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/orderAccess.inline.hpp" 36 #include "runtime/osThread.hpp" 37 #include "runtime/stubRoutines.hpp" 38 #include "runtime/thread.inline.hpp" 39 #include "services/threadService.hpp" 40 #include "trace/tracing.hpp" 41 #include "trace/traceMacros.hpp" 42 #include "utilities/dtrace.hpp" 43 #include "utilities/macros.hpp" 44 #include "utilities/preserveException.hpp" 45 #ifdef TARGET_OS_FAMILY_linux 46 # include "os_linux.inline.hpp" 47 #endif 48 #ifdef TARGET_OS_FAMILY_solaris 49 # include "os_solaris.inline.hpp" 50 #endif 51 #ifdef TARGET_OS_FAMILY_windows 52 # include "os_windows.inline.hpp" 53 #endif 54 #ifdef TARGET_OS_FAMILY_bsd 55 # include "os_bsd.inline.hpp" 56 #endif 57 58 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 59 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 60 #define ATTR __attribute__((noinline)) 61 #else 62 #define ATTR 63 #endif 64 65 66 #ifdef DTRACE_ENABLED 67 68 // Only bother with this argument setup if dtrace is available 69 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 70 71 72 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 73 char* bytes = NULL; \ 74 int len = 0; \ 75 jlong jtid = SharedRuntime::get_java_tid(thread); \ 76 Symbol* klassname = ((oop)obj)->klass()->name(); \ 77 if (klassname != NULL) { \ 78 bytes = (char*)klassname->bytes(); \ 79 len = klassname->utf8_length(); \ 80 } 81 82 #ifndef USDT2 83 84 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 85 jlong, uintptr_t, char*, int); 86 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 87 jlong, uintptr_t, char*, int); 88 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 89 jlong, uintptr_t, char*, int); 90 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 91 jlong, uintptr_t, char*, int); 92 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 93 jlong, uintptr_t, char*, int); 94 95 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 96 { \ 97 if (DTraceMonitorProbes) { \ 98 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 99 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 100 (monitor), bytes, len, (millis)); \ 101 } \ 102 } 103 104 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 105 { \ 106 if (DTraceMonitorProbes) { \ 107 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 108 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 109 (uintptr_t)(monitor), bytes, len); \ 110 } \ 111 } 112 113 #else /* USDT2 */ 114 115 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 116 { \ 117 if (DTraceMonitorProbes) { \ 118 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 119 HOTSPOT_MONITOR_WAIT(jtid, \ 120 (monitor), bytes, len, (millis)); \ 121 } \ 122 } 123 124 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 125 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 126 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 127 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 128 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 129 130 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 131 { \ 132 if (DTraceMonitorProbes) { \ 133 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 134 HOTSPOT_MONITOR_##probe(jtid, \ 135 (uintptr_t)(monitor), bytes, len); \ 136 } \ 137 } 138 139 #endif /* USDT2 */ 140 #else // ndef DTRACE_ENABLED 141 142 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 143 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 144 145 #endif // ndef DTRACE_ENABLED 146 147 // Tunables ... 148 // The knob* variables are effectively final. Once set they should 149 // never be modified hence. Consider using __read_mostly with GCC. 150 151 int ObjectMonitor::Knob_Verbose = 0 ; 152 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 153 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 154 static int Knob_HandOff = 0 ; 155 static int Knob_ReportSettings = 0 ; 156 157 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 158 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 159 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 160 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 161 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 162 static int Knob_SpinEarly = 1 ; 163 static int Knob_SuccEnabled = 1 ; // futile wake throttling 164 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 165 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 166 static int Knob_Bonus = 100 ; // spin success bonus 167 static int Knob_BonusB = 100 ; // spin success bonus 168 static int Knob_Penalty = 200 ; // spin failure penalty 169 static int Knob_Poverty = 1000 ; 170 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 171 static int Knob_FixedSpin = 0 ; 172 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 173 static int Knob_UsePause = 1 ; 174 static int Knob_ExitPolicy = 0 ; 175 static int Knob_PreSpin = 10 ; // 20-100 likely better 176 static int Knob_ResetEvent = 0 ; 177 static int BackOffMask = 0 ; 178 179 static int Knob_FastHSSEC = 0 ; 180 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 181 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 182 static volatile int InitDone = 0 ; 183 184 #define TrySpin TrySpin_VaryDuration 185 186 // ----------------------------------------------------------------------------- 187 // Theory of operations -- Monitors lists, thread residency, etc: 188 // 189 // * A thread acquires ownership of a monitor by successfully 190 // CAS()ing the _owner field from null to non-null. 191 // 192 // * Invariant: A thread appears on at most one monitor list -- 193 // cxq, EntryList or WaitSet -- at any one time. 194 // 195 // * Contending threads "push" themselves onto the cxq with CAS 196 // and then spin/park. 197 // 198 // * After a contending thread eventually acquires the lock it must 199 // dequeue itself from either the EntryList or the cxq. 200 // 201 // * The exiting thread identifies and unparks an "heir presumptive" 202 // tentative successor thread on the EntryList. Critically, the 203 // exiting thread doesn't unlink the successor thread from the EntryList. 204 // After having been unparked, the wakee will recontend for ownership of 205 // the monitor. The successor (wakee) will either acquire the lock or 206 // re-park itself. 207 // 208 // Succession is provided for by a policy of competitive handoff. 209 // The exiting thread does _not_ grant or pass ownership to the 210 // successor thread. (This is also referred to as "handoff" succession"). 211 // Instead the exiting thread releases ownership and possibly wakes 212 // a successor, so the successor can (re)compete for ownership of the lock. 213 // If the EntryList is empty but the cxq is populated the exiting 214 // thread will drain the cxq into the EntryList. It does so by 215 // by detaching the cxq (installing null with CAS) and folding 216 // the threads from the cxq into the EntryList. The EntryList is 217 // doubly linked, while the cxq is singly linked because of the 218 // CAS-based "push" used to enqueue recently arrived threads (RATs). 219 // 220 // * Concurrency invariants: 221 // 222 // -- only the monitor owner may access or mutate the EntryList. 223 // The mutex property of the monitor itself protects the EntryList 224 // from concurrent interference. 225 // -- Only the monitor owner may detach the cxq. 226 // 227 // * The monitor entry list operations avoid locks, but strictly speaking 228 // they're not lock-free. Enter is lock-free, exit is not. 229 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 230 // 231 // * The cxq can have multiple concurrent "pushers" but only one concurrent 232 // detaching thread. This mechanism is immune from the ABA corruption. 233 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 234 // 235 // * Taken together, the cxq and the EntryList constitute or form a 236 // single logical queue of threads stalled trying to acquire the lock. 237 // We use two distinct lists to improve the odds of a constant-time 238 // dequeue operation after acquisition (in the ::enter() epilog) and 239 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 240 // A key desideratum is to minimize queue & monitor metadata manipulation 241 // that occurs while holding the monitor lock -- that is, we want to 242 // minimize monitor lock holds times. Note that even a small amount of 243 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 244 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 245 // locks and monitor metadata. 246 // 247 // Cxq points to the the set of Recently Arrived Threads attempting entry. 248 // Because we push threads onto _cxq with CAS, the RATs must take the form of 249 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 250 // the unlocking thread notices that EntryList is null but _cxq is != null. 251 // 252 // The EntryList is ordered by the prevailing queue discipline and 253 // can be organized in any convenient fashion, such as a doubly-linked list or 254 // a circular doubly-linked list. Critically, we want insert and delete operations 255 // to operate in constant-time. If we need a priority queue then something akin 256 // to Solaris' sleepq would work nicely. Viz., 257 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 258 // Queue discipline is enforced at ::exit() time, when the unlocking thread 259 // drains the cxq into the EntryList, and orders or reorders the threads on the 260 // EntryList accordingly. 261 // 262 // Barring "lock barging", this mechanism provides fair cyclic ordering, 263 // somewhat similar to an elevator-scan. 264 // 265 // * The monitor synchronization subsystem avoids the use of native 266 // synchronization primitives except for the narrow platform-specific 267 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 268 // the semantics of park-unpark. Put another way, this monitor implementation 269 // depends only on atomic operations and park-unpark. The monitor subsystem 270 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 271 // underlying OS manages the READY<->RUN transitions. 272 // 273 // * Waiting threads reside on the WaitSet list -- wait() puts 274 // the caller onto the WaitSet. 275 // 276 // * notify() or notifyAll() simply transfers threads from the WaitSet to 277 // either the EntryList or cxq. Subsequent exit() operations will 278 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 279 // it's likely the notifyee would simply impale itself on the lock held 280 // by the notifier. 281 // 282 // * An interesting alternative is to encode cxq as (List,LockByte) where 283 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 284 // variable, like _recursions, in the scheme. The threads or Events that form 285 // the list would have to be aligned in 256-byte addresses. A thread would 286 // try to acquire the lock or enqueue itself with CAS, but exiting threads 287 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 288 // Note that is is *not* word-tearing, but it does presume that full-word 289 // CAS operations are coherent with intermix with STB operations. That's true 290 // on most common processors. 291 // 292 // * See also http://blogs.sun.com/dave 293 294 295 // ----------------------------------------------------------------------------- 296 // Enter support 297 298 bool ObjectMonitor::try_enter(Thread* THREAD) { 299 if (THREAD != _owner) { 300 if (THREAD->is_lock_owned ((address)_owner)) { 301 assert(_recursions == 0, "internal state error"); 302 _owner = THREAD ; 303 _recursions = 1 ; 304 OwnerIsThread = 1 ; 305 return true; 306 } 307 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 308 return false; 309 } 310 return true; 311 } else { 312 _recursions++; 313 return true; 314 } 315 } 316 317 void ATTR ObjectMonitor::enter(TRAPS) { 318 // The following code is ordered to check the most common cases first 319 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 320 Thread * const Self = THREAD ; 321 void * cur ; 322 323 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 324 if (cur == NULL) { 325 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 326 assert (_recursions == 0 , "invariant") ; 327 assert (_owner == Self, "invariant") ; 328 // CONSIDER: set or assert OwnerIsThread == 1 329 return ; 330 } 331 332 if (cur == Self) { 333 // TODO-FIXME: check for integer overflow! BUGID 6557169. 