EntryList ; 1168 if (w != NULL) { 1169 // I'd like to write: guarantee (w->_thread != Self). 1170 // But in practice an exiting thread may find itself on the EntryList. 1171 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1172 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1173 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1174 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1175 // release the lock "O". T2 resumes immediately after the ST of null into 1176 // _owner, above. T2 notices that the EntryList is populated, so it 1177 // reacquires the lock and then finds itself on the EntryList. 1178 // Given all that, we have to tolerate the circumstance where "w" is 1179 // associated with Self. 1180 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1181 ExitEpilog (Self, w) ; 1182 return ; 1183 } 1184 1185 // If we find that both _cxq and EntryList are null then just 1186 // re-run the exit protocol from the top. 1187 w = _cxq ; 1188 if (w == NULL) continue ; 1189 1190 // Drain _cxq into EntryList - bulk transfer. 1191 // First, detach _cxq. 1192 // The following loop is tantamount to: w = swap (&cxq, NULL) 1193 for (;;) { 1194 assert (w != NULL, "Invariant") ; 1195 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1196 if (u == w) break ; 1197 w = u ; 1198 } 1199 TEVENT (Inflated exit - drain cxq into EntryList) ; 1200 1201 assert (w != NULL , "invariant") ; 1202 assert (_EntryList == NULL , "invariant") ; 1203 1204 // Convert the LIFO SLL anchored by _cxq into a DLL. 1205 // The list reorganization step operates in O(LENGTH(w)) time. 1206 // It's critical that this step operate quickly as 1207 // "Self" still holds the outer-lock, restricting parallelism 1208 // and effectively lengthening the critical section. 1209 // Invariant: s chases t chases u. 1210 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1211 // we have faster access to the tail. 1212 1213 if (QMode == 1) { 1214 // QMode == 1 : drain cxq to EntryList, reversing order 1215 // We also reverse the order of the list. 1216 ObjectWaiter * s = NULL ; 1217 ObjectWaiter * t = w ; 1218 ObjectWaiter * u = NULL ; 1219 while (t != NULL) { 1220 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1221 t->TState = ObjectWaiter::TS_ENTER ; 1222 u = t->_next ; 1223 t->_prev = u ; 1224 t->_next = s ; 1225 s = t; 1226 t = u ; 1227 } 1228 _EntryList = s ; 1229 assert (s != NULL, "invariant") ; 1230 } else { 1231 // QMode == 0 or QMode == 2 1232 _EntryList = w ; 1233 ObjectWaiter * q = NULL ; 1234 ObjectWaiter * p ; 1235 for (p = w ; p != NULL ; p = p->_next) { 1236 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1237 p->TState = ObjectWaiter::TS_ENTER ; 1238 p->_prev = q ; 1239 q = p ; 1240 } 1241 } 1242 1243 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1244 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1245 1246 // See if we can abdicate to a spinner instead of waking a thread. 1247 // A primary goal of the implementation is to reduce the 1248 // context-switch rate. 1249 if (_succ != NULL) continue; 1250 1251 w = _EntryList ; 1252 if (w != NULL) { 1253 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1254 ExitEpilog (Self, w) ; 1255 return ; 1256 } 1257 } 1258 } 1259 1260 // ExitSuspendEquivalent: 1261 // A faster alternate to handle_special_suspend_equivalent_condition() 1262 // 1263 // handle_special_suspend_equivalent_condition() unconditionally 1264 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1265 // operations have high latency. Note that in ::enter() we call HSSEC 1266 // while holding the monitor, so we effectively lengthen the critical sections. 1267 // 1268 // There are a number of possible solutions: 1269 // 1270 // A. To ameliorate the problem we might also defer state transitions 1271 // to as late as possible -- just prior to parking. 1272 // Given that, we'd call HSSEC after having returned from park(), 1273 // but before attempting to acquire the monitor. This is only a 1274 // partial solution. It avoids calling HSSEC while holding the 1275 // monitor (good), but it still increases successor reacquisition latency -- 1276 // the interval between unparking a successor and the time the successor 1277 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1278 // If we use this technique we can also avoid EnterI()-exit() loop 1279 // in ::enter() where we iteratively drop the lock and then attempt 1280 // to reacquire it after suspending. 1281 // 1282 // B. In the future we might fold all the suspend bits into a 1283 // composite per-thread suspend flag and then update it with CAS(). 1284 // Alternately, a Dekker-like mechanism with multiple variables 1285 // would suffice: 1286 // ST Self->_suspend_equivalent = false 1287 // MEMBAR 1288 // LD Self_>_suspend_flags 1289 // 1290 1291 1292 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1293 int Mode = Knob_FastHSSEC ; 1294 if (Mode && !jSelf->is_external_suspend()) { 1295 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1296 jSelf->clear_suspend_equivalent() ; 1297 if (2 == Mode) OrderAccess::storeload() ; 1298 if (!jSelf->is_external_suspend()) return false ; 1299 // We raced a suspension -- fall thru into the slow path 1300 TEVENT (ExitSuspendEquivalent - raced) ; 1301 jSelf->set_suspend_equivalent() ; 1302 } 1303 return jSelf->handle_special_suspend_equivalent_condition() ; 1304 } 1305 1306 1307 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1308 assert (_owner == Self, "invariant") ; 1309 1310 // Exit protocol: 1311 // 1. ST _succ = wakee 1312 // 2. membar #loadstore|#storestore; 1313 // 2. ST _owner = NULL 1314 // 3. unpark(wakee) 1315 1316 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1317 ParkEvent * Trigger = Wakee->_event ; 1318 1319 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1320 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1321 // out-of-scope (non-extant). 1322 Wakee = NULL ; 1323 1324 // Drop the lock 1325 OrderAccess::release_store_ptr (&_owner, NULL) ; 1326 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1327 1328 if (SafepointSynchronize::do_call_back()) { 1329 TEVENT (unpark before SAFEPOINT) ; 1330 } 1331 1332 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1333 Trigger->unpark() ; 1334 1335 // Maintain stats and report events to JVMTI 1336 if (ObjectMonitor::_sync_Parks != NULL) { 1337 ObjectMonitor::_sync_Parks->inc() ; 1338 } 1339 } 1340 1341 1342 // ----------------------------------------------------------------------------- 1343 // Class Loader deadlock handling. 