1 /* 2 * Copyright (c) 1998, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "jfr/support/jfrThreadId.hpp" 29 #include "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "oops/markWord.hpp" 34 #include "oops/oop.inline.hpp" 35 #include "runtime/atomic.hpp" 36 #include "runtime/handles.inline.hpp" 37 #include "runtime/interfaceSupport.inline.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "runtime/objectMonitor.hpp" 40 #include "runtime/objectMonitor.inline.hpp" 41 #include "runtime/orderAccess.hpp" 42 #include "runtime/osThread.hpp" 43 #include "runtime/safepointMechanism.inline.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/stubRoutines.hpp" 46 #include "runtime/thread.inline.hpp" 47 #include "services/threadService.hpp" 48 #include "utilities/dtrace.hpp" 49 #include "utilities/macros.hpp" 50 #include "utilities/preserveException.hpp" 51 #if INCLUDE_JFR 52 #include "jfr/support/jfrFlush.hpp" 53 #endif 54 55 #ifdef DTRACE_ENABLED 56 57 // Only bother with this argument setup if dtrace is available 58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95 #else // ndef DTRACE_ENABLED 96 97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100 #endif // ndef DTRACE_ENABLED 101 102 // Tunables ... 103 // The knob* variables are effectively final. Once set they should 104 // never be modified hence. Consider using __read_mostly with GCC. 105 106 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 107 108 static int Knob_Bonus = 100; // spin success bonus 109 static int Knob_BonusB = 100; // spin success bonus 110 static int Knob_Penalty = 200; // spin failure penalty 111 static int Knob_Poverty = 1000; 112 static int Knob_FixedSpin = 0; 113 static int Knob_PreSpin = 10; // 20-100 likely better 114 115 DEBUG_ONLY(static volatile bool InitDone = false;) 116 117 // ----------------------------------------------------------------------------- 118 // Theory of operations -- Monitors lists, thread residency, etc: 119 // 120 // * A thread acquires ownership of a monitor by successfully 121 // CAS()ing the _owner field from null to non-null. 122 // 123 // * Invariant: A thread appears on at most one monitor list -- 124 // cxq, EntryList or WaitSet -- at any one time. 125 // 126 // * Contending threads "push" themselves onto the cxq with CAS 127 // and then spin/park. 128 // 129 // * After a contending thread eventually acquires the lock it must 130 // dequeue itself from either the EntryList or the cxq. 131 // 132 // * The exiting thread identifies and unparks an "heir presumptive" 133 // tentative successor thread on the EntryList. Critically, the 134 // exiting thread doesn't unlink the successor thread from the EntryList. 135 // After having been unparked, the wakee will recontend for ownership of 136 // the monitor. The successor (wakee) will either acquire the lock or 137 // re-park itself. 138 // 139 // Succession is provided for by a policy of competitive handoff. 140 // The exiting thread does _not_ grant or pass ownership to the 141 // successor thread. (This is also referred to as "handoff" succession"). 142 // Instead the exiting thread releases ownership and possibly wakes 143 // a successor, so the successor can (re)compete for ownership of the lock. 144 // If the EntryList is empty but the cxq is populated the exiting 145 // thread will drain the cxq into the EntryList. It does so by 146 // by detaching the cxq (installing null with CAS) and folding 147 // the threads from the cxq into the EntryList. The EntryList is 148 // doubly linked, while the cxq is singly linked because of the 149 // CAS-based "push" used to enqueue recently arrived threads (RATs). 150 // 151 // * Concurrency invariants: 152 // 153 // -- only the monitor owner may access or mutate the EntryList. 154 // The mutex property of the monitor itself protects the EntryList 155 // from concurrent interference. 156 // -- Only the monitor owner may detach the cxq. 157 // 158 // * The monitor entry list operations avoid locks, but strictly speaking 159 // they're not lock-free. Enter is lock-free, exit is not. 160 // For a description of 'Methods and apparatus providing non-blocking access 161 // to a resource,' see U.S. Pat. No. 7844973. 162 // 163 // * The cxq can have multiple concurrent "pushers" but only one concurrent 164 // detaching thread. This mechanism is immune from the ABA corruption. 165 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 166 // 167 // * Taken together, the cxq and the EntryList constitute or form a 168 // single logical queue of threads stalled trying to acquire the lock. 169 // We use two distinct lists to improve the odds of a constant-time 170 // dequeue operation after acquisition (in the ::enter() epilogue) and 171 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 172 // A key desideratum is to minimize queue & monitor metadata manipulation 173 // that occurs while holding the monitor lock -- that is, we want to 174 // minimize monitor lock holds times. Note that even a small amount of 175 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 176 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 177 // locks and monitor metadata. 178 // 179 // Cxq points to the set of Recently Arrived Threads attempting entry. 180 // Because we push threads onto _cxq with CAS, the RATs must take the form of 181 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 182 // the unlocking thread notices that EntryList is null but _cxq is != null. 183 // 184 // The EntryList is ordered by the prevailing queue discipline and 185 // can be organized in any convenient fashion, such as a doubly-linked list or 186 // a circular doubly-linked list. Critically, we want insert and delete operations 187 // to operate in constant-time. If we need a priority queue then something akin 188 // to Solaris' sleepq would work nicely. Viz., 189 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 190 // Queue discipline is enforced at ::exit() time, when the unlocking thread 191 // drains the cxq into the EntryList, and orders or reorders the threads on the 192 // EntryList accordingly. 193 // 194 // Barring "lock barging", this mechanism provides fair cyclic ordering, 195 // somewhat similar to an elevator-scan. 196 // 197 // * The monitor synchronization subsystem avoids the use of native 198 // synchronization primitives except for the narrow platform-specific 199 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 200 // the semantics of park-unpark. Put another way, this monitor implementation 201 // depends only on atomic operations and park-unpark. The monitor subsystem 202 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 203 // underlying OS manages the READY<->RUN transitions. 204 // 205 // * Waiting threads reside on the WaitSet list -- wait() puts 206 // the caller onto the WaitSet. 207 // 208 // * notify() or notifyAll() simply transfers threads from the WaitSet to 209 // either the EntryList or cxq. Subsequent exit() operations will 210 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 211 // it's likely the notifyee would simply impale itself on the lock held 212 // by the notifier. 213 // 214 // * An interesting alternative is to encode cxq as (List,LockByte) where 215 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 216 // variable, like _recursions, in the scheme. The threads or Events that form 217 // the list would have to be aligned in 256-byte addresses. A thread would 218 // try to acquire the lock or enqueue itself with CAS, but exiting threads 219 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 220 // Note that is is *not* word-tearing, but it does presume that full-word 221 // CAS operations are coherent with intermix with STB operations. That's true 222 // on most common processors. 223 // 224 // * See also http://blogs.sun.com/dave 225 226 227 void* ObjectMonitor::operator new (size_t size) throw() { 228 return AllocateHeap(size, mtInternal); 229 } 230 void* ObjectMonitor::operator new[] (size_t size) throw() { 231 return operator new (size); 232 } 233 void ObjectMonitor::operator delete(void* p) { 234 FreeHeap(p); 235 } 236 void ObjectMonitor::operator delete[] (void *p) { 237 operator delete(p); 238 } 239 240 // ----------------------------------------------------------------------------- 241 // Enter support 242 243 void ObjectMonitor::enter(TRAPS) { 244 jint l_ref_count = ref_count(); 245 ADIM_guarantee(l_ref_count > 0, "must be positive: l_ref_count=%d, ref_count=%d", l_ref_count, ref_count()); 246 247 // The following code is ordered to check the most common cases first 248 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 249 Thread * const Self = THREAD; 250 251 void* cur = try_set_owner_from(NULL, Self); 252 if (cur == NULL) { 253 assert(_recursions == 0, "invariant"); 254 return; 255 } 256 257 if (cur == Self) { 258 // TODO-FIXME: check for integer overflow! BUGID 6557169. 259 _recursions++; 260 return; 261 } 262 263 if (Self->is_lock_owned((address)cur)) { 264 assert(_recursions == 0, "internal state error"); 265 _recursions = 1; 266 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 267 return; 268 } 269 270 if (AsyncDeflateIdleMonitors && 271 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 272 // The deflation protocol finished the first part (setting owner), 273 // but it failed the second part (making ref_count negative) and 274 // bailed. Acquired the monitor. 275 assert(_recursions == 0, "invariant"); 276 return; 277 } 278 279 // We've encountered genuine contention. 280 assert(Self->_Stalled == 0, "invariant"); 281 Self->_Stalled = intptr_t(this); 282 283 // Try one round of spinning *before* enqueueing Self 284 // and before going through the awkward and expensive state 285 // transitions. The following spin is strictly optional ... 286 // Note that if we acquire the monitor from an initial spin 287 // we forgo posting JVMTI events and firing DTRACE probes. 288 if (TrySpin(Self) > 0) { 289 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 290 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 291 assert(((oop)object())->mark() == markWord::encode(this), 292 "object mark must match encoded this: mark=" INTPTR_FORMAT 293 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 294 markWord::encode(this).value()); 295 Self->_Stalled = 0; 296 return; 297 } 298 299 assert(_owner != Self, "invariant"); 300 assert(_succ != Self, "invariant"); 301 assert(Self->is_Java_thread(), "invariant"); 302 JavaThread * jt = (JavaThread *) Self; 303 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 304 assert(jt->thread_state() != _thread_blocked, "invariant"); 305 assert(this->object() != NULL, "invariant"); 306 assert(_contentions >= 0, "must not be negative: contentions=%d", _contentions); 307 308 // Keep track of contention for JVM/TI and M&M queries. 