334 _recursions ++ ; 335 return ; 336 } 337 338 if (Self->is_lock_owned ((address)cur)) { 339 assert (_recursions == 0, "internal state error"); 340 _recursions = 1 ; 341 // Commute owner from a thread-specific on-stack BasicLockObject address to 342 // a full-fledged "Thread *". 343 _owner = Self ; 344 OwnerIsThread = 1 ; 345 return ; 346 } 347 348 // We've encountered genuine contention. 349 assert (Self->_Stalled == 0, "invariant") ; 350 Self->_Stalled = intptr_t(this) ; 351 352 // Try one round of spinning *before* enqueueing Self 353 // and before going through the awkward and expensive state 354 // transitions. The following spin is strictly optional ... 355 // Note that if we acquire the monitor from an initial spin 356 // we forgo posting JVMTI events and firing DTRACE probes. 357 if (Knob_SpinEarly && TrySpin (Self) > 0) { 358 assert (_owner == Self , "invariant") ; 359 assert (_recursions == 0 , "invariant") ; 360 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 361 Self->_Stalled = 0 ; 362 return ; 363 } 364 365 assert (_owner != Self , "invariant") ; 366 assert (_succ != Self , "invariant") ; 367 assert (Self->is_Java_thread() , "invariant") ; 368 JavaThread * jt = (JavaThread *) Self ; 369 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 370 assert (jt->thread_state() != _thread_blocked , "invariant") ; 371 assert (this->object() != NULL , "invariant") ; 372 assert (_count >= 0, "invariant") ; 373 374 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 375 // Ensure the object-monitor relationship remains stable while there's contention. 376 Atomic::inc_ptr(&_count); 377 378 EventJavaMonitorEnter event; 379 380 { // Change java thread status to indicate blocked on monitor enter. 381 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 382 383 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 384 if (JvmtiExport::should_post_monitor_contended_enter()) { 385 JvmtiExport::post_monitor_contended_enter(jt, this); 386 387 // The current thread does not yet own the monitor and does not 388 // yet appear on any queues that would get it made the successor. 389 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 390 // handler cannot accidentally consume an unpark() meant for the 391 // ParkEvent associated with this ObjectMonitor. 392 } 393 394 OSThreadContendState osts(Self->osthread()); 395 ThreadBlockInVM tbivm(jt); 396 397 Self->set_current_pending_monitor(this); 398 399 // TODO-FIXME: change the following for(;;) loop to straight-line code. 400 for (;;) { 401 jt->set_suspend_equivalent(); 402 // cleared by handle_special_suspend_equivalent_condition() 403 // or java_suspend_self() 404 405 EnterI (THREAD) ; 406 407 if (!ExitSuspendEquivalent(jt)) break ; 408 409 // 410 // We have acquired the contended monitor, but while we were 411 // waiting another thread suspended us. We don't want to enter 412 // the monitor while suspended because that would surprise the 413 // thread that suspended us. 414 // 415 _recursions = 0 ; 416 _succ = NULL ; 417 exit (false, Self) ; 418 419 jt->java_suspend_self(); 420 } 421 Self->set_current_pending_monitor(NULL); 422 423 // We cleared the pending monitor info since we've just gotten past 424 // the enter-check-for-suspend dance and we now own the monitor free 425 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 426 // destructor can go to a safepoint at the end of this block. If we 427 // do a thread dump during that safepoint, then this thread will show 428 // as having "-locked" the monitor, but the OS and java.lang.Thread 429 // states will still report that the thread is blocked trying to 430 // acquire it. 431 } 432 433 Atomic::dec_ptr(&_count); 434 assert (_count >= 0, "invariant") ; 435 Self->_Stalled = 0 ; 436 437 // Must either set _recursions = 0 or ASSERT _recursions == 0. 438 assert (_recursions == 0 , "invariant") ; 439 assert (_owner == Self , "invariant") ; 440 assert (_succ != Self , "invariant") ; 441 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 442 443 // The thread -- now the owner -- is back in vm mode. 444 // Report the glorious news via TI,DTrace and jvmstat. 445 // The probe effect is non-trivial. All the reportage occurs 446 // while we hold the monitor, increasing the length of the critical 447 // section. Amdahl's parallel speedup law comes vividly into play. 448 // 449 // Another option might be to aggregate the events (thread local or 450 // per-monitor aggregation) and defer reporting until a more opportune 451 // time -- such as next time some thread encounters contention but has 452 // yet to acquire the lock. While spinning that thread could 453 // spinning we could increment JVMStat counters, etc. 454 455 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 456 if (JvmtiExport::should_post_monitor_contended_entered()) { 457 JvmtiExport::post_monitor_contended_entered(jt, this); 458 459 // The current thread already owns the monitor and is not going to 460 // call park() for the remainder of the monitor enter protocol. So 461 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 462 // event handler consumed an unpark() issued by the thread that 463 // just exited the monitor. 464 } 465 466 if (event.should_commit()) { 467 event.set_klass(((oop)this->object())->klass()); 468 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 469 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 470 event.commit(); 471 } 472 473 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 474 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 475 } 476 } 477 478 479 // Caveat: TryLock() is not necessarily serializing if it returns failure. 480 // Callers must compensate as needed. 481 482 int ObjectMonitor::TryLock (Thread * Self) { 483 for (;;) { 484 void * own = _owner ; 485 if (own != NULL) return 0 ; 486 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 487 // Either guarantee _recursions == 0 or set _recursions = 0. 488 assert (_recursions == 0, "invariant") ; 489 assert (_owner == Self, "invariant") ; 490 // CONSIDER: set or assert that OwnerIsThread == 1 491 return 1 ; 492 } 493 // The lock had been free momentarily, but we lost the race to the lock. 494 // Interference -- the CAS failed. 495 // We can either return -1 or retry. 496 // Retry doesn't make as much sense because the lock was just acquired. 497 if (true) return -1 ; 498 } 499 } 500 501 void ATTR ObjectMonitor::EnterI (TRAPS) { 502 Thread * Self = THREAD ; 503 assert (Self->is_Java_thread(), "invariant") ; 504 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 505 506 // Try the lock - TATAS 507 if (TryLock (Self) > 0) { 508 assert (_succ != Self , "invariant") ; 509 assert (_owner == Self , "invariant") ; 510 assert (_Responsible != Self , "invariant") ; 511 return ; 512 } 513 514 DeferredInitialize () ; 515 516 // We try one round of spinning *before* enqueueing Self. 517 // 518 // If the _owner is ready but OFFPROC we could use a YieldTo() 519 // operation to donate the remainder of this thread's quantum 520 // to the owner. This has subtle but beneficial affinity 521 // effects. 522 523 if (TrySpin (Self) > 0) { 524 assert (_owner == Self , "invariant") ; 525 assert (_succ != Self , "invariant") ; 526 assert (_Responsible != Self , "invariant") ; 527 return ; 528 } 529 530 // The Spin failed -- Enqueue and park the thread ... 531 assert (_succ != Self , "invariant") ; 532 assert (_owner != Self , "invariant") ; 533 assert (_Responsible != Self , "invariant") ; 534 535 // Enqueue "Self" on ObjectMonitor's _cxq. 536 // 537 // Node acts as a proxy for Self. 538 // As an aside, if were to ever rewrite the synchronization code mostly 539 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 540 // Java objects. This would avoid awkward lifecycle and liveness issues, 541 // as well as eliminate a subset of ABA issues. 542 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 543 // 544 545 ObjectWaiter node(Self) ; 546 Self->_ParkEvent->reset() ; 547 node._prev = (ObjectWaiter *) 0xBAD ; 548 node.TState = ObjectWaiter::TS_CXQ ; 549 550 // Push "Self" onto the front of the _cxq. 551 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 552 // Note that spinning tends to reduce the rate at which threads 553 // enqueue and dequeue on EntryList|cxq. 554 ObjectWaiter * nxt ; 555 for (;;) { 556 node._next = nxt = _cxq ; 557 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 558 559 // Interference - the CAS failed because _cxq changed. Just retry. 560 // As an optional optimization we retry the lock. 561 if (TryLock (Self) > 0) { 562 assert (_succ != Self , "invariant") ; 563 assert (_owner == Self , "invariant") ; 564 assert (_Responsible != Self , "invariant") ; 565 return ; 566 } 567 } 568 569 // Check for cxq|EntryList edge transition to non-null. This indicates 570 // the onset of contention. While contention persists exiting threads 571 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 572 // operations revert to the faster 1-0 mode. This enter operation may interleave 573 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 574 // arrange for one of the contending thread to use a timed park() operations 575 // to detect and recover from the race. (Stranding is form of progress failure 576 // where the monitor is unlocked but all the contending threads remain parked). 577 // That is, at least one of the contended threads will periodically poll _owner. 578 // One of the contending threads will become the designated "Responsible" thread. 579 // The Responsible thread uses a timed park instead of a normal indefinite park 580 // operation -- it periodically wakes and checks for and recovers from potential 581 // strandings admitted by 1-0 exit operations. We need at most one Responsible 582 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 583 // be responsible for a monitor. 584 // 585 // Currently, one of the contended threads takes on the added role of "Responsible". 586 // A viable alternative would be to use a dedicated "stranding checker" thread 587 // that periodically iterated over all the threads (or active monitors) and unparked 588 // successors where there was risk of stranding. This would help eliminate the 589 // timer scalability issues we see on some platforms as we'd only have one thread 590 // -- the checker -- parked on a timer. 591 592 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 593 // Try to assume the role of responsible thread for the monitor. 594 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 595 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 596 } 597 598 // The lock have been released while this thread was occupied queueing 599 // itself onto _cxq. To close the race and avoid "stranding" and 600 // progress-liveness failure we must resample-retry _owner before parking. 601 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 602 // In this case the ST-MEMBAR is accomplished with CAS(). 603 // 604 // TODO: Defer all thread state transitions until park-time. 605 // Since state transitions are heavy and inefficient we'd like 606 // to defer the state transitions until absolutely necessary, 607 // and in doing so avoid some transitions ... 608 609 TEVENT (Inflated enter - Contention) ; 610 int nWakeups = 0 ; 611 int RecheckInterval = 1 ; 612 613 for (;;) { 614 615 if (TryLock (Self) > 0) break ; 616 assert (_owner != Self, "invariant") ; 617 618 if ((SyncFlags & 2) && _Responsible == NULL) { 619 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 620 } 621 622 // park self 623 if (_Responsible == Self || (SyncFlags & 1)) { 624 TEVENT (Inflated enter - park TIMED) ; 625 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 626 // Increase the RecheckInterval, but clamp the value. 627 RecheckInterval *= 8 ; 628 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 629 } else { 630 TEVENT (Inflated enter - park UNTIMED) ; 631 Self->_ParkEvent->park() ; 632 } 633 634 if (TryLock(Self) > 0) break ; 635 636 // The lock is still contested. 637 // Keep a tally of the # of futile wakeups. 638 // Note that the counter is not protected by a lock or updated by atomics. 639 // That is by design - we trade "lossy" counters which are exposed to 640 // races during updates for a lower probe effect. 