1 /* 2 * Copyright (c) 1998, 2015, 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 "jfr/jfrEvents.hpp" 28 #include "jfr/support/jfrThreadId.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "oops/markOop.hpp" 31 #include "oops/oop.inline.hpp" 32 #include "runtime/handles.inline.hpp" 33 #include "runtime/interfaceSupport.hpp" 34 #include "runtime/mutexLocker.hpp" 35 #include "runtime/objectMonitor.hpp" 36 #include "runtime/objectMonitor.inline.hpp" 37 #include "runtime/orderAccess.inline.hpp" 38 #include "runtime/osThread.hpp" 39 #include "runtime/stubRoutines.hpp" 40 #include "runtime/thread.inline.hpp" 41 #include "services/threadService.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 #if INCLUDE_JFR 58 #include "jfr/support/jfrFlush.hpp" 59 #endif 60 61 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 62 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 63 #define ATTR __attribute__((noinline)) 64 #else 65 #define ATTR 66 #endif 67 68 69 #ifdef DTRACE_ENABLED 70 71 // Only bother with this argument setup if dtrace is available 72 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 73 74 75 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 76 char* bytes = NULL; \ 77 int len = 0; \ 78 jlong jtid = SharedRuntime::get_java_tid(thread); \ 79 Symbol* klassname = ((oop)obj)->klass()->name(); \ 80 if (klassname != NULL) { \ 81 bytes = (char*)klassname->bytes(); \ 82 len = klassname->utf8_length(); \ 83 } 84 85 #ifndef USDT2 86 87 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 88 jlong, uintptr_t, char*, int); 89 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 90 jlong, uintptr_t, char*, int); 91 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 92 jlong, uintptr_t, char*, int); 93 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 94 jlong, uintptr_t, char*, int); 95 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 96 jlong, uintptr_t, char*, int); 97 98 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 99 { \ 100 if (DTraceMonitorProbes) { \ 101 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 102 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 103 (monitor), bytes, len, (millis)); \ 104 } \ 105 } 106 107 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 108 { \ 109 if (DTraceMonitorProbes) { \ 110 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 111 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 112 (uintptr_t)(monitor), bytes, len); \ 113 } \ 114 } 115 116 #else /* USDT2 */ 117 118 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 119 { \ 120 if (DTraceMonitorProbes) { \ 121 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 122 HOTSPOT_MONITOR_WAIT(jtid, \ 123 (monitor), bytes, len, (millis)); \ 124 } \ 125 } 126 127 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 128 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 129 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 130 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 131 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 132 133 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 134 { \ 135 if (DTraceMonitorProbes) { \ 136 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 137 HOTSPOT_MONITOR_##probe(jtid, \ 138 (uintptr_t)(monitor), bytes, len); \ 139 } \ 140 } 141 142 #endif /* USDT2 */ 143 #else // ndef DTRACE_ENABLED 144 145 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 146 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 147 148 #endif // ndef DTRACE_ENABLED 149 150 // Tunables ... 151 // The knob* variables are effectively final. Once set they should 152 // never be modified hence. Consider using __read_mostly with GCC. 153 154 int ObjectMonitor::Knob_Verbose = 0 ; 155 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 156 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 157 static int Knob_HandOff = 0 ; 158 static int Knob_ReportSettings = 0 ; 159 160 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 161 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 162 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 163 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 164 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 165 static int Knob_SpinEarly = 1 ; 166 static int Knob_SuccEnabled = 1 ; // futile wake throttling 167 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 168 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 169 static int Knob_Bonus = 100 ; // spin success bonus 170 static int Knob_BonusB = 100 ; // spin success bonus 171 static int Knob_Penalty = 200 ; // spin failure penalty 172 static int Knob_Poverty = 1000 ; 173 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 174 static int Knob_FixedSpin = 0 ; 175 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 176 static int Knob_UsePause = 1 ; 177 static int Knob_ExitPolicy = 0 ; 178 static int Knob_PreSpin = 10 ; // 20-100 likely better 179 static int Knob_ResetEvent = 0 ; 180 static int BackOffMask = 0 ; 181 182 static int Knob_FastHSSEC = 0 ; 183 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 184 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 185 static volatile int InitDone = 0 ; 186 187 #define TrySpin TrySpin_VaryDuration 188 189 // ----------------------------------------------------------------------------- 190 // Theory of operations -- Monitors lists, thread residency, etc: 191 // 192 // * A thread acquires ownership of a monitor by successfully 193 // CAS()ing the _owner field from null to non-null. 194 // 195 // * Invariant: A thread appears on at most one monitor list -- 196 // cxq, EntryList or WaitSet -- at any one time. 197 // 198 // * Contending threads "push" themselves onto the cxq with CAS 199 // and then spin/park. 200 // 201 // * After a contending thread eventually acquires the lock it must 202 // dequeue itself from either the EntryList or the cxq. 203 // 204 // * The exiting thread identifies and unparks an "heir presumptive" 205 // tentative successor thread on the EntryList. Critically, the 206 // exiting thread doesn't unlink the successor thread from the EntryList. 207 // After having been unparked, the wakee will recontend for ownership of 208 // the monitor. The successor (wakee) will either acquire the lock or 209 // re-park itself. 210 // 211 // Succession is provided for by a policy of competitive handoff. 212 // The exiting thread does _not_ grant or pass ownership to the 213 // successor thread. (This is also referred to as "handoff" succession"). 214 // Instead the exiting thread releases ownership and possibly wakes 215 // a successor, so the successor can (re)compete for ownership of the lock. 216 // If the EntryList is empty but the cxq is populated the exiting 217 // thread will drain the cxq into the EntryList. It does so by 218 // by detaching the cxq (installing null with CAS) and folding 219 // the threads from the cxq into the EntryList. The EntryList is 220 // doubly linked, while the cxq is singly linked because of the 221 // CAS-based "push" used to enqueue recently arrived threads (RATs). 222 // 223 // * Concurrency invariants: 224 // 225 // -- only the monitor owner may access or mutate the EntryList. 226 // The mutex property of the monitor itself protects the EntryList 227 // from concurrent interference. 228 // -- Only the monitor owner may detach the cxq. 229 // 230 // * The monitor entry list operations avoid locks, but strictly speaking 231 // they're not lock-free. Enter is lock-free, exit is not. 232 // For a description of 'Methods and apparatus providing non-blocking access 233 // to a resource,' see U.S. Pat. No. 7844973. 234 // 235 // * The cxq can have multiple concurrent "pushers" but only one concurrent 236 // detaching thread. This mechanism is immune from the ABA corruption. 237 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 238 // 239 // * Taken together, the cxq and the EntryList constitute or form a 240 // single logical queue of threads stalled trying to acquire the lock. 241 // We use two distinct lists to improve the odds of a constant-time 242 // dequeue operation after acquisition (in the ::enter() epilog) and 243 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 244 // A key desideratum is to minimize queue & monitor metadata manipulation 245 // that occurs while holding the monitor lock -- that is, we want to 246 // minimize monitor lock holds times. Note that even a small amount of 247 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 248 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 249 // locks and monitor metadata. 250 // 251 // Cxq points to the the set of Recently Arrived Threads attempting entry. 252 // Because we push threads onto _cxq with CAS, the RATs must take the form of 253 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 254 // the unlocking thread notices that EntryList is null but _cxq is != null. 255 // 256 // The EntryList is ordered by the prevailing queue discipline and 257 // can be organized in any convenient fashion, such as a doubly-linked list or 258 // a circular doubly-linked list. Critically, we want insert and delete operations 259 // to operate in constant-time. If we need a priority queue then something akin 260 // to Solaris' sleepq would work nicely. Viz., 261 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 262 // Queue discipline is enforced at ::exit() time, when the unlocking thread 263 // drains the cxq into the EntryList, and orders or reorders the threads on the 264 // EntryList accordingly. 265 // 266 // Barring "lock barging", this mechanism provides fair cyclic ordering, 267 // somewhat similar to an elevator-scan. 268 // 269 // * The monitor synchronization subsystem avoids the use of native 270 // synchronization primitives except for the narrow platform-specific 271 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 272 // the semantics of park-unpark. Put another way, this monitor implementation 273 // depends only on atomic operations and park-unpark. The monitor subsystem 274 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 275 // underlying OS manages the READY<->RUN transitions. 276 // 277 // * Waiting threads reside on the WaitSet list -- wait() puts 278 // the caller onto the WaitSet. 279 // 280 // * notify() or notifyAll() simply transfers threads from the WaitSet to 281 // either the EntryList or cxq. Subsequent exit() operations will 282 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 283 // it's likely the notifyee would simply impale itself on the lock held 284 // by the notifier. 285 // 286 // * An interesting alternative is to encode cxq as (List,LockByte) where 287 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 288 // variable, like _recursions, in the scheme. The threads or Events that form 289 // the list would have to be aligned in 256-byte addresses. A thread would 290 // try to acquire the lock or enqueue itself with CAS, but exiting threads 291 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 292 // Note that is is *not* word-tearing, but it does presume that full-word 293 // CAS operations are coherent with intermix with STB operations. That's true 294 // on most common processors. 295 // 296 // * See also http://blogs.sun.com/dave 297 298 299 // ----------------------------------------------------------------------------- 300 // Enter support 301 302 bool ObjectMonitor::try_enter(Thread* THREAD) { 303 if (THREAD != _owner) { 304 if (THREAD->is_lock_owned ((address)_owner)) { 305 assert(_recursions == 0, "internal state error"); 306 _owner = THREAD ; 307 _recursions = 1 ; 308 OwnerIsThread = 1 ; 309 return true; 310 } 311 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 312 return false; 313 } 314 return true; 315 } else { 316 _recursions++; 317 return true; 318 } 319 } 320 321 void ATTR ObjectMonitor::enter(TRAPS) { 322 // The following code is ordered to check the most common cases first 323 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 324 Thread * const Self = THREAD ; 325 void * cur ; 326 327 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 328 if (cur == NULL) { 329 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 330 assert (_recursions == 0 , "invariant") ; 331 assert (_owner == Self, "invariant") ; 332 // CONSIDER: set or assert OwnerIsThread == 1 333 return ; 334 } 335 336 if (cur == Self) { 337 // TODO-FIXME: check for integer overflow! BUGID 6557169. 