309 Atomic::inc(&_contentions); 310 311 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 312 EventJavaMonitorEnter event; 313 if (event.should_commit()) { 314 event.set_monitorClass(((oop)this->object())->klass()); 315 event.set_address((uintptr_t)(this->object_addr())); 316 } 317 318 { // Change java thread status to indicate blocked on monitor enter. 319 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 320 321 Self->set_current_pending_monitor(this); 322 323 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 324 if (JvmtiExport::should_post_monitor_contended_enter()) { 325 JvmtiExport::post_monitor_contended_enter(jt, this); 326 327 // The current thread does not yet own the monitor and does not 328 // yet appear on any queues that would get it made the successor. 329 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 330 // handler cannot accidentally consume an unpark() meant for the 331 // ParkEvent associated with this ObjectMonitor. 332 } 333 334 OSThreadContendState osts(Self->osthread()); 335 ThreadBlockInVM tbivm(jt); 336 337 // TODO-FIXME: change the following for(;;) loop to straight-line code. 338 for (;;) { 339 jt->set_suspend_equivalent(); 340 // cleared by handle_special_suspend_equivalent_condition() 341 // or java_suspend_self() 342 343 EnterI(THREAD); 344 345 if (!ExitSuspendEquivalent(jt)) break; 346 347 // We have acquired the contended monitor, but while we were 348 // waiting another thread suspended us. We don't want to enter 349 // the monitor while suspended because that would surprise the 350 // thread that suspended us. 351 // 352 _recursions = 0; 353 _succ = NULL; 354 exit(false, Self); 355 356 jt->java_suspend_self(); 357 } 358 Self->set_current_pending_monitor(NULL); 359 360 // We cleared the pending monitor info since we've just gotten past 361 // the enter-check-for-suspend dance and we now own the monitor free 362 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 363 // destructor can go to a safepoint at the end of this block. If we 364 // do a thread dump during that safepoint, then this thread will show 365 // as having "-locked" the monitor, but the OS and java.lang.Thread 366 // states will still report that the thread is blocked trying to 367 // acquire it. 368 } 369 370 Atomic::dec(&_contentions); 371 assert(_contentions >= 0, "must not be negative: contentions=%d", _contentions); 372 Self->_Stalled = 0; 373 374 // Must either set _recursions = 0 or ASSERT _recursions == 0. 375 assert(_recursions == 0, "invariant"); 376 assert(_owner == Self, "invariant"); 377 assert(_succ != Self, "invariant"); 378 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 379 380 // The thread -- now the owner -- is back in vm mode. 381 // Report the glorious news via TI,DTrace and jvmstat. 382 // The probe effect is non-trivial. All the reportage occurs 383 // while we hold the monitor, increasing the length of the critical 384 // section. Amdahl's parallel speedup law comes vividly into play. 385 // 386 // Another option might be to aggregate the events (thread local or 387 // per-monitor aggregation) and defer reporting until a more opportune 388 // time -- such as next time some thread encounters contention but has 389 // yet to acquire the lock. While spinning that thread could 390 // spinning we could increment JVMStat counters, etc. 391 392 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 393 if (JvmtiExport::should_post_monitor_contended_entered()) { 394 JvmtiExport::post_monitor_contended_entered(jt, this); 395 396 // The current thread already owns the monitor and is not going to 397 // call park() for the remainder of the monitor enter protocol. So 398 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 399 // event handler consumed an unpark() issued by the thread that 400 // just exited the monitor. 401 } 402 if (event.should_commit()) { 403 event.set_previousOwner((uintptr_t)_previous_owner_tid); 404 event.commit(); 405 } 406 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 407 } 408 409 // Caveat: TryLock() is not necessarily serializing if it returns failure. 410 // Callers must compensate as needed. 411 412 int ObjectMonitor::TryLock(Thread * Self) { 413 void * own = _owner; 414 if (own != NULL) return 0; 415 if (try_set_owner_from(NULL, Self) == NULL) { 416 assert(_recursions == 0, "invariant"); 417 return 1; 418 } 419 // The lock had been free momentarily, but we lost the race to the lock. 420 // Interference -- the CAS failed. 421 // We can either return -1 or retry. 422 // Retry doesn't make as much sense because the lock was just acquired. 423 return -1; 424 } 425 426 // Install the displaced mark word (dmw) of a deflating ObjectMonitor 427 // into the header of the object associated with the monitor. This 428 // idempotent method is called by a thread that is deflating a 429 // monitor and by other threads that have detected a race with the 430 // deflation process. 431 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) { 432 // This function must only be called when (owner == DEFLATER_MARKER 433 // && ref_count <= 0), but we can't guarantee that here because 434 // those values could change when the ObjectMonitor gets moved from 435 // the global free list to a per-thread free list. 436 437 guarantee(obj != NULL, "must be non-NULL"); 438 if (object() != obj) { 439 // ObjectMonitor's object ref no longer refers to the target object 440 // so the object's header has already been restored. 441 return; 442 } 443 444 markWord dmw = header(); 445 if (dmw.value() == 0) { 446 // ObjectMonitor's header/dmw has been cleared so the ObjectMonitor 447 // has been deflated and taken off the global free list. 448 return; 449 } 450 451 // A non-NULL dmw has to be neutral (not locked and not marked). 452 ADIM_guarantee(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 453 454 // Install displaced mark word if the object's header still points 455 // to this ObjectMonitor. All racing callers to this function will 456 // reach this point, but only one can win. 457 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 458 if (res != markWord::encode(this)) { 459 // This should be rare so log at the Info level when it happens. 460 log_info(monitorinflation)("install_displaced_markword_in_object: " 461 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 462 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 463 dmw.value(), markWord::encode(this).value(), 464 res.value()); 465 } 466 467 // Note: It does not matter which thread restored the header/dmw 468 // into the object's header. The thread deflating the monitor just 469 // wanted the object's header restored and it is. The threads that 470 // detected a race with the deflation process also wanted the 471 // object's header restored before they retry their operation and 472 // because it is restored they will only retry once. 473 } 474 475 // Convert the fields used by is_busy() to a string that can be 476 // used for diagnostic output. 477 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 478 ss->print("is_busy: contentions=%d, waiters=%d, ", _contentions, _waiters); 479 if (!AsyncDeflateIdleMonitors) { 480 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 481 } else if (_owner != DEFLATER_MARKER) { 482 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 483 } else { 484 // We report NULL instead of DEFLATER_MARKER here because is_busy() 485 // ignores DEFLATER_MARKER values. 486 ss->print("owner=" INTPTR_FORMAT, NULL); 487 } 488 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq), 489 p2i(_EntryList)); 490 return ss->base(); 491 } 492 493 #define MAX_RECHECK_INTERVAL 1000 494 495 void ObjectMonitor::EnterI(TRAPS) { 496 jint l_ref_count = ref_count(); 497 ADIM_guarantee(l_ref_count > 0, "must be positive: l_ref_count=%d, ref_count=%d", l_ref_count, ref_count()); 498 499 Thread * const Self = THREAD; 500 assert(Self->is_Java_thread(), "invariant"); 501 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 502 503 // Try the lock - TATAS 504 if (TryLock (Self) > 0) { 505 assert(_succ != Self, "invariant"); 506 assert(_owner == Self, "invariant"); 507 assert(_Responsible != Self, "invariant"); 508 return; 509 } 510 511 if (AsyncDeflateIdleMonitors && 512 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 513 // The deflation protocol finished the first part (setting owner), 514 // but it failed the second part (making ref_count negative) and 515 // bailed. Acquired the monitor. 516 assert(_succ != Self, "invariant"); 517 assert(_Responsible != Self, "invariant"); 518 return; 519 } 520 521 assert(InitDone, "Unexpectedly not initialized"); 522 523 // We try one round of spinning *before* enqueueing Self. 524 // 525 // If the _owner is ready but OFFPROC we could use a YieldTo() 526 // operation to donate the remainder of this thread's quantum 527 // to the owner. This has subtle but beneficial affinity 528 // effects. 529 530 if (TrySpin(Self) > 0) { 531 assert(_owner == Self, "invariant"); 532 assert(_succ != Self, "invariant"); 533 assert(_Responsible != Self, "invariant"); 534 return; 535 } 536 537 // The Spin failed -- Enqueue and park the thread ... 538 assert(_succ != Self, "invariant"); 539 assert(_owner != Self, "invariant"); 540 assert(_Responsible != Self, "invariant"); 541 542 // Enqueue "Self" on ObjectMonitor's _cxq. 543 // 544 // Node acts as a proxy for Self. 545 // As an aside, if were to ever rewrite the synchronization code mostly 546 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 547 // Java objects. This would avoid awkward lifecycle and liveness issues, 548 // as well as eliminate a subset of ABA issues. 549 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 550 551 ObjectWaiter node(Self); 552 Self->_ParkEvent->reset(); 553 node._prev = (ObjectWaiter *) 0xBAD; 554 node.TState = ObjectWaiter::TS_CXQ; 555 556 // Push "Self" onto the front of the _cxq. 557 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 558 // Note that spinning tends to reduce the rate at which threads 559 // enqueue and dequeue on EntryList|cxq. 560 ObjectWaiter * nxt; 561 for (;;) { 562 node._next = nxt = _cxq; 563 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 564 565 // Interference - the CAS failed because _cxq changed. Just retry. 566 // As an optional optimization we retry the lock. 567 if (TryLock (Self) > 0) { 568 assert(_succ != Self, "invariant"); 569 assert(_owner == Self, "invariant"); 570 assert(_Responsible != Self, "invariant"); 571 return; 572 } 573 } 574 575 // Check for cxq|EntryList edge transition to non-null. This indicates 576 // the onset of contention. While contention persists exiting threads 577 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 578 // operations revert to the faster 1-0 mode. This enter operation may interleave 579 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 580 // arrange for one of the contending thread to use a timed park() operations 581 // to detect and recover from the race. (Stranding is form of progress failure 582 // where the monitor is unlocked but all the contending threads remain parked). 