1 /* 2 * Copyright (c) 1998, 2014, 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 // The current thread does not yet own the monitor and does not 390 // yet appear on any queues that would get it made the successor. 391 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 392 // handler cannot accidentally consume an unpark() meant for the 393 // ParkEvent associated with this ObjectMonitor. 394 } 395 396 OSThreadContendState osts(Self->osthread()); 397 ThreadBlockInVM tbivm(jt); 398 399 Self->set_current_pending_monitor(this); 400 401 // TODO-FIXME: change the following for(;;) loop to straight-line code. 402 for (;;) { 403 jt->set_suspend_equivalent(); 404 // cleared by handle_special_suspend_equivalent_condition() 405 // or java_suspend_self() 406 407 EnterI (THREAD) ; 408 409 if (!ExitSuspendEquivalent(jt)) break ; 410 411 // 412 // We have acquired the contended monitor, but while we were 413 // waiting another thread suspended us. We don't want to enter 414 // the monitor while suspended because that would surprise the 415 // thread that suspended us. 416 // 417 _recursions = 0 ; 418 _succ = NULL ; 419 exit (false, Self) ; 420 421 jt->java_suspend_self(); 422 } 423 Self->set_current_pending_monitor(NULL); 424 } 425 426 Atomic::dec_ptr(&_count); 427 assert (_count >= 0, "invariant") ; 428 Self->_Stalled = 0 ; 429 430 // Must either set _recursions = 0 or ASSERT _recursions == 0. 431 assert (_recursions == 0 , "invariant") ; 432 assert (_owner == Self , "invariant") ; 433 assert (_succ != Self , "invariant") ; 434 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 435 436 // The thread -- now the owner -- is back in vm mode. 437 // Report the glorious news via TI,DTrace and jvmstat. 438 // The probe effect is non-trivial. All the reportage occurs 439 // while we hold the monitor, increasing the length of the critical 440 // section. Amdahl's parallel speedup law comes vividly into play. 441 // 442 // Another option might be to aggregate the events (thread local or 443 // per-monitor aggregation) and defer reporting until a more opportune 444 // time -- such as next time some thread encounters contention but has 445 // yet to acquire the lock. While spinning that thread could 446 // spinning we could increment JVMStat counters, etc. 447 448 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 449 if (JvmtiExport::should_post_monitor_contended_entered()) { 450 JvmtiExport::post_monitor_contended_entered(jt, this); 451 452 // The current thread already owns the monitor and is not going to 453 // call park() for the remainder of the monitor enter protocol. So 454 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 455 // event handler consumed an unpark() issued by the thread that 456 // just exited the monitor. 457 } 458 459 if (event.should_commit()) { 460 event.set_klass(((oop)this->object())->klass()); 461 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 462 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 463 event.commit(); 464 } 465 466 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 467 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 468 } 469 } 470 471 472 // Caveat: TryLock() is not necessarily serializing if it returns failure. 473 // Callers must compensate as needed. 474 475 int ObjectMonitor::TryLock (Thread * Self) { 476 for (;;) { 477 void * own = _owner ; 478 if (own != NULL) return 0 ; 479 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 480 // Either guarantee _recursions == 0 or set _recursions = 0. 481 assert (_recursions == 0, "invariant") ; 482 assert (_owner == Self, "invariant") ; 483 // CONSIDER: set or assert that OwnerIsThread == 1 484 return 1 ; 485 } 486 // The lock had been free momentarily, but we lost the race to the lock. 487 // Interference -- the CAS failed. 488 // We can either return -1 or retry. 489 // Retry doesn't make as much sense because the lock was just acquired. 490 if (true) return -1 ; 491 } 492 } 493 494 void ATTR ObjectMonitor::EnterI (TRAPS) { 495 Thread * Self = THREAD ; 496 assert (Self->is_Java_thread(), "invariant") ; 497 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 498 499 // Try the lock - TATAS 500 if (TryLock (Self) > 0) { 501 assert (_succ != Self , "invariant") ; 502 assert (_owner == Self , "invariant") ; 503 assert (_Responsible != Self , "invariant") ; 504 return ; 505 } 506 507 DeferredInitialize () ; 508 509 // We try one round of spinning *before* enqueueing Self. 510 // 511 // If the _owner is ready but OFFPROC we could use a YieldTo() 512 // operation to donate the remainder of this thread's quantum 513 // to the owner. This has subtle but beneficial affinity 514 // effects. 515 516 if (TrySpin (Self) > 0) { 517 assert (_owner == Self , "invariant") ; 518 assert (_succ != Self , "invariant") ; 519 assert (_Responsible != Self , "invariant") ; 520 return ; 521 } 522 523 // The Spin failed -- Enqueue and park the thread ... 524 assert (_succ != Self , "invariant") ; 525 assert (_owner != Self , "invariant") ; 526 assert (_Responsible != Self , "invariant") ; 527 528 // Enqueue "Self" on ObjectMonitor's _cxq. 529 // 530 // Node acts as a proxy for Self. 531 // As an aside, if were to ever rewrite the synchronization code mostly 532 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 533 // Java objects. This would avoid awkward lifecycle and liveness issues, 534 // as well as eliminate a subset of ABA issues. 535 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 536 // 537 538 ObjectWaiter node(Self) ; 539 Self->_ParkEvent->reset() ; 540 node._prev = (ObjectWaiter *) 0xBAD ; 541 node.TState = ObjectWaiter::TS_CXQ ; 542 543 // Push "Self" onto the front of the _cxq. 544 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 545 // Note that spinning tends to reduce the rate at which threads 546 // enqueue and dequeue on EntryList|cxq. 547 ObjectWaiter * nxt ; 548 for (;;) { 549 node._next = nxt = _cxq ; 550 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 551 552 // Interference - the CAS failed because _cxq changed. Just retry. 553 // As an optional optimization we retry the lock. 554 if (TryLock (Self) > 0) { 555 assert (_succ != Self , "invariant") ; 556 assert (_owner == Self , "invariant") ; 557 assert (_Responsible != Self , "invariant") ; 558 return ; 559 } 560 } 561 562 // Check for cxq|EntryList edge transition to non-null. This indicates 563 // the onset of contention. While contention persists exiting threads 564 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 565 // operations revert to the faster 1-0 mode. This enter operation may interleave 566 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 567 // arrange for one of the contending thread to use a timed park() operations 568 // to detect and recover from the race. (Stranding is form of progress failure 569 // where the monitor is unlocked but all the contending threads remain parked). 570 // That is, at least one of the contended threads will periodically poll _owner. 571 // One of the contending threads will become the designated "Responsible" thread. 572 // The Responsible thread uses a timed park instead of a normal indefinite park 573 // operation -- it periodically wakes and checks for and recovers from potential 574 // strandings admitted by 1-0 exit operations. We need at most one Responsible 575 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 576 // be responsible for a monitor. 577 // 578 // Currently, one of the contended threads takes on the added role of "Responsible". 579 // A viable alternative would be to use a dedicated "stranding checker" thread 580 // that periodically iterated over all the threads (or active monitors) and unparked 581 // successors where there was risk of stranding. This would help eliminate the 582 // timer scalability issues we see on some platforms as we'd only have one thread 583 // -- the checker -- parked on a timer. 584 585 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 586 // Try to assume the role of responsible thread for the monitor. 587 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 588 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 589 } 590 591 // The lock have been released while this thread was occupied queueing 592 // itself onto _cxq. To close the race and avoid "stranding" and 593 // progress-liveness failure we must resample-retry _owner before parking. 594 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 595 // In this case the ST-MEMBAR is accomplished with CAS(). 596 // 597 // TODO: Defer all thread state transitions until park-time. 598 // Since state transitions are heavy and inefficient we'd like 599 // to defer the state transitions until absolutely necessary, 600 // and in doing so avoid some transitions ... 601 602 TEVENT (Inflated enter - Contention) ; 603 int nWakeups = 0 ; 604 int RecheckInterval = 1 ; 605 606 for (;;) { 607 608 if (TryLock (Self) > 0) break ; 609 assert (_owner != Self, "invariant") ; 610 611 if ((SyncFlags & 2) && _Responsible == NULL) { 612 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 613 } 614 615 // park self 616 if (_Responsible == Self || (SyncFlags & 1)) { 617 TEVENT (Inflated enter - park TIMED) ; 618 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 619 // Increase the RecheckInterval, but clamp the value. 620 RecheckInterval *= 8 ; 621 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 622 } else { 623 TEVENT (Inflated enter - park UNTIMED) ; 624 Self->_ParkEvent->park() ; 625 } 626 627 if (TryLock(Self) > 0) break ; 628 629 // The lock is still contested. 630 // Keep a tally of the # of futile wakeups. 631 // Note that the counter is not protected by a lock or updated by atomics. 632 // That is by design - we trade "lossy" counters which are exposed to 633 // races during updates for a lower probe effect. 634 TEVENT (Inflated enter - Futile wakeup) ; 635 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 636 ObjectMonitor::_sync_FutileWakeups->inc() ; 637 } 638 ++ nWakeups ; 639 640 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 641 // We can defer clearing _succ until after the spin completes 642 // TrySpin() must tolerate being called with _succ == Self. 643 // Try yet another round of adaptive spinning. 