338 _recursions ++ ; 339 return ; 340 } 341 342 if (Self->is_lock_owned ((address)cur)) { 343 assert (_recursions == 0, "internal state error"); 344 _recursions = 1 ; 345 // Commute owner from a thread-specific on-stack BasicLockObject address to 346 // a full-fledged "Thread *". 347 _owner = Self ; 348 OwnerIsThread = 1 ; 349 return ; 350 } 351 352 // We've encountered genuine contention. 353 assert (Self->_Stalled == 0, "invariant") ; 354 Self->_Stalled = intptr_t(this) ; 355 356 // Try one round of spinning *before* enqueueing Self 357 // and before going through the awkward and expensive state 358 // transitions. The following spin is strictly optional ... 359 // Note that if we acquire the monitor from an initial spin 360 // we forgo posting JVMTI events and firing DTRACE probes. 361 if (Knob_SpinEarly && TrySpin (Self) > 0) { 362 assert (_owner == Self , "invariant") ; 363 assert (_recursions == 0 , "invariant") ; 364 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 365 Self->_Stalled = 0 ; 366 return ; 367 } 368 369 assert (_owner != Self , "invariant") ; 370 assert (_succ != Self , "invariant") ; 371 assert (Self->is_Java_thread() , "invariant") ; 372 JavaThread * jt = (JavaThread *) Self ; 373 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 374 assert (jt->thread_state() != _thread_blocked , "invariant") ; 375 assert (this->object() != NULL , "invariant") ; 376 assert (_count >= 0, "invariant") ; 377 378 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 379 // Ensure the object-monitor relationship remains stable while there's contention. 380 Atomic::inc_ptr(&_count); 381 382 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 383 EventJavaMonitorEnter event; 384 if (event.should_commit()) { 385 event.set_monitorClass(((oop)this->object())->klass()); 386 event.set_address((uintptr_t)(this->object_addr())); 387 } 388 389 { // Change java thread status to indicate blocked on monitor enter. 390 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 391 392 Self->set_current_pending_monitor(this); 393 394 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 395 if (JvmtiExport::should_post_monitor_contended_enter()) { 396 JvmtiExport::post_monitor_contended_enter(jt, this); 397 398 // The current thread does not yet own the monitor and does not 399 // yet appear on any queues that would get it made the successor. 400 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 401 // handler cannot accidentally consume an unpark() meant for the 402 // ParkEvent associated with this ObjectMonitor. 403 } 404 405 OSThreadContendState osts(Self->osthread()); 406 ThreadBlockInVM tbivm(jt); 407 408 // TODO-FIXME: change the following for(;;) loop to straight-line code. 409 for (;;) { 410 jt->set_suspend_equivalent(); 411 // cleared by handle_special_suspend_equivalent_condition() 412 // or java_suspend_self() 413 414 EnterI (THREAD) ; 415 416 if (!ExitSuspendEquivalent(jt)) break ; 417 418 // 419 // We have acquired the contended monitor, but while we were 420 // waiting another thread suspended us. We don't want to enter 421 // the monitor while suspended because that would surprise the 422 // thread that suspended us. 423 // 424 _recursions = 0 ; 425 _succ = NULL ; 426 exit (false, Self) ; 427 428 jt->java_suspend_self(); 429 } 430 Self->set_current_pending_monitor(NULL); 431 432 // We cleared the pending monitor info since we've just gotten past 433 // the enter-check-for-suspend dance and we now own the monitor free 434 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 435 // destructor can go to a safepoint at the end of this block. If we 436 // do a thread dump during that safepoint, then this thread will show 437 // as having "-locked" the monitor, but the OS and java.lang.Thread 438 // states will still report that the thread is blocked trying to 439 // acquire it. 440 } 441 442 Atomic::dec_ptr(&_count); 443 assert (_count >= 0, "invariant") ; 444 Self->_Stalled = 0 ; 445 446 // Must either set _recursions = 0 or ASSERT _recursions == 0. 447 assert (_recursions == 0 , "invariant") ; 448 assert (_owner == Self , "invariant") ; 449 assert (_succ != Self , "invariant") ; 450 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 451 452 // The thread -- now the owner -- is back in vm mode. 453 // Report the glorious news via TI,DTrace and jvmstat. 454 // The probe effect is non-trivial. All the reportage occurs 455 // while we hold the monitor, increasing the length of the critical 456 // section. Amdahl's parallel speedup law comes vividly into play. 457 // 458 // Another option might be to aggregate the events (thread local or 459 // per-monitor aggregation) and defer reporting until a more opportune 460 // time -- such as next time some thread encounters contention but has 461 // yet to acquire the lock. While spinning that thread could 462 // spinning we could increment JVMStat counters, etc. 463 464 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 465 if (JvmtiExport::should_post_monitor_contended_entered()) { 466 JvmtiExport::post_monitor_contended_entered(jt, this); 467 468 // The current thread already owns the monitor and is not going to 469 // call park() for the remainder of the monitor enter protocol. So 470 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 471 // event handler consumed an unpark() issued by the thread that 472 // just exited the monitor. 473 } 474 475 if (event.should_commit()) { 476 event.set_previousOwner((uintptr_t)_previous_owner_tid); 477 event.commit(); 478 } 479 480 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 481 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 482 } 483 } 484 485 486 // Caveat: TryLock() is not necessarily serializing if it returns failure. 487 // Callers must compensate as needed. 488 489 int ObjectMonitor::TryLock (Thread * Self) { 490 for (;;) { 491 void * own = _owner ; 492 if (own != NULL) return 0 ; 493 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 494 // Either guarantee _recursions == 0 or set _recursions = 0. 495 assert (_recursions == 0, "invariant") ; 496 assert (_owner == Self, "invariant") ; 497 // CONSIDER: set or assert that OwnerIsThread == 1 498 return 1 ; 499 } 500 // The lock had been free momentarily, but we lost the race to the lock. 501 // Interference -- the CAS failed. 502 // We can either return -1 or retry. 503 // Retry doesn't make as much sense because the lock was just acquired. 504 if (true) return -1 ; 505 } 506 } 507 508 void ATTR ObjectMonitor::EnterI (TRAPS) { 509 Thread * Self = THREAD ; 510 assert (Self->is_Java_thread(), "invariant") ; 511 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 512 513 // Try the lock - TATAS 514 if (TryLock (Self) > 0) { 515 assert (_succ != Self , "invariant") ; 516 assert (_owner == Self , "invariant") ; 517 assert (_Responsible != Self , "invariant") ; 518 return ; 519 } 520 521 DeferredInitialize () ; 522 523 // We try one round of spinning *before* enqueueing Self. 524 // 525 // If the _owner is ready but OFFPROC we could use a YieldTo() 526 // operation to donate the remainder of this thread's quantum 527 // to the owner. This has subtle but beneficial affinity 528 // effects. 529 530 if (TrySpin (Self) > 0) { 531 assert (_owner == Self , "invariant") ; 532 assert (_succ != Self , "invariant") ; 533 assert (_Responsible != Self , "invariant") ; 534 return ; 535 } 536 537 // The Spin failed -- Enqueue and park the thread ... 538 assert (_succ != Self , "invariant") ; 539 assert (_owner != Self , "invariant") ; 540 assert (_Responsible != Self , "invariant") ; 541 542 // Enqueue "Self" on ObjectMonitor's _cxq. 543 // 544 // Node acts as a proxy for Self. 545 // As an aside, if were to ever rewrite the synchronization code mostly 546 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 547 // Java objects. This would avoid awkward lifecycle and liveness issues, 548 // as well as eliminate a subset of ABA issues. 549 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 550 // 551 552 ObjectWaiter node(Self) ; 553 Self->_ParkEvent->reset() ; 554 node._prev = (ObjectWaiter *) 0xBAD ; 555 node.TState = ObjectWaiter::TS_CXQ ; 556 557 // Push "Self" onto the front of the _cxq. 558 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 559 // Note that spinning tends to reduce the rate at which threads 560 // enqueue and dequeue on EntryList|cxq. 561 ObjectWaiter * nxt ; 562 for (;;) { 563 node._next = nxt = _cxq ; 564 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 565 566 // Interference - the CAS failed because _cxq changed. Just retry. 567 // As an optional optimization we retry the lock. 568 if (TryLock (Self) > 0) { 569 assert (_succ != Self , "invariant") ; 570 assert (_owner == Self , "invariant") ; 571 assert (_Responsible != Self , "invariant") ; 572 return ; 573 } 574 } 575 576 // Check for cxq|EntryList edge transition to non-null. This indicates 577 // the onset of contention. While contention persists exiting threads 578 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 579 // operations revert to the faster 1-0 mode. This enter operation may interleave 580 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 581 // arrange for one of the contending thread to use a timed park() operations 582 // to detect and recover from the race. (Stranding is form of progress failure 583 // where the monitor is unlocked but all the contending threads remain parked). 584 // That is, at least one of the contended threads will periodically poll _owner. 585 // One of the contending threads will become the designated "Responsible" thread. 586 // The Responsible thread uses a timed park instead of a normal indefinite park 587 // operation -- it periodically wakes and checks for and recovers from potential 588 // strandings admitted by 1-0 exit operations. We need at most one Responsible 589 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 590 // be responsible for a monitor. 591 // 592 // Currently, one of the contended threads takes on the added role of "Responsible". 593 // A viable alternative would be to use a dedicated "stranding checker" thread 594 // that periodically iterated over all the threads (or active monitors) and unparked 595 // successors where there was risk of stranding. This would help eliminate the 596 // timer scalability issues we see on some platforms as we'd only have one thread 597 // -- the checker -- parked on a timer. 598 599 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 600 // Try to assume the role of responsible thread for the monitor. 601 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 602 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 603 } 604 605 // The lock have been released while this thread was occupied queueing 606 // itself onto _cxq. To close the race and avoid "stranding" and 607 // progress-liveness failure we must resample-retry _owner before parking. 608 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 609 // In this case the ST-MEMBAR is accomplished with CAS(). 610 // 611 // TODO: Defer all thread state transitions until park-time. 612 // Since state transitions are heavy and inefficient we'd like 613 // to defer the state transitions until absolutely necessary, 614 // and in doing so avoid some transitions ... 615 616 TEVENT (Inflated enter - Contention) ; 617 int nWakeups = 0 ; 618 int RecheckInterval = 1 ; 619 620 for (;;) { 621 622 if (TryLock (Self) > 0) break ; 623 assert (_owner != Self, "invariant") ; 624 625 if ((SyncFlags & 2) && _Responsible == NULL) { 626 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 627 } 628 629 // park self 630 if (_Responsible == Self || (SyncFlags & 1)) { 631 TEVENT (Inflated enter - park TIMED) ; 632 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 633 // Increase the RecheckInterval, but clamp the value. 634 RecheckInterval *= 8 ; 635 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 636 } else { 637 TEVENT (Inflated enter - park UNTIMED) ; 638 Self->_ParkEvent->park() ; 639 } 640 641 if (TryLock(Self) > 0) break ; 642 643 // The lock is still contested. 644 // Keep a tally of the # of futile wakeups. 645 // Note that the counter is not protected by a lock or updated by atomics. 646 // That is by design - we trade "lossy" counters which are exposed to 647 // races during updates for a lower probe effect. 648 TEVENT (Inflated enter - Futile wakeup) ; 649 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 650 ObjectMonitor::_sync_FutileWakeups->inc() ; 651 } 652 ++ nWakeups ; 653 654 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 655 // We can defer clearing _succ until after the spin completes 656 // TrySpin() must tolerate being called with _succ == Self. 657 // Try yet another round of adaptive spinning. 