583 // That is, at least one of the contended threads will periodically poll _owner. 584 // One of the contending threads will become the designated "Responsible" thread. 585 // The Responsible thread uses a timed park instead of a normal indefinite park 586 // operation -- it periodically wakes and checks for and recovers from potential 587 // strandings admitted by 1-0 exit operations. We need at most one Responsible 588 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 589 // be responsible for a monitor. 590 // 591 // Currently, one of the contended threads takes on the added role of "Responsible". 592 // A viable alternative would be to use a dedicated "stranding checker" thread 593 // that periodically iterated over all the threads (or active monitors) and unparked 594 // successors where there was risk of stranding. This would help eliminate the 595 // timer scalability issues we see on some platforms as we'd only have one thread 596 // -- the checker -- parked on a timer. 597 598 if (nxt == NULL && _EntryList == NULL) { 599 // Try to assume the role of responsible thread for the monitor. 600 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 601 Atomic::replace_if_null(&_Responsible, Self); 602 } 603 604 // The lock might have been released while this thread was occupied queueing 605 // itself onto _cxq. To close the race and avoid "stranding" and 606 // progress-liveness failure we must resample-retry _owner before parking. 607 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 608 // In this case the ST-MEMBAR is accomplished with CAS(). 609 // 610 // TODO: Defer all thread state transitions until park-time. 611 // Since state transitions are heavy and inefficient we'd like 612 // to defer the state transitions until absolutely necessary, 613 // and in doing so avoid some transitions ... 614 615 int nWakeups = 0; 616 int recheckInterval = 1; 617 618 for (;;) { 619 620 if (TryLock(Self) > 0) break; 621 assert(_owner != Self, "invariant"); 622 623 // park self 624 if (_Responsible == Self) { 625 Self->_ParkEvent->park((jlong) recheckInterval); 626 // Increase the recheckInterval, but clamp the value. 627 recheckInterval *= 8; 628 if (recheckInterval > MAX_RECHECK_INTERVAL) { 629 recheckInterval = MAX_RECHECK_INTERVAL; 630 } 631 } else { 632 Self->_ParkEvent->park(); 633 } 634 635 if (TryLock(Self) > 0) break; 636 637 if (AsyncDeflateIdleMonitors && 638 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 639 // The deflation protocol finished the first part (setting owner), 640 // but it failed the second part (making ref_count negative) and 641 // bailed. Acquired the monitor. 642 break; 643 } 644 645 // The lock is still contested. 646 // Keep a tally of the # of futile wakeups. 647 // Note that the counter is not protected by a lock or updated by atomics. 648 // That is by design - we trade "lossy" counters which are exposed to 649 // races during updates for a lower probe effect. 650 651 // This PerfData object can be used in parallel with a safepoint. 652 // See the work around in PerfDataManager::destroy(). 653 OM_PERFDATA_OP(FutileWakeups, inc()); 654 ++nWakeups; 655 656 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 657 // We can defer clearing _succ until after the spin completes 658 // TrySpin() must tolerate being called with _succ == Self. 659 // Try yet another round of adaptive spinning. 660 if (TrySpin(Self) > 0) break; 661 662 // We can find that we were unpark()ed and redesignated _succ while 663 // we were spinning. That's harmless. If we iterate and call park(), 664 // park() will consume the event and return immediately and we'll 665 // just spin again. This pattern can repeat, leaving _succ to simply 666 // spin on a CPU. 667 668 if (_succ == Self) _succ = NULL; 669 670 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 671 OrderAccess::fence(); 672 } 673 674 // Egress : 675 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 676 // Normally we'll find Self on the EntryList . 677 // From the perspective of the lock owner (this thread), the 678 // EntryList is stable and cxq is prepend-only. 679 // The head of cxq is volatile but the interior is stable. 680 // In addition, Self.TState is stable. 681 682 assert(_owner == Self, "invariant"); 683 assert(object() != NULL, "invariant"); 684 // I'd like to write: 685 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 686 // but as we're at a safepoint that's not safe. 687 688 UnlinkAfterAcquire(Self, &node); 689 if (_succ == Self) _succ = NULL; 690 691 assert(_succ != Self, "invariant"); 692 if (_Responsible == Self) { 693 _Responsible = NULL; 694 OrderAccess::fence(); // Dekker pivot-point 695 696 // We may leave threads on cxq|EntryList without a designated 697 // "Responsible" thread. This is benign. When this thread subsequently 698 // exits the monitor it can "see" such preexisting "old" threads -- 699 // threads that arrived on the cxq|EntryList before the fence, above -- 700 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 701 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 702 // non-null and elect a new "Responsible" timer thread. 703 // 704 // This thread executes: 705 // ST Responsible=null; MEMBAR (in enter epilogue - here) 706 // LD cxq|EntryList (in subsequent exit) 707 // 708 // Entering threads in the slow/contended path execute: 709 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 710 // The (ST cxq; MEMBAR) is accomplished with CAS(). 711 // 712 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 713 // exit operation from floating above the ST Responsible=null. 714 } 715 716 // We've acquired ownership with CAS(). 717 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 718 // But since the CAS() this thread may have also stored into _succ, 719 // EntryList, cxq or Responsible. These meta-data updates must be 720 // visible __before this thread subsequently drops the lock. 721 // Consider what could occur if we didn't enforce this constraint -- 722 // STs to monitor meta-data and user-data could reorder with (become 723 // visible after) the ST in exit that drops ownership of the lock. 724 // Some other thread could then acquire the lock, but observe inconsistent 725 // or old monitor meta-data and heap data. That violates the JMM. 726 // To that end, the 1-0 exit() operation must have at least STST|LDST 727 // "release" barrier semantics. Specifically, there must be at least a 728 // STST|LDST barrier in exit() before the ST of null into _owner that drops 729 // the lock. The barrier ensures that changes to monitor meta-data and data 730 // protected by the lock will be visible before we release the lock, and 731 // therefore before some other thread (CPU) has a chance to acquire the lock. 732 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 733 // 734 // Critically, any prior STs to _succ or EntryList must be visible before 735 // the ST of null into _owner in the *subsequent* (following) corresponding 736 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 737 // execute a serializing instruction. 738 739 return; 740 } 741 742 // ReenterI() is a specialized inline form of the latter half of the 743 // contended slow-path from EnterI(). We use ReenterI() only for 744 // monitor reentry in wait(). 745 // 746 // In the future we should reconcile EnterI() and ReenterI(). 747 748 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 749 jint l_ref_count = ref_count(); 750 ADIM_guarantee(l_ref_count > 0, "must be positive: l_ref_count=%d, ref_count=%d", l_ref_count, ref_count()); 751 752 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() == markWord::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 if (AsyncDeflateIdleMonitors && 770 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 771 // The deflation protocol finished the first part (setting owner), 772 // but it failed the second part (making ref_count negative) and 773 // bailed. Acquired the monitor. 774 break; 775 } 776 777 // State transition wrappers around park() ... 778 // ReenterI() wisely defers state transitions until 779 // it's clear we must park the thread. 780 { 781 OSThreadContendState osts(Self->osthread()); 782 ThreadBlockInVM tbivm(jt); 783 784 // cleared by handle_special_suspend_equivalent_condition() 785 // or java_suspend_self() 786 jt->set_suspend_equivalent(); 787 Self->_ParkEvent->park(); 788 789 // were we externally suspended while we were waiting? 790 for (;;) { 791 if (!ExitSuspendEquivalent(jt)) break; 792 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 793 jt->java_suspend_self(); 794 jt->set_suspend_equivalent(); 795 } 796 } 797 798 // Try again, but just so we distinguish between futile wakeups and 799 // successful wakeups. The following test isn't algorithmically 800 // necessary, but it helps us maintain sensible statistics. 801 if (TryLock(Self) > 0) break; 802 803 // The lock is still contested. 804 // Keep a tally of the # of futile wakeups. 805 // Note that the counter is not protected by a lock or updated by atomics. 806 // That is by design - we trade "lossy" counters which are exposed to 807 // races during updates for a lower probe effect. 808 ++nWakeups; 809 810 // Assuming this is not a spurious wakeup we'll normally 811 // find that _succ == Self. 812 if (_succ == Self) _succ = NULL; 813 814 // Invariant: after clearing _succ a contending thread 815 // *must* retry _owner before parking. 816 OrderAccess::fence(); 817 818 // This PerfData object can be used in parallel with a safepoint. 819 // See the work around in PerfDataManager::destroy(). 820 OM_PERFDATA_OP(FutileWakeups, inc()); 821 } 822 823 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 824 // Normally we'll find Self on the EntryList. 825 // Unlinking from the EntryList is constant-time and atomic-free. 826 // From the perspective of the lock owner (this thread), the 827 // EntryList is stable and cxq is prepend-only. 828 // The head of cxq is volatile but the interior is stable. 829 // In addition, Self.TState is stable. 830 831 assert(_owner == Self, "invariant"); 832 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 833 UnlinkAfterAcquire(Self, SelfNode); 834 if (_succ == Self) _succ = NULL; 835 assert(_succ != Self, "invariant"); 836 SelfNode->TState = ObjectWaiter::TS_RUN; 837 OrderAccess::fence(); // see comments at the end of EnterI() 838 } 839 840 // By convention we unlink a contending thread from EntryList|cxq immediately 841 // after the thread acquires the lock in ::enter(). Equally, we could defer 842 // unlinking the thread until ::exit()-time. 843 844 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 845 assert(_owner == Self, "invariant"); 846 assert(SelfNode->_thread == Self, "invariant"); 847 848 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 849 // Normal case: remove Self from the DLL EntryList . 850 // This is a constant-time operation. 