1344 // 1345 // complete_exit exits a lock returning recursion count 1346 // complete_exit/reenter operate as a wait without waiting 1347 // complete_exit requires an inflated monitor 1348 // The _owner field is not always the Thread addr even with an 1349 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1350 // thread due to contention. 1351 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1352 Thread * const Self = THREAD; 1353 assert(Self->is_Java_thread(), "Must be Java thread!"); 1354 JavaThread *jt = (JavaThread *)THREAD; 1355 1356 DeferredInitialize(); 1357 1358 if (THREAD != _owner) { 1359 if (THREAD->is_lock_owned ((address)_owner)) { 1360 assert(_recursions == 0, "internal state error"); 1361 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1362 _recursions = 0 ; 1363 OwnerIsThread = 1 ; 1364 } 1365 } 1366 1367 guarantee(Self == _owner, "complete_exit not owner"); 1368 intptr_t save = _recursions; // record the old recursion count 1369 _recursions = 0; // set the recursion level to be 0 1370 exit (true, Self) ; // exit the monitor 1371 guarantee (_owner != Self, "invariant"); 1372 return save; 1373 } 1374 1375 // reenter() enters a lock and sets recursion count 1376 // complete_exit/reenter operate as a wait without waiting 1377 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1378 Thread * const Self = THREAD; 1379 assert(Self->is_Java_thread(), "Must be Java thread!"); 1380 JavaThread *jt = (JavaThread *)THREAD; 1381 1382 guarantee(_owner != Self, "reenter already owner"); 1383 enter (THREAD); // enter the monitor 1384 guarantee (_recursions == 0, "reenter recursion"); 1385 _recursions = recursions; 1386 return; 1387 } 1388 1389 1390 // ----------------------------------------------------------------------------- 1391 // A macro is used below because there may already be a pending 1392 // exception which should not abort the execution of the routines 1393 // which use this (which is why we don't put this into check_slow and 1394 // call it with a CHECK argument). 1395 1396 #define CHECK_OWNER() \ 1397 do { \ 1398 if (THREAD != _owner) { \ 1399 if (THREAD->is_lock_owned((address) _owner)) { \ 1400 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1401 _recursions = 0; \ 1402 OwnerIsThread = 1 ; \ 1403 } else { \ 1404 TEVENT (Throw IMSX) ; \ 1405 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1406 } \ 1407 } \ 1408 } while (false) 1409 1410 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1411 // TODO-FIXME: remove check_slow() -- it's likely dead. 1412 1413 void ObjectMonitor::check_slow(TRAPS) { 1414 TEVENT (check_slow - throw IMSX) ; 1415 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1416 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1417 } 1418 1419 static int Adjust (volatile int * adr, int dx) { 1420 int v ; 1421 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1422 return v ; 1423 } 1424 1425 // helper method for posting a monitor wait event 1426 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1427 jlong notifier_tid, 1428 jlong timeout, 1429 bool timedout) { 1430 event->set_klass(((oop)this->object())->klass()); 1431 event->set_timeout((TYPE_ULONG)timeout); 1432 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1433 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1434 event->set_timedOut((TYPE_BOOLEAN)timedout); 1435 event->commit(); 1436 } 1437 1438 // ----------------------------------------------------------------------------- 1439 // Wait/Notify/NotifyAll 1440 // 1441 // Note: a subset of changes to ObjectMonitor::wait() 1442 // will need to be replicated in complete_exit above 1443 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1444 Thread * const Self = THREAD ; 1445 assert(Self->is_Java_thread(), "Must be Java thread!"); 1446 JavaThread *jt = (JavaThread *)THREAD; 1447 1448 DeferredInitialize () ; 1449 1450 // Throw IMSX or IEX. 1451 CHECK_OWNER(); 1452 1453 EventJavaMonitorWait event; 1454 1455 // check for a pending interrupt 1456 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1457 // post monitor waited event. Note that this is past-tense, we are done waiting. 1458 if (JvmtiExport::should_post_monitor_waited()) { 1459 // Note: 'false' parameter is passed here because the 1460 // wait was not timed out due to thread interrupt. 1461 JvmtiExport::post_monitor_waited(jt, this, false); 1462 } 1463 if (event.should_commit()) { 1464 post_monitor_wait_event(&event, 0, millis, false); 1465 } 1466 TEVENT (Wait - Throw IEX) ; 1467 THROW(vmSymbols::java_lang_InterruptedException()); 1468 return ; 1469 } 1470 1471 TEVENT (Wait) ; 1472 1473 assert (Self->_Stalled == 0, "invariant") ; 1474 Self->_Stalled = intptr_t(this) ; 1475 jt->set_current_waiting_monitor(this); 1476 1477 // create a node to be put into the queue 1478 // Critically, after we reset() the event but prior to park(), we must check 1479 // for a pending interrupt. 1480 ObjectWaiter node(Self); 1481 node.TState = ObjectWaiter::TS_WAIT ; 1482 Self->_ParkEvent->reset() ; 1483 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1484 1485 // Enter the waiting queue, which is a circular doubly linked list in this case 1486 // but it could be a priority queue or any data structure. 1487 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1488 // by the the owner of the monitor *except* in the case where park() 1489 // returns because of a timeout of interrupt. Contention is exceptionally rare 1490 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1491 1492 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1493 AddWaiter (&node) ; 1494 Thread::SpinRelease (&_WaitSetLock) ; 1495 1496 if ((SyncFlags & 4) == 0) { 1497 _Responsible = NULL ; 1498 } 1499 intptr_t save = _recursions; // record the old recursion count 1500 _waiters++; // increment the number of waiters 1501 _recursions = 0; // set the recursion level to be 1 1502 exit (true, Self) ; // exit the monitor 1503 guarantee (_owner != Self, "invariant") ; 1504 1505 // As soon as the ObjectMonitor's ownership is dropped in the exit() 1506 // call above, another thread can enter() the ObjectMonitor, do the 1507 // notify(), and exit() the ObjectMonitor. If the other thread's 1508 // exit() call chooses this thread as the successor and the unpark() 1509 // call happens to occur while this thread is posting a 1510 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event 1511 // handler using RawMonitors and consuming the unpark(). 