644 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 645 646 // We can find that we were unpark()ed and redesignated _succ while 647 // we were spinning. That's harmless. If we iterate and call park(), 648 // park() will consume the event and return immediately and we'll 649 // just spin again. This pattern can repeat, leaving _succ to simply 650 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 651 // Alternately, we can sample fired() here, and if set, forgo spinning 652 // in the next iteration. 653 654 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 655 Self->_ParkEvent->reset() ; 656 OrderAccess::fence() ; 657 } 658 if (_succ == Self) _succ = NULL ; 659 660 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 661 OrderAccess::fence() ; 662 } 663 664 // Egress : 665 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 666 // Normally we'll find Self on the EntryList . 667 // From the perspective of the lock owner (this thread), the 668 // EntryList is stable and cxq is prepend-only. 669 // The head of cxq is volatile but the interior is stable. 670 // In addition, Self.TState is stable. 671 672 assert (_owner == Self , "invariant") ; 673 assert (object() != NULL , "invariant") ; 674 // I'd like to write: 675 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 676 // but as we're at a safepoint that's not safe. 677 678 UnlinkAfterAcquire (Self, &node) ; 679 if (_succ == Self) _succ = NULL ; 680 681 assert (_succ != Self, "invariant") ; 682 if (_Responsible == Self) { 683 _Responsible = NULL ; 684 OrderAccess::fence(); // Dekker pivot-point 685 686 // We may leave threads on cxq|EntryList without a designated 687 // "Responsible" thread. This is benign. When this thread subsequently 688 // exits the monitor it can "see" such preexisting "old" threads -- 689 // threads that arrived on the cxq|EntryList before the fence, above -- 690 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 691 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 692 // non-null and elect a new "Responsible" timer thread. 693 // 694 // This thread executes: 695 // ST Responsible=null; MEMBAR (in enter epilog - here) 696 // LD cxq|EntryList (in subsequent exit) 697 // 698 // Entering threads in the slow/contended path execute: 699 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 700 // The (ST cxq; MEMBAR) is accomplished with CAS(). 701 // 702 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 703 // exit operation from floating above the ST Responsible=null. 704 } 705 706 // We've acquired ownership with CAS(). 707 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 708 // But since the CAS() this thread may have also stored into _succ, 709 // EntryList, cxq or Responsible. These meta-data updates must be 710 // visible __before this thread subsequently drops the lock. 711 // Consider what could occur if we didn't enforce this constraint -- 712 // STs to monitor meta-data and user-data could reorder with (become 713 // visible after) the ST in exit that drops ownership of the lock. 714 // Some other thread could then acquire the lock, but observe inconsistent 715 // or old monitor meta-data and heap data. That violates the JMM. 716 // To that end, the 1-0 exit() operation must have at least STST|LDST 717 // "release" barrier semantics. Specifically, there must be at least a 718 // STST|LDST barrier in exit() before the ST of null into _owner that drops 719 // the lock. The barrier ensures that changes to monitor meta-data and data 720 // protected by the lock will be visible before we release the lock, and 721 // therefore before some other thread (CPU) has a chance to acquire the lock. 722 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 723 // 724 // Critically, any prior STs to _succ or EntryList must be visible before 725 // the ST of null into _owner in the *subsequent* (following) corresponding 726 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 727 // execute a serializing instruction. 728 729 if (SyncFlags & 8) { 730 OrderAccess::fence() ; 731 } 732 return ; 733 } 734 735 // ReenterI() is a specialized inline form of the latter half of the 736 // contended slow-path from EnterI(). We use ReenterI() only for 737 // monitor reentry in wait(). 738 // 739 // In the future we should reconcile EnterI() and ReenterI(), adding 740 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 741 // loop accordingly. 742 743 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 744 assert (Self != NULL , "invariant") ; 745 assert (SelfNode != NULL , "invariant") ; 746 assert (SelfNode->_thread == Self , "invariant") ; 747 assert (_waiters > 0 , "invariant") ; 748 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 749 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 750 JavaThread * jt = (JavaThread *) Self ; 751 752 int nWakeups = 0 ; 753 for (;;) { 754 ObjectWaiter::TStates v = SelfNode->TState ; 755 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 756 assert (_owner != Self, "invariant") ; 757 758 if (TryLock (Self) > 0) break ; 759 if (TrySpin (Self) > 0) break ; 760 761 TEVENT (Wait Reentry - parking) ; 762 763 // State transition wrappers around park() ... 764 // ReenterI() wisely defers state transitions until 765 // it's clear we must park the thread. 766 { 767 OSThreadContendState osts(Self->osthread()); 768 ThreadBlockInVM tbivm(jt); 769 770 // cleared by handle_special_suspend_equivalent_condition() 771 // or java_suspend_self() 772 jt->set_suspend_equivalent(); 773 if (SyncFlags & 1) { 774 Self->_ParkEvent->park ((jlong)1000) ; 775 } else { 776 Self->_ParkEvent->park () ; 777 } 778 779 // were we externally suspended while we were waiting? 780 for (;;) { 781 if (!ExitSuspendEquivalent (jt)) break ; 782 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 783 jt->java_suspend_self(); 784 jt->set_suspend_equivalent(); 785 } 786 } 787 788 // Try again, but just so we distinguish between futile wakeups and 789 // successful wakeups. The following test isn't algorithmically 790 // necessary, but it helps us maintain sensible statistics. 791 if (TryLock(Self) > 0) break ; 792 793 // The lock is still contested. 794 // Keep a tally of the # of futile wakeups. 795 // Note that the counter is not protected by a lock or updated by atomics. 796 // That is by design - we trade "lossy" counters which are exposed to 797 // races during updates for a lower probe effect. 798 TEVENT (Wait Reentry - futile wakeup) ; 799 ++ nWakeups ; 800 801 // Assuming this is not a spurious wakeup we'll normally 802 // find that _succ == Self. 803 if (_succ == Self) _succ = NULL ; 804 805 // Invariant: after clearing _succ a contending thread 806 // *must* retry _owner before parking. 807 OrderAccess::fence() ; 808 809 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 810 ObjectMonitor::_sync_FutileWakeups->inc() ; 811 } 812 } 813 814 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 815 // Normally we'll find Self on the EntryList. 816 // Unlinking from the EntryList is constant-time and atomic-free. 817 // From the perspective of the lock owner (this thread), the 818 // EntryList is stable and cxq is prepend-only. 819 // The head of cxq is volatile but the interior is stable. 820 // In addition, Self.TState is stable. 821 822 assert (_owner == Self, "invariant") ; 823 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 824 UnlinkAfterAcquire (Self, SelfNode) ; 825 if (_succ == Self) _succ = NULL ; 826 assert (_succ != Self, "invariant") ; 827 SelfNode->TState = ObjectWaiter::TS_RUN ; 828 OrderAccess::fence() ; // see comments at the end of EnterI() 829 } 830 831 // after the thread acquires the lock in ::enter(). Equally, we could defer 832 // unlinking the thread until ::exit()-time. 833 834 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 835 { 836 assert (_owner == Self, "invariant") ; 837 assert (SelfNode->_thread == Self, "invariant") ; 838 839 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 840 // Normal case: remove Self from the DLL EntryList . 841 // This is a constant-time operation. 842 ObjectWaiter * nxt = SelfNode->_next ; 843 ObjectWaiter * prv = SelfNode->_prev ; 844 if (nxt != NULL) nxt->_prev = prv ; 845 if (prv != NULL) prv->_next = nxt ; 846 if (SelfNode == _EntryList ) _EntryList = nxt ; 847 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 848 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 849 TEVENT (Unlink from EntryList) ; 850 } else { 851 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 852 // Inopportune interleaving -- Self is still on the cxq. 853 // This usually means the enqueue of self raced an exiting thread. 854 // Normally we'll find Self near the front of the cxq, so 855 // dequeueing is typically fast. If needbe we can accelerate 856 // this with some MCS/CHL-like bidirectional list hints and advisory 857 // back-links so dequeueing from the interior will normally operate 858 // in constant-time. 859 // Dequeue Self from either the head (with CAS) or from the interior 860 // with a linear-time scan and normal non-atomic memory operations. 861 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 862 // and then unlink Self from EntryList. We have to drain eventually, 863 // so it might as well be now. 864 865 ObjectWaiter * v = _cxq ; 866 assert (v != NULL, "invariant") ; 867 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 868 // The CAS above can fail from interference IFF a "RAT" arrived. 869 // In that case Self must be in the interior and can no longer be 870 // at the head of cxq. 871 if (v == SelfNode) { 872 assert (_cxq != v, "invariant") ; 873 v = _cxq ; // CAS above failed - start scan at head of list 874 } 875 ObjectWaiter * p ; 876 ObjectWaiter * q = NULL ; 877 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 878 q = p ; 879 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 880 } 881 assert (v != SelfNode, "invariant") ; 882 assert (p == SelfNode, "Node not found on cxq") ; 883 assert (p != _cxq, "invariant") ; 884 assert (q != NULL, "invariant") ; 885 assert (q->_next == p, "invariant") ; 886 q->_next = p->_next ; 887 } 888 TEVENT (Unlink from cxq) ; 889 } 890 891 // Diagnostic hygiene ... 892 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 893 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 894 SelfNode->TState = ObjectWaiter::TS_RUN ; 895 } 896 897 // ----------------------------------------------------------------------------- 898 // Exit support 899 // 900 // exit() 901 // ~~~~~~ 902 // Note that the collector can't reclaim the objectMonitor or deflate 903 // the object out from underneath the thread calling ::exit() as the 904 // thread calling ::exit() never transitions to a stable state. 