851 ObjectWaiter * nxt = SelfNode->_next; 852 ObjectWaiter * prv = SelfNode->_prev; 853 if (nxt != NULL) nxt->_prev = prv; 854 if (prv != NULL) prv->_next = nxt; 855 if (SelfNode == _EntryList) _EntryList = nxt; 856 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 857 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 858 } else { 859 assert(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(&_cxq, v, SelfNode->_next) != 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 } 897 898 #ifdef ASSERT 899 // Diagnostic hygiene ... 900 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 901 SelfNode->_next = (ObjectWaiter *) 0xBAD; 902 SelfNode->TState = ObjectWaiter::TS_RUN; 903 #endif 904 } 905 906 // ----------------------------------------------------------------------------- 907 // Exit support 908 // 909 // exit() 910 // ~~~~~~ 911 // Note that the collector can't reclaim the objectMonitor or deflate 912 // the object out from underneath the thread calling ::exit() as the 913 // thread calling ::exit() never transitions to a stable state. 914 // This inhibits GC, which in turn inhibits asynchronous (and 915 // inopportune) reclamation of "this". 916 // 917 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 918 // There's one exception to the claim above, however. EnterI() can call 919 // exit() to drop a lock if the acquirer has been externally suspended. 920 // In that case exit() is called with _thread_state == _thread_blocked, 921 // but the monitor's ref_count is > 0, which inhibits reclamation. 922 // 923 // 1-0 exit 924 // ~~~~~~~~ 925 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 926 // the fast-path operators have been optimized so the common ::exit() 927 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 928 // The code emitted by fast_unlock() elides the usual MEMBAR. This 929 // greatly improves latency -- MEMBAR and CAS having considerable local 930 // latency on modern processors -- but at the cost of "stranding". Absent the 931 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 932 // ::enter() path, resulting in the entering thread being stranding 933 // and a progress-liveness failure. Stranding is extremely rare. 934 // We use timers (timed park operations) & periodic polling to detect 935 // and recover from stranding. Potentially stranded threads periodically 936 // wake up and poll the lock. See the usage of the _Responsible variable. 937 // 938 // The CAS() in enter provides for safety and exclusion, while the CAS or 939 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 940 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 941 // We detect and recover from stranding with timers. 942 // 943 // If a thread transiently strands it'll park until (a) another 944 // thread acquires the lock and then drops the lock, at which time the 945 // exiting thread will notice and unpark the stranded thread, or, (b) 946 // the timer expires. If the lock is high traffic then the stranding latency 947 // will be low due to (a). If the lock is low traffic then the odds of 948 // stranding are lower, although the worst-case stranding latency 949 // is longer. Critically, we don't want to put excessive load in the 950 // platform's timer subsystem. We want to minimize both the timer injection 951 // rate (timers created/sec) as well as the number of timers active at 952 // any one time. (more precisely, we want to minimize timer-seconds, which is 953 // the integral of the # of active timers at any instant over time). 954 // Both impinge on OS scalability. Given that, at most one thread parked on 955 // a monitor will use a timer. 956 // 957 // There is also the risk of a futile wake-up. If we drop the lock 958 // another thread can reacquire the lock immediately, and we can 959 // then wake a thread unnecessarily. This is benign, and we've 960 // structured the code so the windows are short and the frequency 961 // of such futile wakups is low. 962 963 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 964 Thread* const Self = THREAD; 965 void* cur = Atomic::load(&_owner); 966 if (THREAD != cur) { 967 if (THREAD->is_lock_owned((address)cur)) { 968 assert(_recursions == 0, "invariant"); 969 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 970 _recursions = 0; 971 } else { 972 // Apparent unbalanced locking ... 973 // Naively we'd like to throw IllegalMonitorStateException. 974 // As a practical matter we can neither allocate nor throw an 975 // exception as ::exit() can be called from leaf routines. 976 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 977 // Upon deeper reflection, however, in a properly run JVM the only 978 // way we should encounter this situation is in the presence of 979 // unbalanced JNI locking. TODO: CheckJNICalls. 980 // See also: CR4414101 981 #ifdef ASSERT 982 LogStreamHandle(Error, monitorinflation) lsh; 983 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 984 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 985 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 986 print_debug_style_on(&lsh); 987 #endif 988 assert(false, "Non-balanced monitor enter/exit!"); 989 return; 990 } 991 } 992 993 if (_recursions != 0) { 994 _recursions--; // this is simple recursive enter 995 return; 996 } 997 998 // Invariant: after setting Responsible=null an thread must execute 999 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1000 _Responsible = NULL; 1001 1002 #if INCLUDE_JFR 1003 // get the owner's thread id for the MonitorEnter event 1004 // if it is enabled and the thread isn't suspended 1005 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1006 _previous_owner_tid = JFR_THREAD_ID(Self); 1007 } 1008 #endif 1009 1010 for (;;) { 1011 assert(THREAD == _owner, "invariant"); 1012 1013 // Drop the lock. 1014 // release semantics: prior loads and stores from within the critical section 1015 // must not float (reorder) past the following store that drops the lock. 1016 // Uses a storeload to separate release_store(owner) from the 1017 // successor check. The try_set_owner() below uses cmpxchg() so 1018 // we get the fence down there. 1019 release_clear_owner(Self); 1020 OrderAccess::storeload(); 1021 1022 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1023 return; 1024 } 1025 // Other threads are blocked trying to acquire the lock. 1026 1027 // Normally the exiting thread is responsible for ensuring succession, 1028 // but if other successors are ready or other entering threads are spinning 1029 // then this thread can simply store NULL into _owner and exit without 1030 // waking a successor. The existence of spinners or ready successors 1031 // guarantees proper succession (liveness). Responsibility passes to the 1032 // ready or running successors. The exiting thread delegates the duty. 1033 // More precisely, if a successor already exists this thread is absolved 1034 // of the responsibility of waking (unparking) one. 1035 // 1036 // The _succ variable is critical to reducing futile wakeup frequency. 1037 // _succ identifies the "heir presumptive" thread that has been made 1038 // ready (unparked) but that has not yet run. We need only one such 1039 // successor thread to guarantee progress. 1040 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1041 // section 3.3 "Futile Wakeup Throttling" for details. 1042 // 1043 // Note that spinners in Enter() also set _succ non-null. 1044 // In the current implementation spinners opportunistically set 1045 // _succ so that exiting threads might avoid waking a successor. 1046 // Another less appealing alternative would be for the exiting thread 1047 // to drop the lock and then spin briefly to see if a spinner managed 1048 // to acquire the lock. If so, the exiting thread could exit 1049 // immediately without waking a successor, otherwise the exiting 1050 // thread would need to dequeue and wake a successor. 1051 // (Note that we'd need to make the post-drop spin short, but no 1052 // shorter than the worst-case round-trip cache-line migration time. 1053 // The dropped lock needs to become visible to the spinner, and then 1054 // the acquisition of the lock by the spinner must become visible to 1055 // the exiting thread). 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 (try_set_owner_from(NULL, Self) != NULL) { 1064 return; 1065 } 1066 1067 guarantee(_owner == THREAD, "invariant"); 1068 1069 ObjectWaiter * w = NULL; 1070 1071 w = _EntryList; 1072 if (w != NULL) { 1073 // I'd like to write: guarantee (w->_thread != Self). 1074 // But in practice an exiting thread may find itself on the EntryList. 1075 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1076 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1077 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1078 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1079 // release the lock "O". T2 resumes immediately after the ST of null into 1080 // _owner, above. T2 notices that the EntryList is populated, so it 1081 // reacquires the lock and then finds itself on the EntryList. 1082 // Given all that, we have to tolerate the circumstance where "w" is 1083 // associated with Self. 1084 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1085 ExitEpilog(Self, w); 1086 return; 1087 } 1088 1089 // If we find that both _cxq and EntryList are null then just 1090 // re-run the exit protocol from the top. 1091 w = _cxq; 1092 if (w == NULL) continue; 1093 1094 // Drain _cxq into EntryList - bulk transfer. 1095 // First, detach _cxq. 1096 // The following loop is tantamount to: w = swap(&cxq, NULL) 1097 for (;;) { 1098 assert(w != NULL, "Invariant"); 1099 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL); 1100 if (u == w) break; 1101 w = u; 1102 } 1103 1104 assert(w != NULL, "invariant"); 1105 assert(_EntryList == NULL, "invariant"); 1106 1107 // Convert the LIFO SLL anchored by _cxq into a DLL. 1108 // The list reorganization step operates in O(LENGTH(w)) time. 1109 // It's critical that this step operate quickly as 1110 // "Self" still holds the outer-lock, restricting parallelism 1111 // and effectively lengthening the critical section. 1112 // Invariant: s chases t chases u. 1113 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1114 // we have faster access to the tail. 1115 1116 _EntryList = w; 1117 ObjectWaiter * q = NULL; 1118 ObjectWaiter * p; 1119 for (p = w; p != NULL; p = p->_next) { 1120 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1121 p->TState = ObjectWaiter::TS_ENTER; 1122 p->_prev = q; 1123 q = p; 1124 } 1125 1126 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1127 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1128 1129 // See if we can abdicate to a spinner instead of waking a thread. 1130 // A primary goal of the implementation is to reduce the 1131 // context-switch rate. 1132 if (_succ != NULL) continue; 1133 1134 w = _EntryList; 1135 if (w != NULL) { 1136 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1137 ExitEpilog(Self, w); 1138 return; 1139 } 1140 } 1141 } 1142 1143 // ExitSuspendEquivalent: 1144 // A faster alternate to handle_special_suspend_equivalent_condition() 1145 // 1146 // handle_special_suspend_equivalent_condition() unconditionally 1147 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1148 // operations have high latency. Note that in ::enter() we call HSSEC 1149 // while holding the monitor, so we effectively lengthen the critical sections. 