1 /* 2 * Copyright (c) 1998, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "jfr/support/jfrThreadId.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "memory/resourceArea.hpp" 31 #include "oops/markWord.hpp" 32 #include "oops/oop.inline.hpp" 33 #include "runtime/atomic.hpp" 34 #include "runtime/handles.inline.hpp" 35 #include "runtime/interfaceSupport.inline.hpp" 36 #include "runtime/mutexLocker.hpp" 37 #include "runtime/objectMonitor.hpp" 38 #include "runtime/objectMonitor.inline.hpp" 39 #include "runtime/orderAccess.hpp" 40 #include "runtime/osThread.hpp" 41 #include "runtime/safepointMechanism.inline.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 #include "runtime/stubRoutines.hpp" 44 #include "runtime/thread.inline.hpp" 45 #include "services/threadService.hpp" 46 #include "utilities/dtrace.hpp" 47 #include "utilities/macros.hpp" 48 #include "utilities/preserveException.hpp" 49 #if INCLUDE_JFR 50 #include "jfr/support/jfrFlush.hpp" 51 #endif 52 53 #ifdef DTRACE_ENABLED 54 55 // Only bother with this argument setup if dtrace is available 56 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 57 58 59 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 60 char* bytes = NULL; \ 61 int len = 0; \ 62 jlong jtid = SharedRuntime::get_java_tid(thread); \ 63 Symbol* klassname = ((oop)obj)->klass()->name(); \ 64 if (klassname != NULL) { \ 65 bytes = (char*)klassname->bytes(); \ 66 len = klassname->utf8_length(); \ 67 } 68 69 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 70 { \ 71 if (DTraceMonitorProbes) { \ 72 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 73 HOTSPOT_MONITOR_WAIT(jtid, \ 74 (monitor), bytes, len, (millis)); \ 75 } \ 76 } 77 78 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 79 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 80 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 81 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 82 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 83 84 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 85 { \ 86 if (DTraceMonitorProbes) { \ 87 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 88 HOTSPOT_MONITOR_##probe(jtid, \ 89 (uintptr_t)(monitor), bytes, len); \ 90 } \ 91 } 92 93 #else // ndef DTRACE_ENABLED 94 95 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 96 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 97 98 #endif // ndef DTRACE_ENABLED 99 100 // Tunables ... 101 // The knob* variables are effectively final. Once set they should 102 // never be modified hence. Consider using __read_mostly with GCC. 103 104 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 105 106 static int Knob_Bonus = 100; // spin success bonus 107 static int Knob_BonusB = 100; // spin success bonus 108 static int Knob_Penalty = 200; // spin failure penalty 109 static int Knob_Poverty = 1000; 110 static int Knob_FixedSpin = 0; 111 static int Knob_PreSpin = 10; // 20-100 likely better 112 113 DEBUG_ONLY(static volatile bool InitDone = false;) 114 115 // ----------------------------------------------------------------------------- 116 // Theory of operations -- Monitors lists, thread residency, etc: 117 // 118 // * A thread acquires ownership of a monitor by successfully 119 // CAS()ing the _owner field from null to non-null. 120 // 121 // * Invariant: A thread appears on at most one monitor list -- 122 // cxq, EntryList or WaitSet -- at any one time. 123 // 124 // * Contending threads "push" themselves onto the cxq with CAS 125 // and then spin/park. 126 // 127 // * After a contending thread eventually acquires the lock it must 128 // dequeue itself from either the EntryList or the cxq. 129 // 130 // * The exiting thread identifies and unparks an "heir presumptive" 131 // tentative successor thread on the EntryList. Critically, the 132 // exiting thread doesn't unlink the successor thread from the EntryList. 133 // After having been unparked, the wakee will recontend for ownership of 134 // the monitor. The successor (wakee) will either acquire the lock or 135 // re-park itself. 136 // 137 // Succession is provided for by a policy of competitive handoff. 138 // The exiting thread does _not_ grant or pass ownership to the 139 // successor thread. (This is also referred to as "handoff" succession"). 140 // Instead the exiting thread releases ownership and possibly wakes 141 // a successor, so the successor can (re)compete for ownership of the lock. 142 // If the EntryList is empty but the cxq is populated the exiting 143 // thread will drain the cxq into the EntryList. It does so by 144 // by detaching the cxq (installing null with CAS) and folding 145 // the threads from the cxq into the EntryList. The EntryList is 146 // doubly linked, while the cxq is singly linked because of the 147 // CAS-based "push" used to enqueue recently arrived threads (RATs). 