658 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 659 660 // We can find that we were unpark()ed and redesignated _succ while 661 // we were spinning. That's harmless. If we iterate and call park(), 662 // park() will consume the event and return immediately and we'll 663 // just spin again. This pattern can repeat, leaving _succ to simply 664 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 665 // Alternately, we can sample fired() here, and if set, forgo spinning 666 // in the next iteration. 667 668 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 669 Self->_ParkEvent->reset() ; 670 OrderAccess::fence() ; 671 } 672 if (_succ == Self) _succ = NULL ; 673 674 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 675 OrderAccess::fence() ; 676 } 677 678 // Egress : 679 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 680 // Normally we'll find Self on the EntryList . 681 // From the perspective of the lock owner (this thread), the 682 // EntryList is stable and cxq is prepend-only. 683 // The head of cxq is volatile but the interior is stable. 684 // In addition, Self.TState is stable. 685 686 assert (_owner == Self , "invariant") ; 687 assert (object() != NULL , "invariant") ; 688 // I'd like to write: 689 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 690 // but as we're at a safepoint that's not safe. 691 692 UnlinkAfterAcquire (Self, &node) ; 693 if (_succ == Self) _succ = NULL ; 694 695 assert (_succ != Self, "invariant") ; 696 if (_Responsible == Self) { 697 _Responsible = NULL ; 698 OrderAccess::fence(); // Dekker pivot-point 699 700 // We may leave threads on cxq|EntryList without a designated 701 // "Responsible" thread. This is benign. When this thread subsequently 702 // exits the monitor it can "see" such preexisting "old" threads -- 703 // threads that arrived on the cxq|EntryList before the fence, above -- 704 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 705 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 706 // non-null and elect a new "Responsible" timer thread. 707 // 708 // This thread executes: 709 // ST Responsible=null; MEMBAR (in enter epilog - here) 710 // LD cxq|EntryList (in subsequent exit) 711 // 712 // Entering threads in the slow/contended path execute: 713 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 714 // The (ST cxq; MEMBAR) is accomplished with CAS(). 715 // 716 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 717 // exit operation from floating above the ST Responsible=null. 718 } 719 720 // We've acquired ownership with CAS(). 721 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 722 // But since the CAS() this thread may have also stored into _succ, 723 // EntryList, cxq or Responsible. These meta-data updates must be 724 // visible __before this thread subsequently drops the lock. 725 // Consider what could occur if we didn't enforce this constraint -- 726 // STs to monitor meta-data and user-data could reorder with (become 727 // visible after) the ST in exit that drops ownership of the lock. 728 // Some other thread could then acquire the lock, but observe inconsistent 729 // or old monitor meta-data and heap data. That violates the JMM. 730 // To that end, the 1-0 exit() operation must have at least STST|LDST 731 // "release" barrier semantics. Specifically, there must be at least a 732 // STST|LDST barrier in exit() before the ST of null into _owner that drops 733 // the lock. The barrier ensures that changes to monitor meta-data and data 734 // protected by the lock will be visible before we release the lock, and 735 // therefore before some other thread (CPU) has a chance to acquire the lock. 736 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 737 // 738 // Critically, any prior STs to _succ or EntryList must be visible before 739 // the ST of null into _owner in the *subsequent* (following) corresponding 740 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 741 // execute a serializing instruction. 742 743 if (SyncFlags & 8) { 744 OrderAccess::fence() ; 745 } 746 return ; 747 } 748 749 // ReenterI() is a specialized inline form of the latter half of the 750 // contended slow-path from EnterI(). We use ReenterI() only for 751 // monitor reentry in wait(). 752 // 753 // In the future we should reconcile EnterI() and ReenterI(), adding 754 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 755 // loop accordingly. 756 757 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 758 assert (Self != NULL , "invariant") ; 759 assert (SelfNode != NULL , "invariant") ; 760 assert (SelfNode->_thread == Self , "invariant") ; 761 assert (_waiters > 0 , "invariant") ; 762 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 763 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 764 JavaThread * jt = (JavaThread *) Self ; 765 766 int nWakeups = 0 ; 767 for (;;) { 768 ObjectWaiter::TStates v = SelfNode->TState ; 769 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 770 assert (_owner != Self, "invariant") ; 771 772 if (TryLock (Self) > 0) break ; 773 if (TrySpin (Self) > 0) break ; 774 775 TEVENT (Wait Reentry - parking) ; 776 777 // State transition wrappers around park() ... 778 // ReenterI() wisely defers state transitions until 779 // it's clear we must park the thread. 780 { 781 OSThreadContendState osts(Self->osthread()); 782 ThreadBlockInVM tbivm(jt); 783 784 // cleared by handle_special_suspend_equivalent_condition() 785 // or java_suspend_self() 786 jt->set_suspend_equivalent(); 787 if (SyncFlags & 1) { 788 Self->_ParkEvent->park ((jlong)1000) ; 789 } else { 790 Self->_ParkEvent->park () ; 791 } 792 793 // were we externally suspended while we were waiting? 794 for (;;) { 795 if (!ExitSuspendEquivalent (jt)) break ; 796 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 797 jt->java_suspend_self(); 798 jt->set_suspend_equivalent(); 799 } 800 } 801 802 // Try again, but just so we distinguish between futile wakeups and 803 // successful wakeups. The following test isn't algorithmically 804 // necessary, but it helps us maintain sensible statistics. 805 if (TryLock(Self) > 0) break ; 806 807 // The lock is still contested. 808 // Keep a tally of the # of futile wakeups. 809 // Note that the counter is not protected by a lock or updated by atomics. 810 // That is by design - we trade "lossy" counters which are exposed to 811 // races during updates for a lower probe effect. 812 TEVENT (Wait Reentry - futile wakeup) ; 813 ++ nWakeups ; 814 815 // Assuming this is not a spurious wakeup we'll normally 816 // find that _succ == Self. 817 if (_succ == Self) _succ = NULL ; 818 819 // Invariant: after clearing _succ a contending thread 820 // *must* retry _owner before parking. 821 OrderAccess::fence() ; 822 823 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 824 ObjectMonitor::_sync_FutileWakeups->inc() ; 825 } 826 } 827 828 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 829 // Normally we'll find Self on the EntryList. 830 // Unlinking from the EntryList is constant-time and atomic-free. 831 // From the perspective of the lock owner (this thread), the 832 // EntryList is stable and cxq is prepend-only. 833 // The head of cxq is volatile but the interior is stable. 834 // In addition, Self.TState is stable. 835 836 assert (_owner == Self, "invariant") ; 837 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 838 UnlinkAfterAcquire (Self, SelfNode) ; 839 if (_succ == Self) _succ = NULL ; 840 assert (_succ != Self, "invariant") ; 841 SelfNode->TState = ObjectWaiter::TS_RUN ; 842 OrderAccess::fence() ; // see comments at the end of EnterI() 843 } 844 845 // after the thread acquires the lock in ::enter(). Equally, we could defer 846 // unlinking the thread until ::exit()-time. 847 848 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 849 { 850 assert (_owner == Self, "invariant") ; 851 assert (SelfNode->_thread == Self, "invariant") ; 852 853 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 854 // Normal case: remove Self from the DLL EntryList . 855 // This is a constant-time operation. 856 ObjectWaiter * nxt = SelfNode->_next ; 857 ObjectWaiter * prv = SelfNode->_prev ; 858 if (nxt != NULL) nxt->_prev = prv ; 859 if (prv != NULL) prv->_next = nxt ; 860 if (SelfNode == _EntryList ) _EntryList = nxt ; 861 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 862 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 863 TEVENT (Unlink from EntryList) ; 864 } else { 865 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 866 // Inopportune interleaving -- Self is still on the cxq. 867 // This usually means the enqueue of self raced an exiting thread. 868 // Normally we'll find Self near the front of the cxq, so 869 // dequeueing is typically fast. If needbe we can accelerate 870 // this with some MCS/CHL-like bidirectional list hints and advisory 871 // back-links so dequeueing from the interior will normally operate 872 // in constant-time. 873 // Dequeue Self from either the head (with CAS) or from the interior 874 // with a linear-time scan and normal non-atomic memory operations. 875 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 876 // and then unlink Self from EntryList. We have to drain eventually, 877 // so it might as well be now. 878 879 ObjectWaiter * v = _cxq ; 880 assert (v != NULL, "invariant") ; 881 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 882 // The CAS above can fail from interference IFF a "RAT" arrived. 883 // In that case Self must be in the interior and can no longer be 884 // at the head of cxq. 885 if (v == SelfNode) { 886 assert (_cxq != v, "invariant") ; 887 v = _cxq ; // CAS above failed - start scan at head of list 888 } 889 ObjectWaiter * p ; 890 ObjectWaiter * q = NULL ; 891 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 892 q = p ; 893 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 894 } 895 assert (v != SelfNode, "invariant") ; 896 assert (p == SelfNode, "Node not found on cxq") ; 897 assert (p != _cxq, "invariant") ; 898 assert (q != NULL, "invariant") ; 899 assert (q->_next == p, "invariant") ; 900 q->_next = p->_next ; 901 } 902 TEVENT (Unlink from cxq) ; 903 } 904 905 // Diagnostic hygiene ... 906 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 907 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 908 SelfNode->TState = ObjectWaiter::TS_RUN ; 909 } 910 911 // ----------------------------------------------------------------------------- 912 // Exit support 913 // 914 // exit() 915 // ~~~~~~ 916 // Note that the collector can't reclaim the objectMonitor or deflate 917 // the object out from underneath the thread calling ::exit() as the 918 // thread calling ::exit() never transitions to a stable state. 919 // This inhibits GC, which in turn inhibits asynchronous (and 920 // inopportune) reclamation of "this". 921 // 922 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 923 // There's one exception to the claim above, however. EnterI() can call 924 // exit() to drop a lock if the acquirer has been externally suspended. 925 // In that case exit() is called with _thread_state as _thread_blocked, 926 // but the monitor's _count field is > 0, which inhibits reclamation. 927 // 928 // 1-0 exit 929 // ~~~~~~~~ 930 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 931 // the fast-path operators have been optimized so the common ::exit() 932 // operation is 1-0. See i486.ad fast_unlock(), for instance. 933 // The code emitted by fast_unlock() elides the usual MEMBAR. This 934 // greatly improves latency -- MEMBAR and CAS having considerable local 935 // latency on modern processors -- but at the cost of "stranding". Absent the 936 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 937 // ::enter() path, resulting in the entering thread being stranding 938 // and a progress-liveness failure. Stranding is extremely rare. 939 // We use timers (timed park operations) & periodic polling to detect 940 // and recover from stranding. Potentially stranded threads periodically 941 // wake up and poll the lock. See the usage of the _Responsible variable. 942 // 943 // The CAS() in enter provides for safety and exclusion, while the CAS or 944 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 945 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 946 // We detect and recover from stranding with timers. 