1150 // 1151 // There are a number of possible solutions: 1152 // 1153 // A. To ameliorate the problem we might also defer state transitions 1154 // to as late as possible -- just prior to parking. 1155 // Given that, we'd call HSSEC after having returned from park(), 1156 // but before attempting to acquire the monitor. This is only a 1157 // partial solution. It avoids calling HSSEC while holding the 1158 // monitor (good), but it still increases successor reacquisition latency -- 1159 // the interval between unparking a successor and the time the successor 1160 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1161 // If we use this technique we can also avoid EnterI()-exit() loop 1162 // in ::enter() where we iteratively drop the lock and then attempt 1163 // to reacquire it after suspending. 1164 // 1165 // B. In the future we might fold all the suspend bits into a 1166 // composite per-thread suspend flag and then update it with CAS(). 1167 // Alternately, a Dekker-like mechanism with multiple variables 1168 // would suffice: 1169 // ST Self->_suspend_equivalent = false 1170 // MEMBAR 1171 // LD Self_>_suspend_flags 1172 1173 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1174 return jSelf->handle_special_suspend_equivalent_condition(); 1175 } 1176 1177 1178 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1179 assert(_owner == Self, "invariant"); 1180 1181 // Exit protocol: 1182 // 1. ST _succ = wakee 1183 // 2. membar #loadstore|#storestore; 1184 // 2. ST _owner = NULL 1185 // 3. unpark(wakee) 1186 1187 _succ = Wakee->_thread; 1188 ParkEvent * Trigger = Wakee->_event; 1189 1190 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1191 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1192 // out-of-scope (non-extant). 1193 Wakee = NULL; 1194 1195 // Drop the lock. 1196 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1197 release_clear_owner(Self); 1198 OrderAccess::fence(); 1199 1200 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1201 Trigger->unpark(); 1202 1203 // Maintain stats and report events to JVMTI 1204 OM_PERFDATA_OP(Parks, inc()); 1205 } 1206 1207 1208 // ----------------------------------------------------------------------------- 1209 // Class Loader deadlock handling. 1210 // 1211 // complete_exit exits a lock returning recursion count 1212 // complete_exit/reenter operate as a wait without waiting 1213 // complete_exit requires an inflated monitor 1214 // The _owner field is not always the Thread addr even with an 1215 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1216 // thread due to contention. 1217 intx ObjectMonitor::complete_exit(TRAPS) { 1218 Thread * const Self = THREAD; 1219 assert(Self->is_Java_thread(), "Must be Java thread!"); 1220 JavaThread *jt = (JavaThread *)THREAD; 1221 1222 assert(InitDone, "Unexpectedly not initialized"); 1223 1224 void* cur = Atomic::load(&_owner); 1225 if (THREAD != cur) { 1226 if (THREAD->is_lock_owned((address)cur)) { 1227 assert(_recursions == 0, "internal state error"); 1228 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 1229 _recursions = 0; 1230 } 1231 } 1232 1233 guarantee(Self == _owner, "complete_exit not owner"); 1234 intx save = _recursions; // record the old recursion count 1235 _recursions = 0; // set the recursion level to be 0 1236 exit(true, Self); // exit the monitor 1237 guarantee(_owner != Self, "invariant"); 1238 return save; 1239 } 1240 1241 // reenter() enters a lock and sets recursion count 1242 // complete_exit/reenter operate as a wait without waiting 1243 void ObjectMonitor::reenter(intx recursions, TRAPS) { 1244 Thread * const Self = THREAD; 1245 assert(Self->is_Java_thread(), "Must be Java thread!"); 1246 JavaThread *jt = (JavaThread *)THREAD; 1247 1248 guarantee(_owner != Self, "reenter already owner"); 1249 enter(THREAD); 1250 // Entered the monitor. 1251 guarantee(_recursions == 0, "reenter recursion"); 1252 _recursions = recursions; 1253 } 1254 1255 // Checks that the current THREAD owns this monitor and causes an 1256 // immediate return if it doesn't. We don't use the CHECK macro 1257 // because we want the IMSE to be the only exception that is thrown 1258 // from the call site when false is returned. Any other pending 1259 // exception is ignored. 1260 #define CHECK_OWNER() \ 1261 do { \ 1262 if (!check_owner(THREAD)) { \ 1263 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1264 return; \ 1265 } \ 1266 } while (false) 1267 1268 // Returns true if the specified thread owns the ObjectMonitor. 1269 // Otherwise returns false and throws IllegalMonitorStateException 1270 // (IMSE). If there is a pending exception and the specified thread 1271 // is not the owner, that exception will be replaced by the IMSE. 1272 bool ObjectMonitor::check_owner(Thread* THREAD) { 1273 void* cur = Atomic::load(&_owner); 1274 if (cur == THREAD) { 1275 return true; 1276 } 1277 if (THREAD->is_lock_owned((address)cur)) { 1278 set_owner_from_BasicLock(cur, THREAD); // Convert from BasicLock* to Thread*. 1279 _recursions = 0; 1280 return true; 1281 } 1282 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1283 "current thread is not owner", false); 1284 } 1285 1286 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1287 ObjectMonitor* monitor, 1288 jlong notifier_tid, 1289 jlong timeout, 1290 bool timedout) { 1291 assert(event != NULL, "invariant"); 1292 assert(monitor != NULL, "invariant"); 1293 event->set_monitorClass(((oop)monitor->object())->klass()); 1294 event->set_timeout(timeout); 1295 event->set_address((uintptr_t)monitor->object_addr()); 1296 event->set_notifier(notifier_tid); 1297 event->set_timedOut(timedout); 1298 event->commit(); 1299 } 1300 1301 // ----------------------------------------------------------------------------- 1302 // Wait/Notify/NotifyAll 1303 // 1304 // Note: a subset of changes to ObjectMonitor::wait() 1305 // will need to be replicated in complete_exit 1306 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1307 Thread * const Self = THREAD; 1308 assert(Self->is_Java_thread(), "Must be Java thread!"); 1309 JavaThread *jt = (JavaThread *)THREAD; 1310 1311 assert(InitDone, "Unexpectedly not initialized"); 1312 1313 CHECK_OWNER(); // Throws IMSE if not owner. 1314 1315 EventJavaMonitorWait event; 1316 1317 // check for a pending interrupt 1318 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1319 // post monitor waited event. Note that this is past-tense, we are done waiting. 1320 if (JvmtiExport::should_post_monitor_waited()) { 1321 // Note: 'false' parameter is passed here because the 1322 // wait was not timed out due to thread interrupt. 1323 JvmtiExport::post_monitor_waited(jt, this, false); 1324 1325 // In this short circuit of the monitor wait protocol, the 1326 // current thread never drops ownership of the monitor and 1327 // never gets added to the wait queue so the current thread 1328 // cannot be made the successor. This means that the 1329 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1330 // consume an unpark() meant for the ParkEvent associated with 1331 // this ObjectMonitor. 1332 } 1333 if (event.should_commit()) { 1334 post_monitor_wait_event(&event, this, 0, millis, false); 1335 } 1336 THROW(vmSymbols::java_lang_InterruptedException()); 1337 return; 1338 } 1339 1340 assert(Self->_Stalled == 0, "invariant"); 1341 Self->_Stalled = intptr_t(this); 1342 jt->set_current_waiting_monitor(this); 1343 1344 // create a node to be put into the queue 1345 // Critically, after we reset() the event but prior to park(), we must check 1346 // for a pending interrupt. 1347 ObjectWaiter node(Self); 1348 node.TState = ObjectWaiter::TS_WAIT; 1349 Self->_ParkEvent->reset(); 1350 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1351 1352 // Enter the waiting queue, which is a circular doubly linked list in this case 1353 // but it could be a priority queue or any data structure. 1354 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1355 // by the the owner of the monitor *except* in the case where park() 1356 // returns because of a timeout of interrupt. Contention is exceptionally rare 1357 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1358 1359 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1360 AddWaiter(&node); 1361 Thread::SpinRelease(&_WaitSetLock); 1362 1363 _Responsible = NULL; 1364 1365 intx save = _recursions; // record the old recursion count 1366 _waiters++; // increment the number of waiters 1367 _recursions = 0; // set the recursion level to be 1 1368 exit(true, Self); // exit the monitor 1369 guarantee(_owner != Self, "invariant"); 1370 1371 // The thread is on the WaitSet list - now park() it. 1372 // On MP systems it's conceivable that a brief spin before we park 1373 // could be profitable. 1374 // 1375 // TODO-FIXME: change the following logic to a loop of the form 1376 // while (!timeout && !interrupted && _notified == 0) park() 1377 1378 int ret = OS_OK; 1379 int WasNotified = 0; 1380 1381 // Need to check interrupt state whilst still _thread_in_vm 1382 bool interrupted = interruptible && jt->is_interrupted(false); 1383 1384 { // State transition wrappers 1385 OSThread* osthread = Self->osthread(); 1386 OSThreadWaitState osts(osthread, true); 1387 { 1388 ThreadBlockInVM tbivm(jt); 1389 // Thread is in thread_blocked state and oop access is unsafe. 1390 jt->set_suspend_equivalent(); 1391 1392 if (interrupted || HAS_PENDING_EXCEPTION) { 1393 // Intentionally empty 1394 } else if (node._notified == 0) { 1395 if (millis <= 0) { 1396 Self->_ParkEvent->park(); 1397 } else { 1398 ret = Self->_ParkEvent->park(millis); 1399 } 1400 } 1401 1402 // were we externally suspended while we were waiting? 1403 if (ExitSuspendEquivalent (jt)) { 1404 // TODO-FIXME: add -- if succ == Self then succ = null. 1405 jt->java_suspend_self(); 1406 } 1407 1408 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1409 1410 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1411 // from the WaitSet to the EntryList. 1412 // See if we need to remove Node from the WaitSet. 1413 // We use double-checked locking to avoid grabbing _WaitSetLock 1414 // if the thread is not on the wait queue. 1415 // 1416 // Note that we don't need a fence before the fetch of TState. 1417 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1418 // written by the is thread. (perhaps the fetch might even be satisfied 1419 // by a look-aside into the processor's own store buffer, although given 1420 // the length of the code path between the prior ST and this load that's 1421 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1422 // then we'll acquire the lock and then re-fetch a fresh TState value. 