1 /* 2 * Copyright (c) 1998, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "oops/markOop.hpp" 29 #include "oops/oop.inline.hpp" 30 #include "runtime/handles.inline.hpp" 31 #include "runtime/interfaceSupport.hpp" 32 #include "runtime/mutexLocker.hpp" 33 #include "runtime/objectMonitor.hpp" 34 #include "runtime/objectMonitor.inline.hpp" 35 #include "runtime/orderAccess.inline.hpp" 36 #include "runtime/osThread.hpp" 37 #include "runtime/stubRoutines.hpp" 38 #include "runtime/thread.inline.hpp" 39 #include "services/threadService.hpp" 40 #include "trace/tracing.hpp" 41 #include "trace/traceMacros.hpp" 42 #include "utilities/dtrace.hpp" 43 #include "utilities/macros.hpp" 44 #include "utilities/preserveException.hpp" 45 #ifdef TARGET_OS_FAMILY_linux 46 # include "os_linux.inline.hpp" 47 #endif 48 #ifdef TARGET_OS_FAMILY_solaris 49 # include "os_solaris.inline.hpp" 50 #endif 51 #ifdef TARGET_OS_FAMILY_windows 52 # include "os_windows.inline.hpp" 53 #endif 54 #ifdef TARGET_OS_FAMILY_bsd 55 # include "os_bsd.inline.hpp" 56 #endif 57 58 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 59 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 60 #define ATTR __attribute__((noinline)) 61 #else 62 #define ATTR 63 #endif 64 65 66 #ifdef DTRACE_ENABLED 67 68 // Only bother with this argument setup if dtrace is available 69 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 70 71 72 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 73 char* bytes = NULL; \ 74 int len = 0; \ 75 jlong jtid = SharedRuntime::get_java_tid(thread); \ 76 Symbol* klassname = ((oop)obj)->klass()->name(); \ 77 if (klassname != NULL) { \ 78 bytes = (char*)klassname->bytes(); \ 79 len = klassname->utf8_length(); \ 80 } 81 82 #ifndef USDT2 83 84 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 85 jlong, uintptr_t, char*, int); 86 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 87 jlong, uintptr_t, char*, int); 88 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 89 jlong, uintptr_t, char*, int); 90 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 91 jlong, uintptr_t, char*, int); 92 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 93 jlong, uintptr_t, char*, int); 94 95 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 96 { \ 97 if (DTraceMonitorProbes) { \ 98 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 99 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 100 (monitor), bytes, len, (millis)); \ 101 } \ 102 } 103 104 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 105 { \ 106 if (DTraceMonitorProbes) { \ 107 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 108 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 109 (uintptr_t)(monitor), bytes, len); \ 110 } \ 111 } 112 113 #else /* USDT2 */ 114 115 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 116 { \ 117 if (DTraceMonitorProbes) { \ 118 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 119 HOTSPOT_MONITOR_WAIT(jtid, \ 120 (monitor), bytes, len, (millis)); \ 121 } \ 122 } 123 124 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 125 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 126 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 127 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 128 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 129 130 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 131 { \ 132 if (DTraceMonitorProbes) { \ 133 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 134 HOTSPOT_MONITOR_##probe(jtid, \ 135 (uintptr_t)(monitor), bytes, len); \ 136 } \ 137 } 138 139 #endif /* USDT2 */ 140 #else // ndef DTRACE_ENABLED 141 142 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 143 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 144 145 #endif // ndef DTRACE_ENABLED 146 147 // Tunables ... 148 // The knob* variables are effectively final. Once set they should 149 // never be modified hence. Consider using __read_mostly with GCC. 150 151 int ObjectMonitor::Knob_Verbose = 0 ; 152 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - 153 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 154 static int Knob_HandOff = 0 ; 155 static int Knob_ReportSettings = 0 ; 156 157 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 158 static int Knob_SpinBackOff = 0 ; // spin-loop backoff 159 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 160 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 161 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 162 static int Knob_SpinEarly = 1 ; 163 static int Knob_SuccEnabled = 1 ; // futile wake throttling 164 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 165 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 166 static int Knob_Bonus = 100 ; // spin success bonus 167 static int Knob_BonusB = 100 ; // spin success bonus 168 static int Knob_Penalty = 200 ; // spin failure penalty 169 static int Knob_Poverty = 1000 ; 170 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 171 static int Knob_FixedSpin = 0 ; 172 static int Knob_OState = 3 ; // Spinner checks thread state of _owner 173 static int Knob_UsePause = 1 ; 174 static int Knob_ExitPolicy = 0 ; 175 static int Knob_PreSpin = 10 ; // 20-100 likely better 176 static int Knob_ResetEvent = 0 ; 177 static int BackOffMask = 0 ; 178 179 static int Knob_FastHSSEC = 0 ; 180 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 181 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 182 static volatile int InitDone = 0 ; 183 184 #define TrySpin TrySpin_VaryDuration 185 186 // ----------------------------------------------------------------------------- 187 // Theory of operations -- Monitors lists, thread residency, etc: 188 // 189 // * A thread acquires ownership of a monitor by successfully 190 // CAS()ing the _owner field from null to non-null. 191 // 192 // * Invariant: A thread appears on at most one monitor list -- 193 // cxq, EntryList or WaitSet -- at any one time. 194 // 195 // * Contending threads "push" themselves onto the cxq with CAS 196 // and then spin/park. 197 // 198 // * After a contending thread eventually acquires the lock it must 199 // dequeue itself from either the EntryList or the cxq. 200 // 201 // * The exiting thread identifies and unparks an "heir presumptive" 202 // tentative successor thread on the EntryList. Critically, the 203 // exiting thread doesn't unlink the successor thread from the EntryList. 204 // After having been unparked, the wakee will recontend for ownership of 205 // the monitor. The successor (wakee) will either acquire the lock or 206 // re-park itself. 207 // 208 // Succession is provided for by a policy of competitive handoff. 209 // The exiting thread does _not_ grant or pass ownership to the 210 // successor thread. (This is also referred to as "handoff" succession"). 211 // Instead the exiting thread releases ownership and possibly wakes 212 // a successor, so the successor can (re)compete for ownership of the lock. 213 // If the EntryList is empty but the cxq is populated the exiting 214 // thread will drain the cxq into the EntryList. It does so by 215 // by detaching the cxq (installing null with CAS) and folding 216 // the threads from the cxq into the EntryList. The EntryList is 217 // doubly linked, while the cxq is singly linked because of the 218 // CAS-based "push" used to enqueue recently arrived threads (RATs). 219 // 220 // * Concurrency invariants: 221 // 222 // -- only the monitor owner may access or mutate the EntryList. 223 // The mutex property of the monitor itself protects the EntryList 224 // from concurrent interference. 225 // -- Only the monitor owner may detach the cxq. 226 // 227 // * The monitor entry list operations avoid locks, but strictly speaking 228 // they're not lock-free. Enter is lock-free, exit is not. 229 // For a description of 'Methods and apparatus providing non-blocking access 230 // to a resource,' see U.S. Pat. No. 7844973. 231 // 232 // * The cxq can have multiple concurrent "pushers" but only one concurrent 233 // detaching thread. This mechanism is immune from the ABA corruption. 234 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 235 // 236 // * Taken together, the cxq and the EntryList constitute or form a 237 // single logical queue of threads stalled trying to acquire the lock. 238 // We use two distinct lists to improve the odds of a constant-time 239 // dequeue operation after acquisition (in the ::enter() epilog) and 240 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 241 // A key desideratum is to minimize queue & monitor metadata manipulation 242 // that occurs while holding the monitor lock -- that is, we want to 243 // minimize monitor lock holds times. Note that even a small amount of 244 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 245 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 246 // locks and monitor metadata. 247 // 248 // Cxq points to the the set of Recently Arrived Threads attempting entry. 249 // Because we push threads onto _cxq with CAS, the RATs must take the form of 250 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 251 // the unlocking thread notices that EntryList is null but _cxq is != null. 252 // 253 // The EntryList is ordered by the prevailing queue discipline and 254 // can be organized in any convenient fashion, such as a doubly-linked list or 255 // a circular doubly-linked list. Critically, we want insert and delete operations 256 // to operate in constant-time. If we need a priority queue then something akin 257 // to Solaris' sleepq would work nicely. Viz., 258 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 259 // Queue discipline is enforced at ::exit() time, when the unlocking thread 260 // drains the cxq into the EntryList, and orders or reorders the threads on the 261 // EntryList accordingly. 262 // 263 // Barring "lock barging", this mechanism provides fair cyclic ordering, 264 // somewhat similar to an elevator-scan. 265 // 266 // * The monitor synchronization subsystem avoids the use of native 267 // synchronization primitives except for the narrow platform-specific 268 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 269 // the semantics of park-unpark. Put another way, this monitor implementation 270 // depends only on atomic operations and park-unpark. The monitor subsystem 271 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 272 // underlying OS manages the READY<->RUN transitions. 273 // 274 // * Waiting threads reside on the WaitSet list -- wait() puts 275 // the caller onto the WaitSet. 276 // 277 // * notify() or notifyAll() simply transfers threads from the WaitSet to 278 // either the EntryList or cxq. Subsequent exit() operations will 279 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 280 // it's likely the notifyee would simply impale itself on the lock held 281 // by the notifier. 282 // 283 // * An interesting alternative is to encode cxq as (List,LockByte) where 284 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 285 // variable, like _recursions, in the scheme. The threads or Events that form 286 // the list would have to be aligned in 256-byte addresses. A thread would 287 // try to acquire the lock or enqueue itself with CAS, but exiting threads 288 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 289 // Note that is is *not* word-tearing, but it does presume that full-word 290 // CAS operations are coherent with intermix with STB operations. That's true 291 // on most common processors. 292 // 293 // * See also http://blogs.sun.com/dave 294 295 296 // ----------------------------------------------------------------------------- 297 // Enter support 298 299 bool ObjectMonitor::try_enter(Thread* THREAD) { 300 if (THREAD != _owner) { 301 if (THREAD->is_lock_owned ((address)_owner)) { 302 assert(_recursions == 0, "internal state error"); 303 _owner = THREAD ; 304 _recursions = 1 ; 305 OwnerIsThread = 1 ; 306 return true; 307 } 308 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 309 return false; 310 } 311 return true; 312 } else { 313 _recursions++; 314 return true; 315 } 316 } 317 318 void ATTR ObjectMonitor::enter(TRAPS) { 319 // The following code is ordered to check the most common cases first 320 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 321 Thread * const Self = THREAD ; 322 void * cur ; 323 324 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 325 if (cur == NULL) { 326 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 327 assert (_recursions == 0 , "invariant") ; 328 assert (_owner == Self, "invariant") ; 329 // CONSIDER: set or assert OwnerIsThread == 1 330 return ; 331 } 332 333 if (cur == Self) { 334 // TODO-FIXME: check for integer overflow! BUGID 6557169. 