947 // 948 // If a thread transiently strands it'll park until (a) another 949 // thread acquires the lock and then drops the lock, at which time the 950 // exiting thread will notice and unpark the stranded thread, or, (b) 951 // the timer expires. If the lock is high traffic then the stranding latency 952 // will be low due to (a). If the lock is low traffic then the odds of 953 // stranding are lower, although the worst-case stranding latency 954 // is longer. Critically, we don't want to put excessive load in the 955 // platform's timer subsystem. We want to minimize both the timer injection 956 // rate (timers created/sec) as well as the number of timers active at 957 // any one time. (more precisely, we want to minimize timer-seconds, which is 958 // the integral of the # of active timers at any instant over time). 959 // Both impinge on OS scalability. Given that, at most one thread parked on 960 // a monitor will use a timer. 961 962 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 963 Thread * Self = THREAD ; 964 if (THREAD != _owner) { 965 if (THREAD->is_lock_owned((address) _owner)) { 966 // Transmute _owner from a BasicLock pointer to a Thread address. 967 // We don't need to hold _mutex for this transition. 968 // Non-null to Non-null is safe as long as all readers can 969 // tolerate either flavor. 970 assert (_recursions == 0, "invariant") ; 971 _owner = THREAD ; 972 _recursions = 0 ; 973 OwnerIsThread = 1 ; 974 } else { 975 // NOTE: we need to handle unbalanced monitor enter/exit 976 // in native code by throwing an exception. 977 // TODO: Throw an IllegalMonitorStateException ? 978 TEVENT (Exit - Throw IMSX) ; 979 assert(false, "Non-balanced monitor enter/exit!"); 980 if (false) { 981 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 982 } 983 return; 984 } 985 } 986 987 if (_recursions != 0) { 988 _recursions--; // this is simple recursive enter 989 TEVENT (Inflated exit - recursive) ; 990 return ; 991 } 992 993 // Invariant: after setting Responsible=null an thread must execute 994 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 995 if ((SyncFlags & 4) == 0) { 996 _Responsible = NULL ; 997 } 998 999 #if INCLUDE_JFR 1000 // get the owner's thread id for the MonitorEnter event 1001 // if it is enabled and the thread isn't suspended 1002 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1003 _previous_owner_tid = JFR_THREAD_ID(Self); 1004 } 1005 #endif 1006 1007 for (;;) { 1008 assert (THREAD == _owner, "invariant") ; 1009 1010 1011 if (Knob_ExitPolicy == 0) { 1012 // release semantics: prior loads and stores from within the critical section 1013 // must not float (reorder) past the following store that drops the lock. 1014 // On SPARC that requires MEMBAR #loadstore|#storestore. 1015 // But of course in TSO #loadstore|#storestore is not required. 1016 // I'd like to write one of the following: 1017 // A. OrderAccess::release() ; _owner = NULL 1018 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 1019 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 1020 // store into a _dummy variable. That store is not needed, but can result 1021 // in massive wasteful coherency traffic on classic SMP systems. 1022 // Instead, I use release_store(), which is implemented as just a simple 1023 // ST on x64, x86 and SPARC. 1024 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1025 OrderAccess::storeload() ; // See if we need to wake a successor 1026 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1027 TEVENT (Inflated exit - simple egress) ; 1028 return ; 1029 } 1030 TEVENT (Inflated exit - complex egress) ; 1031 1032 // Normally the exiting thread is responsible for ensuring succession, 1033 // but if other successors are ready or other entering threads are spinning 1034 // then this thread can simply store NULL into _owner and exit without 1035 // waking a successor. The existence of spinners or ready successors 1036 // guarantees proper succession (liveness). Responsibility passes to the 1037 // ready or running successors. The exiting thread delegates the duty. 1038 // More precisely, if a successor already exists this thread is absolved 1039 // of the responsibility of waking (unparking) one. 1040 // 1041 // The _succ variable is critical to reducing futile wakeup frequency. 1042 // _succ identifies the "heir presumptive" thread that has been made 1043 // ready (unparked) but that has not yet run. We need only one such 1044 // successor thread to guarantee progress. 1045 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1046 // section 3.3 "Futile Wakeup Throttling" for details. 1047 // 1048 // Note that spinners in Enter() also set _succ non-null. 1049 // In the current implementation spinners opportunistically set 1050 // _succ so that exiting threads might avoid waking a successor. 1051 // Another less appealing alternative would be for the exiting thread 1052 // to drop the lock and then spin briefly to see if a spinner managed 1053 // to acquire the lock. If so, the exiting thread could exit 1054 // immediately without waking a successor, otherwise the exiting 1055 // thread would need to dequeue and wake a successor. 1056 // (Note that we'd need to make the post-drop spin short, but no 1057 // shorter than the worst-case round-trip cache-line migration time. 1058 // The dropped lock needs to become visible to the spinner, and then 1059 // the acquisition of the lock by the spinner must become visible to 1060 // the exiting thread). 1061 // 1062 1063 // It appears that an heir-presumptive (successor) must be made ready. 1064 // Only the current lock owner can manipulate the EntryList or 1065 // drain _cxq, so we need to reacquire the lock. If we fail 1066 // to reacquire the lock the responsibility for ensuring succession 1067 // falls to the new owner. 1068 // 1069 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1070 return ; 1071 } 1072 TEVENT (Exit - Reacquired) ; 1073 } else { 1074 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1075 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1076 OrderAccess::storeload() ; 1077 // Ratify the previously observed values. 1078 if (_cxq == NULL || _succ != NULL) { 1079 TEVENT (Inflated exit - simple egress) ; 1080 return ; 1081 } 1082 1083 // inopportune interleaving -- the exiting thread (this thread) 1084 // in the fast-exit path raced an entering thread in the slow-enter 1085 // path. 1086 // We have two choices: 1087 // A. Try to reacquire the lock. 1088 // If the CAS() fails return immediately, otherwise 1089 // we either restart/rerun the exit operation, or simply 1090 // fall-through into the code below which wakes a successor. 1091 // B. If the elements forming the EntryList|cxq are TSM 1092 // we could simply unpark() the lead thread and return 1093 // without having set _succ. 1094 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1095 TEVENT (Inflated exit - reacquired succeeded) ; 1096 return ; 1097 } 1098 TEVENT (Inflated exit - reacquired failed) ; 1099 } else { 1100 TEVENT (Inflated exit - complex egress) ; 1101 } 1102 } 1103 1104 guarantee (_owner == THREAD, "invariant") ; 1105 1106 ObjectWaiter * w = NULL ; 1107 int QMode = Knob_QMode ; 1108 1109 if (QMode == 2 && _cxq != NULL) { 1110 // QMode == 2 : cxq has precedence over EntryList. 1111 // Try to directly wake a successor from the cxq. 1112 // If successful, the successor will need to unlink itself from cxq. 1113 w = _cxq ; 1114 assert (w != NULL, "invariant") ; 1115 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1116 ExitEpilog (Self, w) ; 1117 return ; 1118 } 1119 1120 if (QMode == 3 && _cxq != NULL) { 1121 // Aggressively drain cxq into EntryList at the first opportunity. 1122 // This policy ensure that recently-run threads live at the head of EntryList. 1123 // Drain _cxq into EntryList - bulk transfer. 1124 // First, detach _cxq. 1125 // The following loop is tantamount to: w = swap (&cxq, NULL) 1126 w = _cxq ; 1127 for (;;) { 1128 assert (w != NULL, "Invariant") ; 1129 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1130 if (u == w) break ; 1131 w = u ; 1132 } 1133 assert (w != NULL , "invariant") ; 1134 1135 ObjectWaiter * q = NULL ; 1136 ObjectWaiter * p ; 1137 for (p = w ; p != NULL ; p = p->_next) { 1138 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1139 p->TState = ObjectWaiter::TS_ENTER ; 1140 p->_prev = q ; 1141 q = p ; 1142 } 1143 1144 // Append the RATs to the EntryList 1145 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1146 ObjectWaiter * Tail ; 1147 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1148 if (Tail == NULL) { 1149 _EntryList = w ; 1150 } else { 1151 Tail->_next = w ; 1152 w->_prev = Tail ; 1153 } 1154 1155 // Fall thru into code that tries to wake a successor from EntryList 1156 } 1157 1158 if (QMode == 4 && _cxq != NULL) { 1159 // Aggressively drain cxq into EntryList at the first opportunity. 1160 // This policy ensure that recently-run threads live at the head of EntryList. 1161 1162 // Drain _cxq into EntryList - bulk transfer. 1163 // First, detach _cxq. 1164 // The following loop is tantamount to: w = swap (&cxq, NULL) 1165 w = _cxq ; 1166 for (;;) { 1167 assert (w != NULL, "Invariant") ; 1168 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1169 if (u == w) break ; 1170 w = u ; 1171 } 1172 assert (w != NULL , "invariant") ; 1173 1174 ObjectWaiter * q = NULL ; 1175 ObjectWaiter * p ; 1176 for (p = w ; p != NULL ; p = p->_next) { 1177 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1178 p->TState = ObjectWaiter::TS_ENTER ; 1179 p->_prev = q ; 1180 q = p ; 1181 } 1182 1183 // Prepend the RATs to the EntryList 1184 if (_EntryList != NULL) { 1185 q->_next = _EntryList ; 1186 _EntryList->_prev = q ; 1187 } 1188 _EntryList = w ; 1189 1190 // Fall thru into code that tries to wake a successor from EntryList 1191 } 1192 1193 w = _EntryList ; 1194 if (w != NULL) { 1195 // I'd like to write: guarantee (w->_thread != Self). 1196 // But in practice an exiting thread may find itself on the EntryList. 1197 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1198 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1199 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1200 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1201 // release the lock "O". T2 resumes immediately after the ST of null into 1202 // _owner, above. T2 notices that the EntryList is populated, so it 1203 // reacquires the lock and then finds itself on the EntryList. 1204 // Given all that, we have to tolerate the circumstance where "w" is 1205 // associated with Self. 1206 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1207 ExitEpilog (Self, w) ; 1208 return ; 1209 } 1210 1211 // If we find that both _cxq and EntryList are null then just 1212 // re-run the exit protocol from the top. 1213 w = _cxq ; 1214 if (w == NULL) continue ; 1215 1216 // Drain _cxq into EntryList - bulk transfer. 1217 // First, detach _cxq. 1218 // The following loop is tantamount to: w = swap (&cxq, NULL) 1219 for (;;) { 1220 assert (w != NULL, "Invariant") ; 1221 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1222 if (u == w) break ; 1223 w = u ; 1224 } 1225 TEVENT (Inflated exit - drain cxq into EntryList) ; 1226 1227 assert (w != NULL , "invariant") ; 1228 assert (_EntryList == NULL , "invariant") ; 1229 1230 // Convert the LIFO SLL anchored by _cxq into a DLL. 1231 // The list reorganization step operates in O(LENGTH(w)) time. 1232 // It's critical that this step operate quickly as 1233 // "Self" still holds the outer-lock, restricting parallelism 1234 // and effectively lengthening the critical section. 1235 // Invariant: s chases t chases u. 1236 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1237 // we have faster access to the tail. 1238 1239 if (QMode == 1) { 1240 // QMode == 1 : drain cxq to EntryList, reversing order 1241 // We also reverse the order of the list. 1242 ObjectWaiter * s = NULL ; 1243 ObjectWaiter * t = w ; 1244 ObjectWaiter * u = NULL ; 1245 while (t != NULL) { 1246 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1247 t->TState = ObjectWaiter::TS_ENTER ; 1248 u = t->_next ; 1249 t->_prev = u ; 1250 t->_next = s ; 1251 s = t; 1252 t = u ; 1253 } 1254 _EntryList = s ; 1255 assert (s != NULL, "invariant") ; 1256 } else { 1257 // QMode == 0 or QMode == 2 1258 _EntryList = w ; 1259 ObjectWaiter * q = NULL ; 1260 ObjectWaiter * p ; 1261 for (p = w ; p != NULL ; p = p->_next) { 1262 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1263 p->TState = ObjectWaiter::TS_ENTER ; 1264 p->_prev = q ; 1265 q = p ; 1266 } 1267 } 1268 1269 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1270 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1271 1272 // See if we can abdicate to a spinner instead of waking a thread. 