1512 // 1513 // To avoid the problem, we re-post the event. This does no harm 1514 // even if the original unpark() was not consumed because we are the 1515 // chosen successor for this monitor. 1516 if (node._notified != 0 && _succ == Self) { 1517 node._event->unpark(); 1518 } 1519 1520 // The thread is on the WaitSet list - now park() it. 1521 // On MP systems it's conceivable that a brief spin before we park 1522 // could be profitable. 1523 // 1524 // TODO-FIXME: change the following logic to a loop of the form 1525 // while (!timeout && !interrupted && _notified == 0) park() 1526 1527 int ret = OS_OK ; 1528 int WasNotified = 0 ; 1529 { // State transition wrappers 1530 OSThread* osthread = Self->osthread(); 1531 OSThreadWaitState osts(osthread, true); 1532 { 1533 ThreadBlockInVM tbivm(jt); 1534 // Thread is in thread_blocked state and oop access is unsafe. 1535 jt->set_suspend_equivalent(); 1536 1537 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1538 // Intentionally empty 1539 } else 1540 if (node._notified == 0) { 1541 if (millis <= 0) { 1542 Self->_ParkEvent->park () ; 1543 } else { 1544 ret = Self->_ParkEvent->park (millis) ; 1545 } 1546 } 1547 1548 // were we externally suspended while we were waiting? 1549 if (ExitSuspendEquivalent (jt)) { 1550 // TODO-FIXME: add -- if succ == Self then succ = null. 1551 jt->java_suspend_self(); 1552 } 1553 1554 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1555 1556 1557 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1558 // from the WaitSet to the EntryList. 1559 // See if we need to remove Node from the WaitSet. 1560 // We use double-checked locking to avoid grabbing _WaitSetLock 1561 // if the thread is not on the wait queue. 1562 // 1563 // Note that we don't need a fence before the fetch of TState. 1564 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1565 // written by the is thread. (perhaps the fetch might even be satisfied 1566 // by a look-aside into the processor's own store buffer, although given 1567 // the length of the code path between the prior ST and this load that's 1568 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1569 // then we'll acquire the lock and then re-fetch a fresh TState value. 1570 // That is, we fail toward safety. 1571 1572 if (node.TState == ObjectWaiter::TS_WAIT) { 1573 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1574 if (node.TState == ObjectWaiter::TS_WAIT) { 1575 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1576 assert(node._notified == 0, "invariant"); 1577 node.TState = ObjectWaiter::TS_RUN ; 1578 } 1579 Thread::SpinRelease (&_WaitSetLock) ; 1580 } 1581 1582 // The thread is now either on off-list (TS_RUN), 1583 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1584 // The Node's TState variable is stable from the perspective of this thread. 1585 // No other threads will asynchronously modify TState. 1586 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1587 OrderAccess::loadload() ; 1588 if (_succ == Self) _succ = NULL ; 1589 WasNotified = node._notified ; 1590 1591 // Reentry phase -- reacquire the monitor. 1592 // re-enter contended monitor after object.wait(). 1593 // retain OBJECT_WAIT state until re-enter successfully completes 1594 // Thread state is thread_in_vm and oop access is again safe, 1595 // although the raw address of the object may have changed. 1596 // (Don't cache naked oops over safepoints, of course). 1597 1598 // post monitor waited event. Note that this is past-tense, we are done waiting. 1599 if (JvmtiExport::should_post_monitor_waited()) { 1600 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1601 } 1602 1603 if (event.should_commit()) { 1604 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1605 } 1606 1607 OrderAccess::fence() ; 1608 1609 assert (Self->_Stalled != 0, "invariant") ; 1610 Self->_Stalled = 0 ; 1611 1612 assert (_owner != Self, "invariant") ; 1613 ObjectWaiter::TStates v = node.TState ; 1614 if (v == ObjectWaiter::TS_RUN) { 1615 enter (Self) ; 1616 } else { 1617 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1618 ReenterI (Self, &node) ; 1619 node.wait_reenter_end(this); 1620 } 1621 1622 // Self has reacquired the lock. 1623 // Lifecycle - the node representing Self must not appear on any queues. 1624 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1625 // want residual elements associated with this thread left on any lists. 1626 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1627 assert (_owner == Self, "invariant") ; 1628 assert (_succ != Self , "invariant") ; 1629 } // OSThreadWaitState() 1630 1631 jt->set_current_waiting_monitor(NULL); 1632 1633 guarantee (_recursions == 0, "invariant") ; 1634 _recursions = save; // restore the old recursion count 1635 _waiters--; // decrement the number of waiters 1636 1637 // Verify a few postconditions 1638 assert (_owner == Self , "invariant") ; 1639 assert (_succ != Self , "invariant") ; 1640 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1641 1642 if (SyncFlags & 32) { 1643 OrderAccess::fence() ; 1644 } 1645 1646 // check if the notification happened 1647 if (!WasNotified) { 1648 // no, it could be timeout or Thread.interrupt() or both 1649 // check for interrupt event, otherwise it is timeout 1650 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1651 TEVENT (Wait - throw IEX from epilog) ; 1652 THROW(vmSymbols::java_lang_InterruptedException()); 1653 } 1654 } 1655 1656 // NOTE: Spurious wake up will be consider as timeout. 1657 // Monitor notify has precedence over thread interrupt. 