335 _recursions ++ ; 336 return ; 337 } 338 339 if (Self->is_lock_owned ((address)cur)) { 340 assert (_recursions == 0, "internal state error"); 341 _recursions = 1 ; 342 // Commute owner from a thread-specific on-stack BasicLockObject address to 343 // a full-fledged "Thread *". 344 _owner = Self ; 345 OwnerIsThread = 1 ; 346 return ; 347 } 348 349 // We've encountered genuine contention. 350 assert (Self->_Stalled == 0, "invariant") ; 351 Self->_Stalled = intptr_t(this) ; 352 353 // Try one round of spinning *before* enqueueing Self 354 // and before going through the awkward and expensive state 355 // transitions. The following spin is strictly optional ... 356 // Note that if we acquire the monitor from an initial spin 357 // we forgo posting JVMTI events and firing DTRACE probes. 358 if (Knob_SpinEarly && TrySpin (Self) > 0) { 359 assert (_owner == Self , "invariant") ; 360 assert (_recursions == 0 , "invariant") ; 361 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 362 Self->_Stalled = 0 ; 363 return ; 364 } 365 366 assert (_owner != Self , "invariant") ; 367 assert (_succ != Self , "invariant") ; 368 assert (Self->is_Java_thread() , "invariant") ; 369 JavaThread * jt = (JavaThread *) Self ; 370 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 371 assert (jt->thread_state() != _thread_blocked , "invariant") ; 372 assert (this->object() != NULL , "invariant") ; 373 assert (_count >= 0, "invariant") ; 374 375 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 376 // Ensure the object-monitor relationship remains stable while there's contention. 377 Atomic::inc_ptr(&_count); 378 379 EventJavaMonitorEnter event; 380 381 { // Change java thread status to indicate blocked on monitor enter. 382 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 383 384 Self->set_current_pending_monitor(this); 385 386 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 387 if (JvmtiExport::should_post_monitor_contended_enter()) { 388 JvmtiExport::post_monitor_contended_enter(jt, this); 389 390 // The current thread does not yet own the monitor and does not 391 // yet appear on any queues that would get it made the successor. 392 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 393 // handler cannot accidentally consume an unpark() meant for the 394 // ParkEvent associated with this ObjectMonitor. 395 } 396 397 OSThreadContendState osts(Self->osthread()); 398 ThreadBlockInVM tbivm(jt); 399 400 // TODO-FIXME: change the following for(;;) loop to straight-line code. 401 for (;;) { 402 jt->set_suspend_equivalent(); 403 // cleared by handle_special_suspend_equivalent_condition() 404 // or java_suspend_self() 405 406 EnterI (THREAD) ; 407 408 if (!ExitSuspendEquivalent(jt)) break ; 409 410 // 411 // We have acquired the contended monitor, but while we were 412 // waiting another thread suspended us. We don't want to enter 413 // the monitor while suspended because that would surprise the 414 // thread that suspended us. 415 // 416 _recursions = 0 ; 417 _succ = NULL ; 418 exit (false, Self) ; 419 420 jt->java_suspend_self(); 421 } 422 Self->set_current_pending_monitor(NULL); 423 424 // We cleared the pending monitor info since we've just gotten past 425 // the enter-check-for-suspend dance and we now own the monitor free 426 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 427 // destructor can go to a safepoint at the end of this block. If we 428 // do a thread dump during that safepoint, then this thread will show 429 // as having "-locked" the monitor, but the OS and java.lang.Thread 430 // states will still report that the thread is blocked trying to 431 // acquire it. 432 } 433 434 Atomic::dec_ptr(&_count); 435 assert (_count >= 0, "invariant") ; 436 Self->_Stalled = 0 ; 437 438 // Must either set _recursions = 0 or ASSERT _recursions == 0. 439 assert (_recursions == 0 , "invariant") ; 440 assert (_owner == Self , "invariant") ; 441 assert (_succ != Self , "invariant") ; 442 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 443 444 // The thread -- now the owner -- is back in vm mode. 445 // Report the glorious news via TI,DTrace and jvmstat. 446 // The probe effect is non-trivial. All the reportage occurs 447 // while we hold the monitor, increasing the length of the critical 448 // section. Amdahl's parallel speedup law comes vividly into play. 449 // 450 // Another option might be to aggregate the events (thread local or 451 // per-monitor aggregation) and defer reporting until a more opportune 452 // time -- such as next time some thread encounters contention but has 453 // yet to acquire the lock. While spinning that thread could 454 // spinning we could increment JVMStat counters, etc. 455 456 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 457 if (JvmtiExport::should_post_monitor_contended_entered()) { 458 JvmtiExport::post_monitor_contended_entered(jt, this); 459 460 // The current thread already owns the monitor and is not going to 461 // call park() for the remainder of the monitor enter protocol. So 462 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 463 // event handler consumed an unpark() issued by the thread that 464 // just exited the monitor. 465 } 466 467 if (event.should_commit()) { 468 event.set_monitorClass(((oop)this->object())->klass()); 469 event.set_previousOwner((TYPE_THREAD)_previous_owner_tid); 470 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 471 event.commit(); 472 } 473 474 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 475 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; 476 } 477 } 478 479 480 // Caveat: TryLock() is not necessarily serializing if it returns failure. 481 // Callers must compensate as needed. 482 483 int ObjectMonitor::TryLock (Thread * Self) { 484 for (;;) { 485 void * own = _owner ; 486 if (own != NULL) return 0 ; 487 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 488 // Either guarantee _recursions == 0 or set _recursions = 0. 489 assert (_recursions == 0, "invariant") ; 490 assert (_owner == Self, "invariant") ; 491 // CONSIDER: set or assert that OwnerIsThread == 1 492 return 1 ; 493 } 494 // The lock had been free momentarily, but we lost the race to the lock. 495 // Interference -- the CAS failed. 496 // We can either return -1 or retry. 497 // Retry doesn't make as much sense because the lock was just acquired. 498 if (true) return -1 ; 499 } 500 } 501 502 void ATTR ObjectMonitor::EnterI (TRAPS) { 503 Thread * Self = THREAD ; 504 assert (Self->is_Java_thread(), "invariant") ; 505 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 506 507 // Try the lock - TATAS 508 if (TryLock (Self) > 0) { 509 assert (_succ != Self , "invariant") ; 510 assert (_owner == Self , "invariant") ; 511 assert (_Responsible != Self , "invariant") ; 512 return ; 513 } 514 515 DeferredInitialize () ; 516 517 // We try one round of spinning *before* enqueueing Self. 518 // 519 // If the _owner is ready but OFFPROC we could use a YieldTo() 520 // operation to donate the remainder of this thread's quantum 521 // to the owner. This has subtle but beneficial affinity 522 // effects. 523 524 if (TrySpin (Self) > 0) { 525 assert (_owner == Self , "invariant") ; 526 assert (_succ != Self , "invariant") ; 527 assert (_Responsible != Self , "invariant") ; 528 return ; 529 } 530 531 // The Spin failed -- Enqueue and park the thread ... 532 assert (_succ != Self , "invariant") ; 533 assert (_owner != Self , "invariant") ; 534 assert (_Responsible != Self , "invariant") ; 535 536 // Enqueue "Self" on ObjectMonitor's _cxq. 537 // 538 // Node acts as a proxy for Self. 539 // As an aside, if were to ever rewrite the synchronization code mostly 540 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 541 // Java objects. This would avoid awkward lifecycle and liveness issues, 542 // as well as eliminate a subset of ABA issues. 543 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 544 // 545 546 ObjectWaiter node(Self) ; 547 Self->_ParkEvent->reset() ; 548 node._prev = (ObjectWaiter *) 0xBAD ; 549 node.TState = ObjectWaiter::TS_CXQ ; 550 551 // Push "Self" onto the front of the _cxq. 552 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 553 // Note that spinning tends to reduce the rate at which threads 554 // enqueue and dequeue on EntryList|cxq. 555 ObjectWaiter * nxt ; 556 for (;;) { 557 node._next = nxt = _cxq ; 558 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 559 560 // Interference - the CAS failed because _cxq changed. Just retry. 561 // As an optional optimization we retry the lock. 562 if (TryLock (Self) > 0) { 563 assert (_succ != Self , "invariant") ; 564 assert (_owner == Self , "invariant") ; 565 assert (_Responsible != Self , "invariant") ; 566 return ; 567 } 568 } 569 570 // Check for cxq|EntryList edge transition to non-null. This indicates 571 // the onset of contention. While contention persists exiting threads 572 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 573 // operations revert to the faster 1-0 mode. This enter operation may interleave 574 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 575 // arrange for one of the contending thread to use a timed park() operations 576 // to detect and recover from the race. (Stranding is form of progress failure 577 // where the monitor is unlocked but all the contending threads remain parked). 578 // That is, at least one of the contended threads will periodically poll _owner. 579 // One of the contending threads will become the designated "Responsible" thread. 580 // The Responsible thread uses a timed park instead of a normal indefinite park 581 // operation -- it periodically wakes and checks for and recovers from potential 582 // strandings admitted by 1-0 exit operations. We need at most one Responsible 583 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 584 // be responsible for a monitor. 585 // 586 // Currently, one of the contended threads takes on the added role of "Responsible". 587 // A viable alternative would be to use a dedicated "stranding checker" thread 588 // that periodically iterated over all the threads (or active monitors) and unparked 589 // successors where there was risk of stranding. This would help eliminate the 590 // timer scalability issues we see on some platforms as we'd only have one thread 591 // -- the checker -- parked on a timer. 592 593 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 594 // Try to assume the role of responsible thread for the monitor. 595 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 596 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 597 } 598 599 // The lock have been released while this thread was occupied queueing 600 // itself onto _cxq. To close the race and avoid "stranding" and 601 // progress-liveness failure we must resample-retry _owner before parking. 602 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 603 // In this case the ST-MEMBAR is accomplished with CAS(). 604 // 605 // TODO: Defer all thread state transitions until park-time. 606 // Since state transitions are heavy and inefficient we'd like 607 // to defer the state transitions until absolutely necessary, 608 // and in doing so avoid some transitions ... 609 610 TEVENT (Inflated enter - Contention) ; 611 int nWakeups = 0 ; 612 int RecheckInterval = 1 ; 613 614 for (;;) { 615 616 if (TryLock (Self) > 0) break ; 617 assert (_owner != Self, "invariant") ; 618 619 if ((SyncFlags & 2) && _Responsible == NULL) { 620 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 621 } 622 623 // park self 624 if (_Responsible == Self || (SyncFlags & 1)) { 625 TEVENT (Inflated enter - park TIMED) ; 626 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 627 // Increase the RecheckInterval, but clamp the value. 628 RecheckInterval *= 8 ; 629 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 630 } else { 631 TEVENT (Inflated enter - park UNTIMED) ; 632 Self->_ParkEvent->park() ; 633 } 634 635 if (TryLock(Self) > 0) break ; 636 637 // The lock is still contested. 638 // Keep a tally of the # of futile wakeups. 639 // Note that the counter is not protected by a lock or updated by atomics. 640 // That is by design - we trade "lossy" counters which are exposed to 641 // races during updates for a lower probe effect. 