1273 // A primary goal of the implementation is to reduce the 1274 // context-switch rate. 1275 if (_succ != NULL) continue; 1276 1277 w = _EntryList ; 1278 if (w != NULL) { 1279 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1280 ExitEpilog (Self, w) ; 1281 return ; 1282 } 1283 } 1284 } 1285 1286 // ExitSuspendEquivalent: 1287 // A faster alternate to handle_special_suspend_equivalent_condition() 1288 // 1289 // handle_special_suspend_equivalent_condition() unconditionally 1290 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1291 // operations have high latency. Note that in ::enter() we call HSSEC 1292 // while holding the monitor, so we effectively lengthen the critical sections. 1293 // 1294 // There are a number of possible solutions: 1295 // 1296 // A. To ameliorate the problem we might also defer state transitions 1297 // to as late as possible -- just prior to parking. 1298 // Given that, we'd call HSSEC after having returned from park(), 1299 // but before attempting to acquire the monitor. This is only a 1300 // partial solution. It avoids calling HSSEC while holding the 1301 // monitor (good), but it still increases successor reacquisition latency -- 1302 // the interval between unparking a successor and the time the successor 1303 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1304 // If we use this technique we can also avoid EnterI()-exit() loop 1305 // in ::enter() where we iteratively drop the lock and then attempt 1306 // to reacquire it after suspending. 1307 // 1308 // B. In the future we might fold all the suspend bits into a 1309 // composite per-thread suspend flag and then update it with CAS(). 1310 // Alternately, a Dekker-like mechanism with multiple variables 1311 // would suffice: 1312 // ST Self->_suspend_equivalent = false 1313 // MEMBAR 1314 // LD Self_>_suspend_flags 1315 // 1316 1317 1318 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1319 int Mode = Knob_FastHSSEC ; 1320 if (Mode && !jSelf->is_external_suspend()) { 1321 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1322 jSelf->clear_suspend_equivalent() ; 1323 if (2 == Mode) OrderAccess::storeload() ; 1324 if (!jSelf->is_external_suspend()) return false ; 1325 // We raced a suspension -- fall thru into the slow path 1326 TEVENT (ExitSuspendEquivalent - raced) ; 1327 jSelf->set_suspend_equivalent() ; 1328 } 1329 return jSelf->handle_special_suspend_equivalent_condition() ; 1330 } 1331 1332 1333 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1334 assert (_owner == Self, "invariant") ; 1335 1336 // Exit protocol: 1337 // 1. ST _succ = wakee 1338 // 2. membar #loadstore|#storestore; 1339 // 2. ST _owner = NULL 1340 // 3. unpark(wakee) 1341 1342 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1343 ParkEvent * Trigger = Wakee->_event ; 1344 1345 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1346 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1347 // out-of-scope (non-extant). 1348 Wakee = NULL ; 1349 1350 // Drop the lock 1351 OrderAccess::release_store_ptr (&_owner, NULL) ; 1352 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1353 1354 if (SafepointSynchronize::do_call_back()) { 1355 TEVENT (unpark before SAFEPOINT) ; 1356 } 1357 1358 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1359 Trigger->unpark() ; 1360 1361 // Maintain stats and report events to JVMTI 1362 if (ObjectMonitor::_sync_Parks != NULL) { 1363 ObjectMonitor::_sync_Parks->inc() ; 1364 } 1365 } 1366 1367 1368 // ----------------------------------------------------------------------------- 1369 // Class Loader deadlock handling. 1370 // 1371 // complete_exit exits a lock returning recursion count 1372 // complete_exit/reenter operate as a wait without waiting 1373 // complete_exit requires an inflated monitor 1374 // The _owner field is not always the Thread addr even with an 1375 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1376 // thread due to contention. 1377 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1378 Thread * const Self = THREAD; 1379 assert(Self->is_Java_thread(), "Must be Java thread!"); 1380 JavaThread *jt = (JavaThread *)THREAD; 1381 1382 DeferredInitialize(); 1383 1384 if (THREAD != _owner) { 1385 if (THREAD->is_lock_owned ((address)_owner)) { 1386 assert(_recursions == 0, "internal state error"); 1387 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1388 _recursions = 0 ; 1389 OwnerIsThread = 1 ; 1390 } 1391 } 1392 1393 guarantee(Self == _owner, "complete_exit not owner"); 1394 intptr_t save = _recursions; // record the old recursion count 1395 _recursions = 0; // set the recursion level to be 0 1396 exit (true, Self) ; // exit the monitor 1397 guarantee (_owner != Self, "invariant"); 1398 return save; 1399 } 1400 1401 // reenter() enters a lock and sets recursion count 1402 // complete_exit/reenter operate as a wait without waiting 1403 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1404 Thread * const Self = THREAD; 1405 assert(Self->is_Java_thread(), "Must be Java thread!"); 1406 JavaThread *jt = (JavaThread *)THREAD; 1407 1408 guarantee(_owner != Self, "reenter already owner"); 1409 enter (THREAD); // enter the monitor 1410 guarantee (_recursions == 0, "reenter recursion"); 1411 _recursions = recursions; 1412 return; 1413 } 1414 1415 1416 // ----------------------------------------------------------------------------- 1417 // A macro is used below because there may already be a pending 1418 // exception which should not abort the execution of the routines 1419 // which use this (which is why we don't put this into check_slow and 1420 // call it with a CHECK argument). 1421 1422 #define CHECK_OWNER() \ 1423 do { \ 1424 if (THREAD != _owner) { \ 1425 if (THREAD->is_lock_owned((address) _owner)) { \ 1426 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1427 _recursions = 0; \ 1428 OwnerIsThread = 1 ; \ 1429 } else { \ 1430 TEVENT (Throw IMSX) ; \ 1431 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1432 } \ 1433 } \ 1434 } while (false) 1435 1436 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1437 // TODO-FIXME: remove check_slow() -- it's likely dead. 1438 1439 void ObjectMonitor::check_slow(TRAPS) { 1440 TEVENT (check_slow - throw IMSX) ; 1441 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1442 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1443 } 1444 1445 static int Adjust (volatile int * adr, int dx) { 1446 int v ; 1447 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1448 return v ; 1449 } 1450 1451 // helper method for posting a monitor wait event 1452 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1453 ObjectMonitor* monitor, 1454 jlong notifier_tid, 1455 jlong timeout, 1456 bool timedout) { 1457 assert(monitor != NULL, "invariant"); 1458 event->set_monitorClass(((oop)monitor->object())->klass()); 1459 event->set_timeout(timeout); 1460 event->set_address((uintptr_t)monitor->object_addr()); 1461 event->set_notifier((u8)notifier_tid); 1462 event->set_timedOut(timedout); 1463 event->commit(); 1464 } 1465 1466 // ----------------------------------------------------------------------------- 1467 // Wait/Notify/NotifyAll 1468 // 1469 // Note: a subset of changes to ObjectMonitor::wait() 1470 // will need to be replicated in complete_exit above 1471 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1472 Thread * const Self = THREAD ; 1473 assert(Self->is_Java_thread(), "Must be Java thread!"); 1474 JavaThread *jt = (JavaThread *)THREAD; 1475 1476 DeferredInitialize () ; 1477 1478 // Throw IMSX or IEX. 1479 CHECK_OWNER(); 1480 1481 EventJavaMonitorWait event; 1482 1483 // check for a pending interrupt 1484 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1485 // post monitor waited event. Note that this is past-tense, we are done waiting. 1486 if (JvmtiExport::should_post_monitor_waited()) { 1487 // Note: 'false' parameter is passed here because the 1488 // wait was not timed out due to thread interrupt. 1489 JvmtiExport::post_monitor_waited(jt, this, false); 1490 1491 // In this short circuit of the monitor wait protocol, the 1492 // current thread never drops ownership of the monitor and 1493 // never gets added to the wait queue so the current thread 1494 // cannot be made the successor. This means that the 1495 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1496 // consume an unpark() meant for the ParkEvent associated with 1497 // this ObjectMonitor. 1498 } 1499 if (event.should_commit()) { 1500 post_monitor_wait_event(&event, this, 0, millis, false); 1501 } 1502 TEVENT (Wait - Throw IEX) ; 1503 THROW(vmSymbols::java_lang_InterruptedException()); 1504 return ; 1505 } 1506 1507 TEVENT (Wait) ; 1508 1509 assert (Self->_Stalled == 0, "invariant") ; 1510 Self->_Stalled = intptr_t(this) ; 1511 jt->set_current_waiting_monitor(this); 1512 1513 // create a node to be put into the queue 1514 // Critically, after we reset() the event but prior to park(), we must check 1515 // for a pending interrupt. 1516 ObjectWaiter node(Self); 1517 node.TState = ObjectWaiter::TS_WAIT ; 1518 Self->_ParkEvent->reset() ; 1519 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1520 1521 // Enter the waiting queue, which is a circular doubly linked list in this case 1522 // but it could be a priority queue or any data structure. 1523 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1524 // by the the owner of the monitor *except* in the case where park() 1525 // returns because of a timeout of interrupt. Contention is exceptionally rare 1526 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1527 1528 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1529 AddWaiter (&node) ; 1530 Thread::SpinRelease (&_WaitSetLock) ; 1531 1532 if ((SyncFlags & 4) == 0) { 1533 _Responsible = NULL ; 1534 } 1535 intptr_t save = _recursions; // record the old recursion count 1536 _waiters++; // increment the number of waiters 1537 _recursions = 0; // set the recursion level to be 1 1538 exit (true, Self) ; // exit the monitor 1539 guarantee (_owner != Self, "invariant") ; 1540 1541 // The thread is on the WaitSet list - now park() it. 1542 // On MP systems it's conceivable that a brief spin before we park 1543 // could be profitable. 1544 // 1545 // TODO-FIXME: change the following logic to a loop of the form 1546 // while (!timeout && !interrupted && _notified == 0) park() 1547 1548 int ret = OS_OK ; 1549 int WasNotified = 0 ; 1550 { // State transition wrappers 1551 OSThread* osthread = Self->osthread(); 1552 OSThreadWaitState osts(osthread, true); 1553 { 1554 ThreadBlockInVM tbivm(jt); 1555 // Thread is in thread_blocked state and oop access is unsafe. 1556 jt->set_suspend_equivalent(); 1557 1558 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1559 // Intentionally empty 1560 } else 1561 if (node._notified == 0) { 1562 if (millis <= 0) { 1563 Self->_ParkEvent->park () ; 1564 } else { 1565 ret = Self->_ParkEvent->park (millis) ; 1566 } 1567 } 1568 1569 // were we externally suspended while we were waiting? 1570 if (ExitSuspendEquivalent (jt)) { 1571 // TODO-FIXME: add -- if succ == Self then succ = null. 1572 jt->java_suspend_self(); 1573 } 1574 1575 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1576 1577 1578 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1579 // from the WaitSet to the EntryList. 1580 // See if we need to remove Node from the WaitSet. 1581 // We use double-checked locking to avoid grabbing _WaitSetLock 1582 // if the thread is not on the wait queue. 1583 // 1584 // Note that we don't need a fence before the fetch of TState. 1585 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1586 // written by the is thread. (perhaps the fetch might even be satisfied 1587 // by a look-aside into the processor's own store buffer, although given 1588 // the length of the code path between the prior ST and this load that's 1589 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1590 // then we'll acquire the lock and then re-fetch a fresh TState value. 1591 // That is, we fail toward safety. 1592 1593 if (node.TState == ObjectWaiter::TS_WAIT) { 1594 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1595 if (node.TState == ObjectWaiter::TS_WAIT) { 1596 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1597 assert(node._notified == 0, "invariant"); 1598 node.TState = ObjectWaiter::TS_RUN ; 1599 } 1600 Thread::SpinRelease (&_WaitSetLock) ; 1601 } 1602 1603 // The thread is now either on off-list (TS_RUN), 1604 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1605 // The Node's TState variable is stable from the perspective of this thread. 