905 // This inhibits GC, which in turn inhibits asynchronous (and 906 // inopportune) reclamation of "this". 907 // 908 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 909 // There's one exception to the claim above, however. EnterI() can call 910 // exit() to drop a lock if the acquirer has been externally suspended. 911 // In that case exit() is called with _thread_state as _thread_blocked, 912 // but the monitor's _count field is > 0, which inhibits reclamation. 913 // 914 // 1-0 exit 915 // ~~~~~~~~ 916 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 917 // the fast-path operators have been optimized so the common ::exit() 918 // operation is 1-0. See i486.ad fast_unlock(), for instance. 919 // The code emitted by fast_unlock() elides the usual MEMBAR. This 920 // greatly improves latency -- MEMBAR and CAS having considerable local 921 // latency on modern processors -- but at the cost of "stranding". Absent the 922 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 923 // ::enter() path, resulting in the entering thread being stranding 924 // and a progress-liveness failure. Stranding is extremely rare. 925 // We use timers (timed park operations) & periodic polling to detect 926 // and recover from stranding. Potentially stranded threads periodically 927 // wake up and poll the lock. See the usage of the _Responsible variable. 928 // 929 // The CAS() in enter provides for safety and exclusion, while the CAS or 930 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 931 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 932 // We detect and recover from stranding with timers. 933 // 934 // If a thread transiently strands it'll park until (a) another 935 // thread acquires the lock and then drops the lock, at which time the 936 // exiting thread will notice and unpark the stranded thread, or, (b) 937 // the timer expires. If the lock is high traffic then the stranding latency 938 // will be low due to (a). If the lock is low traffic then the odds of 939 // stranding are lower, although the worst-case stranding latency 940 // is longer. Critically, we don't want to put excessive load in the 941 // platform's timer subsystem. We want to minimize both the timer injection 942 // rate (timers created/sec) as well as the number of timers active at 943 // any one time. (more precisely, we want to minimize timer-seconds, which is 944 // the integral of the # of active timers at any instant over time). 945 // Both impinge on OS scalability. Given that, at most one thread parked on 946 // a monitor will use a timer. 947 948 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 949 Thread * Self = THREAD ; 950 if (THREAD != _owner) { 951 if (THREAD->is_lock_owned((address) _owner)) { 952 // Transmute _owner from a BasicLock pointer to a Thread address. 953 // We don't need to hold _mutex for this transition. 954 // Non-null to Non-null is safe as long as all readers can 955 // tolerate either flavor. 956 assert (_recursions == 0, "invariant") ; 957 _owner = THREAD ; 958 _recursions = 0 ; 959 OwnerIsThread = 1 ; 960 } else { 961 // NOTE: we need to handle unbalanced monitor enter/exit 962 // in native code by throwing an exception. 963 // TODO: Throw an IllegalMonitorStateException ? 964 TEVENT (Exit - Throw IMSX) ; 965 assert(false, "Non-balanced monitor enter/exit!"); 966 if (false) { 967 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 968 } 969 return; 970 } 971 } 972 973 if (_recursions != 0) { 974 _recursions--; // this is simple recursive enter 975 TEVENT (Inflated exit - recursive) ; 976 return ; 977 } 978 979 // Invariant: after setting Responsible=null an thread must execute 980 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 981 if ((SyncFlags & 4) == 0) { 982 _Responsible = NULL ; 983 } 984 985 #if INCLUDE_TRACE 986 // get the owner's thread id for the MonitorEnter event 987 // if it is enabled and the thread isn't suspended 988 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 989 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 990 } 991 #endif 992 993 for (;;) { 994 assert (THREAD == _owner, "invariant") ; 995 996 997 if (Knob_ExitPolicy == 0) { 998 // release semantics: prior loads and stores from within the critical section 999 // must not float (reorder) past the following store that drops the lock. 1000 // On SPARC that requires MEMBAR #loadstore|#storestore. 1001 // But of course in TSO #loadstore|#storestore is not required. 1002 // I'd like to write one of the following: 1003 // A. OrderAccess::release() ; _owner = NULL 1004 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 1005 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 1006 // store into a _dummy variable. That store is not needed, but can result 1007 // in massive wasteful coherency traffic on classic SMP systems. 1008 // Instead, I use release_store(), which is implemented as just a simple 1009 // ST on x64, x86 and SPARC. 1010 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1011 OrderAccess::storeload() ; // See if we need to wake a successor 1012 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1013 TEVENT (Inflated exit - simple egress) ; 1014 return ; 1015 } 1016 TEVENT (Inflated exit - complex egress) ; 1017 1018 // Normally the exiting thread is responsible for ensuring succession, 1019 // but if other successors are ready or other entering threads are spinning 1020 // then this thread can simply store NULL into _owner and exit without 1021 // waking a successor. The existence of spinners or ready successors 1022 // guarantees proper succession (liveness). Responsibility passes to the 1023 // ready or running successors. The exiting thread delegates the duty. 1024 // More precisely, if a successor already exists this thread is absolved 1025 // of the responsibility of waking (unparking) one. 1026 // 1027 // The _succ variable is critical to reducing futile wakeup frequency. 1028 // _succ identifies the "heir presumptive" thread that has been made 1029 // ready (unparked) but that has not yet run. We need only one such 1030 // successor thread to guarantee progress. 1031 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1032 // section 3.3 "Futile Wakeup Throttling" for details. 1033 // 1034 // Note that spinners in Enter() also set _succ non-null. 1035 // In the current implementation spinners opportunistically set 1036 // _succ so that exiting threads might avoid waking a successor. 1037 // Another less appealing alternative would be for the exiting thread 1038 // to drop the lock and then spin briefly to see if a spinner managed 1039 // to acquire the lock. If so, the exiting thread could exit 1040 // immediately without waking a successor, otherwise the exiting 1041 // thread would need to dequeue and wake a successor. 1042 // (Note that we'd need to make the post-drop spin short, but no 1043 // shorter than the worst-case round-trip cache-line migration time. 1044 // The dropped lock needs to become visible to the spinner, and then 1045 // the acquisition of the lock by the spinner must become visible to 1046 // the exiting thread). 1047 // 1048 1049 // It appears that an heir-presumptive (successor) must be made ready. 1050 // Only the current lock owner can manipulate the EntryList or 1051 // drain _cxq, so we need to reacquire the lock. If we fail 1052 // to reacquire the lock the responsibility for ensuring succession 1053 // falls to the new owner. 1054 // 1055 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1056 return ; 1057 } 1058 TEVENT (Exit - Reacquired) ; 1059 } else { 1060 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1061 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1062 OrderAccess::storeload() ; 1063 // Ratify the previously observed values. 1064 if (_cxq == NULL || _succ != NULL) { 1065 TEVENT (Inflated exit - simple egress) ; 1066 return ; 1067 } 1068 1069 // inopportune interleaving -- the exiting thread (this thread) 1070 // in the fast-exit path raced an entering thread in the slow-enter 1071 // path. 1072 // We have two choices: 1073 // A. Try to reacquire the lock. 1074 // If the CAS() fails return immediately, otherwise 1075 // we either restart/rerun the exit operation, or simply 1076 // fall-through into the code below which wakes a successor. 1077 // B. If the elements forming the EntryList|cxq are TSM 1078 // we could simply unpark() the lead thread and return 1079 // without having set _succ. 1080 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1081 TEVENT (Inflated exit - reacquired succeeded) ; 1082 return ; 1083 } 1084 TEVENT (Inflated exit - reacquired failed) ; 1085 } else { 1086 TEVENT (Inflated exit - complex egress) ; 1087 } 1088 } 1089 1090 guarantee (_owner == THREAD, "invariant") ; 1091 1092 ObjectWaiter * w = NULL ; 1093 int QMode = Knob_QMode ; 1094 1095 if (QMode == 2 && _cxq != NULL) { 1096 // QMode == 2 : cxq has precedence over EntryList. 1097 // Try to directly wake a successor from the cxq. 1098 // If successful, the successor will need to unlink itself from cxq. 1099 w = _cxq ; 1100 assert (w != NULL, "invariant") ; 1101 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1102 ExitEpilog (Self, w) ; 1103 return ; 1104 } 1105 1106 if (QMode == 3 && _cxq != NULL) { 1107 // Aggressively drain cxq into EntryList at the first opportunity. 1108 // This policy ensure that recently-run threads live at the head of EntryList. 1109 // Drain _cxq into EntryList - bulk transfer. 1110 // First, detach _cxq. 1111 // The following loop is tantamount to: w = swap (&cxq, NULL) 1112 w = _cxq ; 1113 for (;;) { 1114 assert (w != NULL, "Invariant") ; 1115 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1116 if (u == w) break ; 1117 w = u ; 1118 } 1119 assert (w != NULL , "invariant") ; 1120 1121 ObjectWaiter * q = NULL ; 1122 ObjectWaiter * p ; 1123 for (p = w ; p != NULL ; p = p->_next) { 1124 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1125 p->TState = ObjectWaiter::TS_ENTER ; 1126 p->_prev = q ; 1127 q = p ; 1128 } 1129 1130 // Append the RATs to the EntryList 1131 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1132 ObjectWaiter * Tail ; 1133 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1134 if (Tail == NULL) { 1135 _EntryList = w ; 1136 } else { 1137 Tail->_next = w ; 1138 w->_prev = Tail ; 1139 } 1140 1141 // Fall thru into code that tries to wake a successor from EntryList 1142 } 1143 1144 if (QMode == 4 && _cxq != NULL) { 1145 // Aggressively drain cxq into EntryList at the first opportunity. 