642 TEVENT (Inflated enter - Futile wakeup) ; 643 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 644 ObjectMonitor::_sync_FutileWakeups->inc() ; 645 } 646 ++ nWakeups ; 647 648 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 649 // We can defer clearing _succ until after the spin completes 650 // TrySpin() must tolerate being called with _succ == Self. 651 // Try yet another round of adaptive spinning. 652 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 653 654 // We can find that we were unpark()ed and redesignated _succ while 655 // we were spinning. That's harmless. If we iterate and call park(), 656 // park() will consume the event and return immediately and we'll 657 // just spin again. This pattern can repeat, leaving _succ to simply 658 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 659 // Alternately, we can sample fired() here, and if set, forgo spinning 660 // in the next iteration. 661 662 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 663 Self->_ParkEvent->reset() ; 664 OrderAccess::fence() ; 665 } 666 if (_succ == Self) _succ = NULL ; 667 668 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 669 OrderAccess::fence() ; 670 } 671 672 // Egress : 673 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 674 // Normally we'll find Self on the EntryList . 675 // From the perspective of the lock owner (this thread), the 676 // EntryList is stable and cxq is prepend-only. 677 // The head of cxq is volatile but the interior is stable. 678 // In addition, Self.TState is stable. 679 680 assert (_owner == Self , "invariant") ; 681 assert (object() != NULL , "invariant") ; 682 // I'd like to write: 683 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 684 // but as we're at a safepoint that's not safe. 685 686 UnlinkAfterAcquire (Self, &node) ; 687 if (_succ == Self) _succ = NULL ; 688 689 assert (_succ != Self, "invariant") ; 690 if (_Responsible == Self) { 691 _Responsible = NULL ; 692 OrderAccess::fence(); // Dekker pivot-point 693 694 // We may leave threads on cxq|EntryList without a designated 695 // "Responsible" thread. This is benign. When this thread subsequently 696 // exits the monitor it can "see" such preexisting "old" threads -- 697 // threads that arrived on the cxq|EntryList before the fence, above -- 698 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 699 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 700 // non-null and elect a new "Responsible" timer thread. 701 // 702 // This thread executes: 703 // ST Responsible=null; MEMBAR (in enter epilog - here) 704 // LD cxq|EntryList (in subsequent exit) 705 // 706 // Entering threads in the slow/contended path execute: 707 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 708 // The (ST cxq; MEMBAR) is accomplished with CAS(). 709 // 710 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 711 // exit operation from floating above the ST Responsible=null. 712 } 713 714 // We've acquired ownership with CAS(). 715 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 716 // But since the CAS() this thread may have also stored into _succ, 717 // EntryList, cxq or Responsible. These meta-data updates must be 718 // visible __before this thread subsequently drops the lock. 719 // Consider what could occur if we didn't enforce this constraint -- 720 // STs to monitor meta-data and user-data could reorder with (become 721 // visible after) the ST in exit that drops ownership of the lock. 722 // Some other thread could then acquire the lock, but observe inconsistent 723 // or old monitor meta-data and heap data. That violates the JMM. 724 // To that end, the 1-0 exit() operation must have at least STST|LDST 725 // "release" barrier semantics. Specifically, there must be at least a 726 // STST|LDST barrier in exit() before the ST of null into _owner that drops 727 // the lock. The barrier ensures that changes to monitor meta-data and data 728 // protected by the lock will be visible before we release the lock, and 729 // therefore before some other thread (CPU) has a chance to acquire the lock. 730 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 731 // 732 // Critically, any prior STs to _succ or EntryList must be visible before 733 // the ST of null into _owner in the *subsequent* (following) corresponding 734 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 735 // execute a serializing instruction. 736 737 if (SyncFlags & 8) { 738 OrderAccess::fence() ; 739 } 740 return ; 741 } 742 743 // ReenterI() is a specialized inline form of the latter half of the 744 // contended slow-path from EnterI(). We use ReenterI() only for 745 // monitor reentry in wait(). 746 // 747 // In the future we should reconcile EnterI() and ReenterI(), adding 748 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 749 // loop accordingly. 750 751 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 752 assert (Self != NULL , "invariant") ; 753 assert (SelfNode != NULL , "invariant") ; 754 assert (SelfNode->_thread == Self , "invariant") ; 755 assert (_waiters > 0 , "invariant") ; 756 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 757 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 758 JavaThread * jt = (JavaThread *) Self ; 759 760 int nWakeups = 0 ; 761 for (;;) { 762 ObjectWaiter::TStates v = SelfNode->TState ; 763 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 764 assert (_owner != Self, "invariant") ; 765 766 if (TryLock (Self) > 0) break ; 767 if (TrySpin (Self) > 0) break ; 768 769 TEVENT (Wait Reentry - parking) ; 770 771 // State transition wrappers around park() ... 772 // ReenterI() wisely defers state transitions until 773 // it's clear we must park the thread. 774 { 775 OSThreadContendState osts(Self->osthread()); 776 ThreadBlockInVM tbivm(jt); 777 778 // cleared by handle_special_suspend_equivalent_condition() 779 // or java_suspend_self() 780 jt->set_suspend_equivalent(); 781 if (SyncFlags & 1) { 782 Self->_ParkEvent->park ((jlong)1000) ; 783 } else { 784 Self->_ParkEvent->park () ; 785 } 786 787 // were we externally suspended while we were waiting? 788 for (;;) { 789 if (!ExitSuspendEquivalent (jt)) break ; 790 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 791 jt->java_suspend_self(); 792 jt->set_suspend_equivalent(); 793 } 794 } 795 796 // Try again, but just so we distinguish between futile wakeups and 797 // successful wakeups. The following test isn't algorithmically 798 // necessary, but it helps us maintain sensible statistics. 799 if (TryLock(Self) > 0) break ; 800 801 // The lock is still contested. 802 // Keep a tally of the # of futile wakeups. 803 // Note that the counter is not protected by a lock or updated by atomics. 804 // That is by design - we trade "lossy" counters which are exposed to 805 // races during updates for a lower probe effect. 806 TEVENT (Wait Reentry - futile wakeup) ; 807 ++ nWakeups ; 808 809 // Assuming this is not a spurious wakeup we'll normally 810 // find that _succ == Self. 811 if (_succ == Self) _succ = NULL ; 812 813 // Invariant: after clearing _succ a contending thread 814 // *must* retry _owner before parking. 815 OrderAccess::fence() ; 816 817 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 818 ObjectMonitor::_sync_FutileWakeups->inc() ; 819 } 820 } 821 822 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 823 // Normally we'll find Self on the EntryList. 824 // Unlinking from the EntryList is constant-time and atomic-free. 825 // From the perspective of the lock owner (this thread), the 826 // EntryList is stable and cxq is prepend-only. 827 // The head of cxq is volatile but the interior is stable. 828 // In addition, Self.TState is stable. 829 830 assert (_owner == Self, "invariant") ; 831 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 832 UnlinkAfterAcquire (Self, SelfNode) ; 833 if (_succ == Self) _succ = NULL ; 834 assert (_succ != Self, "invariant") ; 835 SelfNode->TState = ObjectWaiter::TS_RUN ; 836 OrderAccess::fence() ; // see comments at the end of EnterI() 837 } 838 839 // after the thread acquires the lock in ::enter(). Equally, we could defer 840 // unlinking the thread until ::exit()-time. 841 842 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 843 { 844 assert (_owner == Self, "invariant") ; 845 assert (SelfNode->_thread == Self, "invariant") ; 846 847 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 848 // Normal case: remove Self from the DLL EntryList . 849 // This is a constant-time operation. 850 ObjectWaiter * nxt = SelfNode->_next ; 851 ObjectWaiter * prv = SelfNode->_prev ; 852 if (nxt != NULL) nxt->_prev = prv ; 853 if (prv != NULL) prv->_next = nxt ; 854 if (SelfNode == _EntryList ) _EntryList = nxt ; 855 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 856 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 857 TEVENT (Unlink from EntryList) ; 858 } else { 859 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 860 // Inopportune interleaving -- Self is still on the cxq. 861 // This usually means the enqueue of self raced an exiting thread. 862 // Normally we'll find Self near the front of the cxq, so 863 // dequeueing is typically fast. If needbe we can accelerate 864 // this with some MCS/CHL-like bidirectional list hints and advisory 865 // back-links so dequeueing from the interior will normally operate 866 // in constant-time. 867 // Dequeue Self from either the head (with CAS) or from the interior 868 // with a linear-time scan and normal non-atomic memory operations. 869 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 870 // and then unlink Self from EntryList. We have to drain eventually, 871 // so it might as well be now. 872 873 ObjectWaiter * v = _cxq ; 874 assert (v != NULL, "invariant") ; 875 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 876 // The CAS above can fail from interference IFF a "RAT" arrived. 877 // In that case Self must be in the interior and can no longer be 878 // at the head of cxq. 879 if (v == SelfNode) { 880 assert (_cxq != v, "invariant") ; 881 v = _cxq ; // CAS above failed - start scan at head of list 882 } 883 ObjectWaiter * p ; 884 ObjectWaiter * q = NULL ; 885 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 886 q = p ; 887 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 888 } 889 assert (v != SelfNode, "invariant") ; 890 assert (p == SelfNode, "Node not found on cxq") ; 891 assert (p != _cxq, "invariant") ; 892 assert (q != NULL, "invariant") ; 893 assert (q->_next == p, "invariant") ; 894 q->_next = p->_next ; 895 } 896 TEVENT (Unlink from cxq) ; 897 } 898 899 // Diagnostic hygiene ... 900 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 901 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 902 SelfNode->TState = ObjectWaiter::TS_RUN ; 903 } 904 905 // ----------------------------------------------------------------------------- 906 // Exit support 907 // 908 // exit() 909 // ~~~~~~ 910 // Note that the collector can't reclaim the objectMonitor or deflate 911 // the object out from underneath the thread calling ::exit() as the 912 // thread calling ::exit() never transitions to a stable state. 913 // This inhibits GC, which in turn inhibits asynchronous (and 914 // inopportune) reclamation of "this". 915 // 916 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 917 // There's one exception to the claim above, however. EnterI() can call 918 // exit() to drop a lock if the acquirer has been externally suspended. 919 // In that case exit() is called with _thread_state as _thread_blocked, 920 // but the monitor's _count field is > 0, which inhibits reclamation. 921 // 922 // 1-0 exit 923 // ~~~~~~~~ 924 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 925 // the fast-path operators have been optimized so the common ::exit() 926 // operation is 1-0. See i486.ad fast_unlock(), for instance. 927 // The code emitted by fast_unlock() elides the usual MEMBAR. This 928 // greatly improves latency -- MEMBAR and CAS having considerable local 929 // latency on modern processors -- but at the cost of "stranding". Absent the 930 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 931 // ::enter() path, resulting in the entering thread being stranding 932 // and a progress-liveness failure. Stranding is extremely rare. 933 // We use timers (timed park operations) & periodic polling to detect 934 // and recover from stranding. Potentially stranded threads periodically 935 // wake up and poll the lock. See the usage of the _Responsible variable. 936 // 937 // The CAS() in enter provides for safety and exclusion, while the CAS or 938 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 939 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 940 // We detect and recover from stranding with timers. 