1606 // No other threads will asynchronously modify TState. 1607 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1608 OrderAccess::loadload() ; 1609 if (_succ == Self) _succ = NULL ; 1610 WasNotified = node._notified ; 1611 1612 // Reentry phase -- reacquire the monitor. 1613 // re-enter contended monitor after object.wait(). 1614 // retain OBJECT_WAIT state until re-enter successfully completes 1615 // Thread state is thread_in_vm and oop access is again safe, 1616 // although the raw address of the object may have changed. 1617 // (Don't cache naked oops over safepoints, of course). 1618 1619 // post monitor waited event. Note that this is past-tense, we are done waiting. 1620 if (JvmtiExport::should_post_monitor_waited()) { 1621 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1622 1623 if (node._notified != 0 && _succ == Self) { 1624 // In this part of the monitor wait-notify-reenter protocol it 1625 // is possible (and normal) for another thread to do a fastpath 1626 // monitor enter-exit while this thread is still trying to get 1627 // to the reenter portion of the protocol. 1628 // 1629 // The ObjectMonitor was notified and the current thread is 1630 // the successor which also means that an unpark() has already 1631 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1632 // consume the unpark() that was done when the successor was 1633 // set because the same ParkEvent is shared between Java 1634 // monitors and JVM/TI RawMonitors (for now). 1635 // 1636 // We redo the unpark() to ensure forward progress, i.e., we 1637 // don't want all pending threads hanging (parked) with none 1638 // entering the unlocked monitor. 1639 node._event->unpark(); 1640 } 1641 } 1642 1643 if (event.should_commit()) { 1644 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1645 } 1646 1647 OrderAccess::fence() ; 1648 1649 assert (Self->_Stalled != 0, "invariant") ; 1650 Self->_Stalled = 0 ; 1651 1652 assert (_owner != Self, "invariant") ; 1653 ObjectWaiter::TStates v = node.TState ; 1654 if (v == ObjectWaiter::TS_RUN) { 1655 enter (Self) ; 1656 } else { 1657 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1658 ReenterI (Self, &node) ; 1659 node.wait_reenter_end(this); 1660 } 1661 1662 // Self has reacquired the lock. 1663 // Lifecycle - the node representing Self must not appear on any queues. 1664 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1665 // want residual elements associated with this thread left on any lists. 1666 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1667 assert (_owner == Self, "invariant") ; 1668 assert (_succ != Self , "invariant") ; 1669 } // OSThreadWaitState() 1670 1671 jt->set_current_waiting_monitor(NULL); 1672 1673 guarantee (_recursions == 0, "invariant") ; 1674 _recursions = save; // restore the old recursion count 1675 _waiters--; // decrement the number of waiters 1676 1677 // Verify a few postconditions 1678 assert (_owner == Self , "invariant") ; 1679 assert (_succ != Self , "invariant") ; 1680 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1681 1682 if (SyncFlags & 32) { 1683 OrderAccess::fence() ; 1684 } 1685 1686 // check if the notification happened 1687 if (!WasNotified) { 1688 // no, it could be timeout or Thread.interrupt() or both 1689 // check for interrupt event, otherwise it is timeout 1690 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1691 TEVENT (Wait - throw IEX from epilog) ; 1692 THROW(vmSymbols::java_lang_InterruptedException()); 1693 } 1694 } 1695 1696 // NOTE: Spurious wake up will be consider as timeout. 1697 // Monitor notify has precedence over thread interrupt. 1698 } 1699 1700 1701 // Consider: 1702 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1703 // then instead of transferring a thread from the WaitSet to the EntryList 1704 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1705 1706 void ObjectMonitor::notify(TRAPS) { 1707 CHECK_OWNER(); 1708 if (_WaitSet == NULL) { 1709 TEVENT (Empty-Notify) ; 1710 return ; 1711 } 1712 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1713 1714 int Policy = Knob_MoveNotifyee ; 1715 1716 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1717 ObjectWaiter * iterator = DequeueWaiter() ; 1718 if (iterator != NULL) { 1719 TEVENT (Notify1 - Transfer) ; 1720 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1721 guarantee (iterator->_notified == 0, "invariant") ; 1722 if (Policy != 4) { 1723 iterator->TState = ObjectWaiter::TS_ENTER ; 1724 } 1725 iterator->_notified = 1 ; 1726 Thread * Self = THREAD; 1727 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1728 1729 ObjectWaiter * List = _EntryList ; 1730 if (List != NULL) { 1731 assert (List->_prev == NULL, "invariant") ; 1732 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1733 assert (List != iterator, "invariant") ; 1734 } 1735 1736 if (Policy == 0) { // prepend to EntryList 1737 if (List == NULL) { 1738 iterator->_next = iterator->_prev = NULL ; 1739 _EntryList = iterator ; 1740 } else { 1741 List->_prev = iterator ; 1742 iterator->_next = List ; 1743 iterator->_prev = NULL ; 1744 _EntryList = iterator ; 1745 } 1746 } else 1747 if (Policy == 1) { // append to EntryList 1748 if (List == NULL) { 1749 iterator->_next = iterator->_prev = NULL ; 1750 _EntryList = iterator ; 1751 } else { 1752 // CONSIDER: finding the tail currently requires a linear-time walk of 1753 // the EntryList. We can make tail access constant-time by converting to 1754 // a CDLL instead of using our current DLL. 1755 ObjectWaiter * Tail ; 1756 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1757 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1758 Tail->_next = iterator ; 1759 iterator->_prev = Tail ; 1760 iterator->_next = NULL ; 1761 } 1762 } else 1763 if (Policy == 2) { // prepend to cxq 1764 // prepend to cxq 1765 if (List == NULL) { 1766 iterator->_next = iterator->_prev = NULL ; 1767 _EntryList = iterator ; 1768 } else { 1769 iterator->TState = ObjectWaiter::TS_CXQ ; 1770 for (;;) { 1771 ObjectWaiter * Front = _cxq ; 1772 iterator->_next = Front ; 1773 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1774 break ; 1775 } 1776 } 1777 } 1778 } else 1779 if (Policy == 3) { // append to cxq 1780 iterator->TState = ObjectWaiter::TS_CXQ ; 1781 for (;;) { 1782 ObjectWaiter * Tail ; 1783 Tail = _cxq ; 1784 if (Tail == NULL) { 1785 iterator->_next = NULL ; 1786 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1787 break ; 1788 } 1789 } else { 1790 while (Tail->_next != NULL) Tail = Tail->_next ; 1791 Tail->_next = iterator ; 1792 iterator->_prev = Tail ; 1793 iterator->_next = NULL ; 1794 break ; 1795 } 1796 } 1797 } else { 1798 ParkEvent * ev = iterator->_event ; 1799 iterator->TState = ObjectWaiter::TS_RUN ; 1800 OrderAccess::fence() ; 1801 ev->unpark() ; 1802 } 1803 1804 if (Policy < 4) { 1805 iterator->wait_reenter_begin(this); 1806 } 1807 1808 // _WaitSetLock protects the wait queue, not the EntryList. We could 1809 // move the add-to-EntryList operation, above, outside the critical section 1810 // protected by _WaitSetLock. In practice that's not useful. With the 1811 // exception of wait() timeouts and interrupts the monitor owner 1812 // is the only thread that grabs _WaitSetLock. There's almost no contention 1813 // on _WaitSetLock so it's not profitable to reduce the length of the 1814 // critical section. 1815 } 1816 1817 Thread::SpinRelease (&_WaitSetLock) ; 1818 1819 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1820 ObjectMonitor::_sync_Notifications->inc() ; 1821 } 1822 } 1823 1824 1825 void ObjectMonitor::notifyAll(TRAPS) { 1826 CHECK_OWNER(); 1827 ObjectWaiter* iterator; 1828 if (_WaitSet == NULL) { 1829 TEVENT (Empty-NotifyAll) ; 1830 return ; 1831 } 1832 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1833 1834 int Policy = Knob_MoveNotifyee ; 1835 int Tally = 0 ; 1836 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1837 1838 for (;;) { 1839 iterator = DequeueWaiter () ; 1840 if (iterator == NULL) break ; 1841 TEVENT (NotifyAll - Transfer1) ; 1842 ++Tally ; 1843 1844 // Disposition - what might we do with iterator ? 1845 // a. add it directly to the EntryList - either tail or head. 1846 // b. push it onto the front of the _cxq. 1847 // For now we use (a). 1848 1849 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1850 guarantee (iterator->_notified == 0, "invariant") ; 1851 iterator->_notified = 1 ; 1852 Thread * Self = THREAD; 1853 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1854 if (Policy != 4) { 1855 iterator->TState = ObjectWaiter::TS_ENTER ; 1856 } 1857 1858 ObjectWaiter * List = _EntryList ; 1859 if (List != NULL) { 1860 assert (List->_prev == NULL, "invariant") ; 1861 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1862 assert (List != iterator, "invariant") ; 1863 } 1864 1865 if (Policy == 0) { // prepend to EntryList 1866 if (List == NULL) { 1867 iterator->_next = iterator->_prev = NULL ; 1868 _EntryList = iterator ; 1869 } else { 1870 List->_prev = iterator ; 1871 iterator->_next = List ; 1872 iterator->_prev = NULL ; 1873 _EntryList = iterator ; 1874 } 1875 } else 1876 if (Policy == 1) { // append to EntryList 1877 if (List == NULL) { 1878 iterator->_next = iterator->_prev = NULL ; 1879 _EntryList = iterator ; 1880 } else { 1881 // CONSIDER: finding the tail currently requires a linear-time walk of 1882 // the EntryList. We can make tail access constant-time by converting to 1883 // a CDLL instead of using our current DLL. 1884 ObjectWaiter * Tail ; 1885 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1886 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1887 Tail->_next = iterator ; 1888 iterator->_prev = Tail ; 1889 iterator->_next = NULL ; 1890 } 1891 } else 1892 if (Policy == 2) { // prepend to cxq 1893 // prepend to cxq 1894 iterator->TState = ObjectWaiter::TS_CXQ ; 1895 for (;;) { 1896 ObjectWaiter * Front = _cxq ; 1897 iterator->_next = Front ; 1898 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1899 break ; 1900 } 1901 } 1902 } else 1903 if (Policy == 3) { // append to cxq 1904 iterator->TState = ObjectWaiter::TS_CXQ ; 1905 for (;;) { 1906 ObjectWaiter * Tail ; 1907 Tail = _cxq ; 1908 if (Tail == NULL) { 1909 iterator->_next = NULL ; 1910 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1911 break ; 1912 } 1913 } else { 1914 while (Tail->_next != NULL) Tail = Tail->_next ; 1915 Tail->_next = iterator ; 1916 iterator->_prev = Tail ; 1917 iterator->_next = NULL ; 1918 break ; 1919 } 1920 } 1921 } else { 1922 ParkEvent * ev = iterator->_event ; 1923 iterator->TState = ObjectWaiter::TS_RUN ; 1924 OrderAccess::fence() ; 1925 ev->unpark() ; 1926 } 1927 1928 if (Policy < 4) { 1929 iterator->wait_reenter_begin(this); 1930 } 1931 1932 // _WaitSetLock protects the wait queue, not the EntryList. We could 1933 // move the add-to-EntryList operation, above, outside the critical section 1934 // protected by _WaitSetLock. In practice that's not useful. With the 1935 // exception of wait() timeouts and interrupts the monitor owner 1936 // is the only thread that grabs _WaitSetLock. There's almost no contention 1937 // on _WaitSetLock so it's not profitable to reduce the length of the 1938 // critical section. 1939 } 1940 1941 Thread::SpinRelease (&_WaitSetLock) ; 1942 1943 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1944 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1945 } 1946 } 1947 1948 // ----------------------------------------------------------------------------- 1949 // Adaptive Spinning Support 1950 // 1951 // Adaptive spin-then-block - rational spinning 1952 // 1953 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1954 // algorithm. On high order SMP systems it would be better to start with 1955 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1956 // a contending thread could enqueue itself on the cxq and then spin locally 1957 // on a thread-specific variable such as its ParkEvent._Event flag. 1958 // That's left as an exercise for the reader. Note that global spinning is 1959 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1960 // low latency and massive bandwidth. 