148 // 149 // * Concurrency invariants: 150 // 151 // -- only the monitor owner may access or mutate the EntryList. 152 // The mutex property of the monitor itself protects the EntryList 153 // from concurrent interference. 154 // -- Only the monitor owner may detach the cxq. 155 // 156 // * The monitor entry list operations avoid locks, but strictly speaking 157 // they're not lock-free. Enter is lock-free, exit is not. 158 // For a description of 'Methods and apparatus providing non-blocking access 159 // to a resource,' see U.S. Pat. No. 7844973. 160 // 161 // * The cxq can have multiple concurrent "pushers" but only one concurrent 162 // detaching thread. This mechanism is immune from the ABA corruption. 163 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 164 // 165 // * Taken together, the cxq and the EntryList constitute or form a 166 // single logical queue of threads stalled trying to acquire the lock. 167 // We use two distinct lists to improve the odds of a constant-time 168 // dequeue operation after acquisition (in the ::enter() epilogue) and 169 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 170 // A key desideratum is to minimize queue & monitor metadata manipulation 171 // that occurs while holding the monitor lock -- that is, we want to 172 // minimize monitor lock holds times. Note that even a small amount of 173 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 174 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 175 // locks and monitor metadata. 176 // 177 // Cxq points to the set of Recently Arrived Threads attempting entry. 178 // Because we push threads onto _cxq with CAS, the RATs must take the form of 179 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 180 // the unlocking thread notices that EntryList is null but _cxq is != null. 181 // 182 // The EntryList is ordered by the prevailing queue discipline and 183 // can be organized in any convenient fashion, such as a doubly-linked list or 184 // a circular doubly-linked list. Critically, we want insert and delete operations 185 // to operate in constant-time. If we need a priority queue then something akin 186 // to Solaris' sleepq would work nicely. Viz., 187 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 188 // Queue discipline is enforced at ::exit() time, when the unlocking thread 189 // drains the cxq into the EntryList, and orders or reorders the threads on the 190 // EntryList accordingly. 191 // 192 // Barring "lock barging", this mechanism provides fair cyclic ordering, 193 // somewhat similar to an elevator-scan. 194 // 195 // * The monitor synchronization subsystem avoids the use of native 196 // synchronization primitives except for the narrow platform-specific 197 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 198 // the semantics of park-unpark. Put another way, this monitor implementation 199 // depends only on atomic operations and park-unpark. The monitor subsystem 200 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 201 // underlying OS manages the READY<->RUN transitions. 202 // 203 // * Waiting threads reside on the WaitSet list -- wait() puts 204 // the caller onto the WaitSet. 205 // 206 // * notify() or notifyAll() simply transfers threads from the WaitSet to 207 // either the EntryList or cxq. Subsequent exit() operations will 208 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 209 // it's likely the notifyee would simply impale itself on the lock held 210 // by the notifier. 211 // 212 // * An interesting alternative is to encode cxq as (List,LockByte) where 213 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 214 // variable, like _recursions, in the scheme. The threads or Events that form 215 // the list would have to be aligned in 256-byte addresses. A thread would 216 // try to acquire the lock or enqueue itself with CAS, but exiting threads 217 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 218 // Note that is is *not* word-tearing, but it does presume that full-word 219 // CAS operations are coherent with intermix with STB operations. That's true 220 // on most common processors. 