1423 // That is, we fail toward safety. 1424 1425 if (node.TState == ObjectWaiter::TS_WAIT) { 1426 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1427 if (node.TState == ObjectWaiter::TS_WAIT) { 1428 DequeueSpecificWaiter(&node); // unlink from WaitSet 1429 assert(node._notified == 0, "invariant"); 1430 node.TState = ObjectWaiter::TS_RUN; 1431 } 1432 Thread::SpinRelease(&_WaitSetLock); 1433 } 1434 1435 // The thread is now either on off-list (TS_RUN), 1436 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1437 // The Node's TState variable is stable from the perspective of this thread. 1438 // No other threads will asynchronously modify TState. 1439 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1440 OrderAccess::loadload(); 1441 if (_succ == Self) _succ = NULL; 1442 WasNotified = node._notified; 1443 1444 // Reentry phase -- reacquire the monitor. 1445 // re-enter contended monitor after object.wait(). 1446 // retain OBJECT_WAIT state until re-enter successfully completes 1447 // Thread state is thread_in_vm and oop access is again safe, 1448 // although the raw address of the object may have changed. 1449 // (Don't cache naked oops over safepoints, of course). 1450 1451 // post monitor waited event. Note that this is past-tense, we are done waiting. 1452 if (JvmtiExport::should_post_monitor_waited()) { 1453 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1454 1455 if (node._notified != 0 && _succ == Self) { 1456 // In this part of the monitor wait-notify-reenter protocol it 1457 // is possible (and normal) for another thread to do a fastpath 1458 // monitor enter-exit while this thread is still trying to get 1459 // to the reenter portion of the protocol. 1460 // 1461 // The ObjectMonitor was notified and the current thread is 1462 // the successor which also means that an unpark() has already 1463 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1464 // consume the unpark() that was done when the successor was 1465 // set because the same ParkEvent is shared between Java 1466 // monitors and JVM/TI RawMonitors (for now). 1467 // 1468 // We redo the unpark() to ensure forward progress, i.e., we 1469 // don't want all pending threads hanging (parked) with none 1470 // entering the unlocked monitor. 1471 node._event->unpark(); 1472 } 1473 } 1474 1475 if (event.should_commit()) { 1476 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1477 } 1478 1479 OrderAccess::fence(); 1480 1481 assert(Self->_Stalled != 0, "invariant"); 1482 Self->_Stalled = 0; 1483 1484 assert(_owner != Self, "invariant"); 1485 ObjectWaiter::TStates v = node.TState; 1486 if (v == ObjectWaiter::TS_RUN) { 1487 enter(Self); 1488 } else { 1489 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1490 ReenterI(Self, &node); 1491 node.wait_reenter_end(this); 1492 } 1493 1494 // Self has reacquired the lock. 1495 // Lifecycle - the node representing Self must not appear on any queues. 1496 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1497 // want residual elements associated with this thread left on any lists. 1498 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1499 assert(_owner == Self, "invariant"); 1500 assert(_succ != Self, "invariant"); 1501 } // OSThreadWaitState() 1502 1503 jt->set_current_waiting_monitor(NULL); 1504 1505 guarantee(_recursions == 0, "invariant"); 1506 _recursions = save; // restore the old recursion count 1507 _waiters--; // decrement the number of waiters 1508 1509 // Verify a few postconditions 1510 assert(_owner == Self, "invariant"); 1511 assert(_succ != Self, "invariant"); 1512 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1513 1514 // check if the notification happened 1515 if (!WasNotified) { 1516 // no, it could be timeout or Thread.interrupt() or both 1517 // check for interrupt event, otherwise it is timeout 1518 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1519 THROW(vmSymbols::java_lang_InterruptedException()); 1520 } 1521 } 1522 1523 // NOTE: Spurious wake up will be consider as timeout. 1524 // Monitor notify has precedence over thread interrupt. 1525 } 1526 1527 1528 // Consider: 1529 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1530 // then instead of transferring a thread from the WaitSet to the EntryList 1531 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1532 1533 void ObjectMonitor::INotify(Thread * Self) { 1534 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1535 ObjectWaiter * iterator = DequeueWaiter(); 1536 if (iterator != NULL) { 1537 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1538 guarantee(iterator->_notified == 0, "invariant"); 1539 // Disposition - what might we do with iterator ? 1540 // a. add it directly to the EntryList - either tail (policy == 1) 1541 // or head (policy == 0). 1542 // b. push it onto the front of the _cxq (policy == 2). 1543 // For now we use (b). 1544 1545 iterator->TState = ObjectWaiter::TS_ENTER; 1546 1547 iterator->_notified = 1; 1548 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1549 1550 ObjectWaiter * list = _EntryList; 1551 if (list != NULL) { 1552 assert(list->_prev == NULL, "invariant"); 1553 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1554 assert(list != iterator, "invariant"); 1555 } 1556 1557 // prepend to cxq 1558 if (list == NULL) { 1559 iterator->_next = iterator->_prev = NULL; 1560 _EntryList = iterator; 1561 } else { 1562 iterator->TState = ObjectWaiter::TS_CXQ; 1563 for (;;) { 1564 ObjectWaiter * front = _cxq; 1565 iterator->_next = front; 1566 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1567 break; 1568 } 1569 } 1570 } 1571 1572 // _WaitSetLock protects the wait queue, not the EntryList. We could 1573 // move the add-to-EntryList operation, above, outside the critical section 1574 // protected by _WaitSetLock. In practice that's not useful. With the 1575 // exception of wait() timeouts and interrupts the monitor owner 1576 // is the only thread that grabs _WaitSetLock. There's almost no contention 1577 // on _WaitSetLock so it's not profitable to reduce the length of the 1578 // critical section. 1579 1580 iterator->wait_reenter_begin(this); 1581 } 1582 Thread::SpinRelease(&_WaitSetLock); 1583 } 1584 1585 // Consider: a not-uncommon synchronization bug is to use notify() when 1586 // notifyAll() is more appropriate, potentially resulting in stranded 1587 // threads; this is one example of a lost wakeup. A useful diagnostic 1588 // option is to force all notify() operations to behave as notifyAll(). 1589 // 1590 // Note: We can also detect many such problems with a "minimum wait". 1591 // When the "minimum wait" is set to a small non-zero timeout value 1592 // and the program does not hang whereas it did absent "minimum wait", 1593 // that suggests a lost wakeup bug. 1594 1595 void ObjectMonitor::notify(TRAPS) { 1596 CHECK_OWNER(); // Throws IMSE if not owner. 1597 if (_WaitSet == NULL) { 1598 return; 1599 } 1600 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1601 INotify(THREAD); 1602 OM_PERFDATA_OP(Notifications, inc(1)); 1603 } 1604 1605 1606 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1607 // from the waitset to the EntryList. This could be done more efficiently with a 1608 // single bulk transfer but in practice it's not time-critical. Beware too, 1609 // that in prepend-mode we invert the order of the waiters. Let's say that the 1610 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1611 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1612 1613 void ObjectMonitor::notifyAll(TRAPS) { 1614 CHECK_OWNER(); // Throws IMSE if not owner. 1615 if (_WaitSet == NULL) { 1616 return; 1617 } 1618 1619 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1620 int tally = 0; 1621 while (_WaitSet != NULL) { 1622 tally++; 1623 INotify(THREAD); 1624 } 1625 1626 OM_PERFDATA_OP(Notifications, inc(tally)); 1627 } 1628 1629 // ----------------------------------------------------------------------------- 1630 // Adaptive Spinning Support 1631 // 1632 // Adaptive spin-then-block - rational spinning 1633 // 1634 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1635 // algorithm. On high order SMP systems it would be better to start with 1636 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1637 // a contending thread could enqueue itself on the cxq and then spin locally 1638 // on a thread-specific variable such as its ParkEvent._Event flag. 1639 // That's left as an exercise for the reader. Note that global spinning is 1640 // not problematic on Niagara, as the L2 cache serves the interconnect and 1641 // has both low latency and massive bandwidth. 1642 // 1643 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1644 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1645 // (duration) or we can fix the count at approximately the duration of 1646 // a context switch and vary the frequency. Of course we could also 1647 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1648 // For a description of 'Adaptive spin-then-block mutual exclusion in 1649 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1650 // 1651 // This implementation varies the duration "D", where D varies with 1652 // the success rate of recent spin attempts. (D is capped at approximately 1653 // length of a round-trip context switch). The success rate for recent 1654 // spin attempts is a good predictor of the success rate of future spin 1655 // attempts. The mechanism adapts automatically to varying critical 1656 // section length (lock modality), system load and degree of parallelism. 1657 // D is maintained per-monitor in _SpinDuration and is initialized 1658 // optimistically. Spin frequency is fixed at 100%. 1659 // 1660 // Note that _SpinDuration is volatile, but we update it without locks 1661 // or atomics. The code is designed so that _SpinDuration stays within 1662 // a reasonable range even in the presence of races. The arithmetic 1663 // operations on _SpinDuration are closed over the domain of legal values, 1664 // so at worst a race will install and older but still legal value. 1665 // At the very worst this introduces some apparent non-determinism. 1666 // We might spin when we shouldn't or vice-versa, but since the spin 1667 // count are relatively short, even in the worst case, the effect is harmless. 1668 // 1669 // Care must be taken that a low "D" value does not become an 1670 // an absorbing state. Transient spinning failures -- when spinning 1671 // is overall profitable -- should not cause the system to converge 1672 // on low "D" values. We want spinning to be stable and predictable 1673 // and fairly responsive to change and at the same time we don't want 1674 // it to oscillate, become metastable, be "too" non-deterministic, 1675 // or converge on or enter undesirable stable absorbing states. 1676 // 1677 // We implement a feedback-based control system -- using past behavior 1678 // to predict future behavior. We face two issues: (a) if the 1679 // input signal is random then the spin predictor won't provide optimal 1680 // results, and (b) if the signal frequency is too high then the control 1681 // system, which has some natural response lag, will "chase" the signal. 