1658 } 1659 1660 1661 // Consider: 1662 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1663 // then instead of transferring a thread from the WaitSet to the EntryList 1664 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1665 1666 void ObjectMonitor::notify(TRAPS) { 1667 CHECK_OWNER(); 1668 if (_WaitSet == NULL) { 1669 TEVENT (Empty-Notify) ; 1670 return ; 1671 } 1672 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1673 1674 int Policy = Knob_MoveNotifyee ; 1675 1676 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1677 ObjectWaiter * iterator = DequeueWaiter() ; 1678 if (iterator != NULL) { 1679 TEVENT (Notify1 - Transfer) ; 1680 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1681 guarantee (iterator->_notified == 0, "invariant") ; 1682 if (Policy != 4) { 1683 iterator->TState = ObjectWaiter::TS_ENTER ; 1684 } 1685 iterator->_notified = 1 ; 1686 Thread * Self = THREAD; 1687 iterator->_notifier_tid = Self->osthread()->thread_id(); 1688 1689 ObjectWaiter * List = _EntryList ; 1690 if (List != NULL) { 1691 assert (List->_prev == NULL, "invariant") ; 1692 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1693 assert (List != iterator, "invariant") ; 1694 } 1695 1696 if (Policy == 0) { // prepend to EntryList 1697 if (List == NULL) { 1698 iterator->_next = iterator->_prev = NULL ; 1699 _EntryList = iterator ; 1700 } else { 1701 List->_prev = iterator ; 1702 iterator->_next = List ; 1703 iterator->_prev = NULL ; 1704 _EntryList = iterator ; 1705 } 1706 } else 1707 if (Policy == 1) { // append to EntryList 1708 if (List == NULL) { 1709 iterator->_next = iterator->_prev = NULL ; 1710 _EntryList = iterator ; 1711 } else { 1712 // CONSIDER: finding the tail currently requires a linear-time walk of 1713 // the EntryList. We can make tail access constant-time by converting to 1714 // a CDLL instead of using our current DLL. 1715 ObjectWaiter * Tail ; 1716 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1717 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1718 Tail->_next = iterator ; 1719 iterator->_prev = Tail ; 1720 iterator->_next = NULL ; 1721 } 1722 } else 1723 if (Policy == 2) { // prepend to cxq 1724 // prepend to cxq 1725 if (List == NULL) { 1726 iterator->_next = iterator->_prev = NULL ; 1727 _EntryList = iterator ; 1728 } else { 1729 iterator->TState = ObjectWaiter::TS_CXQ ; 1730 for (;;) { 1731 ObjectWaiter * Front = _cxq ; 1732 iterator->_next = Front ; 1733 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1734 break ; 1735 } 1736 } 1737 } 1738 } else 1739 if (Policy == 3) { // append to cxq 1740 iterator->TState = ObjectWaiter::TS_CXQ ; 1741 for (;;) { 1742 ObjectWaiter * Tail ; 1743 Tail = _cxq ; 1744 if (Tail == NULL) { 1745 iterator->_next = NULL ; 1746 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1747 break ; 1748 } 1749 } else { 1750 while (Tail->_next != NULL) Tail = Tail->_next ; 1751 Tail->_next = iterator ; 1752 iterator->_prev = Tail ; 1753 iterator->_next = NULL ; 1754 break ; 1755 } 1756 } 1757 } else { 1758 ParkEvent * ev = iterator->_event ; 1759 iterator->TState = ObjectWaiter::TS_RUN ; 1760 OrderAccess::fence() ; 1761 ev->unpark() ; 1762 } 1763 1764 if (Policy < 4) { 1765 iterator->wait_reenter_begin(this); 1766 } 1767 1768 // _WaitSetLock protects the wait queue, not the EntryList. We could 1769 // move the add-to-EntryList operation, above, outside the critical section 1770 // protected by _WaitSetLock. In practice that's not useful. With the 1771 // exception of wait() timeouts and interrupts the monitor owner 1772 // is the only thread that grabs _WaitSetLock. There's almost no contention 1773 // on _WaitSetLock so it's not profitable to reduce the length of the 1774 // critical section. 1775 } 1776 1777 Thread::SpinRelease (&_WaitSetLock) ; 1778 1779 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1780 ObjectMonitor::_sync_Notifications->inc() ; 1781 } 1782 } 1783 1784 1785 void ObjectMonitor::notifyAll(TRAPS) { 1786 CHECK_OWNER(); 1787 ObjectWaiter* iterator; 1788 if (_WaitSet == NULL) { 1789 TEVENT (Empty-NotifyAll) ; 1790 return ; 1791 } 1792 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1793 1794 int Policy = Knob_MoveNotifyee ; 1795 int Tally = 0 ; 1796 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1797 1798 for (;;) { 1799 iterator = DequeueWaiter () ; 1800 if (iterator == NULL) break ; 1801 TEVENT (NotifyAll - Transfer1) ; 1802 ++Tally ; 1803 1804 // Disposition - what might we do with iterator ? 1805 // a. add it directly to the EntryList - either tail or head. 1806 // b. push it onto the front of the _cxq. 1807 // For now we use (a). 1808 1809 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1810 guarantee (iterator->_notified == 0, "invariant") ; 1811 iterator->_notified = 1 ; 1812 Thread * Self = THREAD; 1813 iterator->_notifier_tid = Self->osthread()->thread_id(); 1814 if (Policy != 4) { 1815 iterator->TState = ObjectWaiter::TS_ENTER ; 1816 } 1817 1818 ObjectWaiter * List = _EntryList ; 1819 if (List != NULL) { 1820 assert (List->_prev == NULL, "invariant") ; 1821 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1822 assert (List != iterator, "invariant") ; 1823 } 1824 1825 if (Policy == 0) { // prepend to EntryList 1826 if (List == NULL) { 1827 iterator->_next = iterator->_prev = NULL ; 1828 _EntryList = iterator ; 1829 } else { 1830 List->_prev = iterator ; 1831 iterator->_next = List ; 1832 iterator->_prev = NULL ; 1833 _EntryList = iterator ; 1834 } 1835 } else 1836 if (Policy == 1) { // append to EntryList 1837 if (List == NULL) { 1838 iterator->_next = iterator->_prev = NULL ; 1839 _EntryList = iterator ; 1840 } else { 1841 // CONSIDER: finding the tail currently requires a linear-time walk of 1842 // the EntryList. We can make tail access constant-time by converting to 1843 // a CDLL instead of using our current DLL. 1844 ObjectWaiter * Tail ; 1845 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1846 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1847 Tail->_next = iterator ; 1848 iterator->_prev = Tail ; 1849 iterator->_next = NULL ; 1850 } 1851 } else 1852 if (Policy == 2) { // prepend to cxq 1853 // prepend to cxq 1854 iterator->TState = ObjectWaiter::TS_CXQ ; 1855 for (;;) { 1856 ObjectWaiter * Front = _cxq ; 1857 iterator->_next = Front ; 1858 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1859 break ; 1860 } 1861 } 1862 } else 1863 if (Policy == 3) { // append to cxq 1864 iterator->TState = ObjectWaiter::TS_CXQ ; 1865 for (;;) { 1866 ObjectWaiter * Tail ; 1867 Tail = _cxq ; 1868 if (Tail == NULL) { 1869 iterator->_next = NULL ; 1870 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1871 break ; 1872 } 1873 } else { 1874 while (Tail->_next != NULL) Tail = Tail->_next ; 1875 Tail->_next = iterator ; 1876 iterator->_prev = Tail ; 1877 iterator->_next = NULL ; 1878 break ; 1879 } 1880 } 1881 } else { 1882 ParkEvent * ev = iterator->_event ; 1883 iterator->TState = ObjectWaiter::TS_RUN ; 1884 OrderAccess::fence() ; 1885 ev->unpark() ; 1886 } 1887 1888 if (Policy < 4) { 1889 iterator->wait_reenter_begin(this); 1890 } 1891 1892 // _WaitSetLock protects the wait queue, not the EntryList. We could 1893 // move the add-to-EntryList operation, above, outside the critical section 1894 // protected by _WaitSetLock. In practice that's not useful. With the 1895 // exception of wait() timeouts and interrupts the monitor owner 1896 // is the only thread that grabs _WaitSetLock. There's almost no contention 1897 // on _WaitSetLock so it's not profitable to reduce the length of the 1898 // critical section. 