1961 // 1962 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1963 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1964 // (duration) or we can fix the count at approximately the duration of 1965 // a context switch and vary the frequency. Of course we could also 1966 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1967 // For a description of 'Adaptive spin-then-block mutual exclusion in 1968 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1969 // 1970 // This implementation varies the duration "D", where D varies with 1971 // the success rate of recent spin attempts. (D is capped at approximately 1972 // length of a round-trip context switch). The success rate for recent 1973 // spin attempts is a good predictor of the success rate of future spin 1974 // attempts. The mechanism adapts automatically to varying critical 1975 // section length (lock modality), system load and degree of parallelism. 1976 // D is maintained per-monitor in _SpinDuration and is initialized 1977 // optimistically. Spin frequency is fixed at 100%. 1978 // 1979 // Note that _SpinDuration is volatile, but we update it without locks 1980 // or atomics. The code is designed so that _SpinDuration stays within 1981 // a reasonable range even in the presence of races. The arithmetic 1982 // operations on _SpinDuration are closed over the domain of legal values, 1983 // so at worst a race will install and older but still legal value. 1984 // At the very worst this introduces some apparent non-determinism. 1985 // We might spin when we shouldn't or vice-versa, but since the spin 1986 // count are relatively short, even in the worst case, the effect is harmless. 1987 // 1988 // Care must be taken that a low "D" value does not become an 1989 // an absorbing state. Transient spinning failures -- when spinning 1990 // is overall profitable -- should not cause the system to converge 1991 // on low "D" values. We want spinning to be stable and predictable 1992 // and fairly responsive to change and at the same time we don't want 1993 // it to oscillate, become metastable, be "too" non-deterministic, 1994 // or converge on or enter undesirable stable absorbing states. 1995 // 1996 // We implement a feedback-based control system -- using past behavior 1997 // to predict future behavior. We face two issues: (a) if the 1998 // input signal is random then the spin predictor won't provide optimal 1999 // results, and (b) if the signal frequency is too high then the control 2000 // system, which has some natural response lag, will "chase" the signal. 2001 // (b) can arise from multimodal lock hold times. Transient preemption 2002 // can also result in apparent bimodal lock hold times. 2003 // Although sub-optimal, neither condition is particularly harmful, as 2004 // in the worst-case we'll spin when we shouldn't or vice-versa. 2005 // The maximum spin duration is rather short so the failure modes aren't bad. 2006 // To be conservative, I've tuned the gain in system to bias toward 2007 // _not spinning. Relatedly, the system can sometimes enter a mode where it 2008 // "rings" or oscillates between spinning and not spinning. This happens 2009 // when spinning is just on the cusp of profitability, however, so the 2010 // situation is not dire. The state is benign -- there's no need to add 2011 // hysteresis control to damp the transition rate between spinning and 2012 // not spinning. 2013 // 2014 2015 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 2016 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 2017 2018 // Spinning: Fixed frequency (100%), vary duration 2019 2020 2021 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 2022 2023 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2024 int ctr = Knob_FixedSpin ; 2025 if (ctr != 0) { 2026 while (--ctr >= 0) { 2027 if (TryLock (Self) > 0) return 1 ; 2028 SpinPause () ; 2029 } 2030 return 0 ; 2031 } 2032 2033 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 2034 if (TryLock(Self) > 0) { 2035 // Increase _SpinDuration ... 2036 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2037 // Raising _SpurDuration to the poverty line is key. 2038 int x = _SpinDuration ; 2039 if (x < Knob_SpinLimit) { 2040 if (x < Knob_Poverty) x = Knob_Poverty ; 2041 _SpinDuration = x + Knob_BonusB ; 2042 } 2043 return 1 ; 2044 } 2045 SpinPause () ; 2046 } 2047 2048 // Admission control - verify preconditions for spinning 2049 // 2050 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2051 // becoming an absorbing state. Put another way, we spin briefly to 2052 // sample, just in case the system load, parallelism, contention, or lock 2053 // modality changed. 2054 // 2055 // Consider the following alternative: 2056 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2057 // spin attempt. "Periodically" might mean after a tally of 2058 // the # of failed spin attempts (or iterations) reaches some threshold. 2059 // This takes us into the realm of 1-out-of-N spinning, where we 2060 // hold the duration constant but vary the frequency. 2061 2062 ctr = _SpinDuration ; 2063 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2064 if (ctr <= 0) return 0 ; 2065 2066 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2067 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2068 TEVENT (Spin abort - notrunnable [TOP]); 2069 return 0 ; 2070 } 2071 2072 int MaxSpin = Knob_MaxSpinners ; 2073 if (MaxSpin >= 0) { 2074 if (_Spinner > MaxSpin) { 2075 TEVENT (Spin abort -- too many spinners) ; 2076 return 0 ; 2077 } 2078 // Slighty racy, but benign ... 2079 Adjust (&_Spinner, 1) ; 2080 } 2081 2082 // We're good to spin ... spin ingress. 2083 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2084 // when preparing to LD...CAS _owner, etc and the CAS is likely 2085 // to succeed. 2086 int hits = 0 ; 2087 int msk = 0 ; 2088 int caspty = Knob_CASPenalty ; 2089 int oxpty = Knob_OXPenalty ; 2090 int sss = Knob_SpinSetSucc ; 2091 if (sss && _succ == NULL ) _succ = Self ; 2092 Thread * prv = NULL ; 2093 2094 // There are three ways to exit the following loop: 2095 // 1. A successful spin where this thread has acquired the lock. 2096 // 2. Spin failure with prejudice 2097 // 3. Spin failure without prejudice 2098 2099 while (--ctr >= 0) { 2100 2101 // Periodic polling -- Check for pending GC 2102 // Threads may spin while they're unsafe. 2103 // We don't want spinning threads to delay the JVM from reaching 2104 // a stop-the-world safepoint or to steal cycles from GC. 2105 // If we detect a pending safepoint we abort in order that 2106 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2107 // this thread, if safe, doesn't steal cycles from GC. 2108 // This is in keeping with the "no loitering in runtime" rule. 2109 // We periodically check to see if there's a safepoint pending. 2110 if ((ctr & 0xFF) == 0) { 2111 if (SafepointSynchronize::do_call_back()) { 2112 TEVENT (Spin: safepoint) ; 2113 goto Abort ; // abrupt spin egress 2114 } 2115 if (Knob_UsePause & 1) SpinPause () ; 2116 2117 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2118 if (hits > 50 && scb != NULL) { 2119 int abend = (*scb)(SpinCallbackArgument, 0) ; 2120 } 2121 } 2122 2123 if (Knob_UsePause & 2) SpinPause() ; 2124 2125 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2126 // This is useful on classic SMP systems, but is of less utility on 2127 // N1-style CMT platforms. 2128 // 2129 // Trade-off: lock acquisition latency vs coherency bandwidth. 2130 // Lock hold times are typically short. A histogram 2131 // of successful spin attempts shows that we usually acquire 2132 // the lock early in the spin. That suggests we want to 2133 // sample _owner frequently in the early phase of the spin, 2134 // but then back-off and sample less frequently as the spin 2135 // progresses. The back-off makes a good citizen on SMP big 2136 // SMP systems. Oversampling _owner can consume excessive 2137 // coherency bandwidth. Relatedly, if we _oversample _owner we 2138 // can inadvertently interfere with the the ST m->owner=null. 2139 // executed by the lock owner. 2140 if (ctr & msk) continue ; 2141 ++hits ; 2142 if ((hits & 0xF) == 0) { 2143 // The 0xF, above, corresponds to the exponent. 2144 // Consider: (msk+1)|msk 2145 msk = ((msk << 2)|3) & BackOffMask ; 2146 } 2147 2148 // Probe _owner with TATAS 2149 // If this thread observes the monitor transition or flicker 2150 // from locked to unlocked to locked, then the odds that this 2151 // thread will acquire the lock in this spin attempt go down 2152 // considerably. The same argument applies if the CAS fails 2153 // or if we observe _owner change from one non-null value to 2154 // another non-null value. In such cases we might abort 2155 // the spin without prejudice or apply a "penalty" to the 2156 // spin count-down variable "ctr", reducing it by 100, say. 2157 2158 Thread * ox = (Thread *) _owner ; 2159 if (ox == NULL) { 2160 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2161 if (ox == NULL) { 2162 // The CAS succeeded -- this thread acquired ownership 2163 // Take care of some bookkeeping to exit spin state. 2164 if (sss && _succ == Self) { 2165 _succ = NULL ; 2166 } 2167 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2168 2169 // Increase _SpinDuration : 2170 // The spin was successful (profitable) so we tend toward 2171 // longer spin attempts in the future. 2172 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2173 // If we acquired the lock early in the spin cycle it 2174 // makes sense to increase _SpinDuration proportionally. 2175 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2176 int x = _SpinDuration ; 2177 if (x < Knob_SpinLimit) { 2178 if (x < Knob_Poverty) x = Knob_Poverty ; 2179 _SpinDuration = x + Knob_Bonus ; 2180 } 2181 return 1 ; 2182 } 2183 2184 // The CAS failed ... we can take any of the following actions: 2185 // * penalize: ctr -= Knob_CASPenalty 2186 // * exit spin with prejudice -- goto Abort; 2187 // * exit spin without prejudice. 2188 // * Since CAS is high-latency, retry again immediately. 2189 prv = ox ; 2190 TEVENT (Spin: cas failed) ; 2191 if (caspty == -2) break ; 2192 if (caspty == -1) goto Abort ; 2193 ctr -= caspty ; 2194 continue ; 2195 } 2196 2197 // Did lock ownership change hands ? 2198 if (ox != prv && prv != NULL ) { 2199 TEVENT (spin: Owner changed) 2200 if (oxpty == -2) break ; 2201 if (oxpty == -1) goto Abort ; 2202 ctr -= oxpty ; 2203 } 2204 prv = ox ; 2205 2206 // Abort the spin if the owner is not executing. 2207 // The owner must be executing in order to drop the lock. 2208 // Spinning while the owner is OFFPROC is idiocy. 2209 // Consider: ctr -= RunnablePenalty ; 2210 if (Knob_OState && NotRunnable (Self, ox)) { 2211 TEVENT (Spin abort - notrunnable); 2212 goto Abort ; 2213 } 2214 if (sss && _succ == NULL ) _succ = Self ; 2215 } 2216 2217 // Spin failed with prejudice -- reduce _SpinDuration. 2218 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2219 // AIMD is globally stable. 2220 TEVENT (Spin failure) ; 2221 { 2222 int x = _SpinDuration ; 2223 if (x > 0) { 2224 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2225 // This is globally sample and tends to damp the response. 2226 x -= Knob_Penalty ; 2227 if (x < 0) x = 0 ; 2228 _SpinDuration = x ; 2229 } 2230 } 2231 2232 Abort: 2233 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2234 if (sss && _succ == Self) { 2235 _succ = NULL ; 2236 // Invariant: after setting succ=null a contending thread 2237 // must recheck-retry _owner before parking. This usually happens 2238 // in the normal usage of TrySpin(), but it's safest 2239 // to make TrySpin() as foolproof as possible. 2240 OrderAccess::fence() ; 2241 if (TryLock(Self) > 0) return 1 ; 2242 } 2243 return 0 ; 2244 } 2245 2246 // NotRunnable() -- informed spinning 2247 // 2248 // Don't bother spinning if the owner is not eligible to drop the lock. 2249 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2250 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2251 // The thread must be runnable in order to drop the lock in timely fashion. 2252 // If the _owner is not runnable then spinning will not likely be 2253 // successful (profitable). 2254 // 2255 // Beware -- the thread referenced by _owner could have died 2256 // so a simply fetch from _owner->_thread_state might trap. 2257 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2258 // Because of the lifecycle issues the schedctl and _thread_state values 2259 // observed by NotRunnable() might be garbage. NotRunnable must 2260 // tolerate this and consider the observed _thread_state value 2261 // as advisory. 2262 // 2263 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2264 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2265 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2266 // with the LSB of _owner. Another option would be to probablistically probe 2267 // the putative _owner->TypeTag value. 