1682 // (b) can arise from multimodal lock hold times. Transient preemption 1683 // can also result in apparent bimodal lock hold times. 1684 // Although sub-optimal, neither condition is particularly harmful, as 1685 // in the worst-case we'll spin when we shouldn't or vice-versa. 1686 // The maximum spin duration is rather short so the failure modes aren't bad. 1687 // To be conservative, I've tuned the gain in system to bias toward 1688 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1689 // "rings" or oscillates between spinning and not spinning. This happens 1690 // when spinning is just on the cusp of profitability, however, so the 1691 // situation is not dire. The state is benign -- there's no need to add 1692 // hysteresis control to damp the transition rate between spinning and 1693 // not spinning. 1694 1695 // Spinning: Fixed frequency (100%), vary duration 1696 int ObjectMonitor::TrySpin(Thread * Self) { 1697 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1698 int ctr = Knob_FixedSpin; 1699 if (ctr != 0) { 1700 while (--ctr >= 0) { 1701 if (TryLock(Self) > 0) return 1; 1702 SpinPause(); 1703 } 1704 return 0; 1705 } 1706 1707 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1708 if (TryLock(Self) > 0) { 1709 // Increase _SpinDuration ... 1710 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1711 // Raising _SpurDuration to the poverty line is key. 1712 int x = _SpinDuration; 1713 if (x < Knob_SpinLimit) { 1714 if (x < Knob_Poverty) x = Knob_Poverty; 1715 _SpinDuration = x + Knob_BonusB; 1716 } 1717 return 1; 1718 } 1719 SpinPause(); 1720 } 1721 1722 // Admission control - verify preconditions for spinning 1723 // 1724 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1725 // becoming an absorbing state. Put another way, we spin briefly to 1726 // sample, just in case the system load, parallelism, contention, or lock 1727 // modality changed. 1728 // 1729 // Consider the following alternative: 1730 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1731 // spin attempt. "Periodically" might mean after a tally of 1732 // the # of failed spin attempts (or iterations) reaches some threshold. 1733 // This takes us into the realm of 1-out-of-N spinning, where we 1734 // hold the duration constant but vary the frequency. 1735 1736 ctr = _SpinDuration; 1737 if (ctr <= 0) return 0; 1738 1739 if (NotRunnable(Self, (Thread *) _owner)) { 1740 return 0; 1741 } 1742 1743 // We're good to spin ... spin ingress. 1744 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1745 // when preparing to LD...CAS _owner, etc and the CAS is likely 1746 // to succeed. 1747 if (_succ == NULL) { 1748 _succ = Self; 1749 } 1750 Thread * prv = NULL; 1751 1752 // There are three ways to exit the following loop: 1753 // 1. A successful spin where this thread has acquired the lock. 1754 // 2. Spin failure with prejudice 1755 // 3. Spin failure without prejudice 1756 1757 while (--ctr >= 0) { 1758 1759 // Periodic polling -- Check for pending GC 1760 // Threads may spin while they're unsafe. 1761 // We don't want spinning threads to delay the JVM from reaching 1762 // a stop-the-world safepoint or to steal cycles from GC. 1763 // If we detect a pending safepoint we abort in order that 1764 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1765 // this thread, if safe, doesn't steal cycles from GC. 1766 // This is in keeping with the "no loitering in runtime" rule. 1767 // We periodically check to see if there's a safepoint pending. 1768 if ((ctr & 0xFF) == 0) { 1769 if (SafepointMechanism::should_block(Self)) { 1770 goto Abort; // abrupt spin egress 1771 } 1772 SpinPause(); 1773 } 1774 1775 // Probe _owner with TATAS 1776 // If this thread observes the monitor transition or flicker 1777 // from locked to unlocked to locked, then the odds that this 1778 // thread will acquire the lock in this spin attempt go down 1779 // considerably. The same argument applies if the CAS fails 1780 // or if we observe _owner change from one non-null value to 1781 // another non-null value. In such cases we might abort 1782 // the spin without prejudice or apply a "penalty" to the 1783 // spin count-down variable "ctr", reducing it by 100, say. 1784 1785 Thread * ox = (Thread *) _owner; 1786 if (ox == NULL) { 1787 ox = (Thread*)try_set_owner_from(NULL, Self); 1788 if (ox == NULL) { 1789 // The CAS succeeded -- this thread acquired ownership 1790 // Take care of some bookkeeping to exit spin state. 1791 if (_succ == Self) { 1792 _succ = NULL; 1793 } 1794 1795 // Increase _SpinDuration : 1796 // The spin was successful (profitable) so we tend toward 1797 // longer spin attempts in the future. 1798 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1799 // If we acquired the lock early in the spin cycle it 1800 // makes sense to increase _SpinDuration proportionally. 1801 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1802 int x = _SpinDuration; 1803 if (x < Knob_SpinLimit) { 1804 if (x < Knob_Poverty) x = Knob_Poverty; 1805 _SpinDuration = x + Knob_Bonus; 1806 } 1807 return 1; 1808 } 1809 1810 // The CAS failed ... we can take any of the following actions: 1811 // * penalize: ctr -= CASPenalty 1812 // * exit spin with prejudice -- goto Abort; 1813 // * exit spin without prejudice. 1814 // * Since CAS is high-latency, retry again immediately. 1815 prv = ox; 1816 goto Abort; 1817 } 1818 1819 // Did lock ownership change hands ? 1820 if (ox != prv && prv != NULL) { 1821 goto Abort; 1822 } 1823 prv = ox; 1824 1825 // Abort the spin if the owner is not executing. 1826 // The owner must be executing in order to drop the lock. 1827 // Spinning while the owner is OFFPROC is idiocy. 1828 // Consider: ctr -= RunnablePenalty ; 1829 if (NotRunnable(Self, ox)) { 1830 goto Abort; 1831 } 1832 if (_succ == NULL) { 1833 _succ = Self; 1834 } 1835 } 1836 1837 // Spin failed with prejudice -- reduce _SpinDuration. 1838 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1839 // AIMD is globally stable. 1840 { 1841 int x = _SpinDuration; 1842 if (x > 0) { 1843 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1844 // This is globally sample and tends to damp the response. 1845 x -= Knob_Penalty; 1846 if (x < 0) x = 0; 1847 _SpinDuration = x; 1848 } 1849 } 1850 1851 Abort: 1852 if (_succ == Self) { 1853 _succ = NULL; 1854 // Invariant: after setting succ=null a contending thread 1855 // must recheck-retry _owner before parking. This usually happens 1856 // in the normal usage of TrySpin(), but it's safest 1857 // to make TrySpin() as foolproof as possible. 1858 OrderAccess::fence(); 1859 if (TryLock(Self) > 0) return 1; 1860 } 1861 return 0; 1862 } 1863 1864 // NotRunnable() -- informed spinning 1865 // 1866 // Don't bother spinning if the owner is not eligible to drop the lock. 1867 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1868 // The thread must be runnable in order to drop the lock in timely fashion. 1869 // If the _owner is not runnable then spinning will not likely be 1870 // successful (profitable). 1871 // 1872 // Beware -- the thread referenced by _owner could have died 1873 // so a simply fetch from _owner->_thread_state might trap. 1874 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1875 // Because of the lifecycle issues, the _thread_state values 1876 // observed by NotRunnable() might be garbage. NotRunnable must 1877 // tolerate this and consider the observed _thread_state value 1878 // as advisory. 1879 // 1880 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1881 // a thread pointer. 1882 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1883 // with the LSB of _owner. Another option would be to probabilistically probe 1884 // the putative _owner->TypeTag value. 1885 // 1886 // Checking _thread_state isn't perfect. Even if the thread is 1887 // in_java it might be blocked on a page-fault or have been preempted 1888 // and sitting on a ready/dispatch queue. 1889 // 1890 // The return value from NotRunnable() is *advisory* -- the 1891 // result is based on sampling and is not necessarily coherent. 1892 // The caller must tolerate false-negative and false-positive errors. 1893 // Spinning, in general, is probabilistic anyway. 1894 1895 1896 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1897 // Check ox->TypeTag == 2BAD. 1898 if (ox == NULL) return 0; 1899 1900 // Avoid transitive spinning ... 1901 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1902 // Immediately after T1 acquires L it's possible that T2, also 1903 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1904 // This occurs transiently after T1 acquired L but before 1905 // T1 managed to clear T1.Stalled. T2 does not need to abort 1906 // its spin in this circumstance. 1907 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1908 1909 if (BlockedOn == 1) return 1; 1910 if (BlockedOn != 0) { 1911 return BlockedOn != intptr_t(this) && _owner == ox; 1912 } 1913 1914 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1915 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1916 // consider also: jst != _thread_in_Java -- but that's overspecific. 1917 return jst == _thread_blocked || jst == _thread_in_native; 1918 } 1919 1920 1921 // ----------------------------------------------------------------------------- 1922 // WaitSet management ... 1923 1924 ObjectWaiter::ObjectWaiter(Thread* thread) { 1925 _next = NULL; 1926 _prev = NULL; 1927 _notified = 0; 1928 _notifier_tid = 0; 1929 TState = TS_RUN; 1930 _thread = thread; 1931 _event = thread->_ParkEvent; 1932 _active = false; 1933 assert(_event != NULL, "invariant"); 1934 } 1935 1936 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1937 JavaThread *jt = (JavaThread *)this->_thread; 1938 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1939 } 1940 1941 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1942 JavaThread *jt = (JavaThread *)this->_thread; 1943 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1944 } 1945 1946 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1947 assert(node != NULL, "should not add NULL node"); 1948 assert(node->_prev == NULL, "node already in list"); 1949 assert(node->_next == NULL, "node already in list"); 1950 // put node at end of queue (circular doubly linked list) 1951 if (_WaitSet == NULL) { 1952 _WaitSet = node; 1953 node->_prev = node; 1954 node->_next = node; 1955 } else { 1956 ObjectWaiter* head = _WaitSet; 1957 ObjectWaiter* tail = head->_prev; 1958 assert(tail->_next == head, "invariant check"); 1959 tail->_next = node; 1960 head->_prev = node; 1961 node->_next = head; 1962 node->_prev = tail; 1963 } 1964 } 1965 1966 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1967 // dequeue the very first waiter 1968 ObjectWaiter* waiter = _WaitSet; 1969 if (waiter) { 1970 DequeueSpecificWaiter(waiter); 1971 } 1972 return waiter; 1973 } 1974 1975 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1976 assert(node != NULL, "should not dequeue NULL node"); 1977 assert(node->_prev != NULL, "node already removed from list"); 1978 assert(node->_next != NULL, "node already removed from list"); 1979 // when the waiter has woken up because of interrupt, 1980 // timeout or other spurious wake-up, dequeue the 1981 // waiter from waiting list 1982 ObjectWaiter* next = node->_next; 1983 if (next == node) { 1984 assert(node->_prev == node, "invariant check"); 1985 _WaitSet = NULL; 1986 } else { 1987 ObjectWaiter* prev = node->_prev; 1988 assert(prev->_next == node, "invariant check"); 1989 assert(next->_prev == node, "invariant check"); 1990 next->_prev = prev; 1991 prev->_next = next; 1992 if (_WaitSet == node) { 1993 _WaitSet = next; 1994 } 1995 } 1996 node->_next = NULL; 1997 node->_prev = NULL; 1998 } 1999 2000 // ----------------------------------------------------------------------------- 2001 // PerfData support 2002 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2003 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2004 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2005 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2006 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2007 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2008 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2009 2010 // One-shot global initialization for the sync subsystem. 