941 // 942 // If a thread transiently strands it'll park until (a) another 943 // thread acquires the lock and then drops the lock, at which time the 944 // exiting thread will notice and unpark the stranded thread, or, (b) 945 // the timer expires. If the lock is high traffic then the stranding latency 946 // will be low due to (a). If the lock is low traffic then the odds of 947 // stranding are lower, although the worst-case stranding latency 948 // is longer. Critically, we don't want to put excessive load in the 949 // platform's timer subsystem. We want to minimize both the timer injection 950 // rate (timers created/sec) as well as the number of timers active at 951 // any one time. (more precisely, we want to minimize timer-seconds, which is 952 // the integral of the # of active timers at any instant over time). 953 // Both impinge on OS scalability. Given that, at most one thread parked on 954 // a monitor will use a timer. 955 956 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 957 Thread * Self = THREAD ; 958 if (THREAD != _owner) { 959 if (THREAD->is_lock_owned((address) _owner)) { 960 // Transmute _owner from a BasicLock pointer to a Thread address. 961 // We don't need to hold _mutex for this transition. 962 // Non-null to Non-null is safe as long as all readers can 963 // tolerate either flavor. 964 assert (_recursions == 0, "invariant") ; 965 _owner = THREAD ; 966 _recursions = 0 ; 967 OwnerIsThread = 1 ; 968 } else { 969 // NOTE: we need to handle unbalanced monitor enter/exit 970 // in native code by throwing an exception. 971 // TODO: Throw an IllegalMonitorStateException ? 972 TEVENT (Exit - Throw IMSX) ; 973 assert(false, "Non-balanced monitor enter/exit!"); 974 if (false) { 975 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 976 } 977 return; 978 } 979 } 980 981 if (_recursions != 0) { 982 _recursions--; // this is simple recursive enter 983 TEVENT (Inflated exit - recursive) ; 984 return ; 985 } 986 987 // Invariant: after setting Responsible=null an thread must execute 988 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 989 if ((SyncFlags & 4) == 0) { 990 _Responsible = NULL ; 991 } 992 993 #if INCLUDE_TRACE 994 // get the owner's thread id for the MonitorEnter event 995 // if it is enabled and the thread isn't suspended 996 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 997 _previous_owner_tid = THREAD_TRACE_ID(Self); 998 } 999 #endif 1000 1001 for (;;) { 1002 assert (THREAD == _owner, "invariant") ; 1003 1004 1005 if (Knob_ExitPolicy == 0) { 1006 // release semantics: prior loads and stores from within the critical section 1007 // must not float (reorder) past the following store that drops the lock. 1008 // On SPARC that requires MEMBAR #loadstore|#storestore. 1009 // But of course in TSO #loadstore|#storestore is not required. 1010 // I'd like to write one of the following: 1011 // A. OrderAccess::release() ; _owner = NULL 1012 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 1013 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 1014 // store into a _dummy variable. That store is not needed, but can result 1015 // in massive wasteful coherency traffic on classic SMP systems. 1016 // Instead, I use release_store(), which is implemented as just a simple 1017 // ST on x64, x86 and SPARC. 1018 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1019 OrderAccess::storeload() ; // See if we need to wake a successor 1020 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1021 TEVENT (Inflated exit - simple egress) ; 1022 return ; 1023 } 1024 TEVENT (Inflated exit - complex egress) ; 1025 1026 // Normally the exiting thread is responsible for ensuring succession, 1027 // but if other successors are ready or other entering threads are spinning 1028 // then this thread can simply store NULL into _owner and exit without 1029 // waking a successor. The existence of spinners or ready successors 1030 // guarantees proper succession (liveness). Responsibility passes to the 1031 // ready or running successors. The exiting thread delegates the duty. 1032 // More precisely, if a successor already exists this thread is absolved 1033 // of the responsibility of waking (unparking) one. 1034 // 1035 // The _succ variable is critical to reducing futile wakeup frequency. 1036 // _succ identifies the "heir presumptive" thread that has been made 1037 // ready (unparked) but that has not yet run. We need only one such 1038 // successor thread to guarantee progress. 1039 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1040 // section 3.3 "Futile Wakeup Throttling" for details. 1041 // 1042 // Note that spinners in Enter() also set _succ non-null. 1043 // In the current implementation spinners opportunistically set 1044 // _succ so that exiting threads might avoid waking a successor. 1045 // Another less appealing alternative would be for the exiting thread 1046 // to drop the lock and then spin briefly to see if a spinner managed 1047 // to acquire the lock. If so, the exiting thread could exit 1048 // immediately without waking a successor, otherwise the exiting 1049 // thread would need to dequeue and wake a successor. 1050 // (Note that we'd need to make the post-drop spin short, but no 1051 // shorter than the worst-case round-trip cache-line migration time. 1052 // The dropped lock needs to become visible to the spinner, and then 1053 // the acquisition of the lock by the spinner must become visible to 1054 // the exiting thread). 1055 // 1056 1057 // It appears that an heir-presumptive (successor) must be made ready. 1058 // Only the current lock owner can manipulate the EntryList or 1059 // drain _cxq, so we need to reacquire the lock. If we fail 1060 // to reacquire the lock the responsibility for ensuring succession 1061 // falls to the new owner. 1062 // 1063 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1064 return ; 1065 } 1066 TEVENT (Exit - Reacquired) ; 1067 } else { 1068 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1069 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 1070 OrderAccess::storeload() ; 1071 // Ratify the previously observed values. 1072 if (_cxq == NULL || _succ != NULL) { 1073 TEVENT (Inflated exit - simple egress) ; 1074 return ; 1075 } 1076 1077 // inopportune interleaving -- the exiting thread (this thread) 1078 // in the fast-exit path raced an entering thread in the slow-enter 1079 // path. 1080 // We have two choices: 1081 // A. Try to reacquire the lock. 1082 // If the CAS() fails return immediately, otherwise 1083 // we either restart/rerun the exit operation, or simply 1084 // fall-through into the code below which wakes a successor. 1085 // B. If the elements forming the EntryList|cxq are TSM 1086 // we could simply unpark() the lead thread and return 1087 // without having set _succ. 1088 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1089 TEVENT (Inflated exit - reacquired succeeded) ; 1090 return ; 1091 } 1092 TEVENT (Inflated exit - reacquired failed) ; 1093 } else { 1094 TEVENT (Inflated exit - complex egress) ; 1095 } 1096 } 1097 1098 guarantee (_owner == THREAD, "invariant") ; 1099 1100 ObjectWaiter * w = NULL ; 1101 int QMode = Knob_QMode ; 1102 1103 if (QMode == 2 && _cxq != NULL) { 1104 // QMode == 2 : cxq has precedence over EntryList. 1105 // Try to directly wake a successor from the cxq. 1106 // If successful, the successor will need to unlink itself from cxq. 1107 w = _cxq ; 1108 assert (w != NULL, "invariant") ; 1109 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1110 ExitEpilog (Self, w) ; 1111 return ; 1112 } 1113 1114 if (QMode == 3 && _cxq != NULL) { 1115 // Aggressively drain cxq into EntryList at the first opportunity. 1116 // This policy ensure that recently-run threads live at the head of EntryList. 1117 // Drain _cxq into EntryList - bulk transfer. 1118 // First, detach _cxq. 1119 // The following loop is tantamount to: w = swap (&cxq, NULL) 1120 w = _cxq ; 1121 for (;;) { 1122 assert (w != NULL, "Invariant") ; 1123 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1124 if (u == w) break ; 1125 w = u ; 1126 } 1127 assert (w != NULL , "invariant") ; 1128 1129 ObjectWaiter * q = NULL ; 1130 ObjectWaiter * p ; 1131 for (p = w ; p != NULL ; p = p->_next) { 1132 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1133 p->TState = ObjectWaiter::TS_ENTER ; 1134 p->_prev = q ; 1135 q = p ; 1136 } 1137 1138 // Append the RATs to the EntryList 1139 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1140 ObjectWaiter * Tail ; 1141 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 1142 if (Tail == NULL) { 1143 _EntryList = w ; 1144 } else { 1145 Tail->_next = w ; 1146 w->_prev = Tail ; 1147 } 1148 1149 // Fall thru into code that tries to wake a successor from EntryList 1150 } 1151 1152 if (QMode == 4 && _cxq != NULL) { 1153 // Aggressively drain cxq into EntryList at the first opportunity. 1154 // This policy ensure that recently-run threads live at the head of EntryList. 1155 1156 // Drain _cxq into EntryList - bulk transfer. 1157 // First, detach _cxq. 1158 // The following loop is tantamount to: w = swap (&cxq, NULL) 1159 w = _cxq ; 1160 for (;;) { 1161 assert (w != NULL, "Invariant") ; 1162 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1163 if (u == w) break ; 1164 w = u ; 1165 } 1166 assert (w != NULL , "invariant") ; 1167 1168 ObjectWaiter * q = NULL ; 1169 ObjectWaiter * p ; 1170 for (p = w ; p != NULL ; p = p->_next) { 1171 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1172 p->TState = ObjectWaiter::TS_ENTER ; 1173 p->_prev = q ; 1174 q = p ; 1175 } 1176 1177 // Prepend the RATs to the EntryList 1178 if (_EntryList != NULL) { 1179 q->_next = _EntryList ; 1180 _EntryList->_prev = q ; 1181 } 1182 _EntryList = w ; 1183 1184 // Fall thru into code that tries to wake a successor from EntryList 1185 } 1186 1187 w = _EntryList ; 1188 if (w != NULL) { 1189 // I'd like to write: guarantee (w->_thread != Self). 1190 // But in practice an exiting thread may find itself on the EntryList. 1191 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1192 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1193 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1194 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1195 // release the lock "O". T2 resumes immediately after the ST of null into 1196 // _owner, above. T2 notices that the EntryList is populated, so it 1197 // reacquires the lock and then finds itself on the EntryList. 1198 // Given all that, we have to tolerate the circumstance where "w" is 1199 // associated with Self. 1200 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1201 ExitEpilog (Self, w) ; 1202 return ; 1203 } 1204 1205 // If we find that both _cxq and EntryList are null then just 1206 // re-run the exit protocol from the top. 1207 w = _cxq ; 1208 if (w == NULL) continue ; 1209 1210 // Drain _cxq into EntryList - bulk transfer. 1211 // First, detach _cxq. 1212 // The following loop is tantamount to: w = swap (&cxq, NULL) 1213 for (;;) { 1214 assert (w != NULL, "Invariant") ; 1215 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 1216 if (u == w) break ; 1217 w = u ; 1218 } 1219 TEVENT (Inflated exit - drain cxq into EntryList) ; 1220 1221 assert (w != NULL , "invariant") ; 1222 assert (_EntryList == NULL , "invariant") ; 1223 1224 // Convert the LIFO SLL anchored by _cxq into a DLL. 1225 // The list reorganization step operates in O(LENGTH(w)) time. 1226 // It's critical that this step operate quickly as 1227 // "Self" still holds the outer-lock, restricting parallelism 1228 // and effectively lengthening the critical section. 1229 // Invariant: s chases t chases u. 1230 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1231 // we have faster access to the tail. 1232 1233 if (QMode == 1) { 1234 // QMode == 1 : drain cxq to EntryList, reversing order 1235 // We also reverse the order of the list. 1236 ObjectWaiter * s = NULL ; 1237 ObjectWaiter * t = w ; 1238 ObjectWaiter * u = NULL ; 1239 while (t != NULL) { 1240 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 1241 t->TState = ObjectWaiter::TS_ENTER ; 1242 u = t->_next ; 1243 t->_prev = u ; 1244 t->_next = s ; 1245 s = t; 1246 t = u ; 1247 } 1248 _EntryList = s ; 1249 assert (s != NULL, "invariant") ; 1250 } else { 1251 // QMode == 0 or QMode == 2 1252 _EntryList = w ; 1253 ObjectWaiter * q = NULL ; 1254 ObjectWaiter * p ; 1255 for (p = w ; p != NULL ; p = p->_next) { 1256 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 1257 p->TState = ObjectWaiter::TS_ENTER ; 1258 p->_prev = q ; 1259 q = p ; 1260 } 1261 } 1262 1263 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1264 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1265 1266 // See if we can abdicate to a spinner instead of waking a thread. 