1899 } 1900 1901 Thread::SpinRelease (&_WaitSetLock) ; 1902 1903 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1904 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1905 } 1906 } 1907 1908 // ----------------------------------------------------------------------------- 1909 // Adaptive Spinning Support 1910 // 1911 // Adaptive spin-then-block - rational spinning 1912 // 1913 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1914 // algorithm. On high order SMP systems it would be better to start with 1915 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1916 // a contending thread could enqueue itself on the cxq and then spin locally 1917 // on a thread-specific variable such as its ParkEvent._Event flag. 1918 // That's left as an exercise for the reader. Note that global spinning is 1919 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1920 // low latency and massive bandwidth. 1921 // 1922 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1923 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1924 // (duration) or we can fix the count at approximately the duration of 1925 // a context switch and vary the frequency. Of course we could also 1926 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1927 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1928 // 1929 // This implementation varies the duration "D", where D varies with 1930 // the success rate of recent spin attempts. (D is capped at approximately 1931 // length of a round-trip context switch). The success rate for recent 1932 // spin attempts is a good predictor of the success rate of future spin 1933 // attempts. The mechanism adapts automatically to varying critical 1934 // section length (lock modality), system load and degree of parallelism. 1935 // D is maintained per-monitor in _SpinDuration and is initialized 1936 // optimistically. Spin frequency is fixed at 100%. 1937 // 1938 // Note that _SpinDuration is volatile, but we update it without locks 1939 // or atomics. The code is designed so that _SpinDuration stays within 1940 // a reasonable range even in the presence of races. The arithmetic 1941 // operations on _SpinDuration are closed over the domain of legal values, 1942 // so at worst a race will install and older but still legal value. 1943 // At the very worst this introduces some apparent non-determinism. 1944 // We might spin when we shouldn't or vice-versa, but since the spin 1945 // count are relatively short, even in the worst case, the effect is harmless. 1946 // 1947 // Care must be taken that a low "D" value does not become an 1948 // an absorbing state. Transient spinning failures -- when spinning 1949 // is overall profitable -- should not cause the system to converge 1950 // on low "D" values. We want spinning to be stable and predictable 1951 // and fairly responsive to change and at the same time we don't want 1952 // it to oscillate, become metastable, be "too" non-deterministic, 1953 // or converge on or enter undesirable stable absorbing states. 1954 // 1955 // We implement a feedback-based control system -- using past behavior 1956 // to predict future behavior. We face two issues: (a) if the 1957 // input signal is random then the spin predictor won't provide optimal 1958 // results, and (b) if the signal frequency is too high then the control 1959 // system, which has some natural response lag, will "chase" the signal. 1960 // (b) can arise from multimodal lock hold times. Transient preemption 1961 // can also result in apparent bimodal lock hold times. 1962 // Although sub-optimal, neither condition is particularly harmful, as 1963 // in the worst-case we'll spin when we shouldn't or vice-versa. 1964 // The maximum spin duration is rather short so the failure modes aren't bad. 1965 // To be conservative, I've tuned the gain in system to bias toward 1966 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1967 // "rings" or oscillates between spinning and not spinning. This happens 1968 // when spinning is just on the cusp of profitability, however, so the 1969 // situation is not dire. The state is benign -- there's no need to add 1970 // hysteresis control to damp the transition rate between spinning and 1971 // not spinning. 1972 // 1973 1974 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 1975 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 1976 1977 // Spinning: Fixed frequency (100%), vary duration 1978 1979 1980 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1981 1982 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1983 int ctr = Knob_FixedSpin ; 1984 if (ctr != 0) { 1985 while (--ctr >= 0) { 1986 if (TryLock (Self) > 0) return 1 ; 1987 SpinPause () ; 1988 } 1989 return 0 ; 1990 } 1991 1992 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 1993 if (TryLock(Self) > 0) { 1994 // Increase _SpinDuration ... 1995 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1996 // Raising _SpurDuration to the poverty line is key. 1997 int x = _SpinDuration ; 1998 if (x < Knob_SpinLimit) { 1999 if (x < Knob_Poverty) x = Knob_Poverty ; 2000 _SpinDuration = x + Knob_BonusB ; 2001 } 2002 return 1 ; 2003 } 2004 SpinPause () ; 2005 } 2006 2007 // Admission control - verify preconditions for spinning 2008 // 2009 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2010 // becoming an absorbing state. Put another way, we spin briefly to 2011 // sample, just in case the system load, parallelism, contention, or lock 2012 // modality changed. 