221 // 222 // * See also http://blogs.sun.com/dave 223 224 225 void* ObjectMonitor::operator new (size_t size) throw() { 226 return AllocateHeap(size, mtInternal); 227 } 228 void* ObjectMonitor::operator new[] (size_t size) throw() { 229 return operator new (size); 230 } 231 void ObjectMonitor::operator delete(void* p) { 232 FreeHeap(p); 233 } 234 void ObjectMonitor::operator delete[] (void *p) { 235 operator delete(p); 236 } 237 238 // ----------------------------------------------------------------------------- 239 // Enter support 240 241 void ObjectMonitor::enter(TRAPS) { 242 ADIM_guarantee(ref_count() > 0, "must be positive: ref_count=%d", ref_count()); 243 244 // The following code is ordered to check the most common cases first 245 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 246 Thread * const Self = THREAD; 247 248 void* cur = try_set_owner_from(Self, NULL); 249 if (cur == NULL) { 250 assert(_recursions == 0, "invariant"); 251 return; 252 } 253 254 if (cur == Self) { 255 // TODO-FIXME: check for integer overflow! BUGID 6557169. 256 _recursions++; 257 return; 258 } 259 260 if (Self->is_lock_owned ((address)cur)) { 261 assert(_recursions == 0, "internal state error"); 262 _recursions = 1; 263 set_owner_from_BasicLock(Self, cur); // Convert from BasicLock* to Thread*. 264 return; 265 } 266 267 if (AsyncDeflateIdleMonitors && 268 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 269 // The deflation protocol finished the first part (setting owner), 270 // but it failed the second part (making ref_count negative) and 271 // bailed. Or the ObjectMonitor was async deflated and reused. 272 // Acquired the monitor. 273 assert(_recursions == 0, "invariant"); 274 return; 275 } 276 277 // We've encountered genuine contention. 278 assert(Self->_Stalled == 0, "invariant"); 279 Self->_Stalled = intptr_t(this); 280 281 // Try one round of spinning *before* enqueueing Self 282 // and before going through the awkward and expensive state 283 // transitions. The following spin is strictly optional ... 284 // Note that if we acquire the monitor from an initial spin 285 // we forgo posting JVMTI events and firing DTRACE probes. 286 if (TrySpin(Self) > 0) { 287 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 288 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 289 assert(((oop)object())->mark() == markWord::encode(this), 290 "object mark must match encoded this: mark=" INTPTR_FORMAT 291 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 292 markWord::encode(this).value()); 293 Self->_Stalled = 0; 294 return; 295 } 296 297 assert(_owner != Self, "invariant"); 298 assert(_succ != Self, "invariant"); 299 assert(Self->is_Java_thread(), "invariant"); 300 JavaThread * jt = (JavaThread *) Self; 301 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 302 assert(jt->thread_state() != _thread_blocked, "invariant"); 303 assert(AsyncDeflateIdleMonitors || this->object() != NULL, "invariant"); 304 assert(_contentions >= 0, "must not be negative: contentions=%d", _contentions); 305 306 // Prevent deflation. See ObjectSynchronizer::deflate_monitor(), 307 // ObjectSynchronizer::deflate_monitor_using_JT() and is_busy(). 308 // Ensure the object <-> monitor relationship remains stable while 309 // there's contention. 310 Atomic::add(1, &_contentions); 311 312 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 313 EventJavaMonitorEnter event; 314 if (event.should_commit()) { 315 event.set_monitorClass(((oop)this->object())->klass()); 316 event.set_address((uintptr_t)(this->object_addr())); 317 } 318 319 { // Change java thread status to indicate blocked on monitor enter. 320 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 321 322 Self->set_current_pending_monitor(this); 323 324 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 325 if (JvmtiExport::should_post_monitor_contended_enter()) { 326 JvmtiExport::post_monitor_contended_enter(jt, this); 327 328 // The current thread does not yet own the monitor and does not 329 // yet appear on any queues that would get it made the successor. 330 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 331 // handler cannot accidentally consume an unpark() meant for the 332 // ParkEvent associated with this ObjectMonitor. 333 } 334 335 OSThreadContendState osts(Self->osthread()); 336 ThreadBlockInVM tbivm(jt); 337 338 // TODO-FIXME: change the following for(;;) loop to straight-line code. 339 for (;;) { 340 jt->set_suspend_equivalent(); 341 // cleared by handle_special_suspend_equivalent_condition() 342 // or java_suspend_self() 343 344 EnterI(THREAD); 345 346 if (!ExitSuspendEquivalent(jt)) break; 347 348 // We have acquired the contended monitor, but while we were 349 // waiting another thread suspended us. We don't want to enter 350 // the monitor while suspended because that would surprise the 351 // thread that suspended us. 352 // 353 _recursions = 0; 354 _succ = NULL; 355 exit(false, Self); 356 357 jt->java_suspend_self(); 358 } 359 Self->set_current_pending_monitor(NULL); 360 361 // We cleared the pending monitor info since we've just gotten past 362 // the enter-check-for-suspend dance and we now own the monitor free 363 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 364 // destructor can go to a safepoint at the end of this block. If we 365 // do a thread dump during that safepoint, then this thread will show 366 // as having "-locked" the monitor, but the OS and java.lang.Thread 367 // states will still report that the thread is blocked trying to 368 // acquire it. 369 } 370 371 Atomic::dec(&_contentions); 372 assert(_contentions >= 0, "must not be negative: contentions=%d", _contentions); 373 Self->_Stalled = 0; 374 375 // Must either set _recursions = 0 or ASSERT _recursions == 0. 