641 TEVENT (Inflated enter - Futile wakeup) ; 642 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 643 ObjectMonitor::_sync_FutileWakeups->inc() ; 644 } 645 ++ nWakeups ; 646 647 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 648 // We can defer clearing _succ until after the spin completes 649 // TrySpin() must tolerate being called with _succ == Self. 650 // Try yet another round of adaptive spinning. 651 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 652 653 // We can find that we were unpark()ed and redesignated _succ while 654 // we were spinning. That's harmless. If we iterate and call park(), 655 // park() will consume the event and return immediately and we'll 656 // just spin again. This pattern can repeat, leaving _succ to simply 657 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 658 // Alternately, we can sample fired() here, and if set, forgo spinning 659 // in the next iteration. 660 661 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 662 Self->_ParkEvent->reset() ; 663 OrderAccess::fence() ; 664 } 665 if (_succ == Self) _succ = NULL ; 666 667 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 668 OrderAccess::fence() ; 669 } 670 671 // Egress : 672 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 673 // Normally we'll find Self on the EntryList . 674 // From the perspective of the lock owner (this thread), the 675 // EntryList is stable and cxq is prepend-only. 676 // The head of cxq is volatile but the interior is stable. 677 // In addition, Self.TState is stable. 678 679 assert (_owner == Self , "invariant") ; 680 assert (object() != NULL , "invariant") ; 681 // I'd like to write: 682 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 683 // but as we're at a safepoint that's not safe. 684 685 UnlinkAfterAcquire (Self, &node) ; 686 if (_succ == Self) _succ = NULL ; 687 688 assert (_succ != Self, "invariant") ; 689 if (_Responsible == Self) { 690 _Responsible = NULL ; 691 OrderAccess::fence(); // Dekker pivot-point 692 693 // We may leave threads on cxq|EntryList without a designated 694 // "Responsible" thread. This is benign. When this thread subsequently 695 // exits the monitor it can "see" such preexisting "old" threads -- 696 // threads that arrived on the cxq|EntryList before the fence, above -- 697 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 698 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 699 // non-null and elect a new "Responsible" timer thread. 700 // 701 // This thread executes: 702 // ST Responsible=null; MEMBAR (in enter epilog - here) 703 // LD cxq|EntryList (in subsequent exit) 704 // 705 // Entering threads in the slow/contended path execute: 706 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 707 // The (ST cxq; MEMBAR) is accomplished with CAS(). 708 // 709 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 710 // exit operation from floating above the ST Responsible=null. 711 } 712 713 // We've acquired ownership with CAS(). 714 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 715 // But since the CAS() this thread may have also stored into _succ, 716 // EntryList, cxq or Responsible. These meta-data updates must be 717 // visible __before this thread subsequently drops the lock. 718 // Consider what could occur if we didn't enforce this constraint -- 719 // STs to monitor meta-data and user-data could reorder with (become 720 // visible after) the ST in exit that drops ownership of the lock. 721 // Some other thread could then acquire the lock, but observe inconsistent 722 // or old monitor meta-data and heap data. That violates the JMM. 723 // To that end, the 1-0 exit() operation must have at least STST|LDST 724 // "release" barrier semantics. Specifically, there must be at least a 725 // STST|LDST barrier in exit() before the ST of null into _owner that drops 726 // the lock. The barrier ensures that changes to monitor meta-data and data 727 // protected by the lock will be visible before we release the lock, and 728 // therefore before some other thread (CPU) has a chance to acquire the lock. 729 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 730 // 731 // Critically, any prior STs to _succ or EntryList must be visible before 732 // the ST of null into _owner in the *subsequent* (following) corresponding 733 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 734 // execute a serializing instruction. 735 736 if (SyncFlags & 8) { 737 OrderAccess::fence() ; 738 } 739 return ; 740 } 741 742 // ReenterI() is a specialized inline form of the latter half of the 743 // contended slow-path from EnterI(). We use ReenterI() only for 744 // monitor reentry in wait(). 745 // 746 // In the future we should reconcile EnterI() and ReenterI(), adding 747 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 748 // loop accordingly. 749 750 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 751 assert (Self != NULL , "invariant") ; 752 assert (SelfNode != NULL , "invariant") ; 753 assert (SelfNode->_thread == Self , "invariant") ; 754 assert (_waiters > 0 , "invariant") ; 755 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 756 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 757 JavaThread * jt = (JavaThread *) Self ; 758 759 int nWakeups = 0 ; 760 for (;;) { 761 ObjectWaiter::TStates v = SelfNode->TState ; 762 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 763 assert (_owner != Self, "invariant") ; 764 765 if (TryLock (Self) > 0) break ; 766 if (TrySpin (Self) > 0) break ; 767 768 TEVENT (Wait Reentry - parking) ; 769 770 // State transition wrappers around park() ... 771 // ReenterI() wisely defers state transitions until 772 // it's clear we must park the thread. 773 { 774 OSThreadContendState osts(Self->osthread()); 775 ThreadBlockInVM tbivm(jt); 776 777 // cleared by handle_special_suspend_equivalent_condition() 778 // or java_suspend_self() 779 jt->set_suspend_equivalent(); 780 if (SyncFlags & 1) { 781 Self->_ParkEvent->park ((jlong)1000) ; 782 } else { 783 Self->_ParkEvent->park () ; 784 } 785 786 // were we externally suspended while we were waiting? 787 for (;;) { 788 if (!ExitSuspendEquivalent (jt)) break ; 789 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 790 jt->java_suspend_self(); 791 jt->set_suspend_equivalent(); 792 } 793 } 794 795 // Try again, but just so we distinguish between futile wakeups and 796 // successful wakeups. The following test isn't algorithmically 797 // necessary, but it helps us maintain sensible statistics. 798 if (TryLock(Self) > 0) break ; 799 800 // The lock is still contested. 801 // Keep a tally of the # of futile wakeups. 802 // Note that the counter is not protected by a lock or updated by atomics. 803 // That is by design - we trade "lossy" counters which are exposed to 804 // races during updates for a lower probe effect. 805 TEVENT (Wait Reentry - futile wakeup) ; 806 ++ nWakeups ; 807 808 // Assuming this is not a spurious wakeup we'll normally 809 // find that _succ == Self. 810 if (_succ == Self) _succ = NULL ; 811 812 // Invariant: after clearing _succ a contending thread 813 // *must* retry _owner before parking. 814 OrderAccess::fence() ; 815 816 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 817 ObjectMonitor::_sync_FutileWakeups->inc() ; 818 } 819 } 820 821 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 822 // Normally we'll find Self on the EntryList. 823 // Unlinking from the EntryList is constant-time and atomic-free. 824 // From the perspective of the lock owner (this thread), the 825 // EntryList is stable and cxq is prepend-only. 826 // The head of cxq is volatile but the interior is stable. 827 // In addition, Self.TState is stable. 828 829 assert (_owner == Self, "invariant") ; 830 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 831 UnlinkAfterAcquire (Self, SelfNode) ; 832 if (_succ == Self) _succ = NULL ; 833 assert (_succ != Self, "invariant") ; 834 SelfNode->TState = ObjectWaiter::TS_RUN ; 835 OrderAccess::fence() ; // see comments at the end of EnterI() 836 } 837 838 // after the thread acquires the lock in ::enter(). Equally, we could defer 839 // unlinking the thread until ::exit()-time. 840 841 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 842 { 843 assert (_owner == Self, "invariant") ; 844 assert (SelfNode->_thread == Self, "invariant") ; 845 846 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 847 // Normal case: remove Self from the DLL EntryList . 848 // This is a constant-time operation. 849 ObjectWaiter * nxt = SelfNode->_next ; 850 ObjectWaiter * prv = SelfNode->_prev ; 851 if (nxt != NULL) nxt->_prev = prv ; 852 if (prv != NULL) prv->_next = nxt ; 853 if (SelfNode == _EntryList ) _EntryList = nxt ; 854 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 855 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 856 TEVENT (Unlink from EntryList) ; 857 } else { 858 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 859 // Inopportune interleaving -- Self is still on the cxq. 860 // This usually means the enqueue of self raced an exiting thread. 861 // Normally we'll find Self near the front of the cxq, so 862 // dequeueing is typically fast. If needbe we can accelerate 863 // this with some MCS/CHL-like bidirectional list hints and advisory 864 // back-links so dequeueing from the interior will normally operate 865 // in constant-time. 866 // Dequeue Self from either the head (with CAS) or from the interior 867 // with a linear-time scan and normal non-atomic memory operations. 868 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 869 // and then unlink Self from EntryList. We have to drain eventually, 870 // so it might as well be now. 871 872 ObjectWaiter * v = _cxq ; 873 assert (v != NULL, "invariant") ; 874 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 875 // The CAS above can fail from interference IFF a "RAT" arrived. 876 // In that case Self must be in the interior and can no longer be 877 // at the head of cxq. 878 if (v == SelfNode) { 879 assert (_cxq != v, "invariant") ; 880 v = _cxq ; // CAS above failed - start scan at head of list 881 } 882 ObjectWaiter * p ; 883 ObjectWaiter * q = NULL ; 884 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 885 q = p ; 886 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 887 } 888 assert (v != SelfNode, "invariant") ; 889 assert (p == SelfNode, "Node not found on cxq") ; 890 assert (p != _cxq, "invariant") ; 891 assert (q != NULL, "invariant") ; 892 assert (q->_next == p, "invariant") ; 893 q->_next = p->_next ; 894 } 895 TEVENT (Unlink from cxq) ; 896 } 897 898 // Diagnostic hygiene ... 899 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 900 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 901 SelfNode->TState = ObjectWaiter::TS_RUN ; 902 } 903 904 // ----------------------------------------------------------------------------- 905 // Exit support 906 // 907 // exit() 908 // ~~~~~~ 909 // Note that the collector can't reclaim the objectMonitor or deflate 910 // the object out from underneath the thread calling ::exit() as the 911 // thread calling ::exit() never transitions to a stable state. 912 // This inhibits GC, which in turn inhibits asynchronous (and 913 // inopportune) reclamation of "this". 914 // 915 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 916 // There's one exception to the claim above, however. EnterI() can call 917 // exit() to drop a lock if the acquirer has been externally suspended. 918 // In that case exit() is called with _thread_state as _thread_blocked, 919 // but the monitor's _count field is > 0, which inhibits reclamation. 920 // 921 // 1-0 exit 922 // ~~~~~~~~ 923 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 924 // the fast-path operators have been optimized so the common ::exit() 925 // operation is 1-0. See i486.ad fast_unlock(), for instance. 926 // The code emitted by fast_unlock() elides the usual MEMBAR. This 927 // greatly improves latency -- MEMBAR and CAS having considerable local 928 // latency on modern processors -- but at the cost of "stranding". Absent the 929 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 930 // ::enter() path, resulting in the entering thread being stranding 931 // and a progress-liveness failure. Stranding is extremely rare. 932 // We use timers (timed park operations) & periodic polling to detect 933 // and recover from stranding. Potentially stranded threads periodically 934 // wake up and poll the lock. See the usage of the _Responsible variable. 