1146 // This policy ensure that recently-run threads live at the head of EntryList. 1147 1148 // Drain _cxq into EntryList - bulk transfer. 1149 // First, detach _cxq. 1150 // The following loop is tantamount to: w = swap (&cxq, NULL) 1151 w = _cxq ; 1152 for (;;) { 1153 assert (w != NULL, "Invariant") ; 1154 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1155 if (u == w) break ; 1156 w = u ; 1157 } 1158 assert (w != NULL , "invariant") ; 1159 1160 ObjectWaiter * q = NULL ; 1161 ObjectWaiter * p ; 1162 for (p = w ; p != NULL ; p = p->_next) { 1163 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1164 p->TState = ObjectWaiter::TS_ENTER ; 1165 p->_prev = q ; 1166 q = p ; 1167 } 1168 1169 // Prepend the RATs to the EntryList 1170 if (_EntryList != NULL) { 1171 q->_next = _EntryList ; 1172 _EntryList->_prev = q ; 1173 } 1174 _EntryList = w ; 1175 1176 // Fall thru into code that tries to wake a successor from EntryList 1177 } 1178 1179 w = _EntryList ; 1180 if (w != NULL) { 1181 // I'd like to write: guarantee (w->_thread != Self). 1182 // But in practice an exiting thread may find itself on the EntryList. 1183 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1184 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1185 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1186 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1187 // release the lock "O". T2 resumes immediately after the ST of null into 1188 // _owner, above. T2 notices that the EntryList is populated, so it 1189 // reacquires the lock and then finds itself on the EntryList. 1190 // Given all that, we have to tolerate the circumstance where "w" is 1191 // associated with Self. 1192 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1193 ExitEpilog (Self, w) ; 1194 return ; 1195 } 1196 1197 // If we find that both _cxq and EntryList are null then just 1198 // re-run the exit protocol from the top. 1199 w = _cxq ; 1200 if (w == NULL) continue ; 1201 1202 // Drain _cxq into EntryList - bulk transfer. 1203 // First, detach _cxq. 1204 // The following loop is tantamount to: w = swap (&cxq, NULL) 1205 for (;;) { 1206 assert (w != NULL, "Invariant") ; 1207 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1208 if (u == w) break ; 1209 w = u ; 1210 } 1211 TEVENT (Inflated exit - drain cxq into EntryList) ; 1212 1213 assert (w != NULL , "invariant") ; 1214 assert (_EntryList == NULL , "invariant") ; 1215 1216 // Convert the LIFO SLL anchored by _cxq into a DLL. 1217 // The list reorganization step operates in O(LENGTH(w)) time. 1218 // It's critical that this step operate quickly as 1219 // "Self" still holds the outer-lock, restricting parallelism 1220 // and effectively lengthening the critical section. 1221 // Invariant: s chases t chases u. 1222 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1223 // we have faster access to the tail. 1224 1225 if (QMode == 1) { 1226 // QMode == 1 : drain cxq to EntryList, reversing order 1227 // We also reverse the order of the list. 1228 ObjectWaiter * s = NULL ; 1229 ObjectWaiter * t = w ; 1230 ObjectWaiter * u = NULL ; 1231 while (t != NULL) { 1232 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1233 t->TState = ObjectWaiter::TS_ENTER ; 1234 u = t->_next ; 1235 t->_prev = u ; 1236 t->_next = s ; 1237 s = t; 1238 t = u ; 1239 } 1240 _EntryList = s ; 1241 assert (s != NULL, "invariant") ; 1242 } else { 1243 // QMode == 0 or QMode == 2 1244 _EntryList = w ; 1245 ObjectWaiter * q = NULL ; 1246 ObjectWaiter * p ; 1247 for (p = w ; p != NULL ; p = p->_next) { 1248 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1249 p->TState = ObjectWaiter::TS_ENTER ; 1250 p->_prev = q ; 1251 q = p ; 1252 } 1253 } 1254 1255 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1256 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1257 1258 // See if we can abdicate to a spinner instead of waking a thread. 1259 // A primary goal of the implementation is to reduce the 1260 // context-switch rate. 1261 if (_succ != NULL) continue; 1262 1263 w = _EntryList ; 1264 if (w != NULL) { 1265 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1266 ExitEpilog (Self, w) ; 1267 return ; 1268 } 1269 } 1270 } 1271 1272 // ExitSuspendEquivalent: 1273 // A faster alternate to handle_special_suspend_equivalent_condition() 1274 // 1275 // handle_special_suspend_equivalent_condition() unconditionally 1276 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1277 // operations have high latency. Note that in ::enter() we call HSSEC 1278 // while holding the monitor, so we effectively lengthen the critical sections. 1279 // 1280 // There are a number of possible solutions: 1281 // 1282 // A. To ameliorate the problem we might also defer state transitions 1283 // to as late as possible -- just prior to parking. 1284 // Given that, we'd call HSSEC after having returned from park(), 1285 // but before attempting to acquire the monitor. This is only a 1286 // partial solution. It avoids calling HSSEC while holding the 1287 // monitor (good), but it still increases successor reacquisition latency -- 1288 // the interval between unparking a successor and the time the successor 1289 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1290 // If we use this technique we can also avoid EnterI()-exit() loop 1291 // in ::enter() where we iteratively drop the lock and then attempt 1292 // to reacquire it after suspending. 1293 // 1294 // B. In the future we might fold all the suspend bits into a 1295 // composite per-thread suspend flag and then update it with CAS(). 1296 // Alternately, a Dekker-like mechanism with multiple variables 1297 // would suffice: 1298 // ST Self->_suspend_equivalent = false 1299 // MEMBAR 1300 // LD Self_>_suspend_flags 1301 // 1302 1303 1304 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1305 int Mode = Knob_FastHSSEC ; 1306 if (Mode && !jSelf->is_external_suspend()) { 1307 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1308 jSelf->clear_suspend_equivalent() ; 1309 if (2 == Mode) OrderAccess::storeload() ; 1310 if (!jSelf->is_external_suspend()) return false ; 1311 // We raced a suspension -- fall thru into the slow path 1312 TEVENT (ExitSuspendEquivalent - raced) ; 1313 jSelf->set_suspend_equivalent() ; 1314 } 1315 return jSelf->handle_special_suspend_equivalent_condition() ; 1316 } 1317 1318 1319 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1320 assert (_owner == Self, "invariant") ; 1321 1322 // Exit protocol: 1323 // 1. ST _succ = wakee 1324 // 2. membar #loadstore|#storestore; 1325 // 2. ST _owner = NULL 1326 // 3. unpark(wakee) 1327 1328 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1329 ParkEvent * Trigger = Wakee->_event ; 1330 1331 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1332 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1333 // out-of-scope (non-extant). 1334 Wakee = NULL ; 1335 1336 // Drop the lock 1337 OrderAccess::release_store_ptr (&_owner, NULL) ; 1338 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1339 1340 if (SafepointSynchronize::do_call_back()) { 1341 TEVENT (unpark before SAFEPOINT) ; 1342 } 1343 1344 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1345 Trigger->unpark() ; 1346 1347 // Maintain stats and report events to JVMTI 1348 if (ObjectMonitor::_sync_Parks != NULL) { 1349 ObjectMonitor::_sync_Parks->inc() ; 1350 } 1351 } 1352 1353 1354 // ----------------------------------------------------------------------------- 1355 // Class Loader deadlock handling. 1356 // 1357 // complete_exit exits a lock returning recursion count 1358 // complete_exit/reenter operate as a wait without waiting 1359 // complete_exit requires an inflated monitor 1360 // The _owner field is not always the Thread addr even with an 1361 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1362 // thread due to contention. 1363 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1364 Thread * const Self = THREAD; 1365 assert(Self->is_Java_thread(), "Must be Java thread!"); 1366 JavaThread *jt = (JavaThread *)THREAD; 1367 1368 DeferredInitialize(); 1369 1370 if (THREAD != _owner) { 1371 if (THREAD->is_lock_owned ((address)_owner)) { 1372 assert(_recursions == 0, "internal state error"); 1373 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1374 _recursions = 0 ; 1375 OwnerIsThread = 1 ; 1376 } 1377 } 1378 1379 guarantee(Self == _owner, "complete_exit not owner"); 1380 intptr_t save = _recursions; // record the old recursion count 1381 _recursions = 0; // set the recursion level to be 0 1382 exit (true, Self) ; // exit the monitor 1383 guarantee (_owner != Self, "invariant"); 1384 return save; 1385 } 1386 1387 // reenter() enters a lock and sets recursion count 1388 // complete_exit/reenter operate as a wait without waiting 1389 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1390 Thread * const Self = THREAD; 1391 assert(Self->is_Java_thread(), "Must be Java thread!"); 1392 JavaThread *jt = (JavaThread *)THREAD; 1393 1394 guarantee(_owner != Self, "reenter already owner"); 1395 enter (THREAD); // enter the monitor 1396 guarantee (_recursions == 0, "reenter recursion"); 1397 _recursions = recursions; 1398 return; 1399 } 1400 1401 1402 // ----------------------------------------------------------------------------- 1403 // A macro is used below because there may already be a pending 1404 // exception which should not abort the execution of the routines 1405 // which use this (which is why we don't put this into check_slow and 1406 // call it with a CHECK argument). 