2011 // We could also defer initialization and initialize on-demand 2012 // the first time we call ObjectSynchronizer::inflate(). 2013 // Initialization would be protected - like so many things - by 2014 // the MonitorCache_lock. 2015 2016 void ObjectMonitor::Initialize() { 2017 assert(!InitDone, "invariant"); 2018 2019 if (!os::is_MP()) { 2020 Knob_SpinLimit = 0; 2021 Knob_PreSpin = 0; 2022 Knob_FixedSpin = -1; 2023 } 2024 2025 if (UsePerfData) { 2026 EXCEPTION_MARK; 2027 #define NEWPERFCOUNTER(n) \ 2028 { \ 2029 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2030 CHECK); \ 2031 } 2032 #define NEWPERFVARIABLE(n) \ 2033 { \ 2034 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2035 CHECK); \ 2036 } 2037 NEWPERFCOUNTER(_sync_Inflations); 2038 NEWPERFCOUNTER(_sync_Deflations); 2039 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2040 NEWPERFCOUNTER(_sync_FutileWakeups); 2041 NEWPERFCOUNTER(_sync_Parks); 2042 NEWPERFCOUNTER(_sync_Notifications); 2043 NEWPERFVARIABLE(_sync_MonExtant); 2044 #undef NEWPERFCOUNTER 2045 #undef NEWPERFVARIABLE 2046 } 2047 2048 DEBUG_ONLY(InitDone = true;) 2049 } 2050 2051 ObjectMonitorHandle::~ObjectMonitorHandle() { 2052 if (_om_ptr != NULL) { 2053 _om_ptr->dec_ref_count(); 2054 _om_ptr = NULL; 2055 } 2056 } 2057 2058 // Save the ObjectMonitor* associated with the specified markWord and 2059 // increment the ref_count. This function should only be called if 2060 // the caller has verified mark.has_monitor() == true. The object 2061 // parameter is needed to verify that ObjectMonitor* has not been 2062 // deflated and reused for another object. 2063 // 2064 // This function returns true if the ObjectMonitor* has been safely 2065 // saved. This function returns false if we have lost a race with 2066 // async deflation; the caller should retry as appropriate. 2067 // 2068 bool ObjectMonitorHandle::save_om_ptr(oop object, markWord mark) { 2069 // is_marked() is a superset of has_monitor() so make sure we 2070 // are called with the proper markWord value. 2071 guarantee(mark.has_monitor() && !mark.is_marked(), "sanity check: mark=" 2072 INTPTR_FORMAT, mark.value()); 2073 2074 ObjectMonitor* om_ptr = mark.monitor(); 2075 om_ptr->inc_ref_count(); 2076 2077 if (AsyncDeflateIdleMonitors) { 2078 // Race here if monitor is not owned! The above ref_count bump 2079 // will cause subsequent async deflation to skip it. However, 2080 // previous or concurrent async deflation is a race. 2081 if (om_ptr->owner_is_DEFLATER_MARKER() && om_ptr->ref_count() <= 0) { 2082 // Async deflation is in progress and our ref_count increment 2083 // above lost the race to async deflation. Attempt to restore 2084 // the header/dmw to the object's header so that we only retry 2085 // once if the deflater thread happens to be slow. 2086 om_ptr->install_displaced_markword_in_object(object); 2087 om_ptr->dec_ref_count(); 2088 return false; 2089 } 2090 if (om_ptr->ref_count() <= 0) { 2091 // Async deflation is in the process of bailing out, but has not 2092 // yet restored the ref_count field so we return false to force 2093 // a retry. We want a positive ref_count value for a true return. 2094 om_ptr->dec_ref_count(); 2095 return false; 2096 } 2097 // The ObjectMonitor could have been deflated and reused for 2098 // another object before we bumped the ref_count so make sure 2099 // our object still refers to this ObjectMonitor. 2100 // Note: With handshakes after deflation is this race even 2101 // possible anymore? 2102 const markWord tmp = object->mark(); 2103 if (!tmp.has_monitor() || tmp.monitor() != om_ptr) { 2104 // Async deflation and reuse won the race so we have to retry. 2105 // Skip object header restoration since that's already done. 2106 om_ptr->dec_ref_count(); 2107 return false; 2108 } 2109 } 2110 2111 ADIM_guarantee(_om_ptr == NULL, "sanity check: _om_ptr=" INTPTR_FORMAT, 2112 p2i(_om_ptr)); 2113 _om_ptr = om_ptr; 2114 return true; 2115 } 2116 2117 // For internal use by ObjectSynchronizer::inflate(). 2118 // This function is only used when we don't have to worry about async 2119 // deflation of the specified ObjectMonitor*. 2120 void ObjectMonitorHandle::set_om_ptr(ObjectMonitor* om_ptr) { 2121 if (_om_ptr == NULL) { 2122 ADIM_guarantee(om_ptr != NULL, "cannot clear an unset om_ptr"); 2123 om_ptr->inc_ref_count(); 2124 _om_ptr = om_ptr; 2125 } else { 2126 ADIM_guarantee(om_ptr == NULL, "can only clear a set om_ptr"); 2127 _om_ptr->dec_ref_count(); 2128 _om_ptr = NULL; 2129 } 2130 } 2131 2132 // Save the specified ObjectMonitor* if it is safe, i.e., not being 2133 // async deflated. 2134 // 2135 // This function returns true if the ObjectMonitor* has been safely 2136 // saved. This function returns false if the specified ObjectMonitor* 2137 // is NULL or if we have lost a race with async deflation; the caller 2138 // can retry as appropriate. 2139 bool ObjectMonitorHandle::save_om_ptr_if_safe(ObjectMonitor* om_ptr) { 2140 if (om_ptr == NULL) { 2141 return false; // Nothing to save if input is NULL 2142 } 2143 2144 om_ptr->inc_ref_count(); 2145 2146 if (AsyncDeflateIdleMonitors) { 2147 if (om_ptr->owner_is_DEFLATER_MARKER() && om_ptr->ref_count() <= 0) { 2148 // Async deflation is in progress and our ref_count increment 2149 // above lost the race to async deflation. 2150 om_ptr->dec_ref_count(); 2151 return false; 2152 } 2153 if (om_ptr->ref_count() <= 0) { 2154 // Async deflation is in the process of bailing out, but has not 2155 // yet restored the ref_count field so we return false to force 2156 // a retry. We want a positive ref_count value for a true return. 2157 om_ptr->dec_ref_count(); 2158 return false; 2159 } 2160 // Unlike save_om_ptr(), we don't have context to determine if 2161 // the ObjectMonitor has been deflated and reused for another 2162 // object. 2163 } 2164 2165 ADIM_guarantee(_om_ptr == NULL, "sanity check: _om_ptr=" INTPTR_FORMAT, 2166 p2i(_om_ptr)); 2167 _om_ptr = om_ptr; 2168 return true; 2169 } 2170 2171 // Unset the _om_ptr field and decrement the ref_count field. 2172 void ObjectMonitorHandle::unset_om_ptr() { 2173 ADIM_guarantee(_om_ptr != NULL, "_om_ptr must not be NULL"); 2174 _om_ptr->dec_ref_count(); 2175 _om_ptr = NULL; 2176 } 2177 2178 void ObjectMonitor::print_on(outputStream* st) const { 2179 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2180 st->print("{contentions=0x%08x,waiters=0x%08x" 2181 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2182 contentions(), waiters(), recursions(), 2183 p2i(owner())); 2184 } 2185 void ObjectMonitor::print() const { print_on(tty); } 2186 2187 #ifdef ASSERT 2188 // Print the ObjectMonitor like a debugger would: 2189 // 2190 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2191 // _header = 0x0000000000000001 2192 // _object = 0x000000070ff45fd0 2193 // _allocation_state = Old 2194 // _pad_buf0 = { 2195 // [0] = '\0' 2196 // ... 2197 // [43] = '\0' 2198 // } 2199 // _owner = 0x0000000000000000 2200 // _previous_owner_tid = 0 2201 // _pad_buf1 = { 2202 // [0] = '\0' 2203 // ... 2204 // [47] = '\0' 2205 // } 2206 // _ref_count = 1 2207 // _pad_buf2 = { 2208 // [0] = '\0' 2209 // ... 2210 // [47] = '\0' 2211 // } 2212 // _next_om = 0x0000000000000000 2213 // _recursions = 0 2214 // _EntryList = 0x0000000000000000 2215 // _cxq = 0x0000000000000000 2216 // _succ = 0x0000000000000000 2217 // _Responsible = 0x0000000000000000 2218 // _Spinner = 0 2219 // _SpinDuration = 5000 2220 // _contentions = 0 2221 // _WaitSet = 0x0000700009756248 2222 // _waiters = 1 2223 // _WaitSetLock = 0 2224 // } 2225 // 2226 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2227 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2228 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2229 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 2230 st->print(" _allocation_state = "); 2231 if (is_free()) { 2232 st->print("Free"); 2233 } else if (is_old()) { 2234 st->print("Old"); 2235 } else if (is_new()) { 2236 st->print("New"); 2237 } else { 2238 st->print("unknown=%d", _allocation_state); 2239 } 2240 st->cr(); 2241 st->print_cr(" _pad_buf0 = {"); 2242 st->print_cr(" [0] = '\\0'"); 2243 st->print_cr(" ..."); 2244 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2245 st->print_cr(" }"); 2246 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 2247 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 2248 st->print_cr(" _pad_buf1 = {"); 2249 st->print_cr(" [0] = '\\0'"); 2250 st->print_cr(" ..."); 2251 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2252 st->print_cr(" }"); 2253 st->print_cr(" _ref_count = %d", ref_count()); 2254 st->print_cr(" _pad_buf2 = {"); 2255 st->print_cr(" [0] = '\\0'"); 2256 st->print_cr(" ..."); 2257 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf2) - 1); 2258 st->print_cr(" }"); 2259 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2260 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2261 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2262 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2263 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2264 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2265 st->print_cr(" _Spinner = %d", _Spinner); 2266 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2267 st->print_cr(" _contentions = %d", _contentions); 2268 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2269 st->print_cr(" _waiters = %d", _waiters); 2270 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2271 st->print_cr("}"); 2272 } 2273 #endif