935 // 936 // The CAS() in enter provides for safety and exclusion, while the CAS or 937 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 938 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 939 // We detect and recover from stranding with timers. 940 // 941 // If a thread transiently strands it'll park until (a) another 942 // thread acquires the lock and then drops the lock, at which time the 943 // exiting thread will notice and unpark the stranded thread, or, (b) 944 // the timer expires. If the lock is high traffic then the stranding latency 945 // will be low due to (a). If the lock is low traffic then the odds of 946 // stranding are lower, although the worst-case stranding latency 947 // is longer. Critically, we don't want to put excessive load in the 948 // platform's timer subsystem. We want to minimize both the timer injection 949 // rate (timers created/sec) as well as the number of timers active at 950 // any one time. (more precisely, we want to minimize timer-seconds, which is 951 // the integral of the # of active timers at any instant over time). 952 // Both impinge on OS scalability. Given that, at most one thread parked on 953 // a monitor will use a timer. 954 955 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 956 Thread * Self = THREAD ; 957 if (THREAD != _owner) { 958 if (THREAD->is_lock_owned((address) _owner)) { 959 // Transmute _owner from a BasicLock pointer to a Thread address. 960 // We don't need to hold _mutex for this transition. 961 // Non-null to Non-null is safe as long as all readers can 962 // tolerate either flavor. 963 assert (_recursions == 0, "invariant") ; 964 _owner = THREAD ; 965 _recursions = 0 ; 966 OwnerIsThread = 1 ; 967 } else { 968 // NOTE: we need to handle unbalanced monitor enter/exit 969 // in native code by throwing an exception. 970 // TODO: Throw an IllegalMonitorStateException ? 971 TEVENT (Exit - Throw IMSX) ; 972 assert(false, "Non-balanced monitor enter/exit!"); 973 if (false) { 974 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 975 } 976 return; 977 } 978 } 979 980 if (_recursions != 0) { 981 _recursions--; // this is simple recursive enter 982 TEVENT (Inflated exit - recursive) ; 983 return ; 984 } 985 986 // Invariant: after setting Responsible=null an thread must execute 987 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 988 if ((SyncFlags & 4) == 0) { 989 _Responsible = NULL ; 990 } 991 992 #if INCLUDE_TRACE 993 // get the owner's thread id for the MonitorEnter event 994 // if it is enabled and the thread isn't suspended 995 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 996 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 997 } 998 #endif 999 1000 for (;;) { 1001 assert (THREAD == _owner, "invariant") ; 1002 1003 1004 if (Knob_ExitPolicy == 0) { 1005 // release semantics: prior loads and stores from within the critical section 1006 // must not float (reorder) past the following store that drops the lock. 1007 // On SPARC that requires MEMBAR #loadstore|#storestore. 1008 // But of course in TSO #loadstore|#storestore is not required. 1009 // I'd like to write one of the following: 1010 // A. OrderAccess::release() ; _owner = NULL 1011 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 1012 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 1013 // store into a _dummy variable. That store is not needed, but can result 1014 // in massive wasteful coherency traffic on classic SMP systems. 1015 // Instead, I use release_store(), which is implemented as just a simple 1016 // ST on x64, x86 and SPARC. 1017 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1018 OrderAccess::storeload() ; // See if we need to wake a successor 1019 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1020 TEVENT (Inflated exit - simple egress) ; 1021 return ; 1022 } 1023 TEVENT (Inflated exit - complex egress) ; 1024 1025 // Normally the exiting thread is responsible for ensuring succession, 1026 // but if other successors are ready or other entering threads are spinning 1027 // then this thread can simply store NULL into _owner and exit without 1028 // waking a successor. The existence of spinners or ready successors 1029 // guarantees proper succession (liveness). Responsibility passes to the 1030 // ready or running successors. The exiting thread delegates the duty. 1031 // More precisely, if a successor already exists this thread is absolved 1032 // of the responsibility of waking (unparking) one. 1033 // 1034 // The _succ variable is critical to reducing futile wakeup frequency. 1035 // _succ identifies the "heir presumptive" thread that has been made 1036 // ready (unparked) but that has not yet run. We need only one such 1037 // successor thread to guarantee progress. 1038 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1039 // section 3.3 "Futile Wakeup Throttling" for details. 1040 // 1041 // Note that spinners in Enter() also set _succ non-null. 1042 // In the current implementation spinners opportunistically set 1043 // _succ so that exiting threads might avoid waking a successor. 1044 // Another less appealing alternative would be for the exiting thread 1045 // to drop the lock and then spin briefly to see if a spinner managed 1046 // to acquire the lock. If so, the exiting thread could exit 1047 // immediately without waking a successor, otherwise the exiting 1048 // thread would need to dequeue and wake a successor. 1049 // (Note that we'd need to make the post-drop spin short, but no 1050 // shorter than the worst-case round-trip cache-line migration time. 1051 // The dropped lock needs to become visible to the spinner, and then 1052 // the acquisition of the lock by the spinner must become visible to 1053 // the exiting thread). 1054 // 1055 1056 // It appears that an heir-presumptive (successor) must be made ready. 1057 // Only the current lock owner can manipulate the EntryList or 1058 // drain _cxq, so we need to reacquire the lock. If we fail 1059 // to reacquire the lock the responsibility for ensuring succession 1060 // falls to the new owner. 1061 // 1062 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1063 return ; 1064 } 1065 TEVENT (Exit - Reacquired) ; 1066 } else { 1067 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1068 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1069 OrderAccess::storeload() ; 1070 // Ratify the previously observed values. 1071 if (_cxq == NULL || _succ != NULL) { 1072 TEVENT (Inflated exit - simple egress) ; 1073 return ; 1074 } 1075 1076 // inopportune interleaving -- the exiting thread (this thread) 1077 // in the fast-exit path raced an entering thread in the slow-enter 1078 // path. 1079 // We have two choices: 1080 // A. Try to reacquire the lock. 1081 // If the CAS() fails return immediately, otherwise 1082 // we either restart/rerun the exit operation, or simply 1083 // fall-through into the code below which wakes a successor. 1084 // B. If the elements forming the EntryList|cxq are TSM 1085 // we could simply unpark() the lead thread and return 1086 // without having set _succ. 1087 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1088 TEVENT (Inflated exit - reacquired succeeded) ; 1089 return ; 1090 } 1091 TEVENT (Inflated exit - reacquired failed) ; 1092 } else { 1093 TEVENT (Inflated exit - complex egress) ; 1094 } 1095 } 1096 1097 guarantee (_owner == THREAD, "invariant") ; 1098 1099 ObjectWaiter * w = NULL ; 1100 int QMode = Knob_QMode ; 1101 1102 if (QMode == 2 && _cxq != NULL) { 1103 // QMode == 2 : cxq has precedence over EntryList. 1104 // Try to directly wake a successor from the cxq. 1105 // If successful, the successor will need to unlink itself from cxq. 1106 w = _cxq ; 1107 assert (w != NULL, "invariant") ; 1108 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1109 ExitEpilog (Self, w) ; 1110 return ; 1111 } 1112 1113 if (QMode == 3 && _cxq != NULL) { 1114 // Aggressively drain cxq into EntryList at the first opportunity. 1115 // This policy ensure that recently-run threads live at the head of EntryList. 1116 // Drain _cxq into EntryList - bulk transfer. 1117 // First, detach _cxq. 1118 // The following loop is tantamount to: w = swap (&cxq, NULL) 1119 w = _cxq ; 1120 for (;;) { 1121 assert (w != NULL, "Invariant") ; 1122 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1123 if (u == w) break ; 1124 w = u ; 1125 } 1126 assert (w != NULL , "invariant") ; 1127 1128 ObjectWaiter * q = NULL ; 1129 ObjectWaiter * p ; 1130 for (p = w ; p != NULL ; p = p->_next) { 1131 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1132 p->TState = ObjectWaiter::TS_ENTER ; 1133 p->_prev = q ; 1134 q = p ; 1135 } 1136 1137 // Append the RATs to the EntryList 1138 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1139 ObjectWaiter * Tail ; 1140 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1141 if (Tail == NULL) { 1142 _EntryList = w ; 1143 } else { 1144 Tail->_next = w ; 1145 w->_prev = Tail ; 1146 } 1147 1148 // Fall thru into code that tries to wake a successor from EntryList 1149 } 1150 1151 if (QMode == 4 && _cxq != NULL) { 1152 // Aggressively drain cxq into EntryList at the first opportunity. 1153 // This policy ensure that recently-run threads live at the head of EntryList. 1154 1155 // Drain _cxq into EntryList - bulk transfer. 1156 // First, detach _cxq. 1157 // The following loop is tantamount to: w = swap (&cxq, NULL) 1158 w = _cxq ; 1159 for (;;) { 1160 assert (w != NULL, "Invariant") ; 1161 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1162 if (u == w) break ; 1163 w = u ; 1164 } 1165 assert (w != NULL , "invariant") ; 1166 1167 ObjectWaiter * q = NULL ; 1168 ObjectWaiter * p ; 1169 for (p = w ; p != NULL ; p = p->_next) { 1170 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1171 p->TState = ObjectWaiter::TS_ENTER ; 1172 p->_prev = q ; 1173 q = p ; 1174 } 1175 1176 // Prepend the RATs to the EntryList 1177 if (_EntryList != NULL) { 1178 q->_next = _EntryList ; 1179 _EntryList->_prev = q ; 1180 } 1181 _EntryList = w ; 1182 1183 // Fall thru into code that tries to wake a successor from EntryList 1184 } 1185 1186 w = _EntryList ; 1187 if (w != NULL) { 1188 // I'd like to write: guarantee (w->_thread != Self). 1189 // But in practice an exiting thread may find itself on the EntryList. 1190 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1191 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1192 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1193 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1194 // release the lock "O". T2 resumes immediately after the ST of null into 1195 // _owner, above. T2 notices that the EntryList is populated, so it 1196 // reacquires the lock and then finds itself on the EntryList. 1197 // Given all that, we have to tolerate the circumstance where "w" is 1198 // associated with Self. 1199 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1200 ExitEpilog (Self, w) ; 1201 return ; 1202 } 1203 1204 // If we find that both _cxq and EntryList are null then just 1205 // re-run the exit protocol from the top. 1206 w = _cxq ; 1207 if (w == NULL) continue ; 1208 1209 // Drain _cxq into EntryList - bulk transfer. 1210 // First, detach _cxq. 1211 // The following loop is tantamount to: w = swap (&cxq, NULL) 1212 for (;;) { 1213 assert (w != NULL, "Invariant") ; 1214 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1215 if (u == w) break ; 1216 w = u ; 1217 } 1218 TEVENT (Inflated exit - drain cxq into EntryList) ; 1219 1220 assert (w != NULL , "invariant") ; 1221 assert (_EntryList == NULL , "invariant") ; 1222 1223 // Convert the LIFO SLL anchored by _cxq into a DLL. 1224 // The list reorganization step operates in O(LENGTH(w)) time. 1225 // It's critical that this step operate quickly as 1226 // "Self" still holds the outer-lock, restricting parallelism 1227 // and effectively lengthening the critical section. 