1267 // A primary goal of the implementation is to reduce the 1268 // context-switch rate. 1269 if (_succ != NULL) continue; 1270 1271 w = _EntryList ; 1272 if (w != NULL) { 1273 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1274 ExitEpilog (Self, w) ; 1275 return ; 1276 } 1277 } 1278 } 1279 1280 // ExitSuspendEquivalent: 1281 // A faster alternate to handle_special_suspend_equivalent_condition() 1282 // 1283 // handle_special_suspend_equivalent_condition() unconditionally 1284 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1285 // operations have high latency. Note that in ::enter() we call HSSEC 1286 // while holding the monitor, so we effectively lengthen the critical sections. 1287 // 1288 // There are a number of possible solutions: 1289 // 1290 // A. To ameliorate the problem we might also defer state transitions 1291 // to as late as possible -- just prior to parking. 1292 // Given that, we'd call HSSEC after having returned from park(), 1293 // but before attempting to acquire the monitor. This is only a 1294 // partial solution. It avoids calling HSSEC while holding the 1295 // monitor (good), but it still increases successor reacquisition latency -- 1296 // the interval between unparking a successor and the time the successor 1297 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1298 // If we use this technique we can also avoid EnterI()-exit() loop 1299 // in ::enter() where we iteratively drop the lock and then attempt 1300 // to reacquire it after suspending. 1301 // 1302 // B. In the future we might fold all the suspend bits into a 1303 // composite per-thread suspend flag and then update it with CAS(). 1304 // Alternately, a Dekker-like mechanism with multiple variables 1305 // would suffice: 1306 // ST Self->_suspend_equivalent = false 1307 // MEMBAR 1308 // LD Self_>_suspend_flags 1309 // 1310 1311 1312 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1313 int Mode = Knob_FastHSSEC ; 1314 if (Mode && !jSelf->is_external_suspend()) { 1315 assert (jSelf->is_suspend_equivalent(), "invariant") ; 1316 jSelf->clear_suspend_equivalent() ; 1317 if (2 == Mode) OrderAccess::storeload() ; 1318 if (!jSelf->is_external_suspend()) return false ; 1319 // We raced a suspension -- fall thru into the slow path 1320 TEVENT (ExitSuspendEquivalent - raced) ; 1321 jSelf->set_suspend_equivalent() ; 1322 } 1323 return jSelf->handle_special_suspend_equivalent_condition() ; 1324 } 1325 1326 1327 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1328 assert (_owner == Self, "invariant") ; 1329 1330 // Exit protocol: 1331 // 1. ST _succ = wakee 1332 // 2. membar #loadstore|#storestore; 1333 // 2. ST _owner = NULL 1334 // 3. unpark(wakee) 1335 1336 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 1337 ParkEvent * Trigger = Wakee->_event ; 1338 1339 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1340 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1341 // out-of-scope (non-extant). 1342 Wakee = NULL ; 1343 1344 // Drop the lock 1345 OrderAccess::release_store_ptr (&_owner, NULL) ; 1346 OrderAccess::fence() ; // ST _owner vs LD in unpark() 1347 1348 if (SafepointSynchronize::do_call_back()) { 1349 TEVENT (unpark before SAFEPOINT) ; 1350 } 1351 1352 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1353 Trigger->unpark() ; 1354 1355 // Maintain stats and report events to JVMTI 1356 if (ObjectMonitor::_sync_Parks != NULL) { 1357 ObjectMonitor::_sync_Parks->inc() ; 1358 } 1359 } 1360 1361 1362 // ----------------------------------------------------------------------------- 1363 // Class Loader deadlock handling. 1364 // 1365 // complete_exit exits a lock returning recursion count 1366 // complete_exit/reenter operate as a wait without waiting 1367 // complete_exit requires an inflated monitor 1368 // The _owner field is not always the Thread addr even with an 1369 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1370 // thread due to contention. 1371 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1372 Thread * const Self = THREAD; 1373 assert(Self->is_Java_thread(), "Must be Java thread!"); 1374 JavaThread *jt = (JavaThread *)THREAD; 1375 1376 DeferredInitialize(); 1377 1378 if (THREAD != _owner) { 1379 if (THREAD->is_lock_owned ((address)_owner)) { 1380 assert(_recursions == 0, "internal state error"); 1381 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 1382 _recursions = 0 ; 1383 OwnerIsThread = 1 ; 1384 } 1385 } 1386 1387 guarantee(Self == _owner, "complete_exit not owner"); 1388 intptr_t save = _recursions; // record the old recursion count 1389 _recursions = 0; // set the recursion level to be 0 1390 exit (true, Self) ; // exit the monitor 1391 guarantee (_owner != Self, "invariant"); 1392 return save; 1393 } 1394 1395 // reenter() enters a lock and sets recursion count 1396 // complete_exit/reenter operate as a wait without waiting 1397 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1398 Thread * const Self = THREAD; 1399 assert(Self->is_Java_thread(), "Must be Java thread!"); 1400 JavaThread *jt = (JavaThread *)THREAD; 1401 1402 guarantee(_owner != Self, "reenter already owner"); 1403 enter (THREAD); // enter the monitor 1404 guarantee (_recursions == 0, "reenter recursion"); 1405 _recursions = recursions; 1406 return; 1407 } 1408 1409 1410 // ----------------------------------------------------------------------------- 1411 // A macro is used below because there may already be a pending 1412 // exception which should not abort the execution of the routines 1413 // which use this (which is why we don't put this into check_slow and 1414 // call it with a CHECK argument). 1415 1416 #define CHECK_OWNER() \ 1417 do { \ 1418 if (THREAD != _owner) { \ 1419 if (THREAD->is_lock_owned((address) _owner)) { \ 1420 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1421 _recursions = 0; \ 1422 OwnerIsThread = 1 ; \ 1423 } else { \ 1424 TEVENT (Throw IMSX) ; \ 1425 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1426 } \ 1427 } \ 1428 } while (false) 1429 1430 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1431 // TODO-FIXME: remove check_slow() -- it's likely dead. 1432 1433 void ObjectMonitor::check_slow(TRAPS) { 1434 TEVENT (check_slow - throw IMSX) ; 1435 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1436 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1437 } 1438 1439 static int Adjust (volatile int * adr, int dx) { 1440 int v ; 1441 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 1442 return v ; 1443 } 1444 1445 // helper method for posting a monitor wait event 1446 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1447 jlong notifier_tid, 1448 jlong timeout, 1449 bool timedout) { 1450 assert(event != NULL, "invariant"); 1451 event->set_monitorClass(((oop)this->object())->klass()); 1452 event->set_timeout(timeout); 1453 event->set_address((TYPE_ADDRESS)this->object_addr()); 1454 event->set_notifier(notifier_tid); 1455 event->set_timedOut(timedout); 1456 event->commit(); 1457 } 1458 1459 // ----------------------------------------------------------------------------- 1460 // Wait/Notify/NotifyAll 1461 // 1462 // Note: a subset of changes to ObjectMonitor::wait() 1463 // will need to be replicated in complete_exit above 1464 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1465 Thread * const Self = THREAD ; 1466 assert(Self->is_Java_thread(), "Must be Java thread!"); 1467 JavaThread *jt = (JavaThread *)THREAD; 1468 1469 DeferredInitialize () ; 1470 1471 // Throw IMSX or IEX. 1472 CHECK_OWNER(); 1473 1474 EventJavaMonitorWait event; 1475 1476 // check for a pending interrupt 1477 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1478 // post monitor waited event. Note that this is past-tense, we are done waiting. 1479 if (JvmtiExport::should_post_monitor_waited()) { 1480 // Note: 'false' parameter is passed here because the 1481 // wait was not timed out due to thread interrupt. 1482 JvmtiExport::post_monitor_waited(jt, this, false); 1483 1484 // In this short circuit of the monitor wait protocol, the 1485 // current thread never drops ownership of the monitor and 1486 // never gets added to the wait queue so the current thread 1487 // cannot be made the successor. This means that the 1488 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1489 // consume an unpark() meant for the ParkEvent associated with 1490 // this ObjectMonitor. 1491 } 1492 if (event.should_commit()) { 1493 post_monitor_wait_event(&event, 0, millis, false); 1494 } 1495 TEVENT (Wait - Throw IEX) ; 1496 THROW(vmSymbols::java_lang_InterruptedException()); 1497 return ; 1498 } 1499 1500 TEVENT (Wait) ; 1501 1502 assert (Self->_Stalled == 0, "invariant") ; 1503 Self->_Stalled = intptr_t(this) ; 1504 jt->set_current_waiting_monitor(this); 1505 1506 // create a node to be put into the queue 1507 // Critically, after we reset() the event but prior to park(), we must check 1508 // for a pending interrupt. 1509 ObjectWaiter node(Self); 1510 node.TState = ObjectWaiter::TS_WAIT ; 1511 Self->_ParkEvent->reset() ; 1512 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1513 1514 // Enter the waiting queue, which is a circular doubly linked list in this case 1515 // but it could be a priority queue or any data structure. 1516 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1517 // by the the owner of the monitor *except* in the case where park() 1518 // returns because of a timeout of interrupt. Contention is exceptionally rare 1519 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1520 1521 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 1522 AddWaiter (&node) ; 1523 Thread::SpinRelease (&_WaitSetLock) ; 1524 1525 if ((SyncFlags & 4) == 0) { 1526 _Responsible = NULL ; 1527 } 1528 intptr_t save = _recursions; // record the old recursion count 1529 _waiters++; // increment the number of waiters 1530 _recursions = 0; // set the recursion level to be 1 1531 exit (true, Self) ; // exit the monitor 1532 guarantee (_owner != Self, "invariant") ; 1533 1534 // The thread is on the WaitSet list - now park() it. 1535 // On MP systems it's conceivable that a brief spin before we park 1536 // could be profitable. 1537 // 1538 // TODO-FIXME: change the following logic to a loop of the form 1539 // while (!timeout && !interrupted && _notified == 0) park() 1540 1541 int ret = OS_OK ; 1542 int WasNotified = 0 ; 1543 { // State transition wrappers 1544 OSThread* osthread = Self->osthread(); 1545 OSThreadWaitState osts(osthread, true); 1546 { 1547 ThreadBlockInVM tbivm(jt); 1548 // Thread is in thread_blocked state and oop access is unsafe. 1549 jt->set_suspend_equivalent(); 1550 1551 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1552 // Intentionally empty 1553 } else 1554 if (node._notified == 0) { 1555 if (millis <= 0) { 1556 Self->_ParkEvent->park () ; 1557 } else { 1558 ret = Self->_ParkEvent->park (millis) ; 1559 } 1560 } 1561 1562 // were we externally suspended while we were waiting? 1563 if (ExitSuspendEquivalent (jt)) { 1564 // TODO-FIXME: add -- if succ == Self then succ = null. 1565 jt->java_suspend_self(); 1566 } 1567 1568 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1569 1570 1571 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1572 // from the WaitSet to the EntryList. 1573 // See if we need to remove Node from the WaitSet. 1574 // We use double-checked locking to avoid grabbing _WaitSetLock 1575 // if the thread is not on the wait queue. 1576 // 1577 // Note that we don't need a fence before the fetch of TState. 1578 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1579 // written by the is thread. (perhaps the fetch might even be satisfied 1580 // by a look-aside into the processor's own store buffer, although given 1581 // the length of the code path between the prior ST and this load that's 1582 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1583 // then we'll acquire the lock and then re-fetch a fresh TState value. 1584 // That is, we fail toward safety. 1585 1586 if (node.TState == ObjectWaiter::TS_WAIT) { 1587 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 1588 if (node.TState == ObjectWaiter::TS_WAIT) { 1589 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 1590 assert(node._notified == 0, "invariant"); 1591 node.TState = ObjectWaiter::TS_RUN ; 1592 } 1593 Thread::SpinRelease (&_WaitSetLock) ; 1594 } 1595 1596 // The thread is now either on off-list (TS_RUN), 1597 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1598 // The Node's TState variable is stable from the perspective of this thread. 