1407 1408 #define CHECK_OWNER() \ 1409 do { \ 1410 if (THREAD != _owner) { \ 1411 if (THREAD->is_lock_owned((address) _owner)) { \ 1412 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1413 _recursions = 0; \ 1414 OwnerIsThread = 1 ; \ 1415 } else { \ 1416 TEVENT (Throw IMSX) ; \ 1417 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1418 } \ 1419 } \ 1420 } while (false) 1421 1422 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1423 // TODO-FIXME: remove check_slow() -- it's likely dead. 1424 1425 void ObjectMonitor::check_slow(TRAPS) { 1426 TEVENT (check_slow - throw IMSX) ; 1427 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1428 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1429 } 1430 1431 static int Adjust (volatile int * adr, int dx) { 1432 int v ; 1433 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1434 return v ; 1435 } 1436 1437 // helper method for posting a monitor wait event 1438 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1439 jlong notifier_tid, 1440 jlong timeout, 1441 bool timedout) { 1442 event->set_klass(((oop)this->object())->klass()); 1443 event->set_timeout((TYPE_ULONG)timeout); 1444 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1445 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1446 event->set_timedOut((TYPE_BOOLEAN)timedout); 1447 event->commit(); 1448 } 1449 1450 // ----------------------------------------------------------------------------- 1451 // Wait/Notify/NotifyAll 1452 // 1453 // Note: a subset of changes to ObjectMonitor::wait() 1454 // will need to be replicated in complete_exit above 1455 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1456 Thread * const Self = THREAD ; 1457 assert(Self->is_Java_thread(), "Must be Java thread!"); 1458 JavaThread *jt = (JavaThread *)THREAD; 1459 1460 DeferredInitialize () ; 1461 1462 // Throw IMSX or IEX. 1463 CHECK_OWNER(); 1464 1465 EventJavaMonitorWait event; 1466 1467 // check for a pending interrupt 1468 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1469 // post monitor waited event. Note that this is past-tense, we are done waiting. 1470 if (JvmtiExport::should_post_monitor_waited()) { 1471 // Note: 'false' parameter is passed here because the 1472 // wait was not timed out due to thread interrupt. 1473 JvmtiExport::post_monitor_waited(jt, this, false); 1474 1475 // In this short circuit of the monitor wait protocol, the 1476 // current thread never drops ownership of the monitor and 1477 // never gets added to the wait queue so the current thread 1478 // cannot be made the successor. This means that the 1479 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1480 // consume an unpark() meant for the ParkEvent associated with 1481 // this ObjectMonitor. 1482 } 1483 if (event.should_commit()) { 1484 post_monitor_wait_event(&event, 0, millis, false); 1485 } 1486 TEVENT (Wait - Throw IEX) ; 1487 THROW(vmSymbols::java_lang_InterruptedException()); 1488 return ; 1489 } 1490 1491 TEVENT (Wait) ; 1492 1493 assert (Self->_Stalled == 0, "invariant") ; 1494 Self->_Stalled = intptr_t(this) ; 1495 jt->set_current_waiting_monitor(this); 1496 1497 // create a node to be put into the queue 1498 // Critically, after we reset() the event but prior to park(), we must check 1499 // for a pending interrupt. 1500 ObjectWaiter node(Self); 1501 node.TState = ObjectWaiter::TS_WAIT ; 1502 Self->_ParkEvent->reset() ; 1503 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1504 1505 // Enter the waiting queue, which is a circular doubly linked list in this case 1506 // but it could be a priority queue or any data structure. 1507 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1508 // by the the owner of the monitor *except* in the case where park() 1509 // returns because of a timeout of interrupt. Contention is exceptionally rare 1510 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1511 1512 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1513 AddWaiter (&node) ; 1514 Thread::SpinRelease (&_WaitSetLock) ; 1515 1516 if ((SyncFlags & 4) == 0) { 1517 _Responsible = NULL ; 1518 } 1519 intptr_t save = _recursions; // record the old recursion count 1520 _waiters++; // increment the number of waiters 1521 _recursions = 0; // set the recursion level to be 1 1522 exit (true, Self) ; // exit the monitor 1523 guarantee (_owner != Self, "invariant") ; 1524 1525 // The thread is on the WaitSet list - now park() it. 1526 // On MP systems it's conceivable that a brief spin before we park 1527 // could be profitable. 1528 // 1529 // TODO-FIXME: change the following logic to a loop of the form 1530 // while (!timeout && !interrupted && _notified == 0) park() 1531 1532 int ret = OS_OK ; 1533 int WasNotified = 0 ; 1534 { // State transition wrappers 1535 OSThread* osthread = Self->osthread(); 1536 OSThreadWaitState osts(osthread, true); 1537 { 1538 ThreadBlockInVM tbivm(jt); 1539 // Thread is in thread_blocked state and oop access is unsafe. 1540 jt->set_suspend_equivalent(); 1541 1542 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1543 // Intentionally empty 1544 } else 1545 if (node._notified == 0) { 1546 if (millis <= 0) { 1547 Self->_ParkEvent->park () ; 1548 } else { 1549 ret = Self->_ParkEvent->park (millis) ; 1550 } 1551 } 1552 1553 // were we externally suspended while we were waiting? 1554 if (ExitSuspendEquivalent (jt)) { 1555 // TODO-FIXME: add -- if succ == Self then succ = null. 1556 jt->java_suspend_self(); 1557 } 1558 1559 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1560 1561 1562 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1563 // from the WaitSet to the EntryList. 1564 // See if we need to remove Node from the WaitSet. 1565 // We use double-checked locking to avoid grabbing _WaitSetLock 1566 // if the thread is not on the wait queue. 1567 // 1568 // Note that we don't need a fence before the fetch of TState. 1569 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1570 // written by the is thread. (perhaps the fetch might even be satisfied 1571 // by a look-aside into the processor's own store buffer, although given 1572 // the length of the code path between the prior ST and this load that's 1573 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1574 // then we'll acquire the lock and then re-fetch a fresh TState value. 1575 // That is, we fail toward safety. 1576 1577 if (node.TState == ObjectWaiter::TS_WAIT) { 1578 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1579 if (node.TState == ObjectWaiter::TS_WAIT) { 1580 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1581 assert(node._notified == 0, "invariant"); 1582 node.TState = ObjectWaiter::TS_RUN ; 1583 } 1584 Thread::SpinRelease (&_WaitSetLock) ; 1585 } 1586 1587 // The thread is now either on off-list (TS_RUN), 1588 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1589 // The Node's TState variable is stable from the perspective of this thread. 1590 // No other threads will asynchronously modify TState. 1591 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1592 OrderAccess::loadload() ; 1593 if (_succ == Self) _succ = NULL ; 1594 WasNotified = node._notified ; 1595 1596 // Reentry phase -- reacquire the monitor. 1597 // re-enter contended monitor after object.wait(). 1598 // retain OBJECT_WAIT state until re-enter successfully completes 1599 // Thread state is thread_in_vm and oop access is again safe, 1600 // although the raw address of the object may have changed. 1601 // (Don't cache naked oops over safepoints, of course). 1602 1603 // post monitor waited event. Note that this is past-tense, we are done waiting. 1604 if (JvmtiExport::should_post_monitor_waited()) { 1605 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1606 } 1607 1608 // Without the fix for 8028280, it is possible for the above call: 1609 // 1610 // Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1611 // 1612 // to consume the unpark() that was done when the successor was set. 1613 // The solution for this very rare possibility is to redo the unpark() 1614 // outside of the JvmtiExport::should_post_monitor_waited() check. 1615 // 1616 if (node._notified != 0 && _succ == Self) { 1617 // In this part of the monitor wait-notify-reenter protocol it 1618 // is possible (and normal) for another thread to do a fastpath 1619 // monitor enter-exit while this thread is still trying to get 1620 // to the reenter portion of the protocol. 1621 // 1622 // The ObjectMonitor was notified and the current thread is 1623 // the successor which also means that an unpark() has already 1624 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1625 // consume the unpark() that was done when the successor was 1626 // set because the same ParkEvent is shared between Java 1627 // monitors and JVM/TI RawMonitors (for now). 1628 // 1629 // We redo the unpark() to ensure forward progress, i.e., we 1630 // don't want all pending threads hanging (parked) with none 1631 // entering the unlocked monitor. 1632 node._event->unpark(); 1633 } 1634 1635 if (event.should_commit()) { 1636 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1637 } 1638 1639 OrderAccess::fence() ; 1640 1641 assert (Self->_Stalled != 0, "invariant") ; 1642 Self->_Stalled = 0 ; 1643 1644 assert (_owner != Self, "invariant") ; 1645 ObjectWaiter::TStates v = node.TState ; 1646 if (v == ObjectWaiter::TS_RUN) { 1647 enter (Self) ; 1648 } else { 1649 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1650 ReenterI (Self, &node) ; 1651 node.wait_reenter_end(this); 1652 } 1653 1654 // Self has reacquired the lock. 1655 // Lifecycle - the node representing Self must not appear on any queues. 1656 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1657 // want residual elements associated with this thread left on any lists. 1658 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1659 assert (_owner == Self, "invariant") ; 1660 assert (_succ != Self , "invariant") ; 1661 } // OSThreadWaitState() 1662 1663 jt->set_current_waiting_monitor(NULL); 1664 1665 guarantee (_recursions == 0, "invariant") ; 1666 _recursions = save; // restore the old recursion count 1667 _waiters--; // decrement the number of waiters 1668 1669 // Verify a few postconditions 1670 assert (_owner == Self , "invariant") ; 1671 assert (_succ != Self , "invariant") ; 1672 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1673 1674 if (SyncFlags & 32) { 1675 OrderAccess::fence() ; 1676 } 1677 1678 // check if the notification happened 1679 if (!WasNotified) { 1680 // no, it could be timeout or Thread.interrupt() or both 1681 // check for interrupt event, otherwise it is timeout 1682 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1683 TEVENT (Wait - throw IEX from epilog) ; 1684 THROW(vmSymbols::java_lang_InterruptedException()); 1685 } 1686 } 1687 1688 // NOTE: Spurious wake up will be consider as timeout. 1689 // Monitor notify has precedence over thread interrupt. 