2013 // 2014 // Consider the following alternative: 2015 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2016 // spin attempt. "Periodically" might mean after a tally of 2017 // the # of failed spin attempts (or iterations) reaches some threshold. 2018 // This takes us into the realm of 1-out-of-N spinning, where we 2019 // hold the duration constant but vary the frequency. 2020 2021 ctr = _SpinDuration ; 2022 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2023 if (ctr <= 0) return 0 ; 2024 2025 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2026 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2027 TEVENT (Spin abort - notrunnable [TOP]); 2028 return 0 ; 2029 } 2030 2031 int MaxSpin = Knob_MaxSpinners ; 2032 if (MaxSpin >= 0) { 2033 if (_Spinner > MaxSpin) { 2034 TEVENT (Spin abort -- too many spinners) ; 2035 return 0 ; 2036 } 2037 // Slighty racy, but benign ... 2038 Adjust (&_Spinner, 1) ; 2039 } 2040 2041 // We're good to spin ... spin ingress. 2042 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2043 // when preparing to LD...CAS _owner, etc and the CAS is likely 2044 // to succeed. 2045 int hits = 0 ; 2046 int msk = 0 ; 2047 int caspty = Knob_CASPenalty ; 2048 int oxpty = Knob_OXPenalty ; 2049 int sss = Knob_SpinSetSucc ; 2050 if (sss && _succ == NULL ) _succ = Self ; 2051 Thread * prv = NULL ; 2052 2053 // There are three ways to exit the following loop: 2054 // 1. A successful spin where this thread has acquired the lock. 2055 // 2. Spin failure with prejudice 2056 // 3. Spin failure without prejudice 2057 2058 while (--ctr >= 0) { 2059 2060 // Periodic polling -- Check for pending GC 2061 // Threads may spin while they're unsafe. 2062 // We don't want spinning threads to delay the JVM from reaching 2063 // a stop-the-world safepoint or to steal cycles from GC. 2064 // If we detect a pending safepoint we abort in order that 2065 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2066 // this thread, if safe, doesn't steal cycles from GC. 2067 // This is in keeping with the "no loitering in runtime" rule. 2068 // We periodically check to see if there's a safepoint pending. 2069 if ((ctr & 0xFF) == 0) { 2070 if (SafepointSynchronize::do_call_back()) { 2071 TEVENT (Spin: safepoint) ; 2072 goto Abort ; // abrupt spin egress 2073 } 2074 if (Knob_UsePause & 1) SpinPause () ; 2075 2076 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2077 if (hits > 50 && scb != NULL) { 2078 int abend = (*scb)(SpinCallbackArgument, 0) ; 2079 } 2080 } 2081 2082 if (Knob_UsePause & 2) SpinPause() ; 2083 2084 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2085 // This is useful on classic SMP systems, but is of less utility on 2086 // N1-style CMT platforms. 2087 // 2088 // Trade-off: lock acquisition latency vs coherency bandwidth. 2089 // Lock hold times are typically short. A histogram 2090 // of successful spin attempts shows that we usually acquire 2091 // the lock early in the spin. That suggests we want to 2092 // sample _owner frequently in the early phase of the spin, 2093 // but then back-off and sample less frequently as the spin 2094 // progresses. The back-off makes a good citizen on SMP big 2095 // SMP systems. Oversampling _owner can consume excessive 2096 // coherency bandwidth. Relatedly, if we _oversample _owner we 2097 // can inadvertently interfere with the the ST m->owner=null. 2098 // executed by the lock owner. 2099 if (ctr & msk) continue ; 2100 ++hits ; 2101 if ((hits & 0xF) == 0) { 2102 // The 0xF, above, corresponds to the exponent. 2103 // Consider: (msk+1)|msk 2104 msk = ((msk << 2)|3) & BackOffMask ; 2105 } 2106 2107 // Probe _owner with TATAS 2108 // If this thread observes the monitor transition or flicker 2109 // from locked to unlocked to locked, then the odds that this 2110 // thread will acquire the lock in this spin attempt go down 2111 // considerably. The same argument applies if the CAS fails 2112 // or if we observe _owner change from one non-null value to 2113 // another non-null value. In such cases we might abort 2114 // the spin without prejudice or apply a "penalty" to the 2115 // spin count-down variable "ctr", reducing it by 100, say. 2116 2117 Thread * ox = (Thread *) _owner ; 2118 if (ox == NULL) { 2119 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2120 if (ox == NULL) { 2121 // The CAS succeeded -- this thread acquired ownership 2122 // Take care of some bookkeeping to exit spin state. 2123 if (sss && _succ == Self) { 2124 _succ = NULL ; 2125 } 2126 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2127 2128 // Increase _SpinDuration : 2129 // The spin was successful (profitable) so we tend toward 2130 // longer spin attempts in the future. 2131 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2132 // If we acquired the lock early in the spin cycle it 2133 // makes sense to increase _SpinDuration proportionally. 2134 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2135 int x = _SpinDuration ; 2136 if (x < Knob_SpinLimit) { 2137 if (x < Knob_Poverty) x = Knob_Poverty ; 2138 _SpinDuration = x + Knob_Bonus ; 2139 } 2140 return 1 ; 2141 } 2142 2143 // The CAS failed ... we can take any of the following actions: 2144 // * penalize: ctr -= Knob_CASPenalty 2145 // * exit spin with prejudice -- goto Abort; 2146 // * exit spin without prejudice. 2147 // * Since CAS is high-latency, retry again immediately. 2148 prv = ox ; 2149 TEVENT (Spin: cas failed) ; 2150 if (caspty == -2) break ; 2151 if (caspty == -1) goto Abort ; 2152 ctr -= caspty ; 2153 continue ; 2154 } 2155 2156 // Did lock ownership change hands ? 2157 if (ox != prv && prv != NULL ) { 2158 TEVENT (spin: Owner changed) 2159 if (oxpty == -2) break ; 2160 if (oxpty == -1) goto Abort ; 2161 ctr -= oxpty ; 2162 } 2163 prv = ox ; 2164 2165 // Abort the spin if the owner is not executing. 2166 // The owner must be executing in order to drop the lock. 2167 // Spinning while the owner is OFFPROC is idiocy. 2168 // Consider: ctr -= RunnablePenalty ; 2169 if (Knob_OState && NotRunnable (Self, ox)) { 2170 TEVENT (Spin abort - notrunnable); 2171 goto Abort ; 2172 } 2173 if (sss && _succ == NULL ) _succ = Self ; 2174 } 2175 2176 // Spin failed with prejudice -- reduce _SpinDuration. 