1228 // Invariant: s chases t chases u. 1229 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1230 // we have faster access to the tail. 1231 1232 if (QMode == 1) { 1233 // QMode == 1 : drain cxq to EntryList, reversing order 1234 // We also reverse the order of the list. 1235 ObjectWaiter * s = NULL ; 1236 ObjectWaiter * t = w ; 1237 ObjectWaiter * u = NULL ; 1238 while (t != NULL) { 1239 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1240 t->TState = ObjectWaiter::TS_ENTER ; 1241 u = t->_next ; 1242 t->_prev = u ; 1243 t->_next = s ; 1244 s = t; 1245 t = u ; 1246 } 1247 _EntryList = s ; 1248 assert (s != NULL, "invariant") ; 1249 } else { 1250 // QMode == 0 or QMode == 2 1251 _EntryList = w ; 1252 ObjectWaiter * q = NULL ; 1253 ObjectWaiter * p ; 1254 for (p = w ; p != NULL ; p = p->_next) { 1255 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1256 p->TState = ObjectWaiter::TS_ENTER ; 1257 p->_prev = q ; 1258 q = p ; 1259 } 1260 } 1261 1262 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1263 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1264 1265 // See if we can abdicate to a spinner instead of waking a thread. 1266 // A primary goal of the implementation is to reduce the 1267 // context-switch rate. 1268 if (_succ != NULL) continue; 1269 1270 w = _EntryList ; 1271 if (w != NULL) { 1272 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1273 ExitEpilog (Self, w) ; 1274 return ; 1275 } 1276 } 1277 } 1278 1279 // ExitSuspendEquivalent: 1280 // A faster alternate to handle_special_suspend_equivalent_condition() 1281 // 1282 // handle_special_suspend_equivalent_condition() unconditionally 1283 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1284 // operations have high latency. Note that in ::enter() we call HSSEC 1285 // while holding the monitor, so we effectively lengthen the critical sections. 1286 // 1287 // There are a number of possible solutions: 1288 // 1289 // A. To ameliorate the problem we might also defer state transitions 1290 // to as late as possible -- just prior to parking. 1291 // Given that, we'd call HSSEC after having returned from park(), 1292 // but before attempting to acquire the monitor. This is only a 1293 // partial solution. It avoids calling HSSEC while holding the 1294 // monitor (good), but it still increases successor reacquisition latency -- 1295 // the interval between unparking a successor and the time the successor 1296 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1297 // If we use this technique we can also avoid EnterI()-exit() loop 1298 // in ::enter() where we iteratively drop the lock and then attempt 1299 // to reacquire it after suspending. 1300 // 1301 // B. In the future we might fold all the suspend bits into a 1302 // composite per-thread suspend flag and then update it with CAS(). 1303 // Alternately, a Dekker-like mechanism with multiple variables 1304 // would suffice: 1305 // ST Self->_suspend_equivalent = false 1306 // MEMBAR 1307 // LD Self_>_suspend_flags 1308 // 1309 1310 1311 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1312 int Mode = Knob_FastHSSEC ; 1313 if (Mode && !jSelf->is_external_suspend()) { 1314 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1315 jSelf->clear_suspend_equivalent() ; 1316 if (2 == Mode) OrderAccess::storeload() ; 1317 if (!jSelf->is_external_suspend()) return false ; 1318 // We raced a suspension -- fall thru into the slow path 1319 TEVENT (ExitSuspendEquivalent - raced) ; 1320 jSelf->set_suspend_equivalent() ; 1321 } 1322 return jSelf->handle_special_suspend_equivalent_condition() ; 1323 } 1324 1325 1326 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1327 assert (_owner == Self, "invariant") ; 1328 1329 // Exit protocol: 1330 // 1. ST _succ = wakee 1331 // 2. membar #loadstore|#storestore; 1332 // 2. ST _owner = NULL 1333 // 3. unpark(wakee) 1334 1335 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1336 ParkEvent * Trigger = Wakee->_event ; 1337 1338 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1339 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1340 // out-of-scope (non-extant). 1341 Wakee = NULL ; 1342 1343 // Drop the lock 1344 OrderAccess::release_store_ptr (&_owner, NULL) ; 1345 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1346 1347 if (SafepointSynchronize::do_call_back()) { 1348 TEVENT (unpark before SAFEPOINT) ; 1349 } 1350 1351 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1352 Trigger->unpark() ; 1353 1354 // Maintain stats and report events to JVMTI 1355 if (ObjectMonitor::_sync_Parks != NULL) { 1356 ObjectMonitor::_sync_Parks->inc() ; 1357 } 1358 } 1359 1360 1361 // ----------------------------------------------------------------------------- 1362 // Class Loader deadlock handling. 1363 // 1364 // complete_exit exits a lock returning recursion count 1365 // complete_exit/reenter operate as a wait without waiting 1366 // complete_exit requires an inflated monitor 1367 // The _owner field is not always the Thread addr even with an 1368 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1369 // thread due to contention. 1370 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1371 Thread * const Self = THREAD; 1372 assert(Self->is_Java_thread(), "Must be Java thread!"); 1373 JavaThread *jt = (JavaThread *)THREAD; 1374 1375 DeferredInitialize(); 1376 1377 if (THREAD != _owner) { 1378 if (THREAD->is_lock_owned ((address)_owner)) { 1379 assert(_recursions == 0, "internal state error"); 1380 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1381 _recursions = 0 ; 1382 OwnerIsThread = 1 ; 1383 } 1384 } 1385 1386 guarantee(Self == _owner, "complete_exit not owner"); 1387 intptr_t save = _recursions; // record the old recursion count 1388 _recursions = 0; // set the recursion level to be 0 1389 exit (true, Self) ; // exit the monitor 1390 guarantee (_owner != Self, "invariant"); 1391 return save; 1392 } 1393 1394 // reenter() enters a lock and sets recursion count 1395 // complete_exit/reenter operate as a wait without waiting 1396 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1397 Thread * const Self = THREAD; 1398 assert(Self->is_Java_thread(), "Must be Java thread!"); 1399 JavaThread *jt = (JavaThread *)THREAD; 1400 1401 guarantee(_owner != Self, "reenter already owner"); 1402 enter (THREAD); // enter the monitor 1403 guarantee (_recursions == 0, "reenter recursion"); 1404 _recursions = recursions; 1405 return; 1406 } 1407 1408 1409 // ----------------------------------------------------------------------------- 1410 // A macro is used below because there may already be a pending 1411 // exception which should not abort the execution of the routines 1412 // which use this (which is why we don't put this into check_slow and 1413 // call it with a CHECK argument). 1414 1415 #define CHECK_OWNER() \ 1416 do { \ 1417 if (THREAD != _owner) { \ 1418 if (THREAD->is_lock_owned((address) _owner)) { \ 1419 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1420 _recursions = 0; \ 1421 OwnerIsThread = 1 ; \ 1422 } else { \ 1423 TEVENT (Throw IMSX) ; \ 1424 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1425 } \ 1426 } \ 1427 } while (false) 1428 1429 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1430 // TODO-FIXME: remove check_slow() -- it's likely dead. 1431 1432 void ObjectMonitor::check_slow(TRAPS) { 1433 TEVENT (check_slow - throw IMSX) ; 1434 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1435 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1436 } 1437 1438 static int Adjust (volatile int * adr, int dx) { 1439 int v ; 1440 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1441 return v ; 1442 } 1443 1444 // helper method for posting a monitor wait event 1445 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1446 jlong notifier_tid, 1447 jlong timeout, 1448 bool timedout) { 1449 event->set_klass(((oop)this->object())->klass()); 1450 event->set_timeout((TYPE_ULONG)timeout); 1451 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1452 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1453 event->set_timedOut((TYPE_BOOLEAN)timedout); 1454 event->commit(); 1455 } 1456 1457 // ----------------------------------------------------------------------------- 1458 // Wait/Notify/NotifyAll 1459 // 1460 // Note: a subset of changes to ObjectMonitor::wait() 1461 // will need to be replicated in complete_exit above 1462 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1463 Thread * const Self = THREAD ; 1464 assert(Self->is_Java_thread(), "Must be Java thread!"); 1465 JavaThread *jt = (JavaThread *)THREAD; 1466 1467 DeferredInitialize () ; 1468 1469 // Throw IMSX or IEX. 1470 CHECK_OWNER(); 1471 1472 EventJavaMonitorWait event; 1473 1474 // check for a pending interrupt 1475 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1476 // post monitor waited event. Note that this is past-tense, we are done waiting. 1477 if (JvmtiExport::should_post_monitor_waited()) { 1478 // Note: 'false' parameter is passed here because the 1479 // wait was not timed out due to thread interrupt. 1480 JvmtiExport::post_monitor_waited(jt, this, false); 1481 1482 // In this short circuit of the monitor wait protocol, the 1483 // current thread never drops ownership of the monitor and 1484 // never gets added to the wait queue so the current thread 1485 // cannot be made the successor. This means that the 1486 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1487 // consume an unpark() meant for the ParkEvent associated with 1488 // this ObjectMonitor. 1489 } 1490 if (event.should_commit()) { 1491 post_monitor_wait_event(&event, 0, millis, false); 1492 } 1493 TEVENT (Wait - Throw IEX) ; 1494 THROW(vmSymbols::java_lang_InterruptedException()); 1495 return ; 1496 } 1497 1498 TEVENT (Wait) ; 1499 1500 assert (Self->_Stalled == 0, "invariant") ; 1501 Self->_Stalled = intptr_t(this) ; 1502 jt->set_current_waiting_monitor(this); 1503 1504 // create a node to be put into the queue 1505 // Critically, after we reset() the event but prior to park(), we must check 1506 // for a pending interrupt. 1507 ObjectWaiter node(Self); 1508 node.TState = ObjectWaiter::TS_WAIT ; 1509 Self->_ParkEvent->reset() ; 1510 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1511 1512 // Enter the waiting queue, which is a circular doubly linked list in this case 1513 // but it could be a priority queue or any data structure. 1514 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1515 // by the the owner of the monitor *except* in the case where park() 1516 // returns because of a timeout of interrupt. Contention is exceptionally rare 1517 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1518 1519 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1520 AddWaiter (&node) ; 1521 Thread::SpinRelease (&_WaitSetLock) ; 1522 1523 if ((SyncFlags & 4) == 0) { 1524 _Responsible = NULL ; 1525 } 1526 intptr_t save = _recursions; // record the old recursion count 1527 _waiters++; // increment the number of waiters 1528 _recursions = 0; // set the recursion level to be 1 1529 exit (true, Self) ; // exit the monitor 1530 guarantee (_owner != Self, "invariant") ; 1531 1532 // The thread is on the WaitSet list - now park() it. 1533 // On MP systems it's conceivable that a brief spin before we park 1534 // could be profitable. 1535 // 1536 // TODO-FIXME: change the following logic to a loop of the form 1537 // while (!timeout && !interrupted && _notified == 0) park() 1538 1539 int ret = OS_OK ; 1540 int WasNotified = 0 ; 1541 { // State transition wrappers 1542 OSThread* osthread = Self->osthread(); 1543 OSThreadWaitState osts(osthread, true); 1544 { 1545 ThreadBlockInVM tbivm(jt); 1546 // Thread is in thread_blocked state and oop access is unsafe. 1547 jt->set_suspend_equivalent(); 1548 1549 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1550 // Intentionally empty 1551 } else 1552 if (node._notified == 0) { 1553 if (millis <= 0) { 1554 Self->_ParkEvent->park () ; 1555 } else { 1556 ret = Self->_ParkEvent->park (millis) ; 1557 } 1558 } 1559 1560 // were we externally suspended while we were waiting? 1561 if (ExitSuspendEquivalent (jt)) { 1562 // TODO-FIXME: add -- if succ == Self then succ = null. 1563 jt->java_suspend_self(); 1564 } 1565 1566 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1567 1568 1569 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1570 // from the WaitSet to the EntryList. 