1690 } 1691 1692 1693 // Consider: 1694 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1695 // then instead of transferring a thread from the WaitSet to the EntryList 1696 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1697 1698 void ObjectMonitor::notify(TRAPS) { 1699 CHECK_OWNER(); 1700 if (_WaitSet == NULL) { 1701 TEVENT (Empty-Notify) ; 1702 return ; 1703 } 1704 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1705 1706 int Policy = Knob_MoveNotifyee ; 1707 1708 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1709 ObjectWaiter * iterator = DequeueWaiter() ; 1710 if (iterator != NULL) { 1711 TEVENT (Notify1 - Transfer) ; 1712 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1713 guarantee (iterator->_notified == 0, "invariant") ; 1714 if (Policy != 4) { 1715 iterator->TState = ObjectWaiter::TS_ENTER ; 1716 } 1717 iterator->_notified = 1 ; 1718 Thread * Self = THREAD; 1719 iterator->_notifier_tid = Self->osthread()->thread_id(); 1720 1721 ObjectWaiter * List = _EntryList ; 1722 if (List != NULL) { 1723 assert (List->_prev == NULL, "invariant") ; 1724 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1725 assert (List != iterator, "invariant") ; 1726 } 1727 1728 if (Policy == 0) { // prepend to EntryList 1729 if (List == NULL) { 1730 iterator->_next = iterator->_prev = NULL ; 1731 _EntryList = iterator ; 1732 } else { 1733 List->_prev = iterator ; 1734 iterator->_next = List ; 1735 iterator->_prev = NULL ; 1736 _EntryList = iterator ; 1737 } 1738 } else 1739 if (Policy == 1) { // append to EntryList 1740 if (List == NULL) { 1741 iterator->_next = iterator->_prev = NULL ; 1742 _EntryList = iterator ; 1743 } else { 1744 // CONSIDER: finding the tail currently requires a linear-time walk of 1745 // the EntryList. We can make tail access constant-time by converting to 1746 // a CDLL instead of using our current DLL. 1747 ObjectWaiter * Tail ; 1748 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1749 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1750 Tail->_next = iterator ; 1751 iterator->_prev = Tail ; 1752 iterator->_next = NULL ; 1753 } 1754 } else 1755 if (Policy == 2) { // prepend to cxq 1756 // prepend to cxq 1757 if (List == NULL) { 1758 iterator->_next = iterator->_prev = NULL ; 1759 _EntryList = iterator ; 1760 } else { 1761 iterator->TState = ObjectWaiter::TS_CXQ ; 1762 for (;;) { 1763 ObjectWaiter * Front = _cxq ; 1764 iterator->_next = Front ; 1765 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1766 break ; 1767 } 1768 } 1769 } 1770 } else 1771 if (Policy == 3) { // append to cxq 1772 iterator->TState = ObjectWaiter::TS_CXQ ; 1773 for (;;) { 1774 ObjectWaiter * Tail ; 1775 Tail = _cxq ; 1776 if (Tail == NULL) { 1777 iterator->_next = NULL ; 1778 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1779 break ; 1780 } 1781 } else { 1782 while (Tail->_next != NULL) Tail = Tail->_next ; 1783 Tail->_next = iterator ; 1784 iterator->_prev = Tail ; 1785 iterator->_next = NULL ; 1786 break ; 1787 } 1788 } 1789 } else { 1790 ParkEvent * ev = iterator->_event ; 1791 iterator->TState = ObjectWaiter::TS_RUN ; 1792 OrderAccess::fence() ; 1793 ev->unpark() ; 1794 } 1795 1796 if (Policy < 4) { 1797 iterator->wait_reenter_begin(this); 1798 } 1799 1800 // _WaitSetLock protects the wait queue, not the EntryList. We could 1801 // move the add-to-EntryList operation, above, outside the critical section 1802 // protected by _WaitSetLock. In practice that's not useful. With the 1803 // exception of wait() timeouts and interrupts the monitor owner 1804 // is the only thread that grabs _WaitSetLock. There's almost no contention 1805 // on _WaitSetLock so it's not profitable to reduce the length of the 1806 // critical section. 1807 } 1808 1809 Thread::SpinRelease (&_WaitSetLock) ; 1810 1811 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1812 ObjectMonitor::_sync_Notifications->inc() ; 1813 } 1814 } 1815 1816 1817 void ObjectMonitor::notifyAll(TRAPS) { 1818 CHECK_OWNER(); 1819 ObjectWaiter* iterator; 1820 if (_WaitSet == NULL) { 1821 TEVENT (Empty-NotifyAll) ; 1822 return ; 1823 } 1824 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1825 1826 int Policy = Knob_MoveNotifyee ; 1827 int Tally = 0 ; 1828 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1829 1830 for (;;) { 1831 iterator = DequeueWaiter () ; 1832 if (iterator == NULL) break ; 1833 TEVENT (NotifyAll - Transfer1) ; 1834 ++Tally ; 1835 1836 // Disposition - what might we do with iterator ? 1837 // a. add it directly to the EntryList - either tail or head. 1838 // b. push it onto the front of the _cxq. 1839 // For now we use (a). 1840 1841 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1842 guarantee (iterator->_notified == 0, "invariant") ; 1843 iterator->_notified = 1 ; 1844 Thread * Self = THREAD; 1845 iterator->_notifier_tid = Self->osthread()->thread_id(); 1846 if (Policy != 4) { 1847 iterator->TState = ObjectWaiter::TS_ENTER ; 1848 } 1849 1850 ObjectWaiter * List = _EntryList ; 1851 if (List != NULL) { 1852 assert (List->_prev == NULL, "invariant") ; 1853 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1854 assert (List != iterator, "invariant") ; 1855 } 1856 1857 if (Policy == 0) { // prepend to EntryList 1858 if (List == NULL) { 1859 iterator->_next = iterator->_prev = NULL ; 1860 _EntryList = iterator ; 1861 } else { 1862 List->_prev = iterator ; 1863 iterator->_next = List ; 1864 iterator->_prev = NULL ; 1865 _EntryList = iterator ; 1866 } 1867 } else 1868 if (Policy == 1) { // append to EntryList 1869 if (List == NULL) { 1870 iterator->_next = iterator->_prev = NULL ; 1871 _EntryList = iterator ; 1872 } else { 1873 // CONSIDER: finding the tail currently requires a linear-time walk of 1874 // the EntryList. We can make tail access constant-time by converting to 1875 // a CDLL instead of using our current DLL. 1876 ObjectWaiter * Tail ; 1877 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1878 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1879 Tail->_next = iterator ; 1880 iterator->_prev = Tail ; 1881 iterator->_next = NULL ; 1882 } 1883 } else 1884 if (Policy == 2) { // prepend to cxq 1885 // prepend to cxq 1886 iterator->TState = ObjectWaiter::TS_CXQ ; 1887 for (;;) { 1888 ObjectWaiter * Front = _cxq ; 1889 iterator->_next = Front ; 1890 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1891 break ; 1892 } 1893 } 1894 } else 1895 if (Policy == 3) { // append to cxq 1896 iterator->TState = ObjectWaiter::TS_CXQ ; 1897 for (;;) { 1898 ObjectWaiter * Tail ; 1899 Tail = _cxq ; 1900 if (Tail == NULL) { 1901 iterator->_next = NULL ; 1902 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1903 break ; 1904 } 1905 } else { 1906 while (Tail->_next != NULL) Tail = Tail->_next ; 1907 Tail->_next = iterator ; 1908 iterator->_prev = Tail ; 1909 iterator->_next = NULL ; 1910 break ; 1911 } 1912 } 1913 } else { 1914 ParkEvent * ev = iterator->_event ; 1915 iterator->TState = ObjectWaiter::TS_RUN ; 1916 OrderAccess::fence() ; 1917 ev->unpark() ; 1918 } 1919 1920 if (Policy < 4) { 1921 iterator->wait_reenter_begin(this); 1922 } 1923 1924 // _WaitSetLock protects the wait queue, not the EntryList. We could 1925 // move the add-to-EntryList operation, above, outside the critical section 1926 // protected by _WaitSetLock. In practice that's not useful. With the 1927 // exception of wait() timeouts and interrupts the monitor owner 1928 // is the only thread that grabs _WaitSetLock. There's almost no contention 1929 // on _WaitSetLock so it's not profitable to reduce the length of the 1930 // critical section. 1931 } 1932 1933 Thread::SpinRelease (&_WaitSetLock) ; 1934 1935 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1936 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1937 } 1938 } 1939 1940 // ----------------------------------------------------------------------------- 1941 // Adaptive Spinning Support 1942 // 1943 // Adaptive spin-then-block - rational spinning 1944 // 1945 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1946 // algorithm. On high order SMP systems it would be better to start with 1947 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1948 // a contending thread could enqueue itself on the cxq and then spin locally 1949 // on a thread-specific variable such as its ParkEvent._Event flag. 1950 // That's left as an exercise for the reader. Note that global spinning is 1951 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1952 // low latency and massive bandwidth. 1953 // 1954 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1955 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1956 // (duration) or we can fix the count at approximately the duration of 1957 // a context switch and vary the frequency. Of course we could also 1958 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1959 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1960 // 1961 // This implementation varies the duration "D", where D varies with 1962 // the success rate of recent spin attempts. (D is capped at approximately 1963 // length of a round-trip context switch). The success rate for recent 1964 // spin attempts is a good predictor of the success rate of future spin 1965 // attempts. The mechanism adapts automatically to varying critical 1966 // section length (lock modality), system load and degree of parallelism. 1967 // D is maintained per-monitor in _SpinDuration and is initialized 1968 // optimistically. Spin frequency is fixed at 100%. 1969 // 1970 // Note that _SpinDuration is volatile, but we update it without locks 1971 // or atomics. The code is designed so that _SpinDuration stays within 1972 // a reasonable range even in the presence of races. The arithmetic 1973 // operations on _SpinDuration are closed over the domain of legal values, 1974 // so at worst a race will install and older but still legal value. 1975 // At the very worst this introduces some apparent non-determinism. 1976 // We might spin when we shouldn't or vice-versa, but since the spin 1977 // count are relatively short, even in the worst case, the effect is harmless. 1978 // 1979 // Care must be taken that a low "D" value does not become an 1980 // an absorbing state. Transient spinning failures -- when spinning 1981 // is overall profitable -- should not cause the system to converge 1982 // on low "D" values. We want spinning to be stable and predictable 1983 // and fairly responsive to change and at the same time we don't want 1984 // it to oscillate, become metastable, be "too" non-deterministic, 1985 // or converge on or enter undesirable stable absorbing states. 1986 // 1987 // We implement a feedback-based control system -- using past behavior 1988 // to predict future behavior. We face two issues: (a) if the 1989 // input signal is random then the spin predictor won't provide optimal 1990 // results, and (b) if the signal frequency is too high then the control 1991 // system, which has some natural response lag, will "chase" the signal. 