2177 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2178 // AIMD is globally stable. 2179 TEVENT (Spin failure) ; 2180 { 2181 int x = _SpinDuration ; 2182 if (x > 0) { 2183 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2184 // This is globally sample and tends to damp the response. 2185 x -= Knob_Penalty ; 2186 if (x < 0) x = 0 ; 2187 _SpinDuration = x ; 2188 } 2189 } 2190 2191 Abort: 2192 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2193 if (sss && _succ == Self) { 2194 _succ = NULL ; 2195 // Invariant: after setting succ=null a contending thread 2196 // must recheck-retry _owner before parking. This usually happens 2197 // in the normal usage of TrySpin(), but it's safest 2198 // to make TrySpin() as foolproof as possible. 2199 OrderAccess::fence() ; 2200 if (TryLock(Self) > 0) return 1 ; 2201 } 2202 return 0 ; 2203 } 2204 2205 // NotRunnable() -- informed spinning 2206 // 2207 // Don't bother spinning if the owner is not eligible to drop the lock. 2208 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2209 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2210 // The thread must be runnable in order to drop the lock in timely fashion. 2211 // If the _owner is not runnable then spinning will not likely be 2212 // successful (profitable). 2213 // 2214 // Beware -- the thread referenced by _owner could have died 2215 // so a simply fetch from _owner->_thread_state might trap. 2216 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2217 // Because of the lifecycle issues the schedctl and _thread_state values 2218 // observed by NotRunnable() might be garbage. NotRunnable must 2219 // tolerate this and consider the observed _thread_state value 2220 // as advisory. 2221 // 2222 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2223 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2224 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2225 // with the LSB of _owner. Another option would be to probablistically probe 2226 // the putative _owner->TypeTag value. 2227 // 2228 // Checking _thread_state isn't perfect. Even if the thread is 2229 // in_java it might be blocked on a page-fault or have been preempted 2230 // and sitting on a ready/dispatch queue. _thread state in conjunction 2231 // with schedctl.sc_state gives us a good picture of what the 2232 // thread is doing, however. 2233 // 2234 // TODO: check schedctl.sc_state. 2235 // We'll need to use SafeFetch32() to read from the schedctl block. 2236 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2237 // 2238 // The return value from NotRunnable() is *advisory* -- the 2239 // result is based on sampling and is not necessarily coherent. 2240 // The caller must tolerate false-negative and false-positive errors. 2241 // Spinning, in general, is probabilistic anyway. 2242 2243 2244 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2245 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2246 if (!OwnerIsThread) return 0 ; 2247 2248 if (ox == NULL) return 0 ; 2249 2250 // Avoid transitive spinning ... 2251 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2252 // Immediately after T1 acquires L it's possible that T2, also 2253 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2254 // This occurs transiently after T1 acquired L but before 2255 // T1 managed to clear T1.Stalled. T2 does not need to abort 2256 // its spin in this circumstance. 2257 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2258 2259 if (BlockedOn == 1) return 1 ; 2260 if (BlockedOn != 0) { 2261 return BlockedOn != intptr_t(this) && _owner == ox ; 2262 } 2263 2264 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2265 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2266 // consider also: jst != _thread_in_Java -- but that's overspecific. 2267 return jst == _thread_blocked || jst == _thread_in_native ; 2268 } 2269 2270 2271 // ----------------------------------------------------------------------------- 2272 // WaitSet management ... 2273 2274 ObjectWaiter::ObjectWaiter(Thread* thread) { 2275 _next = NULL; 2276 _prev = NULL; 2277 _notified = 0; 2278 TState = TS_RUN ; 2279 _thread = thread; 2280 _event = thread->_ParkEvent ; 2281 _active = false; 2282 assert (_event != NULL, "invariant") ; 2283 } 2284 2285 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2286 JavaThread *jt = (JavaThread *)this->_thread; 2287 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2288 } 2289 2290 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2291 JavaThread *jt = (JavaThread *)this->_thread; 2292 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2293 } 2294 2295 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2296 assert(node != NULL, "should not dequeue NULL node"); 2297 assert(node->_prev == NULL, "node already in list"); 2298 assert(node->_next == NULL, "node already in list"); 2299 // put node at end of queue (circular doubly linked list) 2300 if (_WaitSet == NULL) { 2301 _WaitSet = node; 2302 node->_prev = node; 2303 node->_next = node; 2304 } else { 2305 ObjectWaiter* head = _WaitSet ; 2306 ObjectWaiter* tail = head->_prev; 2307 assert(tail->_next == head, "invariant check"); 2308 tail->_next = node; 2309 head->_prev = node; 2310 node->_next = head; 2311 node->_prev = tail; 2312 } 2313 } 2314 2315 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2316 // dequeue the very first waiter 2317 ObjectWaiter* waiter = _WaitSet; 2318 if (waiter) { 2319 DequeueSpecificWaiter(waiter); 2320 } 2321 return waiter; 2322 } 2323 2324 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2325 assert(node != NULL, "should not dequeue NULL node"); 2326 assert(node->_prev != NULL, "node already removed from list"); 2327 assert(node->_next != NULL, "node already removed from list"); 2328 // when the waiter has woken up because of interrupt, 2329 // timeout or other spurious wake-up, dequeue the 2330 // waiter from waiting list 2331 ObjectWaiter* next = node->_next; 2332 if (next == node) { 2333 assert(node->_prev == node, "invariant check"); 2334 _WaitSet = NULL; 2335 } else { 2336 ObjectWaiter* prev = node->_prev; 2337 assert(prev->_next == node, "invariant check"); 2338 assert(next->_prev == node, "invariant check"); 2339 next->_prev = prev; 