376 assert(_recursions == 0, "invariant"); 377 assert(_owner == Self, "invariant"); 378 assert(_succ != Self, "invariant"); 379 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 380 381 // The thread -- now the owner -- is back in vm mode. 382 // Report the glorious news via TI,DTrace and jvmstat. 383 // The probe effect is non-trivial. All the reportage occurs 384 // while we hold the monitor, increasing the length of the critical 385 // section. Amdahl's parallel speedup law comes vividly into play. 386 // 387 // Another option might be to aggregate the events (thread local or 388 // per-monitor aggregation) and defer reporting until a more opportune 389 // time -- such as next time some thread encounters contention but has 390 // yet to acquire the lock. While spinning that thread could 391 // spinning we could increment JVMStat counters, etc. 392 393 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 394 if (JvmtiExport::should_post_monitor_contended_entered()) { 395 JvmtiExport::post_monitor_contended_entered(jt, this); 396 397 // The current thread already owns the monitor and is not going to 398 // call park() for the remainder of the monitor enter protocol. So 399 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 400 // event handler consumed an unpark() issued by the thread that 401 // just exited the monitor. 402 } 403 if (event.should_commit()) { 404 event.set_previousOwner((uintptr_t)_previous_owner_tid); 405 event.commit(); 406 } 407 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 408 } 409 410 // Caveat: TryLock() is not necessarily serializing if it returns failure. 411 // Callers must compensate as needed. 412 413 int ObjectMonitor::TryLock(Thread * Self) { 414 void * own = _owner; 415 if (own != NULL) return 0; 416 if (try_set_owner_from(Self, NULL) == NULL) { 417 assert(_recursions == 0, "invariant"); 418 return 1; 419 } 420 // The lock had been free momentarily, but we lost the race to the lock. 421 // Interference -- the CAS failed. 422 // We can either return -1 or retry. 423 // Retry doesn't make as much sense because the lock was just acquired. 424 return -1; 425 } 426 427 // Install the displaced mark word (dmw) of a deflating ObjectMonitor 428 // into the header of the object associated with the monitor. This 429 // idempotent method is called by a thread that is deflating a 430 // monitor and by other threads that have detected a race with the 431 // deflation process. 432 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) { 433 // This function must only be called when (owner == DEFLATER_MARKER 434 // && ref_count <= 0), but we can't guarantee that here because 435 // those values could change when the ObjectMonitor gets moved from 436 // the global free list to a per-thread free list. 437 438 guarantee(obj != NULL, "must be non-NULL"); 439 if (object() != obj) { 440 // ObjectMonitor's object ref no longer refers to the target object 441 // so the object's header has already been restored. 442 return; 443 } 444 445 markWord dmw = header(); 446 if (dmw.value() == 0) { 447 // ObjectMonitor's header/dmw has been cleared so the ObjectMonitor 448 // has been deflated and taken off the global free list. 449 return; 450 } 451 452 // A non-NULL dmw has to be either neutral (not locked and not marked) 453 // or is already participating in this restoration protocol. 454 assert(dmw.is_neutral() || (dmw.is_marked() && dmw.hash() == 0), 455 "failed precondition: dmw=" INTPTR_FORMAT, dmw.value()); 456 457 markWord marked_dmw = markWord::zero(); 458 if (!dmw.is_marked() && dmw.hash() == 0) { 459 // This dmw has not yet started the restoration protocol so we 460 // mark a copy of the dmw to begin the protocol. 461 // Note: A dmw with a hashcode does not take this code path. 462 marked_dmw = dmw.set_marked(); 463 464 // All of the callers to this function can be racing with each 465 // other trying to update the _header field. 466 dmw = (markWord) Atomic::cmpxchg(marked_dmw, &_header, dmw); 467 if (dmw.value() == 0) { 468 // ObjectMonitor's header/dmw has been cleared so the object's 469 // header has already been restored. 