1571 // See if we need to remove Node from the WaitSet. 1572 // We use double-checked locking to avoid grabbing _WaitSetLock 1573 // if the thread is not on the wait queue. 1574 // 1575 // Note that we don't need a fence before the fetch of TState. 1576 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1577 // written by the is thread. (perhaps the fetch might even be satisfied 1578 // by a look-aside into the processor's own store buffer, although given 1579 // the length of the code path between the prior ST and this load that's 1580 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1581 // then we'll acquire the lock and then re-fetch a fresh TState value. 1582 // That is, we fail toward safety. 1583 1584 if (node.TState == ObjectWaiter::TS_WAIT) { 1585 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1586 if (node.TState == ObjectWaiter::TS_WAIT) { 1587 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1588 assert(node._notified == 0, "invariant"); 1589 node.TState = ObjectWaiter::TS_RUN ; 1590 } 1591 Thread::SpinRelease (&_WaitSetLock) ; 1592 } 1593 1594 // The thread is now either on off-list (TS_RUN), 1595 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1596 // The Node's TState variable is stable from the perspective of this thread. 1597 // No other threads will asynchronously modify TState. 1598 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1599 OrderAccess::loadload() ; 1600 if (_succ == Self) _succ = NULL ; 1601 WasNotified = node._notified ; 1602 1603 // Reentry phase -- reacquire the monitor. 1604 // re-enter contended monitor after object.wait(). 1605 // retain OBJECT_WAIT state until re-enter successfully completes 1606 // Thread state is thread_in_vm and oop access is again safe, 1607 // although the raw address of the object may have changed. 1608 // (Don't cache naked oops over safepoints, of course). 1609 1610 // post monitor waited event. Note that this is past-tense, we are done waiting. 1611 if (JvmtiExport::should_post_monitor_waited()) { 1612 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1613 1614 if (node._notified != 0 && _succ == Self) { 1615 // In this part of the monitor wait-notify-reenter protocol it 1616 // is possible (and normal) for another thread to do a fastpath 1617 // monitor enter-exit while this thread is still trying to get 1618 // to the reenter portion of the protocol. 1619 // 1620 // The ObjectMonitor was notified and the current thread is 1621 // the successor which also means that an unpark() has already 1622 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1623 // consume the unpark() that was done when the successor was 1624 // set because the same ParkEvent is shared between Java 1625 // monitors and JVM/TI RawMonitors (for now). 1626 // 1627 // We redo the unpark() to ensure forward progress, i.e., we 1628 // don't want all pending threads hanging (parked) with none 1629 // entering the unlocked monitor. 1630 node._event->unpark(); 1631 } 1632 } 1633 1634 if (event.should_commit()) { 1635 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1636 } 1637 1638 OrderAccess::fence() ; 1639 1640 assert (Self->_Stalled != 0, "invariant") ; 1641 Self->_Stalled = 0 ; 1642 1643 assert (_owner != Self, "invariant") ; 1644 ObjectWaiter::TStates v = node.TState ; 1645 if (v == ObjectWaiter::TS_RUN) { 1646 enter (Self) ; 1647 } else { 1648 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1649 ReenterI (Self, &node) ; 1650 node.wait_reenter_end(this); 1651 } 1652 1653 // Self has reacquired the lock. 1654 // Lifecycle - the node representing Self must not appear on any queues. 1655 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1656 // want residual elements associated with this thread left on any lists. 1657 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1658 assert (_owner == Self, "invariant") ; 1659 assert (_succ != Self , "invariant") ; 1660 } // OSThreadWaitState() 1661 1662 jt->set_current_waiting_monitor(NULL); 1663 1664 guarantee (_recursions == 0, "invariant") ; 1665 _recursions = save; // restore the old recursion count 1666 _waiters--; // decrement the number of waiters 1667 1668 // Verify a few postconditions 1669 assert (_owner == Self , "invariant") ; 1670 assert (_succ != Self , "invariant") ; 1671 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1672 1673 if (SyncFlags & 32) { 1674 OrderAccess::fence() ; 1675 } 1676 1677 // check if the notification happened 1678 if (!WasNotified) { 1679 // no, it could be timeout or Thread.interrupt() or both 1680 // check for interrupt event, otherwise it is timeout 1681 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1682 TEVENT (Wait - throw IEX from epilog) ; 1683 THROW(vmSymbols::java_lang_InterruptedException()); 1684 } 1685 } 1686 1687 // NOTE: Spurious wake up will be consider as timeout. 1688 // Monitor notify has precedence over thread interrupt. 1689 } 1690 1691 1692 // Consider: 1693 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1694 // then instead of transferring a thread from the WaitSet to the EntryList 1695 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1696 1697 void ObjectMonitor::notify(TRAPS) { 1698 CHECK_OWNER(); 1699 if (_WaitSet == NULL) { 1700 TEVENT (Empty-Notify) ; 1701 return ; 1702 } 1703 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1704 1705 int Policy = Knob_MoveNotifyee ; 1706 1707 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1708 ObjectWaiter * iterator = DequeueWaiter() ; 1709 if (iterator != NULL) { 1710 TEVENT (Notify1 - Transfer) ; 1711 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1712 guarantee (iterator->_notified == 0, "invariant") ; 1713 if (Policy != 4) { 1714 iterator->TState = ObjectWaiter::TS_ENTER ; 1715 } 1716 iterator->_notified = 1 ; 1717 Thread * Self = THREAD; 1718 iterator->_notifier_tid = Self->osthread()->thread_id(); 1719 1720 ObjectWaiter * List = _EntryList ; 1721 if (List != NULL) { 1722 assert (List->_prev == NULL, "invariant") ; 1723 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1724 assert (List != iterator, "invariant") ; 1725 } 1726 1727 if (Policy == 0) { // prepend to EntryList 1728 if (List == NULL) { 1729 iterator->_next = iterator->_prev = NULL ; 1730 _EntryList = iterator ; 1731 } else { 1732 List->_prev = iterator ; 1733 iterator->_next = List ; 1734 iterator->_prev = NULL ; 1735 _EntryList = iterator ; 1736 } 1737 } else 1738 if (Policy == 1) { // append to EntryList 1739 if (List == NULL) { 1740 iterator->_next = iterator->_prev = NULL ; 1741 _EntryList = iterator ; 1742 } else { 1743 // CONSIDER: finding the tail currently requires a linear-time walk of 1744 // the EntryList. We can make tail access constant-time by converting to 1745 // a CDLL instead of using our current DLL. 1746 ObjectWaiter * Tail ; 1747 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1748 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1749 Tail->_next = iterator ; 1750 iterator->_prev = Tail ; 1751 iterator->_next = NULL ; 1752 } 1753 } else 1754 if (Policy == 2) { // prepend to cxq 1755 // prepend to cxq 1756 if (List == NULL) { 1757 iterator->_next = iterator->_prev = NULL ; 1758 _EntryList = iterator ; 1759 } else { 1760 iterator->TState = ObjectWaiter::TS_CXQ ; 1761 for (;;) { 1762 ObjectWaiter * Front = _cxq ; 1763 iterator->_next = Front ; 1764 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1765 break ; 1766 } 1767 } 1768 } 1769 } else 1770 if (Policy == 3) { // append to cxq 1771 iterator->TState = ObjectWaiter::TS_CXQ ; 1772 for (;;) { 1773 ObjectWaiter * Tail ; 1774 Tail = _cxq ; 1775 if (Tail == NULL) { 1776 iterator->_next = NULL ; 1777 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1778 break ; 1779 } 1780 } else { 1781 while (Tail->_next != NULL) Tail = Tail->_next ; 1782 Tail->_next = iterator ; 1783 iterator->_prev = Tail ; 1784 iterator->_next = NULL ; 1785 break ; 1786 } 1787 } 1788 } else { 1789 ParkEvent * ev = iterator->_event ; 1790 iterator->TState = ObjectWaiter::TS_RUN ; 1791 OrderAccess::fence() ; 1792 ev->unpark() ; 1793 } 1794 1795 if (Policy < 4) { 1796 iterator->wait_reenter_begin(this); 1797 } 1798 1799 // _WaitSetLock protects the wait queue, not the EntryList. We could 1800 // move the add-to-EntryList operation, above, outside the critical section 1801 // protected by _WaitSetLock. In practice that's not useful. With the 1802 // exception of wait() timeouts and interrupts the monitor owner 1803 // is the only thread that grabs _WaitSetLock. There's almost no contention 1804 // on _WaitSetLock so it's not profitable to reduce the length of the 1805 // critical section. 1806 } 1807 1808 Thread::SpinRelease (&_WaitSetLock) ; 1809 1810 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1811 ObjectMonitor::_sync_Notifications->inc() ; 1812 } 1813 } 1814 1815 1816 void ObjectMonitor::notifyAll(TRAPS) { 1817 CHECK_OWNER(); 1818 ObjectWaiter* iterator; 1819 if (_WaitSet == NULL) { 1820 TEVENT (Empty-NotifyAll) ; 1821 return ; 1822 } 1823 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1824 1825 int Policy = Knob_MoveNotifyee ; 1826 int Tally = 0 ; 1827 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1828 1829 for (;;) { 1830 iterator = DequeueWaiter () ; 1831 if (iterator == NULL) break ; 1832 TEVENT (NotifyAll - Transfer1) ; 1833 ++Tally ; 1834 1835 // Disposition - what might we do with iterator ? 1836 // a. add it directly to the EntryList - either tail or head. 1837 // b. push it onto the front of the _cxq. 1838 // For now we use (a). 1839 1840 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1841 guarantee (iterator->_notified == 0, "invariant") ; 1842 iterator->_notified = 1 ; 1843 Thread * Self = THREAD; 1844 iterator->_notifier_tid = Self->osthread()->thread_id(); 1845 if (Policy != 4) { 1846 iterator->TState = ObjectWaiter::TS_ENTER ; 1847 } 1848 1849 ObjectWaiter * List = _EntryList ; 1850 if (List != NULL) { 1851 assert (List->_prev == NULL, "invariant") ; 1852 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1853 assert (List != iterator, "invariant") ; 1854 } 1855 1856 if (Policy == 0) { // prepend to EntryList 1857 if (List == NULL) { 1858 iterator->_next = iterator->_prev = NULL ; 1859 _EntryList = iterator ; 1860 } else { 1861 List->_prev = iterator ; 1862 iterator->_next = List ; 1863 iterator->_prev = NULL ; 1864 _EntryList = iterator ; 1865 } 1866 } else 1867 if (Policy == 1) { // append to EntryList 1868 if (List == NULL) { 1869 iterator->_next = iterator->_prev = NULL ; 1870 _EntryList = iterator ; 1871 } else { 1872 // CONSIDER: finding the tail currently requires a linear-time walk of 1873 // the EntryList. We can make tail access constant-time by converting to 1874 // a CDLL instead of using our current DLL. 1875 ObjectWaiter * Tail ; 1876 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1877 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1878 Tail->_next = iterator ; 1879 iterator->_prev = Tail ; 1880 iterator->_next = NULL ; 1881 } 1882 } else 1883 if (Policy == 2) { // prepend to cxq 1884 // prepend to cxq 1885 iterator->TState = ObjectWaiter::TS_CXQ ; 1886 for (;;) { 1887 ObjectWaiter * Front = _cxq ; 1888 iterator->_next = Front ; 1889 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1890 break ; 1891 } 1892 } 1893 } else 1894 if (Policy == 3) { // append to cxq 1895 iterator->TState = ObjectWaiter::TS_CXQ ; 1896 for (;;) { 1897 ObjectWaiter * Tail ; 1898 Tail = _cxq ; 1899 if (Tail == NULL) { 1900 iterator->_next = NULL ; 1901 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1902 break ; 1903 } 1904 } else { 1905 while (Tail->_next != NULL) Tail = Tail->_next ; 1906 Tail->_next = iterator ; 1907 iterator->_prev = Tail ; 1908 iterator->_next = NULL ; 1909 break ; 1910 } 1911 } 1912 } else { 1913 ParkEvent * ev = iterator->_event ; 1914 iterator->TState = ObjectWaiter::TS_RUN ; 1915 OrderAccess::fence() ; 1916 ev->unpark() ; 1917 } 1918 1919 if (Policy < 4) { 1920 iterator->wait_reenter_begin(this); 1921 } 1922 1923 // _WaitSetLock protects the wait queue, not the EntryList. We could 1924 // move the add-to-EntryList operation, above, outside the critical section 1925 // protected by _WaitSetLock. In practice that's not useful. With the 1926 // exception of wait() timeouts and interrupts the monitor owner 1927 // is the only thread that grabs _WaitSetLock. There's almost no contention 1928 // on _WaitSetLock so it's not profitable to reduce the length of the 1929 // critical section. 