1992 // (b) can arise from multimodal lock hold times. Transient preemption 1993 // can also result in apparent bimodal lock hold times. 1994 // Although sub-optimal, neither condition is particularly harmful, as 1995 // in the worst-case we'll spin when we shouldn't or vice-versa. 1996 // The maximum spin duration is rather short so the failure modes aren't bad. 1997 // To be conservative, I've tuned the gain in system to bias toward 1998 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1999 // "rings" or oscillates between spinning and not spinning. This happens 2000 // when spinning is just on the cusp of profitability, however, so the 2001 // situation is not dire. The state is benign -- there's no need to add 2002 // hysteresis control to damp the transition rate between spinning and 2003 // not spinning. 2004 // 2005 2006 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 2007 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 2008 2009 // Spinning: Fixed frequency (100%), vary duration 2010 2011 2012 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 2013 2014 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2015 int ctr = Knob_FixedSpin ; 2016 if (ctr != 0) { 2017 while (--ctr >= 0) { 2018 if (TryLock (Self) > 0) return 1 ; 2019 SpinPause () ; 2020 } 2021 return 0 ; 2022 } 2023 2024 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 2025 if (TryLock(Self) > 0) { 2026 // Increase _SpinDuration ... 2027 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2028 // Raising _SpurDuration to the poverty line is key. 2029 int x = _SpinDuration ; 2030 if (x < Knob_SpinLimit) { 2031 if (x < Knob_Poverty) x = Knob_Poverty ; 2032 _SpinDuration = x + Knob_BonusB ; 2033 } 2034 return 1 ; 2035 } 2036 SpinPause () ; 2037 } 2038 2039 // Admission control - verify preconditions for spinning 2040 // 2041 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2042 // becoming an absorbing state. Put another way, we spin briefly to 2043 // sample, just in case the system load, parallelism, contention, or lock 2044 // modality changed. 2045 // 2046 // Consider the following alternative: 2047 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2048 // spin attempt. "Periodically" might mean after a tally of 2049 // the # of failed spin attempts (or iterations) reaches some threshold. 2050 // This takes us into the realm of 1-out-of-N spinning, where we 2051 // hold the duration constant but vary the frequency. 2052 2053 ctr = _SpinDuration ; 2054 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2055 if (ctr <= 0) return 0 ; 2056 2057 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2058 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2059 TEVENT (Spin abort - notrunnable [TOP]); 2060 return 0 ; 2061 } 2062 2063 int MaxSpin = Knob_MaxSpinners ; 2064 if (MaxSpin >= 0) { 2065 if (_Spinner > MaxSpin) { 2066 TEVENT (Spin abort -- too many spinners) ; 2067 return 0 ; 2068 } 2069 // Slighty racy, but benign ... 2070 Adjust (&_Spinner, 1) ; 2071 } 2072 2073 // We're good to spin ... spin ingress. 2074 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2075 // when preparing to LD...CAS _owner, etc and the CAS is likely 2076 // to succeed. 2077 int hits = 0 ; 2078 int msk = 0 ; 2079 int caspty = Knob_CASPenalty ; 2080 int oxpty = Knob_OXPenalty ; 2081 int sss = Knob_SpinSetSucc ; 2082 if (sss && _succ == NULL ) _succ = Self ; 2083 Thread * prv = NULL ; 2084 2085 // There are three ways to exit the following loop: 2086 // 1. A successful spin where this thread has acquired the lock. 2087 // 2. Spin failure with prejudice 2088 // 3. Spin failure without prejudice 2089 2090 while (--ctr >= 0) { 2091 2092 // Periodic polling -- Check for pending GC 2093 // Threads may spin while they're unsafe. 2094 // We don't want spinning threads to delay the JVM from reaching 2095 // a stop-the-world safepoint or to steal cycles from GC. 2096 // If we detect a pending safepoint we abort in order that 2097 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2098 // this thread, if safe, doesn't steal cycles from GC. 2099 // This is in keeping with the "no loitering in runtime" rule. 2100 // We periodically check to see if there's a safepoint pending. 2101 if ((ctr & 0xFF) == 0) { 2102 if (SafepointSynchronize::do_call_back()) { 2103 TEVENT (Spin: safepoint) ; 2104 goto Abort ; // abrupt spin egress 2105 } 2106 if (Knob_UsePause & 1) SpinPause () ; 2107 2108 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2109 if (hits > 50 && scb != NULL) { 2110 int abend = (*scb)(SpinCallbackArgument, 0) ; 2111 } 2112 } 2113 2114 if (Knob_UsePause & 2) SpinPause() ; 2115 2116 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2117 // This is useful on classic SMP systems, but is of less utility on 2118 // N1-style CMT platforms. 2119 // 2120 // Trade-off: lock acquisition latency vs coherency bandwidth. 2121 // Lock hold times are typically short. A histogram 2122 // of successful spin attempts shows that we usually acquire 2123 // the lock early in the spin. That suggests we want to 2124 // sample _owner frequently in the early phase of the spin, 2125 // but then back-off and sample less frequently as the spin 2126 // progresses. The back-off makes a good citizen on SMP big 2127 // SMP systems. Oversampling _owner can consume excessive 2128 // coherency bandwidth. Relatedly, if we _oversample _owner we 2129 // can inadvertently interfere with the the ST m->owner=null. 2130 // executed by the lock owner. 2131 if (ctr & msk) continue ; 2132 ++hits ; 2133 if ((hits & 0xF) == 0) { 2134 // The 0xF, above, corresponds to the exponent. 2135 // Consider: (msk+1)|msk 2136 msk = ((msk << 2)|3) & BackOffMask ; 2137 } 2138 2139 // Probe _owner with TATAS 2140 // If this thread observes the monitor transition or flicker 2141 // from locked to unlocked to locked, then the odds that this 2142 // thread will acquire the lock in this spin attempt go down 2143 // considerably. The same argument applies if the CAS fails 2144 // or if we observe _owner change from one non-null value to 2145 // another non-null value. In such cases we might abort 2146 // the spin without prejudice or apply a "penalty" to the 2147 // spin count-down variable "ctr", reducing it by 100, say. 2148 2149 Thread * ox = (Thread *) _owner ; 2150 if (ox == NULL) { 2151 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2152 if (ox == NULL) { 2153 // The CAS succeeded -- this thread acquired ownership 2154 // Take care of some bookkeeping to exit spin state. 2155 if (sss && _succ == Self) { 2156 _succ = NULL ; 2157 } 2158 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2159 2160 // Increase _SpinDuration : 2161 // The spin was successful (profitable) so we tend toward 2162 // longer spin attempts in the future. 2163 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2164 // If we acquired the lock early in the spin cycle it 2165 // makes sense to increase _SpinDuration proportionally. 2166 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2167 int x = _SpinDuration ; 2168 if (x < Knob_SpinLimit) { 2169 if (x < Knob_Poverty) x = Knob_Poverty ; 2170 _SpinDuration = x + Knob_Bonus ; 2171 } 2172 return 1 ; 2173 } 2174 2175 // The CAS failed ... we can take any of the following actions: 2176 // * penalize: ctr -= Knob_CASPenalty 2177 // * exit spin with prejudice -- goto Abort; 2178 // * exit spin without prejudice. 2179 // * Since CAS is high-latency, retry again immediately. 2180 prv = ox ; 2181 TEVENT (Spin: cas failed) ; 2182 if (caspty == -2) break ; 2183 if (caspty == -1) goto Abort ; 2184 ctr -= caspty ; 2185 continue ; 2186 } 2187 2188 // Did lock ownership change hands ? 2189 if (ox != prv && prv != NULL ) { 2190 TEVENT (spin: Owner changed) 2191 if (oxpty == -2) break ; 2192 if (oxpty == -1) goto Abort ; 2193 ctr -= oxpty ; 2194 } 2195 prv = ox ; 2196 2197 // Abort the spin if the owner is not executing. 2198 // The owner must be executing in order to drop the lock. 2199 // Spinning while the owner is OFFPROC is idiocy. 2200 // Consider: ctr -= RunnablePenalty ; 2201 if (Knob_OState && NotRunnable (Self, ox)) { 2202 TEVENT (Spin abort - notrunnable); 2203 goto Abort ; 2204 } 2205 if (sss && _succ == NULL ) _succ = Self ; 2206 } 2207 2208 // Spin failed with prejudice -- reduce _SpinDuration. 2209 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2210 // AIMD is globally stable. 2211 TEVENT (Spin failure) ; 2212 { 2213 int x = _SpinDuration ; 2214 if (x > 0) { 2215 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2216 // This is globally sample and tends to damp the response. 2217 x -= Knob_Penalty ; 2218 if (x < 0) x = 0 ; 2219 _SpinDuration = x ; 2220 } 2221 } 2222 2223 Abort: 2224 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2225 if (sss && _succ == Self) { 2226 _succ = NULL ; 2227 // Invariant: after setting succ=null a contending thread 2228 // must recheck-retry _owner before parking. This usually happens 2229 // in the normal usage of TrySpin(), but it's safest 2230 // to make TrySpin() as foolproof as possible. 2231 OrderAccess::fence() ; 2232 if (TryLock(Self) > 0) return 1 ; 2233 } 2234 return 0 ; 2235 } 2236 2237 // NotRunnable() -- informed spinning 2238 // 2239 // Don't bother spinning if the owner is not eligible to drop the lock. 2240 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2241 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2242 // The thread must be runnable in order to drop the lock in timely fashion. 2243 // If the _owner is not runnable then spinning will not likely be 2244 // successful (profitable). 2245 // 2246 // Beware -- the thread referenced by _owner could have died 2247 // so a simply fetch from _owner->_thread_state might trap. 2248 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2249 // Because of the lifecycle issues the schedctl and _thread_state values 2250 // observed by NotRunnable() might be garbage. NotRunnable must 2251 // tolerate this and consider the observed _thread_state value 2252 // as advisory. 2253 // 2254 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2255 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2256 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2257 // with the LSB of _owner. Another option would be to probablistically probe 2258 // the putative _owner->TypeTag value. 