2340 prev->_next = next; 2341 if (_WaitSet == node) { 2342 _WaitSet = next; 2343 } 2344 } 2345 node->_next = NULL; 2346 node->_prev = NULL; 2347 } 2348 2349 // ----------------------------------------------------------------------------- 2350 // PerfData support 2351 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2352 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2353 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2354 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2355 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2356 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2357 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2358 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2359 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2360 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2361 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2362 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2363 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2364 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2365 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2366 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2367 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2368 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2369 2370 // One-shot global initialization for the sync subsystem. 2371 // We could also defer initialization and initialize on-demand 2372 // the first time we call inflate(). Initialization would 2373 // be protected - like so many things - by the MonitorCache_lock. 2374 2375 void ObjectMonitor::Initialize () { 2376 static int InitializationCompleted = 0 ; 2377 assert (InitializationCompleted == 0, "invariant") ; 2378 InitializationCompleted = 1 ; 2379 if (UsePerfData) { 2380 EXCEPTION_MARK ; 2381 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2382 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2383 NEWPERFCOUNTER(_sync_Inflations) ; 2384 NEWPERFCOUNTER(_sync_Deflations) ; 2385 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2386 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2387 NEWPERFCOUNTER(_sync_Parks) ; 2388 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2389 NEWPERFCOUNTER(_sync_Notifications) ; 2390 NEWPERFCOUNTER(_sync_SlowEnter) ; 2391 NEWPERFCOUNTER(_sync_SlowExit) ; 2392 NEWPERFCOUNTER(_sync_SlowNotify) ; 2393 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2394 NEWPERFCOUNTER(_sync_FailedSpins) ; 2395 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2396 NEWPERFCOUNTER(_sync_PrivateA) ; 2397 NEWPERFCOUNTER(_sync_PrivateB) ; 2398 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2399 NEWPERFCOUNTER(_sync_MonScavenged) ; 2400 NEWPERFVARIABLE(_sync_MonExtant) ; 2401 #undef NEWPERFCOUNTER 2402 } 2403 } 2404 2405 2406 // Compile-time asserts 2407 // When possible, it's better to catch errors deterministically at 2408 // compile-time than at runtime. The down-side to using compile-time 2409 // asserts is that error message -- often something about negative array 2410 // indices -- is opaque. 2411 2412 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2413 2414 void ObjectMonitor::ctAsserts() { 2415 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2416 } 2417 2418 2419 static char * kvGet (char * kvList, const char * Key) { 2420 if (kvList == NULL) return NULL ; 2421 size_t n = strlen (Key) ; 2422 char * Search ; 2423 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2424 if (strncmp (Search, Key, n) == 0) { 2425 if (Search[n] == '=') return Search + n + 1 ; 2426 if (Search[n] == 0) return (char *) "1" ; 2427 } 2428 } 2429 return NULL ; 2430 } 2431 2432 static int kvGetInt (char * kvList, const char * Key, int Default) { 2433 char * v = kvGet (kvList, Key) ; 2434 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2435 if (Knob_ReportSettings && v != NULL) { 2436 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2437 ::fflush (stdout) ; 2438 } 2439 return rslt ; 2440 } 2441 2442 void ObjectMonitor::DeferredInitialize () { 2443 if (InitDone > 0) return ; 2444 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2445 while (InitDone != 1) ; 2446 return ; 2447 } 2448 2449 // One-shot global initialization ... 2450 // The initialization is idempotent, so we don't need locks. 2451 // In the future consider doing this via os::init_2(). 2452 // SyncKnobs consist of <Key>=<Value> pairs in the style 2453 // of environment variables. Start by converting ':' to NUL. 2454 2455 if (SyncKnobs == NULL) SyncKnobs = "" ; 2456 2457 size_t sz = strlen (SyncKnobs) ; 2458 char * knobs = (char *) malloc (sz + 2) ; 2459 if (knobs == NULL) { 2460 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; 2461 guarantee (0, "invariant") ; 2462 } 2463 strcpy (knobs, SyncKnobs) ; 2464 knobs[sz+1] = 0 ; 2465 for (char * p = knobs ; *p ; p++) { 2466 if (*p == ':') *p = 0 ; 2467 } 2468 2469 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2470 SETKNOB(ReportSettings) ; 2471 SETKNOB(Verbose) ; 2472 SETKNOB(FixedSpin) ; 2473 SETKNOB(SpinLimit) ; 2474 SETKNOB(SpinBase) ; 2475 SETKNOB(SpinBackOff); 2476 SETKNOB(CASPenalty) ; 2477 SETKNOB(OXPenalty) ; 2478 SETKNOB(LogSpins) ; 2479 SETKNOB(SpinSetSucc) ; 2480 SETKNOB(SuccEnabled) ; 2481 SETKNOB(SuccRestrict) ; 2482 SETKNOB(Penalty) ; 2483 SETKNOB(Bonus) ; 2484 SETKNOB(BonusB) ; 2485 SETKNOB(Poverty) ; 2486 SETKNOB(SpinAfterFutile) ; 2487 SETKNOB(UsePause) ; 2488 SETKNOB(SpinEarly) ; 2489 SETKNOB(OState) ; 2490 SETKNOB(MaxSpinners) ; 2491 SETKNOB(PreSpin) ; 2492 SETKNOB(ExitPolicy) ; 2493 SETKNOB(QMode); 2494 SETKNOB(ResetEvent) ; 2495 SETKNOB(MoveNotifyee) ; 2496 SETKNOB(FastHSSEC) ; 2497 #undef SETKNOB 2498 2499 if (os::is_MP()) { 2500 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2501 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2502 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2503 } else { 2504 Knob_SpinLimit = 0 ; 2505 Knob_SpinBase = 0 ; 2506 Knob_PreSpin = 0 ; 2507 Knob_FixedSpin = -1 ; 2508 } 2509 2510 if (Knob_LogSpins == 0) { 2511 ObjectMonitor::_sync_FailedSpins = NULL ; 2512 } 2513 2514 free (knobs) ; 2515 OrderAccess::fence() ; 2516 InitDone = 1 ; 2517 } 2518 2519 #ifndef PRODUCT 2520 void ObjectMonitor::verify() { 2521 } 2522 2523 void ObjectMonitor::print() { 2524 } 2525 #endif