470 return; 471 } 472 // The _header field is now marked. The winner's 'dmw' variable 473 // contains the original, unmarked header/dmw value and any 474 // losers have a marked header/dmw value that will be cleaned 475 // up below. 476 } 477 478 if (dmw.is_marked()) { 479 // Clear the mark from the header/dmw copy in preparation for 480 // possible restoration from this thread. 481 assert(dmw.hash() == 0, "hashcode must be 0: dmw=" INTPTR_FORMAT, 482 dmw.value()); 483 dmw = dmw.set_unmarked(); 484 } 485 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 486 487 // Install displaced mark word if the object's header still points 488 // to this ObjectMonitor. All racing callers to this function will 489 // reach this point, but only one can win. 490 obj->cas_set_mark(dmw, markWord::encode(this)); 491 492 // Note: It does not matter which thread restored the header/dmw 493 // into the object's header. The thread deflating the monitor just 494 // wanted the object's header restored and it is. The threads that 495 // detected a race with the deflation process also wanted the 496 // object's header restored before they retry their operation and 497 // because it is restored they will only retry once. 498 } 499 500 // Convert the fields used by is_busy() to a string that can be 501 // used for diagnostic output. 502 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 503 ss->print("is_busy: contentions=%d, waiters=%d, ", _contentions, _waiters); 504 if (!AsyncDeflateIdleMonitors) { 505 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 506 } else if (_owner != DEFLATER_MARKER) { 507 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 508 } else { 509 ss->print("owner=" INTPTR_FORMAT, NULL); 510 } 511 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq), 512 p2i(_EntryList)); 513 return ss->base(); 514 } 515 516 #define MAX_RECHECK_INTERVAL 1000 517 518 void ObjectMonitor::EnterI(TRAPS) { 519 ADIM_guarantee(ref_count() > 0, "must be positive: ref_count=%d", ref_count()); 520 521 Thread * const Self = THREAD; 522 assert(Self->is_Java_thread(), "invariant"); 523 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 524 525 // Try the lock - TATAS 526 if (TryLock (Self) > 0) { 527 assert(_succ != Self, "invariant"); 528 assert(_owner == Self, "invariant"); 529 assert(_Responsible != Self, "invariant"); 530 return; 531 } 532 533 if (AsyncDeflateIdleMonitors && 534 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 535 // The deflation protocol finished the first part (setting owner), 536 // but it failed the second part (making ref_count negative) and 537 // bailed. Or the ObjectMonitor was async deflated and reused. 538 // Acquired the monitor. 539 assert(_succ != Self, "invariant"); 540 assert(_Responsible != Self, "invariant"); 541 return; 542 } 543 544 assert(InitDone, "Unexpectedly not initialized"); 545 546 // We try one round of spinning *before* enqueueing Self. 547 // 548 // If the _owner is ready but OFFPROC we could use a YieldTo() 549 // operation to donate the remainder of this thread's quantum 550 // to the owner. This has subtle but beneficial affinity 551 // effects. 552 553 if (TrySpin(Self) > 0) { 554 assert(_owner == Self, "invariant"); 555 assert(_succ != Self, "invariant"); 556 assert(_Responsible != Self, "invariant"); 557 return; 558 } 559 560 // The Spin failed -- Enqueue and park the thread ... 561 assert(_succ != Self, "invariant"); 562 assert(_owner != Self, "invariant"); 563 assert(_Responsible != Self, "invariant"); 564 565 // Enqueue "Self" on ObjectMonitor's _cxq. 566 // 567 // Node acts as a proxy for Self. 568 // As an aside, if were to ever rewrite the synchronization code mostly 569 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 570 // Java objects. This would avoid awkward lifecycle and liveness issues, 571 // as well as eliminate a subset of ABA issues. 572 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 573 574 ObjectWaiter node(Self); 575 Self->_ParkEvent->reset(); 576 node._prev = (ObjectWaiter *) 0xBAD; 577 node.TState = ObjectWaiter::TS_CXQ; 578 579 // Push "Self" onto the front of the _cxq. 580 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 581 // Note that spinning tends to reduce the rate at which threads 582 // enqueue and dequeue on EntryList|cxq. 583 ObjectWaiter * nxt; 584 for (;;) { 585 node._next = nxt = _cxq; 586 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break; 587 588 // Interference - the CAS failed because _cxq changed. Just retry. 