1599 // No other threads will asynchronously modify TState. 1600 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 1601 OrderAccess::loadload() ; 1602 if (_succ == Self) _succ = NULL ; 1603 WasNotified = node._notified ; 1604 1605 // Reentry phase -- reacquire the monitor. 1606 // re-enter contended monitor after object.wait(). 1607 // retain OBJECT_WAIT state until re-enter successfully completes 1608 // Thread state is thread_in_vm and oop access is again safe, 1609 // although the raw address of the object may have changed. 1610 // (Don't cache naked oops over safepoints, of course). 1611 1612 // post monitor waited event. Note that this is past-tense, we are done waiting. 1613 if (JvmtiExport::should_post_monitor_waited()) { 1614 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1615 1616 if (node._notified != 0 && _succ == Self) { 1617 // In this part of the monitor wait-notify-reenter protocol it 1618 // is possible (and normal) for another thread to do a fastpath 1619 // monitor enter-exit while this thread is still trying to get 1620 // to the reenter portion of the protocol. 1621 // 1622 // The ObjectMonitor was notified and the current thread is 1623 // the successor which also means that an unpark() has already 1624 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1625 // consume the unpark() that was done when the successor was 1626 // set because the same ParkEvent is shared between Java 1627 // monitors and JVM/TI RawMonitors (for now). 1628 // 1629 // We redo the unpark() to ensure forward progress, i.e., we 1630 // don't want all pending threads hanging (parked) with none 1631 // entering the unlocked monitor. 1632 node._event->unpark(); 1633 } 1634 } 1635 1636 if (event.should_commit()) { 1637 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1638 } 1639 1640 OrderAccess::fence() ; 1641 1642 assert (Self->_Stalled != 0, "invariant") ; 1643 Self->_Stalled = 0 ; 1644 1645 assert (_owner != Self, "invariant") ; 1646 ObjectWaiter::TStates v = node.TState ; 1647 if (v == ObjectWaiter::TS_RUN) { 1648 enter (Self) ; 1649 } else { 1650 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 1651 ReenterI (Self, &node) ; 1652 node.wait_reenter_end(this); 1653 } 1654 1655 // Self has reacquired the lock. 1656 // Lifecycle - the node representing Self must not appear on any queues. 1657 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1658 // want residual elements associated with this thread left on any lists. 1659 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 1660 assert (_owner == Self, "invariant") ; 1661 assert (_succ != Self , "invariant") ; 1662 } // OSThreadWaitState() 1663 1664 jt->set_current_waiting_monitor(NULL); 1665 1666 guarantee (_recursions == 0, "invariant") ; 1667 _recursions = save; // restore the old recursion count 1668 _waiters--; // decrement the number of waiters 1669 1670 // Verify a few postconditions 1671 assert (_owner == Self , "invariant") ; 1672 assert (_succ != Self , "invariant") ; 1673 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 1674 1675 if (SyncFlags & 32) { 1676 OrderAccess::fence() ; 1677 } 1678 1679 // check if the notification happened 1680 if (!WasNotified) { 1681 // no, it could be timeout or Thread.interrupt() or both 1682 // check for interrupt event, otherwise it is timeout 1683 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1684 TEVENT (Wait - throw IEX from epilog) ; 1685 THROW(vmSymbols::java_lang_InterruptedException()); 1686 } 1687 } 1688 1689 // NOTE: Spurious wake up will be consider as timeout. 1690 // Monitor notify has precedence over thread interrupt. 1691 } 1692 1693 1694 // Consider: 1695 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1696 // then instead of transferring a thread from the WaitSet to the EntryList 1697 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1698 1699 void ObjectMonitor::notify(TRAPS) { 1700 CHECK_OWNER(); 1701 if (_WaitSet == NULL) { 1702 TEVENT (Empty-Notify) ; 1703 return ; 1704 } 1705 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1706 1707 int Policy = Knob_MoveNotifyee ; 1708 1709 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 1710 ObjectWaiter * iterator = DequeueWaiter() ; 1711 if (iterator != NULL) { 1712 TEVENT (Notify1 - Transfer) ; 1713 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1714 guarantee (iterator->_notified == 0, "invariant") ; 1715 if (Policy != 4) { 1716 iterator->TState = ObjectWaiter::TS_ENTER ; 1717 } 1718 iterator->_notified = 1 ; 1719 Thread * Self = THREAD; 1720 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1721 1722 ObjectWaiter * List = _EntryList ; 1723 if (List != NULL) { 1724 assert (List->_prev == NULL, "invariant") ; 1725 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1726 assert (List != iterator, "invariant") ; 1727 } 1728 1729 if (Policy == 0) { // prepend to EntryList 1730 if (List == NULL) { 1731 iterator->_next = iterator->_prev = NULL ; 1732 _EntryList = iterator ; 1733 } else { 1734 List->_prev = iterator ; 1735 iterator->_next = List ; 1736 iterator->_prev = NULL ; 1737 _EntryList = iterator ; 1738 } 1739 } else 1740 if (Policy == 1) { // append to EntryList 1741 if (List == NULL) { 1742 iterator->_next = iterator->_prev = NULL ; 1743 _EntryList = iterator ; 1744 } else { 1745 // CONSIDER: finding the tail currently requires a linear-time walk of 1746 // the EntryList. We can make tail access constant-time by converting to 1747 // a CDLL instead of using our current DLL. 1748 ObjectWaiter * Tail ; 1749 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1750 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1751 Tail->_next = iterator ; 1752 iterator->_prev = Tail ; 1753 iterator->_next = NULL ; 1754 } 1755 } else 1756 if (Policy == 2) { // prepend to cxq 1757 // prepend to cxq 1758 if (List == NULL) { 1759 iterator->_next = iterator->_prev = NULL ; 1760 _EntryList = iterator ; 1761 } else { 1762 iterator->TState = ObjectWaiter::TS_CXQ ; 1763 for (;;) { 1764 ObjectWaiter * Front = _cxq ; 1765 iterator->_next = Front ; 1766 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1767 break ; 1768 } 1769 } 1770 } 1771 } else 1772 if (Policy == 3) { // append to cxq 1773 iterator->TState = ObjectWaiter::TS_CXQ ; 1774 for (;;) { 1775 ObjectWaiter * Tail ; 1776 Tail = _cxq ; 1777 if (Tail == NULL) { 1778 iterator->_next = NULL ; 1779 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1780 break ; 1781 } 1782 } else { 1783 while (Tail->_next != NULL) Tail = Tail->_next ; 1784 Tail->_next = iterator ; 1785 iterator->_prev = Tail ; 1786 iterator->_next = NULL ; 1787 break ; 1788 } 1789 } 1790 } else { 1791 ParkEvent * ev = iterator->_event ; 1792 iterator->TState = ObjectWaiter::TS_RUN ; 1793 OrderAccess::fence() ; 1794 ev->unpark() ; 1795 } 1796 1797 if (Policy < 4) { 1798 iterator->wait_reenter_begin(this); 1799 } 1800 1801 // _WaitSetLock protects the wait queue, not the EntryList. We could 1802 // move the add-to-EntryList operation, above, outside the critical section 1803 // protected by _WaitSetLock. In practice that's not useful. With the 1804 // exception of wait() timeouts and interrupts the monitor owner 1805 // is the only thread that grabs _WaitSetLock. There's almost no contention 1806 // on _WaitSetLock so it's not profitable to reduce the length of the 1807 // critical section. 1808 } 1809 1810 Thread::SpinRelease (&_WaitSetLock) ; 1811 1812 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1813 ObjectMonitor::_sync_Notifications->inc() ; 1814 } 1815 } 1816 1817 1818 void ObjectMonitor::notifyAll(TRAPS) { 1819 CHECK_OWNER(); 1820 ObjectWaiter* iterator; 1821 if (_WaitSet == NULL) { 1822 TEVENT (Empty-NotifyAll) ; 1823 return ; 1824 } 1825 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1826 1827 int Policy = Knob_MoveNotifyee ; 1828 int Tally = 0 ; 1829 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 1830 1831 for (;;) { 1832 iterator = DequeueWaiter () ; 1833 if (iterator == NULL) break ; 1834 TEVENT (NotifyAll - Transfer1) ; 1835 ++Tally ; 1836 1837 // Disposition - what might we do with iterator ? 1838 // a. add it directly to the EntryList - either tail or head. 1839 // b. push it onto the front of the _cxq. 1840 // For now we use (a). 1841 1842 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 1843 guarantee (iterator->_notified == 0, "invariant") ; 1844 iterator->_notified = 1 ; 1845 Thread * Self = THREAD; 1846 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1847 if (Policy != 4) { 1848 iterator->TState = ObjectWaiter::TS_ENTER ; 1849 } 1850 1851 ObjectWaiter * List = _EntryList ; 1852 if (List != NULL) { 1853 assert (List->_prev == NULL, "invariant") ; 1854 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 1855 assert (List != iterator, "invariant") ; 1856 } 1857 1858 if (Policy == 0) { // prepend to EntryList 1859 if (List == NULL) { 1860 iterator->_next = iterator->_prev = NULL ; 1861 _EntryList = iterator ; 1862 } else { 1863 List->_prev = iterator ; 1864 iterator->_next = List ; 1865 iterator->_prev = NULL ; 1866 _EntryList = iterator ; 1867 } 1868 } else 1869 if (Policy == 1) { // append to EntryList 1870 if (List == NULL) { 1871 iterator->_next = iterator->_prev = NULL ; 1872 _EntryList = iterator ; 1873 } else { 1874 // CONSIDER: finding the tail currently requires a linear-time walk of 1875 // the EntryList. We can make tail access constant-time by converting to 1876 // a CDLL instead of using our current DLL. 1877 ObjectWaiter * Tail ; 1878 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 1879 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 1880 Tail->_next = iterator ; 1881 iterator->_prev = Tail ; 1882 iterator->_next = NULL ; 1883 } 1884 } else 1885 if (Policy == 2) { // prepend to cxq 1886 // prepend to cxq 1887 iterator->TState = ObjectWaiter::TS_CXQ ; 1888 for (;;) { 1889 ObjectWaiter * Front = _cxq ; 1890 iterator->_next = Front ; 1891 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1892 break ; 1893 } 1894 } 1895 } else 1896 if (Policy == 3) { // append to cxq 1897 iterator->TState = ObjectWaiter::TS_CXQ ; 1898 for (;;) { 1899 ObjectWaiter * Tail ; 1900 Tail = _cxq ; 1901 if (Tail == NULL) { 1902 iterator->_next = NULL ; 1903 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1904 break ; 1905 } 1906 } else { 1907 while (Tail->_next != NULL) Tail = Tail->_next ; 1908 Tail->_next = iterator ; 1909 iterator->_prev = Tail ; 1910 iterator->_next = NULL ; 1911 break ; 1912 } 1913 } 1914 } else { 1915 ParkEvent * ev = iterator->_event ; 1916 iterator->TState = ObjectWaiter::TS_RUN ; 1917 OrderAccess::fence() ; 1918 ev->unpark() ; 1919 } 1920 1921 if (Policy < 4) { 1922 iterator->wait_reenter_begin(this); 1923 } 1924 1925 // _WaitSetLock protects the wait queue, not the EntryList. We could 1926 // move the add-to-EntryList operation, above, outside the critical section 1927 // protected by _WaitSetLock. In practice that's not useful. With the 1928 // exception of wait() timeouts and interrupts the monitor owner 1929 // is the only thread that grabs _WaitSetLock. There's almost no contention 1930 // on _WaitSetLock so it's not profitable to reduce the length of the 1931 // critical section. 1932 } 1933 1934 Thread::SpinRelease (&_WaitSetLock) ; 1935 1936 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1937 ObjectMonitor::_sync_Notifications->inc(Tally) ; 1938 } 1939 } 1940 1941 // ----------------------------------------------------------------------------- 1942 // Adaptive Spinning Support 1943 // 1944 // Adaptive spin-then-block - rational spinning 1945 // 1946 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1947 // algorithm. On high order SMP systems it would be better to start with 1948 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1949 // a contending thread could enqueue itself on the cxq and then spin locally 1950 // on a thread-specific variable such as its ParkEvent._Event flag. 1951 // That's left as an exercise for the reader. Note that global spinning is 1952 // not problematic on Niagara, as the L2$ serves the interconnect and has both 1953 // low latency and massive bandwidth. 