1930 } 1931 1932 Thread::SpinRelease (&_WaitSetLock) ; 1933 1934 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1935 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1936 } 1937 } 1938 1939 // ----------------------------------------------------------------------------- 1940 // Adaptive Spinning Support 1941 // 1942 // Adaptive spin-then-block - rational spinning 1943 // 1944 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1945 // algorithm. On high order SMP systems it would be better to start with 1946 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1947 // a contending thread could enqueue itself on the cxq and then spin locally 1948 // on a thread-specific variable such as its ParkEvent._Event flag. 1949 // That's left as an exercise for the reader. Note that global spinning is 1950 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1951 // low latency and massive bandwidth. 1952 // 1953 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1954 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1955 // (duration) or we can fix the count at approximately the duration of 1956 // a context switch and vary the frequency. Of course we could also 1957 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1958 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1959 // 1960 // This implementation varies the duration "D", where D varies with 1961 // the success rate of recent spin attempts. (D is capped at approximately 1962 // length of a round-trip context switch). The success rate for recent 1963 // spin attempts is a good predictor of the success rate of future spin 1964 // attempts. The mechanism adapts automatically to varying critical 1965 // section length (lock modality), system load and degree of parallelism. 1966 // D is maintained per-monitor in _SpinDuration and is initialized 1967 // optimistically. Spin frequency is fixed at 100%. 1968 // 1969 // Note that _SpinDuration is volatile, but we update it without locks 1970 // or atomics. The code is designed so that _SpinDuration stays within 1971 // a reasonable range even in the presence of races. The arithmetic 1972 // operations on _SpinDuration are closed over the domain of legal values, 1973 // so at worst a race will install and older but still legal value. 1974 // At the very worst this introduces some apparent non-determinism. 1975 // We might spin when we shouldn't or vice-versa, but since the spin 1976 // count are relatively short, even in the worst case, the effect is harmless. 1977 // 1978 // Care must be taken that a low "D" value does not become an 1979 // an absorbing state. Transient spinning failures -- when spinning 1980 // is overall profitable -- should not cause the system to converge 1981 // on low "D" values. We want spinning to be stable and predictable 1982 // and fairly responsive to change and at the same time we don't want 1983 // it to oscillate, become metastable, be "too" non-deterministic, 1984 // or converge on or enter undesirable stable absorbing states. 1985 // 1986 // We implement a feedback-based control system -- using past behavior 1987 // to predict future behavior. We face two issues: (a) if the 1988 // input signal is random then the spin predictor won't provide optimal 1989 // results, and (b) if the signal frequency is too high then the control 1990 // system, which has some natural response lag, will "chase" the signal. 1991 // (b) can arise from multimodal lock hold times. Transient preemption 1992 // can also result in apparent bimodal lock hold times. 1993 // Although sub-optimal, neither condition is particularly harmful, as 1994 // in the worst-case we'll spin when we shouldn't or vice-versa. 1995 // The maximum spin duration is rather short so the failure modes aren't bad. 1996 // To be conservative, I've tuned the gain in system to bias toward 1997 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1998 // "rings" or oscillates between spinning and not spinning. This happens 1999 // when spinning is just on the cusp of profitability, however, so the 2000 // situation is not dire. The state is benign -- there's no need to add 2001 // hysteresis control to damp the transition rate between spinning and 2002 // not spinning. 2003 // 2004 2005 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 2006 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 2007 2008 // Spinning: Fixed frequency (100%), vary duration 2009 2010 2011 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 2012 2013 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2014 int ctr = Knob_FixedSpin ; 2015 if (ctr != 0) { 2016 while (--ctr >= 0) { 2017 if (TryLock (Self) > 0) return 1 ; 2018 SpinPause () ; 2019 } 2020 return 0 ; 2021 } 2022 2023 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 2024 if (TryLock(Self) > 0) { 2025 // Increase _SpinDuration ... 2026 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2027 // Raising _SpurDuration to the poverty line is key. 2028 int x = _SpinDuration ; 2029 if (x < Knob_SpinLimit) { 2030 if (x < Knob_Poverty) x = Knob_Poverty ; 2031 _SpinDuration = x + Knob_BonusB ; 2032 } 2033 return 1 ; 2034 } 2035 SpinPause () ; 2036 } 2037 2038 // Admission control - verify preconditions for spinning 2039 // 2040 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2041 // becoming an absorbing state. Put another way, we spin briefly to 2042 // sample, just in case the system load, parallelism, contention, or lock 2043 // modality changed. 2044 // 2045 // Consider the following alternative: 2046 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2047 // spin attempt. "Periodically" might mean after a tally of 2048 // the # of failed spin attempts (or iterations) reaches some threshold. 2049 // This takes us into the realm of 1-out-of-N spinning, where we 2050 // hold the duration constant but vary the frequency. 2051 2052 ctr = _SpinDuration ; 2053 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2054 if (ctr <= 0) return 0 ; 2055 2056 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2057 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2058 TEVENT (Spin abort - notrunnable [TOP]); 2059 return 0 ; 2060 } 2061 2062 int MaxSpin = Knob_MaxSpinners ; 2063 if (MaxSpin >= 0) { 2064 if (_Spinner > MaxSpin) { 2065 TEVENT (Spin abort -- too many spinners) ; 2066 return 0 ; 2067 } 2068 // Slighty racy, but benign ... 2069 Adjust (&_Spinner, 1) ; 2070 } 2071 2072 // We're good to spin ... spin ingress. 2073 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2074 // when preparing to LD...CAS _owner, etc and the CAS is likely 2075 // to succeed. 2076 int hits = 0 ; 2077 int msk = 0 ; 2078 int caspty = Knob_CASPenalty ; 2079 int oxpty = Knob_OXPenalty ; 2080 int sss = Knob_SpinSetSucc ; 2081 if (sss && _succ == NULL ) _succ = Self ; 2082 Thread * prv = NULL ; 2083 2084 // There are three ways to exit the following loop: 2085 // 1. A successful spin where this thread has acquired the lock. 2086 // 2. Spin failure with prejudice 2087 // 3. Spin failure without prejudice 2088 2089 while (--ctr >= 0) { 2090 2091 // Periodic polling -- Check for pending GC 2092 // Threads may spin while they're unsafe. 2093 // We don't want spinning threads to delay the JVM from reaching 2094 // a stop-the-world safepoint or to steal cycles from GC. 2095 // If we detect a pending safepoint we abort in order that 2096 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2097 // this thread, if safe, doesn't steal cycles from GC. 2098 // This is in keeping with the "no loitering in runtime" rule. 2099 // We periodically check to see if there's a safepoint pending. 2100 if ((ctr & 0xFF) == 0) { 2101 if (SafepointSynchronize::do_call_back()) { 2102 TEVENT (Spin: safepoint) ; 2103 goto Abort ; // abrupt spin egress 2104 } 2105 if (Knob_UsePause & 1) SpinPause () ; 2106 2107 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2108 if (hits > 50 && scb != NULL) { 2109 int abend = (*scb)(SpinCallbackArgument, 0) ; 2110 } 2111 } 2112 2113 if (Knob_UsePause & 2) SpinPause() ; 2114 2115 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2116 // This is useful on classic SMP systems, but is of less utility on 2117 // N1-style CMT platforms. 2118 // 2119 // Trade-off: lock acquisition latency vs coherency bandwidth. 2120 // Lock hold times are typically short. A histogram 2121 // of successful spin attempts shows that we usually acquire 2122 // the lock early in the spin. That suggests we want to 2123 // sample _owner frequently in the early phase of the spin, 2124 // but then back-off and sample less frequently as the spin 2125 // progresses. The back-off makes a good citizen on SMP big 2126 // SMP systems. Oversampling _owner can consume excessive 2127 // coherency bandwidth. Relatedly, if we _oversample _owner we 2128 // can inadvertently interfere with the the ST m->owner=null. 2129 // executed by the lock owner. 2130 if (ctr & msk) continue ; 2131 ++hits ; 2132 if ((hits & 0xF) == 0) { 2133 // The 0xF, above, corresponds to the exponent. 2134 // Consider: (msk+1)|msk 2135 msk = ((msk << 2)|3) & BackOffMask ; 2136 } 2137 2138 // Probe _owner with TATAS 2139 // If this thread observes the monitor transition or flicker 2140 // from locked to unlocked to locked, then the odds that this 2141 // thread will acquire the lock in this spin attempt go down 2142 // considerably. The same argument applies if the CAS fails 2143 // or if we observe _owner change from one non-null value to 2144 // another non-null value. In such cases we might abort 2145 // the spin without prejudice or apply a "penalty" to the 2146 // spin count-down variable "ctr", reducing it by 100, say. 2147 2148 Thread * ox = (Thread *) _owner ; 2149 if (ox == NULL) { 2150 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2151 if (ox == NULL) { 2152 // The CAS succeeded -- this thread acquired ownership 2153 // Take care of some bookkeeping to exit spin state. 2154 if (sss && _succ == Self) { 2155 _succ = NULL ; 2156 } 2157 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2158 2159 // Increase _SpinDuration : 2160 // The spin was successful (profitable) so we tend toward 2161 // longer spin attempts in the future. 2162 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2163 // If we acquired the lock early in the spin cycle it 2164 // makes sense to increase _SpinDuration proportionally. 2165 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2166 int x = _SpinDuration ; 2167 if (x < Knob_SpinLimit) { 2168 if (x < Knob_Poverty) x = Knob_Poverty ; 2169 _SpinDuration = x + Knob_Bonus ; 2170 } 2171 return 1 ; 2172 } 2173 2174 // The CAS failed ... we can take any of the following actions: 2175 // * penalize: ctr -= Knob_CASPenalty 2176 // * exit spin with prejudice -- goto Abort; 2177 // * exit spin without prejudice. 2178 // * Since CAS is high-latency, retry again immediately. 2179 prv = ox ; 2180 TEVENT (Spin: cas failed) ; 2181 if (caspty == -2) break ; 2182 if (caspty == -1) goto Abort ; 2183 ctr -= caspty ; 2184 continue ; 2185 } 2186 2187 // Did lock ownership change hands ? 2188 if (ox != prv && prv != NULL ) { 2189 TEVENT (spin: Owner changed) 2190 if (oxpty == -2) break ; 2191 if (oxpty == -1) goto Abort ; 2192 ctr -= oxpty ; 2193 } 2194 prv = ox ; 2195 2196 // Abort the spin if the owner is not executing. 2197 // The owner must be executing in order to drop the lock. 2198 // Spinning while the owner is OFFPROC is idiocy. 2199 // Consider: ctr -= RunnablePenalty ; 2200 if (Knob_OState && NotRunnable (Self, ox)) { 2201 TEVENT (Spin abort - notrunnable); 2202 goto Abort ; 2203 } 2204 if (sss && _succ == NULL ) _succ = Self ; 2205 } 2206 2207 // Spin failed with prejudice -- reduce _SpinDuration. 2208 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2209 // AIMD is globally stable. 2210 TEVENT (Spin failure) ; 2211 { 2212 int x = _SpinDuration ; 2213 if (x > 0) { 2214 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2215 // This is globally sample and tends to damp the response. 