2268 // 2269 // Checking _thread_state isn't perfect. Even if the thread is 2270 // in_java it might be blocked on a page-fault or have been preempted 2271 // and sitting on a ready/dispatch queue. _thread state in conjunction 2272 // with schedctl.sc_state gives us a good picture of what the 2273 // thread is doing, however. 2274 // 2275 // TODO: check schedctl.sc_state. 2276 // We'll need to use SafeFetch32() to read from the schedctl block. 2277 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2278 // 2279 // The return value from NotRunnable() is *advisory* -- the 2280 // result is based on sampling and is not necessarily coherent. 2281 // The caller must tolerate false-negative and false-positive errors. 2282 // Spinning, in general, is probabilistic anyway. 2283 2284 2285 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2286 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2287 if (!OwnerIsThread) return 0 ; 2288 2289 if (ox == NULL) return 0 ; 2290 2291 // Avoid transitive spinning ... 2292 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2293 // Immediately after T1 acquires L it's possible that T2, also 2294 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2295 // This occurs transiently after T1 acquired L but before 2296 // T1 managed to clear T1.Stalled. T2 does not need to abort 2297 // its spin in this circumstance. 2298 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2299 2300 if (BlockedOn == 1) return 1 ; 2301 if (BlockedOn != 0) { 2302 return BlockedOn != intptr_t(this) && _owner == ox ; 2303 } 2304 2305 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2306 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2307 // consider also: jst != _thread_in_Java -- but that's overspecific. 2308 return jst == _thread_blocked || jst == _thread_in_native ; 2309 } 2310 2311 2312 // ----------------------------------------------------------------------------- 2313 // WaitSet management ... 2314 2315 ObjectWaiter::ObjectWaiter(Thread* thread) { 2316 _next = NULL; 2317 _prev = NULL; 2318 _notified = 0; 2319 TState = TS_RUN ; 2320 _thread = thread; 2321 _event = thread->_ParkEvent ; 2322 _active = false; 2323 assert (_event != NULL, "invariant") ; 2324 } 2325 2326 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2327 JavaThread *jt = (JavaThread *)this->_thread; 2328 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2329 } 2330 2331 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2332 JavaThread *jt = (JavaThread *)this->_thread; 2333 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2334 } 2335 2336 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2337 assert(node != NULL, "should not dequeue NULL node"); 2338 assert(node->_prev == NULL, "node already in list"); 2339 assert(node->_next == NULL, "node already in list"); 2340 // put node at end of queue (circular doubly linked list) 2341 if (_WaitSet == NULL) { 2342 _WaitSet = node; 2343 node->_prev = node; 2344 node->_next = node; 2345 } else { 2346 ObjectWaiter* head = _WaitSet ; 2347 ObjectWaiter* tail = head->_prev; 2348 assert(tail->_next == head, "invariant check"); 2349 tail->_next = node; 2350 head->_prev = node; 2351 node->_next = head; 2352 node->_prev = tail; 2353 } 2354 } 2355 2356 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2357 // dequeue the very first waiter 2358 ObjectWaiter* waiter = _WaitSet; 2359 if (waiter) { 2360 DequeueSpecificWaiter(waiter); 2361 } 2362 return waiter; 2363 } 2364 2365 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2366 assert(node != NULL, "should not dequeue NULL node"); 2367 assert(node->_prev != NULL, "node already removed from list"); 2368 assert(node->_next != NULL, "node already removed from list"); 2369 // when the waiter has woken up because of interrupt, 2370 // timeout or other spurious wake-up, dequeue the 2371 // waiter from waiting list 2372 ObjectWaiter* next = node->_next; 2373 if (next == node) { 2374 assert(node->_prev == node, "invariant check"); 2375 _WaitSet = NULL; 2376 } else { 2377 ObjectWaiter* prev = node->_prev; 2378 assert(prev->_next == node, "invariant check"); 2379 assert(next->_prev == node, "invariant check"); 2380 next->_prev = prev; 2381 prev->_next = next; 2382 if (_WaitSet == node) { 2383 _WaitSet = next; 2384 } 2385 } 2386 node->_next = NULL; 2387 node->_prev = NULL; 2388 } 2389 2390 // ----------------------------------------------------------------------------- 2391 // PerfData support 2392 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2393 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2394 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2395 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2396 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2397 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2398 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2399 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2400 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2401 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2402 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2403 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2404 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2405 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2406 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2407 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2408 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2409 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2410 2411 // One-shot global initialization for the sync subsystem. 2412 // We could also defer initialization and initialize on-demand 2413 // the first time we call inflate(). Initialization would 2414 // be protected - like so many things - by the MonitorCache_lock. 2415 2416 void ObjectMonitor::Initialize () { 2417 static int InitializationCompleted = 0 ; 2418 assert (InitializationCompleted == 0, "invariant") ; 2419 InitializationCompleted = 1 ; 2420 if (UsePerfData) { 2421 EXCEPTION_MARK ; 2422 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2423 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2424 NEWPERFCOUNTER(_sync_Inflations) ; 2425 NEWPERFCOUNTER(_sync_Deflations) ; 2426 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2427 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2428 NEWPERFCOUNTER(_sync_Parks) ; 2429 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2430 NEWPERFCOUNTER(_sync_Notifications) ; 2431 NEWPERFCOUNTER(_sync_SlowEnter) ; 2432 NEWPERFCOUNTER(_sync_SlowExit) ; 2433 NEWPERFCOUNTER(_sync_SlowNotify) ; 2434 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2435 NEWPERFCOUNTER(_sync_FailedSpins) ; 2436 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2437 NEWPERFCOUNTER(_sync_PrivateA) ; 2438 NEWPERFCOUNTER(_sync_PrivateB) ; 2439 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2440 NEWPERFCOUNTER(_sync_MonScavenged) ; 2441 NEWPERFVARIABLE(_sync_MonExtant) ; 2442 #undef NEWPERFCOUNTER 2443 } 2444 } 2445 2446 2447 // Compile-time asserts 2448 // When possible, it's better to catch errors deterministically at 2449 // compile-time than at runtime. The down-side to using compile-time 2450 // asserts is that error message -- often something about negative array 2451 // indices -- is opaque. 2452 2453 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2454 2455 void ObjectMonitor::ctAsserts() { 2456 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2457 } 2458 2459 2460 static char * kvGet (char * kvList, const char * Key) { 2461 if (kvList == NULL) return NULL ; 2462 size_t n = strlen (Key) ; 2463 char * Search ; 2464 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2465 if (strncmp (Search, Key, n) == 0) { 2466 if (Search[n] == '=') return Search + n + 1 ; 2467 if (Search[n] == 0) return (char *) "1" ; 2468 } 2469 } 2470 return NULL ; 2471 } 2472 2473 static int kvGetInt (char * kvList, const char * Key, int Default) { 2474 char * v = kvGet (kvList, Key) ; 2475 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2476 if (Knob_ReportSettings && v != NULL) { 2477 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2478 ::fflush (stdout) ; 2479 } 2480 return rslt ; 2481 } 2482 2483 void ObjectMonitor::DeferredInitialize () { 2484 if (InitDone > 0) return ; 2485 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2486 while (InitDone != 1) ; 2487 return ; 2488 } 2489 2490 // One-shot global initialization ... 2491 // The initialization is idempotent, so we don't need locks. 2492 // In the future consider doing this via os::init_2(). 2493 // SyncKnobs consist of <Key>=<Value> pairs in the style 2494 // of environment variables. Start by converting ':' to NUL. 2495 2496 if (SyncKnobs == NULL) SyncKnobs = "" ; 2497 2498 size_t sz = strlen (SyncKnobs) ; 2499 char * knobs = (char *) malloc (sz + 2) ; 2500 if (knobs == NULL) { 2501 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; 2502 guarantee (0, "invariant") ; 2503 } 2504 strcpy (knobs, SyncKnobs) ; 2505 knobs[sz+1] = 0 ; 2506 for (char * p = knobs ; *p ; p++) { 2507 if (*p == ':') *p = 0 ; 2508 } 2509 2510 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2511 SETKNOB(ReportSettings) ; 2512 SETKNOB(Verbose) ; 2513 SETKNOB(FixedSpin) ; 2514 SETKNOB(SpinLimit) ; 2515 SETKNOB(SpinBase) ; 2516 SETKNOB(SpinBackOff); 2517 SETKNOB(CASPenalty) ; 2518 SETKNOB(OXPenalty) ; 2519 SETKNOB(LogSpins) ; 2520 SETKNOB(SpinSetSucc) ; 2521 SETKNOB(SuccEnabled) ; 2522 SETKNOB(SuccRestrict) ; 2523 SETKNOB(Penalty) ; 2524 SETKNOB(Bonus) ; 2525 SETKNOB(BonusB) ; 2526 SETKNOB(Poverty) ; 2527 SETKNOB(SpinAfterFutile) ; 2528 SETKNOB(UsePause) ; 2529 SETKNOB(SpinEarly) ; 2530 SETKNOB(OState) ; 2531 SETKNOB(MaxSpinners) ; 2532 SETKNOB(PreSpin) ; 2533 SETKNOB(ExitPolicy) ; 2534 SETKNOB(QMode); 2535 SETKNOB(ResetEvent) ; 2536 SETKNOB(MoveNotifyee) ; 2537 SETKNOB(FastHSSEC) ; 2538 #undef SETKNOB 2539 2540 if (Knob_Verbose) { 2541 sanity_checks(); 2542 } 2543 2544 if (os::is_MP()) { 2545 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2546 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2547 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2548 } else { 2549 Knob_SpinLimit = 0 ; 2550 Knob_SpinBase = 0 ; 2551 Knob_PreSpin = 0 ; 2552 Knob_FixedSpin = -1 ; 2553 } 2554 2555 if (Knob_LogSpins == 0) { 2556 ObjectMonitor::_sync_FailedSpins = NULL ; 2557 } 2558 2559 free (knobs) ; 2560 OrderAccess::fence() ; 2561 InitDone = 1 ; 2562 } 2563 2564 void ObjectMonitor::sanity_checks() { 2565 int error_cnt = 0; 2566 int warning_cnt = 0; 2567 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2568 2569 if (verbose) { 2570 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2571 sizeof(ObjectMonitor)); 2572 } 2573 2574 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2575 if (verbose) { 2576 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2577 } 2578 2579 ObjectMonitor dummy; 2580 u_char *addr_begin = (u_char*)&dummy; 2581 u_char *addr_header = (u_char*)&dummy._header; 2582 u_char *addr_owner = (u_char*)&dummy._owner; 2583 2584 uint offset_header = (uint)(addr_header - addr_begin); 2585 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2586 2587 uint offset_owner = (uint)(addr_owner - addr_begin); 2588 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2589 2590 if ((uint)(addr_header - addr_begin) != 0) { 2591 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2592 error_cnt++; 2593 } 2594 2595 if (cache_line_size != 0) { 2596 // We were able to determine the L1 data cache line size so 2597 // do some cache line specific sanity checks 2598 2599 if ((offset_owner - offset_header) < cache_line_size) { 2600 tty->print_cr("WARNING: the _header and _owner fields are closer " 2601 "than a cache line which permits false sharing."); 2602 warning_cnt++; 2603 } 2604 2605 if ((sizeof(ObjectMonitor) % cache_line_size) != 0) { 2606 tty->print_cr("WARNING: ObjectMonitor size is not a multiple of " 2607 "a cache line which permits false sharing."); 2608 warning_cnt++; 2609 } 2610 } 2611 2612 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2613 &warning_cnt); 2614 2615 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2616 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2617 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2618 } 2619 2620 guarantee(error_cnt == 0, 2621 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2622 } 2623 2624 #ifndef PRODUCT 2625 void ObjectMonitor::verify() { 2626 } 2627 2628 void ObjectMonitor::print() { 2629 } 2630 #endif