2259 // 2260 // Checking _thread_state isn't perfect. Even if the thread is 2261 // in_java it might be blocked on a page-fault or have been preempted 2262 // and sitting on a ready/dispatch queue. _thread state in conjunction 2263 // with schedctl.sc_state gives us a good picture of what the 2264 // thread is doing, however. 2265 // 2266 // TODO: check schedctl.sc_state. 2267 // We'll need to use SafeFetch32() to read from the schedctl block. 2268 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2269 // 2270 // The return value from NotRunnable() is *advisory* -- the 2271 // result is based on sampling and is not necessarily coherent. 2272 // The caller must tolerate false-negative and false-positive errors. 2273 // Spinning, in general, is probabilistic anyway. 2274 2275 2276 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2277 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2278 if (!OwnerIsThread) return 0 ; 2279 2280 if (ox == NULL) return 0 ; 2281 2282 // Avoid transitive spinning ... 2283 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2284 // Immediately after T1 acquires L it's possible that T2, also 2285 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2286 // This occurs transiently after T1 acquired L but before 2287 // T1 managed to clear T1.Stalled. T2 does not need to abort 2288 // its spin in this circumstance. 2289 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2290 2291 if (BlockedOn == 1) return 1 ; 2292 if (BlockedOn != 0) { 2293 return BlockedOn != intptr_t(this) && _owner == ox ; 2294 } 2295 2296 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2297 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2298 // consider also: jst != _thread_in_Java -- but that's overspecific. 2299 return jst == _thread_blocked || jst == _thread_in_native ; 2300 } 2301 2302 2303 // ----------------------------------------------------------------------------- 2304 // WaitSet management ... 2305 2306 ObjectWaiter::ObjectWaiter(Thread* thread) { 2307 _next = NULL; 2308 _prev = NULL; 2309 _notified = 0; 2310 TState = TS_RUN ; 2311 _thread = thread; 2312 _event = thread->_ParkEvent ; 2313 _active = false; 2314 assert (_event != NULL, "invariant") ; 2315 } 2316 2317 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2318 JavaThread *jt = (JavaThread *)this->_thread; 2319 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2320 } 2321 2322 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2323 JavaThread *jt = (JavaThread *)this->_thread; 2324 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2325 } 2326 2327 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2328 assert(node != NULL, "should not dequeue NULL node"); 2329 assert(node->_prev == NULL, "node already in list"); 2330 assert(node->_next == NULL, "node already in list"); 2331 // put node at end of queue (circular doubly linked list) 2332 if (_WaitSet == NULL) { 2333 _WaitSet = node; 2334 node->_prev = node; 2335 node->_next = node; 2336 } else { 2337 ObjectWaiter* head = _WaitSet ; 2338 ObjectWaiter* tail = head->_prev; 2339 assert(tail->_next == head, "invariant check"); 2340 tail->_next = node; 2341 head->_prev = node; 2342 node->_next = head; 2343 node->_prev = tail; 2344 } 2345 } 2346 2347 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2348 // dequeue the very first waiter 2349 ObjectWaiter* waiter = _WaitSet; 2350 if (waiter) { 2351 DequeueSpecificWaiter(waiter); 2352 } 2353 return waiter; 2354 } 2355 2356 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2357 assert(node != NULL, "should not dequeue NULL node"); 2358 assert(node->_prev != NULL, "node already removed from list"); 2359 assert(node->_next != NULL, "node already removed from list"); 2360 // when the waiter has woken up because of interrupt, 2361 // timeout or other spurious wake-up, dequeue the 2362 // waiter from waiting list 2363 ObjectWaiter* next = node->_next; 2364 if (next == node) { 2365 assert(node->_prev == node, "invariant check"); 2366 _WaitSet = NULL; 2367 } else { 2368 ObjectWaiter* prev = node->_prev; 2369 assert(prev->_next == node, "invariant check"); 2370 assert(next->_prev == node, "invariant check"); 2371 next->_prev = prev; 2372 prev->_next = next; 2373 if (_WaitSet == node) { 2374 _WaitSet = next; 2375 } 2376 } 2377 node->_next = NULL; 2378 node->_prev = NULL; 2379 } 2380 2381 // ----------------------------------------------------------------------------- 2382 // PerfData support 2383 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2384 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2385 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2386 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2387 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2388 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2389 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2390 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2391 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2392 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2393 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2394 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2395 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2396 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2397 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2398 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2399 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2400 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2401 2402 // One-shot global initialization for the sync subsystem. 2403 // We could also defer initialization and initialize on-demand 2404 // the first time we call inflate(). Initialization would 2405 // be protected - like so many things - by the MonitorCache_lock. 2406 2407 void ObjectMonitor::Initialize () { 2408 static int InitializationCompleted = 0 ; 2409 assert (InitializationCompleted == 0, "invariant") ; 2410 InitializationCompleted = 1 ; 2411 if (UsePerfData) { 2412 EXCEPTION_MARK ; 2413 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2414 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2415 NEWPERFCOUNTER(_sync_Inflations) ; 2416 NEWPERFCOUNTER(_sync_Deflations) ; 2417 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2418 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2419 NEWPERFCOUNTER(_sync_Parks) ; 2420 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2421 NEWPERFCOUNTER(_sync_Notifications) ; 2422 NEWPERFCOUNTER(_sync_SlowEnter) ; 2423 NEWPERFCOUNTER(_sync_SlowExit) ; 2424 NEWPERFCOUNTER(_sync_SlowNotify) ; 2425 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2426 NEWPERFCOUNTER(_sync_FailedSpins) ; 2427 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2428 NEWPERFCOUNTER(_sync_PrivateA) ; 2429 NEWPERFCOUNTER(_sync_PrivateB) ; 2430 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2431 NEWPERFCOUNTER(_sync_MonScavenged) ; 2432 NEWPERFVARIABLE(_sync_MonExtant) ; 2433 #undef NEWPERFCOUNTER 2434 } 2435 } 2436 2437 2438 // Compile-time asserts 2439 // When possible, it's better to catch errors deterministically at 2440 // compile-time than at runtime. The down-side to using compile-time 2441 // asserts is that error message -- often something about negative array 2442 // indices -- is opaque. 2443 2444 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2445 2446 void ObjectMonitor::ctAsserts() { 2447 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2448 } 2449 2450 2451 static char * kvGet (char * kvList, const char * Key) { 2452 if (kvList == NULL) return NULL ; 2453 size_t n = strlen (Key) ; 2454 char * Search ; 2455 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2456 if (strncmp (Search, Key, n) == 0) { 2457 if (Search[n] == '=') return Search + n + 1 ; 2458 if (Search[n] == 0) return (char *) "1" ; 2459 } 2460 } 2461 return NULL ; 2462 } 2463 2464 static int kvGetInt (char * kvList, const char * Key, int Default) { 2465 char * v = kvGet (kvList, Key) ; 2466 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2467 if (Knob_ReportSettings && v != NULL) { 2468 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2469 ::fflush (stdout) ; 2470 } 2471 return rslt ; 2472 } 2473 2474 void ObjectMonitor::DeferredInitialize () { 2475 if (InitDone > 0) return ; 2476 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2477 while (InitDone != 1) ; 2478 return ; 2479 } 2480 2481 // One-shot global initialization ... 2482 // The initialization is idempotent, so we don't need locks. 2483 // In the future consider doing this via os::init_2(). 2484 // SyncKnobs consist of <Key>=<Value> pairs in the style 2485 // of environment variables. Start by converting ':' to NUL. 2486 2487 if (SyncKnobs == NULL) SyncKnobs = "" ; 2488 2489 size_t sz = strlen (SyncKnobs) ; 2490 char * knobs = (char *) malloc (sz + 2) ; 2491 if (knobs == NULL) { 2492 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; 2493 guarantee (0, "invariant") ; 2494 } 2495 strcpy (knobs, SyncKnobs) ; 2496 knobs[sz+1] = 0 ; 2497 for (char * p = knobs ; *p ; p++) { 2498 if (*p == ':') *p = 0 ; 2499 } 2500 2501 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2502 SETKNOB(ReportSettings) ; 2503 SETKNOB(Verbose) ; 2504 SETKNOB(FixedSpin) ; 2505 SETKNOB(SpinLimit) ; 2506 SETKNOB(SpinBase) ; 2507 SETKNOB(SpinBackOff); 2508 SETKNOB(CASPenalty) ; 2509 SETKNOB(OXPenalty) ; 2510 SETKNOB(LogSpins) ; 2511 SETKNOB(SpinSetSucc) ; 2512 SETKNOB(SuccEnabled) ; 2513 SETKNOB(SuccRestrict) ; 2514 SETKNOB(Penalty) ; 2515 SETKNOB(Bonus) ; 2516 SETKNOB(BonusB) ; 2517 SETKNOB(Poverty) ; 2518 SETKNOB(SpinAfterFutile) ; 2519 SETKNOB(UsePause) ; 2520 SETKNOB(SpinEarly) ; 2521 SETKNOB(OState) ; 2522 SETKNOB(MaxSpinners) ; 2523 SETKNOB(PreSpin) ; 2524 SETKNOB(ExitPolicy) ; 2525 SETKNOB(QMode); 2526 SETKNOB(ResetEvent) ; 2527 SETKNOB(MoveNotifyee) ; 2528 SETKNOB(FastHSSEC) ; 2529 #undef SETKNOB 2530 2531 if (os::is_MP()) { 2532 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2533 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2534 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2535 } else { 2536 Knob_SpinLimit = 0 ; 2537 Knob_SpinBase = 0 ; 2538 Knob_PreSpin = 0 ; 2539 Knob_FixedSpin = -1 ; 2540 } 2541 2542 if (Knob_LogSpins == 0) { 2543 ObjectMonitor::_sync_FailedSpins = NULL ; 2544 } 2545 2546 free (knobs) ; 2547 OrderAccess::fence() ; 2548 InitDone = 1 ; 2549 } 2550 2551 #ifndef PRODUCT 2552 void ObjectMonitor::verify() { 2553 } 2554 2555 void ObjectMonitor::print() { 2556 } 2557 #endif