1954 // 1955 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1956 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1957 // (duration) or we can fix the count at approximately the duration of 1958 // a context switch and vary the frequency. Of course we could also 1959 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1960 // For a description of 'Adaptive spin-then-block mutual exclusion in 1961 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1962 // 1963 // This implementation varies the duration "D", where D varies with 1964 // the success rate of recent spin attempts. (D is capped at approximately 1965 // length of a round-trip context switch). The success rate for recent 1966 // spin attempts is a good predictor of the success rate of future spin 1967 // attempts. The mechanism adapts automatically to varying critical 1968 // section length (lock modality), system load and degree of parallelism. 1969 // D is maintained per-monitor in _SpinDuration and is initialized 1970 // optimistically. Spin frequency is fixed at 100%. 1971 // 1972 // Note that _SpinDuration is volatile, but we update it without locks 1973 // or atomics. The code is designed so that _SpinDuration stays within 1974 // a reasonable range even in the presence of races. The arithmetic 1975 // operations on _SpinDuration are closed over the domain of legal values, 1976 // so at worst a race will install and older but still legal value. 1977 // At the very worst this introduces some apparent non-determinism. 1978 // We might spin when we shouldn't or vice-versa, but since the spin 1979 // count are relatively short, even in the worst case, the effect is harmless. 1980 // 1981 // Care must be taken that a low "D" value does not become an 1982 // an absorbing state. Transient spinning failures -- when spinning 1983 // is overall profitable -- should not cause the system to converge 1984 // on low "D" values. We want spinning to be stable and predictable 1985 // and fairly responsive to change and at the same time we don't want 1986 // it to oscillate, become metastable, be "too" non-deterministic, 1987 // or converge on or enter undesirable stable absorbing states. 1988 // 1989 // We implement a feedback-based control system -- using past behavior 1990 // to predict future behavior. We face two issues: (a) if the 1991 // input signal is random then the spin predictor won't provide optimal 1992 // results, and (b) if the signal frequency is too high then the control 1993 // system, which has some natural response lag, will "chase" the signal. 1994 // (b) can arise from multimodal lock hold times. Transient preemption 1995 // can also result in apparent bimodal lock hold times. 1996 // Although sub-optimal, neither condition is particularly harmful, as 1997 // in the worst-case we'll spin when we shouldn't or vice-versa. 1998 // The maximum spin duration is rather short so the failure modes aren't bad. 1999 // To be conservative, I've tuned the gain in system to bias toward 2000 // _not spinning. Relatedly, the system can sometimes enter a mode where it 2001 // "rings" or oscillates between spinning and not spinning. This happens 2002 // when spinning is just on the cusp of profitability, however, so the 2003 // situation is not dire. The state is benign -- there's no need to add 2004 // hysteresis control to damp the transition rate between spinning and 2005 // not spinning. 2006 // 2007 2008 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 2009 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 2010 2011 // Spinning: Fixed frequency (100%), vary duration 2012 2013 2014 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 2015 2016 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2017 int ctr = Knob_FixedSpin ; 2018 if (ctr != 0) { 2019 while (--ctr >= 0) { 2020 if (TryLock (Self) > 0) return 1 ; 2021 SpinPause () ; 2022 } 2023 return 0 ; 2024 } 2025 2026 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 2027 if (TryLock(Self) > 0) { 2028 // Increase _SpinDuration ... 2029 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2030 // Raising _SpurDuration to the poverty line is key. 2031 int x = _SpinDuration ; 2032 if (x < Knob_SpinLimit) { 2033 if (x < Knob_Poverty) x = Knob_Poverty ; 2034 _SpinDuration = x + Knob_BonusB ; 2035 } 2036 return 1 ; 2037 } 2038 SpinPause () ; 2039 } 2040 2041 // Admission control - verify preconditions for spinning 2042 // 2043 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2044 // becoming an absorbing state. Put another way, we spin briefly to 2045 // sample, just in case the system load, parallelism, contention, or lock 2046 // modality changed. 2047 // 2048 // Consider the following alternative: 2049 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2050 // spin attempt. "Periodically" might mean after a tally of 2051 // the # of failed spin attempts (or iterations) reaches some threshold. 2052 // This takes us into the realm of 1-out-of-N spinning, where we 2053 // hold the duration constant but vary the frequency. 2054 2055 ctr = _SpinDuration ; 2056 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2057 if (ctr <= 0) return 0 ; 2058 2059 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2060 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2061 TEVENT (Spin abort - notrunnable [TOP]); 2062 return 0 ; 2063 } 2064 2065 int MaxSpin = Knob_MaxSpinners ; 2066 if (MaxSpin >= 0) { 2067 if (_Spinner > MaxSpin) { 2068 TEVENT (Spin abort -- too many spinners) ; 2069 return 0 ; 2070 } 2071 // Slighty racy, but benign ... 2072 Adjust (&_Spinner, 1) ; 2073 } 2074 2075 // We're good to spin ... spin ingress. 2076 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2077 // when preparing to LD...CAS _owner, etc and the CAS is likely 2078 // to succeed. 2079 int hits = 0 ; 2080 int msk = 0 ; 2081 int caspty = Knob_CASPenalty ; 2082 int oxpty = Knob_OXPenalty ; 2083 int sss = Knob_SpinSetSucc ; 2084 if (sss && _succ == NULL ) _succ = Self ; 2085 Thread * prv = NULL ; 2086 2087 // There are three ways to exit the following loop: 2088 // 1. A successful spin where this thread has acquired the lock. 2089 // 2. Spin failure with prejudice 2090 // 3. Spin failure without prejudice 2091 2092 while (--ctr >= 0) { 2093 2094 // Periodic polling -- Check for pending GC 2095 // Threads may spin while they're unsafe. 2096 // We don't want spinning threads to delay the JVM from reaching 2097 // a stop-the-world safepoint or to steal cycles from GC. 2098 // If we detect a pending safepoint we abort in order that 2099 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2100 // this thread, if safe, doesn't steal cycles from GC. 2101 // This is in keeping with the "no loitering in runtime" rule. 2102 // We periodically check to see if there's a safepoint pending. 2103 if ((ctr & 0xFF) == 0) { 2104 if (SafepointSynchronize::do_call_back()) { 2105 TEVENT (Spin: safepoint) ; 2106 goto Abort ; // abrupt spin egress 2107 } 2108 if (Knob_UsePause & 1) SpinPause () ; 2109 2110 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2111 if (hits > 50 && scb != NULL) { 2112 int abend = (*scb)(SpinCallbackArgument, 0) ; 2113 } 2114 } 2115 2116 if (Knob_UsePause & 2) SpinPause() ; 2117 2118 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2119 // This is useful on classic SMP systems, but is of less utility on 2120 // N1-style CMT platforms. 2121 // 2122 // Trade-off: lock acquisition latency vs coherency bandwidth. 2123 // Lock hold times are typically short. A histogram 2124 // of successful spin attempts shows that we usually acquire 2125 // the lock early in the spin. That suggests we want to 2126 // sample _owner frequently in the early phase of the spin, 2127 // but then back-off and sample less frequently as the spin 2128 // progresses. The back-off makes a good citizen on SMP big 2129 // SMP systems. Oversampling _owner can consume excessive 2130 // coherency bandwidth. Relatedly, if we _oversample _owner we 2131 // can inadvertently interfere with the the ST m->owner=null. 2132 // executed by the lock owner. 2133 if (ctr & msk) continue ; 2134 ++hits ; 2135 if ((hits & 0xF) == 0) { 2136 // The 0xF, above, corresponds to the exponent. 2137 // Consider: (msk+1)|msk 2138 msk = ((msk << 2)|3) & BackOffMask ; 2139 } 2140 2141 // Probe _owner with TATAS 2142 // If this thread observes the monitor transition or flicker 2143 // from locked to unlocked to locked, then the odds that this 2144 // thread will acquire the lock in this spin attempt go down 2145 // considerably. The same argument applies if the CAS fails 2146 // or if we observe _owner change from one non-null value to 2147 // another non-null value. In such cases we might abort 2148 // the spin without prejudice or apply a "penalty" to the 2149 // spin count-down variable "ctr", reducing it by 100, say. 2150 2151 Thread * ox = (Thread *) _owner ; 2152 if (ox == NULL) { 2153 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2154 if (ox == NULL) { 2155 // The CAS succeeded -- this thread acquired ownership 2156 // Take care of some bookkeeping to exit spin state. 2157 if (sss && _succ == Self) { 2158 _succ = NULL ; 2159 } 2160 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2161 2162 // Increase _SpinDuration : 2163 // The spin was successful (profitable) so we tend toward 2164 // longer spin attempts in the future. 2165 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2166 // If we acquired the lock early in the spin cycle it 2167 // makes sense to increase _SpinDuration proportionally. 2168 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2169 int x = _SpinDuration ; 2170 if (x < Knob_SpinLimit) { 2171 if (x < Knob_Poverty) x = Knob_Poverty ; 2172 _SpinDuration = x + Knob_Bonus ; 2173 } 2174 return 1 ; 2175 } 2176 2177 // The CAS failed ... we can take any of the following actions: 2178 // * penalize: ctr -= Knob_CASPenalty 2179 // * exit spin with prejudice -- goto Abort; 2180 // * exit spin without prejudice. 2181 // * Since CAS is high-latency, retry again immediately. 2182 prv = ox ; 2183 TEVENT (Spin: cas failed) ; 2184 if (caspty == -2) break ; 2185 if (caspty == -1) goto Abort ; 2186 ctr -= caspty ; 2187 continue ; 2188 } 2189 2190 // Did lock ownership change hands ? 2191 if (ox != prv && prv != NULL ) { 2192 TEVENT (spin: Owner changed) 2193 if (oxpty == -2) break ; 2194 if (oxpty == -1) goto Abort ; 2195 ctr -= oxpty ; 2196 } 2197 prv = ox ; 2198 2199 // Abort the spin if the owner is not executing. 2200 // The owner must be executing in order to drop the lock. 2201 // Spinning while the owner is OFFPROC is idiocy. 2202 // Consider: ctr -= RunnablePenalty ; 2203 if (Knob_OState && NotRunnable (Self, ox)) { 2204 TEVENT (Spin abort - notrunnable); 2205 goto Abort ; 2206 } 2207 if (sss && _succ == NULL ) _succ = Self ; 2208 } 2209 2210 // Spin failed with prejudice -- reduce _SpinDuration. 2211 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2212 // AIMD is globally stable. 2213 TEVENT (Spin failure) ; 2214 { 2215 int x = _SpinDuration ; 2216 if (x > 0) { 2217 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2218 // This is globally sample and tends to damp the response. 2219 x -= Knob_Penalty ; 2220 if (x < 0) x = 0 ; 2221 _SpinDuration = x ; 2222 } 2223 } 2224 2225 Abort: 2226 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2227 if (sss && _succ == Self) { 2228 _succ = NULL ; 2229 // Invariant: after setting succ=null a contending thread 2230 // must recheck-retry _owner before parking. This usually happens 2231 // in the normal usage of TrySpin(), but it's safest 2232 // to make TrySpin() as foolproof as possible. 2233 OrderAccess::fence() ; 2234 if (TryLock(Self) > 0) return 1 ; 2235 } 2236 return 0 ; 2237 } 2238 2239 // NotRunnable() -- informed spinning 2240 // 2241 // Don't bother spinning if the owner is not eligible to drop the lock. 2242 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2243 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2244 // The thread must be runnable in order to drop the lock in timely fashion. 2245 // If the _owner is not runnable then spinning will not likely be 2246 // successful (profitable). 2247 // 2248 // Beware -- the thread referenced by _owner could have died 2249 // so a simply fetch from _owner->_thread_state might trap. 2250 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2251 // Because of the lifecycle issues the schedctl and _thread_state values 2252 // observed by NotRunnable() might be garbage. NotRunnable must 2253 // tolerate this and consider the observed _thread_state value 2254 // as advisory. 2255 // 2256 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2257 // a thread pointer. We differentiate the two cases with OwnerIsThread. 2258 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2259 // with the LSB of _owner. Another option would be to probablistically probe 2260 // the putative _owner->TypeTag value. 