2216 x -= Knob_Penalty ; 2217 if (x < 0) x = 0 ; 2218 _SpinDuration = x ; 2219 } 2220 } 2221 2222 Abort: 2223 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2224 if (sss && _succ == Self) { 2225 _succ = NULL ; 2226 // Invariant: after setting succ=null a contending thread 2227 // must recheck-retry _owner before parking. This usually happens 2228 // in the normal usage of TrySpin(), but it's safest 2229 // to make TrySpin() as foolproof as possible. 2230 OrderAccess::fence() ; 2231 if (TryLock(Self) > 0) return 1 ; 2232 } 2233 return 0 ; 2234 } 2235 2236 // NotRunnable() -- informed spinning 2237 // 2238 // Don't bother spinning if the owner is not eligible to drop the lock. 2239 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2240 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2241 // The thread must be runnable in order to drop the lock in timely fashion. 2242 // If the _owner is not runnable then spinning will not likely be 2243 // successful (profitable). 2244 // 2245 // Beware -- the thread referenced by _owner could have died 2246 // so a simply fetch from _owner->_thread_state might trap. 2247 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2248 // Because of the lifecycle issues the schedctl and _thread_state values 2249 // observed by NotRunnable() might be garbage. NotRunnable must 2250 // tolerate this and consider the observed _thread_state value 2251 // as advisory. 2252 // 2253 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2254 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2255 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2256 // with the LSB of _owner. Another option would be to probablistically probe 2257 // the putative _owner->TypeTag value. 2258 // 2259 // Checking _thread_state isn't perfect. Even if the thread is 2260 // in_java it might be blocked on a page-fault or have been preempted 2261 // and sitting on a ready/dispatch queue. _thread state in conjunction 2262 // with schedctl.sc_state gives us a good picture of what the 2263 // thread is doing, however. 2264 // 2265 // TODO: check schedctl.sc_state. 2266 // We'll need to use SafeFetch32() to read from the schedctl block. 2267 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2268 // 2269 // The return value from NotRunnable() is *advisory* -- the 2270 // result is based on sampling and is not necessarily coherent. 2271 // The caller must tolerate false-negative and false-positive errors. 2272 // Spinning, in general, is probabilistic anyway. 2273 2274 2275 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2276 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2277 if (!OwnerIsThread) return 0 ; 2278 2279 if (ox == NULL) return 0 ; 2280 2281 // Avoid transitive spinning ... 2282 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2283 // Immediately after T1 acquires L it's possible that T2, also 2284 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2285 // This occurs transiently after T1 acquired L but before 2286 // T1 managed to clear T1.Stalled. T2 does not need to abort 2287 // its spin in this circumstance. 2288 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2289 2290 if (BlockedOn == 1) return 1 ; 2291 if (BlockedOn != 0) { 2292 return BlockedOn != intptr_t(this) && _owner == ox ; 2293 } 2294 2295 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2296 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2297 // consider also: jst != _thread_in_Java -- but that's overspecific. 2298 return jst == _thread_blocked || jst == _thread_in_native ; 2299 } 2300 2301 2302 // ----------------------------------------------------------------------------- 2303 // WaitSet management ... 2304 2305 ObjectWaiter::ObjectWaiter(Thread* thread) { 2306 _next = NULL; 2307 _prev = NULL; 2308 _notified = 0; 2309 TState = TS_RUN ; 2310 _thread = thread; 2311 _event = thread->_ParkEvent ; 2312 _active = false; 2313 assert (_event != NULL, "invariant") ; 2314 } 2315 2316 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2317 JavaThread *jt = (JavaThread *)this->_thread; 2318 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2319 } 2320 2321 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2322 JavaThread *jt = (JavaThread *)this->_thread; 2323 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2324 } 2325 2326 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2327 assert(node != NULL, "should not dequeue NULL node"); 2328 assert(node->_prev == NULL, "node already in list"); 2329 assert(node->_next == NULL, "node already in list"); 2330 // put node at end of queue (circular doubly linked list) 2331 if (_WaitSet == NULL) { 2332 _WaitSet = node; 2333 node->_prev = node; 2334 node->_next = node; 2335 } else { 2336 ObjectWaiter* head = _WaitSet ; 2337 ObjectWaiter* tail = head->_prev; 2338 assert(tail->_next == head, "invariant check"); 2339 tail->_next = node; 2340 head->_prev = node; 2341 node->_next = head; 2342 node->_prev = tail; 2343 } 2344 } 2345 2346 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2347 // dequeue the very first waiter 2348 ObjectWaiter* waiter = _WaitSet; 2349 if (waiter) { 2350 DequeueSpecificWaiter(waiter); 2351 } 2352 return waiter; 2353 } 2354 2355 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2356 assert(node != NULL, "should not dequeue NULL node"); 2357 assert(node->_prev != NULL, "node already removed from list"); 2358 assert(node->_next != NULL, "node already removed from list"); 2359 // when the waiter has woken up because of interrupt, 2360 // timeout or other spurious wake-up, dequeue the 2361 // waiter from waiting list 2362 ObjectWaiter* next = node->_next; 2363 if (next == node) { 2364 assert(node->_prev == node, "invariant check"); 2365 _WaitSet = NULL; 2366 } else { 2367 ObjectWaiter* prev = node->_prev; 2368 assert(prev->_next == node, "invariant check"); 2369 assert(next->_prev == node, "invariant check"); 2370 next->_prev = prev; 2371 prev->_next = next; 2372 if (_WaitSet == node) { 2373 _WaitSet = next; 2374 } 2375 } 2376 node->_next = NULL; 2377 node->_prev = NULL; 2378 } 2379 2380 // ----------------------------------------------------------------------------- 2381 // PerfData support 2382 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2383 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2384 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2385 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2386 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2387 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2388 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2389 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2390 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2391 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2392 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2393 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2394 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2395 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2396 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2397 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2398 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2399 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2400 2401 // One-shot global initialization for the sync subsystem. 2402 // We could also defer initialization and initialize on-demand 2403 // the first time we call inflate(). Initialization would 2404 // be protected - like so many things - by the MonitorCache_lock. 2405 2406 void ObjectMonitor::Initialize () { 2407 static int InitializationCompleted = 0 ; 2408 assert (InitializationCompleted == 0, "invariant") ; 2409 InitializationCompleted = 1 ; 2410 if (UsePerfData) { 2411 EXCEPTION_MARK ; 2412 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2413 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2414 NEWPERFCOUNTER(_sync_Inflations) ; 2415 NEWPERFCOUNTER(_sync_Deflations) ; 2416 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2417 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2418 NEWPERFCOUNTER(_sync_Parks) ; 2419 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2420 NEWPERFCOUNTER(_sync_Notifications) ; 2421 NEWPERFCOUNTER(_sync_SlowEnter) ; 2422 NEWPERFCOUNTER(_sync_SlowExit) ; 2423 NEWPERFCOUNTER(_sync_SlowNotify) ; 2424 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2425 NEWPERFCOUNTER(_sync_FailedSpins) ; 2426 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2427 NEWPERFCOUNTER(_sync_PrivateA) ; 2428 NEWPERFCOUNTER(_sync_PrivateB) ; 2429 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2430 NEWPERFCOUNTER(_sync_MonScavenged) ; 2431 NEWPERFVARIABLE(_sync_MonExtant) ; 2432 #undef NEWPERFCOUNTER 2433 } 2434 } 2435 2436 2437 // Compile-time asserts 2438 // When possible, it's better to catch errors deterministically at 2439 // compile-time than at runtime. The down-side to using compile-time 2440 // asserts is that error message -- often something about negative array 2441 // indices -- is opaque. 2442 2443 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2444 2445 void ObjectMonitor::ctAsserts() { 2446 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2447 } 2448 2449 2450 static char * kvGet (char * kvList, const char * Key) { 2451 if (kvList == NULL) return NULL ; 2452 size_t n = strlen (Key) ; 2453 char * Search ; 2454 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2455 if (strncmp (Search, Key, n) == 0) { 2456 if (Search[n] == '=') return Search + n + 1 ; 2457 if (Search[n] == 0) return (char *) "1" ; 2458 } 2459 } 2460 return NULL ; 2461 } 2462 2463 static int kvGetInt (char * kvList, const char * Key, int Default) { 2464 char * v = kvGet (kvList, Key) ; 2465 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2466 if (Knob_ReportSettings && v != NULL) { 2467 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2468 ::fflush (stdout) ; 2469 } 2470 return rslt ; 2471 } 2472 2473 void ObjectMonitor::DeferredInitialize () { 2474 if (InitDone > 0) return ; 2475 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2476 while (InitDone != 1) ; 2477 return ; 2478 } 2479 2480 // One-shot global initialization ... 2481 // The initialization is idempotent, so we don't need locks. 2482 // In the future consider doing this via os::init_2(). 2483 // SyncKnobs consist of <Key>=<Value> pairs in the style 2484 // of environment variables. Start by converting ':' to NUL. 2485 2486 if (SyncKnobs == NULL) SyncKnobs = "" ; 2487 2488 size_t sz = strlen (SyncKnobs) ; 2489 char * knobs = (char *) malloc (sz + 2) ; 2490 if (knobs == NULL) { 2491 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; 2492 guarantee (0, "invariant") ; 2493 } 2494 strcpy (knobs, SyncKnobs) ; 2495 knobs[sz+1] = 0 ; 2496 for (char * p = knobs ; *p ; p++) { 2497 if (*p == ':') *p = 0 ; 2498 } 2499 2500 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2501 SETKNOB(ReportSettings) ; 2502 SETKNOB(Verbose) ; 2503 SETKNOB(FixedSpin) ; 2504 SETKNOB(SpinLimit) ; 2505 SETKNOB(SpinBase) ; 2506 SETKNOB(SpinBackOff); 2507 SETKNOB(CASPenalty) ; 2508 SETKNOB(OXPenalty) ; 2509 SETKNOB(LogSpins) ; 2510 SETKNOB(SpinSetSucc) ; 2511 SETKNOB(SuccEnabled) ; 2512 SETKNOB(SuccRestrict) ; 2513 SETKNOB(Penalty) ; 2514 SETKNOB(Bonus) ; 2515 SETKNOB(BonusB) ; 2516 SETKNOB(Poverty) ; 2517 SETKNOB(SpinAfterFutile) ; 2518 SETKNOB(UsePause) ; 2519 SETKNOB(SpinEarly) ; 2520 SETKNOB(OState) ; 2521 SETKNOB(MaxSpinners) ; 2522 SETKNOB(PreSpin) ; 2523 SETKNOB(ExitPolicy) ; 2524 SETKNOB(QMode); 2525 SETKNOB(ResetEvent) ; 2526 SETKNOB(MoveNotifyee) ; 2527 SETKNOB(FastHSSEC) ; 2528 #undef SETKNOB 2529 2530 if (os::is_MP()) { 2531 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2532 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2533 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2534 } else { 2535 Knob_SpinLimit = 0 ; 2536 Knob_SpinBase = 0 ; 2537 Knob_PreSpin = 0 ; 2538 Knob_FixedSpin = -1 ; 2539 } 2540 2541 if (Knob_LogSpins == 0) { 2542 ObjectMonitor::_sync_FailedSpins = NULL ; 2543 } 2544 2545 free (knobs) ; 2546 OrderAccess::fence() ; 2547 InitDone = 1 ; 2548 } 2549 2550 #ifndef PRODUCT 2551 void ObjectMonitor::verify() { 2552 } 2553 2554 void ObjectMonitor::print() { 2555 } 2556 #endif