2261 // 2262 // Checking _thread_state isn't perfect. Even if the thread is 2263 // in_java it might be blocked on a page-fault or have been preempted 2264 // and sitting on a ready/dispatch queue. _thread state in conjunction 2265 // with schedctl.sc_state gives us a good picture of what the 2266 // thread is doing, however. 2267 // 2268 // TODO: check schedctl.sc_state. 2269 // We'll need to use SafeFetch32() to read from the schedctl block. 2270 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2271 // 2272 // The return value from NotRunnable() is *advisory* -- the 2273 // result is based on sampling and is not necessarily coherent. 2274 // The caller must tolerate false-negative and false-positive errors. 2275 // Spinning, in general, is probabilistic anyway. 2276 2277 2278 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2279 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2280 if (!OwnerIsThread) return 0 ; 2281 2282 if (ox == NULL) return 0 ; 2283 2284 // Avoid transitive spinning ... 2285 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2286 // Immediately after T1 acquires L it's possible that T2, also 2287 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2288 // This occurs transiently after T1 acquired L but before 2289 // T1 managed to clear T1.Stalled. T2 does not need to abort 2290 // its spin in this circumstance. 2291 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2292 2293 if (BlockedOn == 1) return 1 ; 2294 if (BlockedOn != 0) { 2295 return BlockedOn != intptr_t(this) && _owner == ox ; 2296 } 2297 2298 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2299 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2300 // consider also: jst != _thread_in_Java -- but that's overspecific. 2301 return jst == _thread_blocked || jst == _thread_in_native ; 2302 } 2303 2304 2305 // ----------------------------------------------------------------------------- 2306 // WaitSet management ... 2307 2308 ObjectWaiter::ObjectWaiter(Thread* thread) { 2309 _next = NULL; 2310 _prev = NULL; 2311 _notified = 0; 2312 TState = TS_RUN ; 2313 _thread = thread; 2314 _event = thread->_ParkEvent ; 2315 _active = false; 2316 assert (_event != NULL, "invariant") ; 2317 } 2318 2319 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2320 JavaThread *jt = (JavaThread *)this->_thread; 2321 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2322 } 2323 2324 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2325 JavaThread *jt = (JavaThread *)this->_thread; 2326 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2327 } 2328 2329 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2330 assert(node != NULL, "should not dequeue NULL node"); 2331 assert(node->_prev == NULL, "node already in list"); 2332 assert(node->_next == NULL, "node already in list"); 2333 // put node at end of queue (circular doubly linked list) 2334 if (_WaitSet == NULL) { 2335 _WaitSet = node; 2336 node->_prev = node; 2337 node->_next = node; 2338 } else { 2339 ObjectWaiter* head = _WaitSet ; 2340 ObjectWaiter* tail = head->_prev; 2341 assert(tail->_next == head, "invariant check"); 2342 tail->_next = node; 2343 head->_prev = node; 2344 node->_next = head; 2345 node->_prev = tail; 2346 } 2347 } 2348 2349 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2350 // dequeue the very first waiter 2351 ObjectWaiter* waiter = _WaitSet; 2352 if (waiter) { 2353 DequeueSpecificWaiter(waiter); 2354 } 2355 return waiter; 2356 } 2357 2358 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2359 assert(node != NULL, "should not dequeue NULL node"); 2360 assert(node->_prev != NULL, "node already removed from list"); 2361 assert(node->_next != NULL, "node already removed from list"); 2362 // when the waiter has woken up because of interrupt, 2363 // timeout or other spurious wake-up, dequeue the 2364 // waiter from waiting list 2365 ObjectWaiter* next = node->_next; 2366 if (next == node) { 2367 assert(node->_prev == node, "invariant check"); 2368 _WaitSet = NULL; 2369 } else { 2370 ObjectWaiter* prev = node->_prev; 2371 assert(prev->_next == node, "invariant check"); 2372 assert(next->_prev == node, "invariant check"); 2373 next->_prev = prev; 2374 prev->_next = next; 2375 if (_WaitSet == node) { 2376 _WaitSet = next; 2377 } 2378 } 2379 node->_next = NULL; 2380 node->_prev = NULL; 2381 } 2382 2383 // ----------------------------------------------------------------------------- 2384 // PerfData support 2385 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; 2386 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; 2387 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; 2388 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; 2389 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; 2390 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; 2391 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; 2392 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; 2393 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; 2394 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; 2395 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; 2396 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; 2397 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; 2398 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; 2399 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; 2400 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; 2401 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; 2402 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; 2403 2404 // One-shot global initialization for the sync subsystem. 2405 // We could also defer initialization and initialize on-demand 2406 // the first time we call inflate(). Initialization would 2407 // be protected - like so many things - by the MonitorCache_lock. 2408 2409 void ObjectMonitor::Initialize () { 2410 static int InitializationCompleted = 0 ; 2411 assert (InitializationCompleted == 0, "invariant") ; 2412 InitializationCompleted = 1 ; 2413 if (UsePerfData) { 2414 EXCEPTION_MARK ; 2415 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2416 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2417 NEWPERFCOUNTER(_sync_Inflations) ; 2418 NEWPERFCOUNTER(_sync_Deflations) ; 2419 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 2420 NEWPERFCOUNTER(_sync_FutileWakeups) ; 2421 NEWPERFCOUNTER(_sync_Parks) ; 2422 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 2423 NEWPERFCOUNTER(_sync_Notifications) ; 2424 NEWPERFCOUNTER(_sync_SlowEnter) ; 2425 NEWPERFCOUNTER(_sync_SlowExit) ; 2426 NEWPERFCOUNTER(_sync_SlowNotify) ; 2427 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 2428 NEWPERFCOUNTER(_sync_FailedSpins) ; 2429 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 2430 NEWPERFCOUNTER(_sync_PrivateA) ; 2431 NEWPERFCOUNTER(_sync_PrivateB) ; 2432 NEWPERFCOUNTER(_sync_MonInCirculation) ; 2433 NEWPERFCOUNTER(_sync_MonScavenged) ; 2434 NEWPERFVARIABLE(_sync_MonExtant) ; 2435 #undef NEWPERFCOUNTER 2436 } 2437 } 2438 2439 2440 // Compile-time asserts 2441 // When possible, it's better to catch errors deterministically at 2442 // compile-time than at runtime. The down-side to using compile-time 2443 // asserts is that error message -- often something about negative array 2444 // indices -- is opaque. 2445 2446 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2447 2448 void ObjectMonitor::ctAsserts() { 2449 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2450 } 2451 2452 2453 static char * kvGet (char * kvList, const char * Key) { 2454 if (kvList == NULL) return NULL ; 2455 size_t n = strlen (Key) ; 2456 char * Search ; 2457 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2458 if (strncmp (Search, Key, n) == 0) { 2459 if (Search[n] == '=') return Search + n + 1 ; 2460 if (Search[n] == 0) return (char *) "1" ; 2461 } 2462 } 2463 return NULL ; 2464 } 2465 2466 static int kvGetInt (char * kvList, const char * Key, int Default) { 2467 char * v = kvGet (kvList, Key) ; 2468 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2469 if (Knob_ReportSettings && v != NULL) { 2470 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2471 ::fflush (stdout) ; 2472 } 2473 return rslt ; 2474 } 2475 2476 void ObjectMonitor::DeferredInitialize () { 2477 if (InitDone > 0) return ; 2478 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2479 while (InitDone != 1) ; 2480 return ; 2481 } 2482 2483 // One-shot global initialization ... 2484 // The initialization is idempotent, so we don't need locks. 2485 // In the future consider doing this via os::init_2(). 2486 // SyncKnobs consist of <Key>=<Value> pairs in the style 2487 // of environment variables. Start by converting ':' to NUL. 2488 2489 if (SyncKnobs == NULL) SyncKnobs = "" ; 2490 2491 size_t sz = strlen (SyncKnobs) ; 2492 char * knobs = (char *) malloc (sz + 2) ; 2493 if (knobs == NULL) { 2494 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; 2495 guarantee (0, "invariant") ; 2496 } 2497 strcpy (knobs, SyncKnobs) ; 2498 knobs[sz+1] = 0 ; 2499 for (char * p = knobs ; *p ; p++) { 2500 if (*p == ':') *p = 0 ; 2501 } 2502 2503 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2504 SETKNOB(ReportSettings) ; 2505 SETKNOB(Verbose) ; 2506 SETKNOB(FixedSpin) ; 2507 SETKNOB(SpinLimit) ; 2508 SETKNOB(SpinBase) ; 2509 SETKNOB(SpinBackOff); 2510 SETKNOB(CASPenalty) ; 2511 SETKNOB(OXPenalty) ; 2512 SETKNOB(LogSpins) ; 2513 SETKNOB(SpinSetSucc) ; 2514 SETKNOB(SuccEnabled) ; 2515 SETKNOB(SuccRestrict) ; 2516 SETKNOB(Penalty) ; 2517 SETKNOB(Bonus) ; 2518 SETKNOB(BonusB) ; 2519 SETKNOB(Poverty) ; 2520 SETKNOB(SpinAfterFutile) ; 2521 SETKNOB(UsePause) ; 2522 SETKNOB(SpinEarly) ; 2523 SETKNOB(OState) ; 2524 SETKNOB(MaxSpinners) ; 2525 SETKNOB(PreSpin) ; 2526 SETKNOB(ExitPolicy) ; 2527 SETKNOB(QMode); 2528 SETKNOB(ResetEvent) ; 2529 SETKNOB(MoveNotifyee) ; 2530 SETKNOB(FastHSSEC) ; 2531 #undef SETKNOB 2532 2533 if (Knob_Verbose) { 2534 sanity_checks(); 2535 } 2536 2537 if (os::is_MP()) { 2538 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2539 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2540 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2541 } else { 2542 Knob_SpinLimit = 0 ; 2543 Knob_SpinBase = 0 ; 2544 Knob_PreSpin = 0 ; 2545 Knob_FixedSpin = -1 ; 2546 } 2547 2548 if (Knob_LogSpins == 0) { 2549 ObjectMonitor::_sync_FailedSpins = NULL ; 2550 } 2551 2552 free (knobs) ; 2553 OrderAccess::fence() ; 2554 InitDone = 1 ; 2555 } 2556 2557 void ObjectMonitor::sanity_checks() { 2558 int error_cnt = 0; 2559 int warning_cnt = 0; 2560 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2561 2562 if (verbose) { 2563 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2564 sizeof(ObjectMonitor)); 2565 } 2566 2567 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2568 if (verbose) { 2569 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2570 } 2571 2572 ObjectMonitor dummy; 2573 u_char *addr_begin = (u_char*)&dummy; 2574 u_char *addr_header = (u_char*)&dummy._header; 2575 u_char *addr_owner = (u_char*)&dummy._owner; 2576 2577 uint offset_header = (uint)(addr_header - addr_begin); 2578 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2579 2580 uint offset_owner = (uint)(addr_owner - addr_begin); 2581 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2582 2583 if ((uint)(addr_header - addr_begin) != 0) { 2584 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2585 error_cnt++; 2586 } 2587 2588 if (cache_line_size != 0) { 2589 // We were able to determine the L1 data cache line size so 2590 // do some cache line specific sanity checks 2591 2592 if ((offset_owner - offset_header) < cache_line_size) { 2593 tty->print_cr("WARNING: the _header and _owner fields are closer " 2594 "than a cache line which permits false sharing."); 2595 warning_cnt++; 2596 } 2597 2598 if ((sizeof(ObjectMonitor) % cache_line_size) != 0) { 2599 tty->print_cr("WARNING: ObjectMonitor size is not a multiple of " 2600 "a cache line which permits false sharing."); 2601 warning_cnt++; 2602 } 2603 } 2604 2605 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2606 &warning_cnt); 2607 2608 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2609 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2610 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2611 } 2612 2613 guarantee(error_cnt == 0, 2614 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2615 } 2616 2617 #ifndef PRODUCT 2618 void ObjectMonitor::verify() { 2619 } 2620 2621 void ObjectMonitor::print() { 2622 } 2623 #endif