589 // As an optional optimization we retry the lock. 590 if (TryLock (Self) > 0) { 591 assert(_succ != Self, "invariant"); 592 assert(_owner == Self, "invariant"); 593 assert(_Responsible != Self, "invariant"); 594 return; 595 } 596 } 597 598 // Check for cxq|EntryList edge transition to non-null. This indicates 599 // the onset of contention. While contention persists exiting threads 600 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 601 // operations revert to the faster 1-0 mode. This enter operation may interleave 602 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 603 // arrange for one of the contending thread to use a timed park() operations 604 // to detect and recover from the race. (Stranding is form of progress failure 605 // where the monitor is unlocked but all the contending threads remain parked). 606 // That is, at least one of the contended threads will periodically poll _owner. 607 // One of the contending threads will become the designated "Responsible" thread. 608 // The Responsible thread uses a timed park instead of a normal indefinite park 609 // operation -- it periodically wakes and checks for and recovers from potential 610 // strandings admitted by 1-0 exit operations. We need at most one Responsible 611 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 612 // be responsible for a monitor. 613 // 614 // Currently, one of the contended threads takes on the added role of "Responsible". 615 // A viable alternative would be to use a dedicated "stranding checker" thread 616 // that periodically iterated over all the threads (or active monitors) and unparked 617 // successors where there was risk of stranding. This would help eliminate the 618 // timer scalability issues we see on some platforms as we'd only have one thread 619 // -- the checker -- parked on a timer. 620 621 if (nxt == NULL && _EntryList == NULL) { 622 // Try to assume the role of responsible thread for the monitor. 623 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 624 Atomic::replace_if_null(Self, &_Responsible); 625 } 626 627 // The lock might have been released while this thread was occupied queueing 628 // itself onto _cxq. To close the race and avoid "stranding" and 629 // progress-liveness failure we must resample-retry _owner before parking. 630 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 631 // In this case the ST-MEMBAR is accomplished with CAS(). 632 // 633 // TODO: Defer all thread state transitions until park-time. 634 // Since state transitions are heavy and inefficient we'd like 635 // to defer the state transitions until absolutely necessary, 636 // and in doing so avoid some transitions ... 637 638 int nWakeups = 0; 639 int recheckInterval = 1; 640 641 for (;;) { 642 643 if (TryLock(Self) > 0) break; 644 assert(_owner != Self, "invariant"); 645 646 // park self 647 if (_Responsible == Self) { 648 Self->_ParkEvent->park((jlong) recheckInterval); 649 // Increase the recheckInterval, but clamp the value. 650 recheckInterval *= 8; 651 if (recheckInterval > MAX_RECHECK_INTERVAL) { 652 recheckInterval = MAX_RECHECK_INTERVAL; 653 } 654 } else { 655 Self->_ParkEvent->park(); 656 } 657 658 if (TryLock(Self) > 0) break; 659 660 if (AsyncDeflateIdleMonitors && 661 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 662 // The deflation protocol finished the first part (setting owner), 663 // but it failed the second part (making ref_count negative) and 664 // bailed. Or the ObjectMonitor was async deflated and reused. 665 // Acquired the monitor. 666 break; 667 } 668 669 // The lock is still contested. 670 // Keep a tally of the # of futile wakeups. 671 // Note that the counter is not protected by a lock or updated by atomics. 672 // That is by design - we trade "lossy" counters which are exposed to 673 // races during updates for a lower probe effect. 674 675 // This PerfData object can be used in parallel with a safepoint. 676 // See the work around in PerfDataManager::destroy(). 677 OM_PERFDATA_OP(FutileWakeups, inc()); 678 ++nWakeups; 679 680 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 681 // We can defer clearing _succ until after the spin completes 682 // TrySpin() must tolerate being called with _succ == Self. 683 // Try yet another round of adaptive spinning. 684 if (TrySpin(Self) > 0) break; 685 686 // We can find that we were unpark()ed and redesignated _succ while 687 // we were spinning. That's harmless. If we iterate and call park(), 688 // park() will consume the event and return immediately and we'll 689 // just spin again. This pattern can repeat, leaving _succ to simply 690 // spin on a CPU. 691 692 if (_succ == Self) _succ = NULL; 693 694 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 695 OrderAccess::fence(); 696 } 697 698 // Egress : 699 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 700 // Normally we'll find Self on the EntryList . 701 // From the perspective of the lock owner (this thread), the 702 // EntryList is stable and cxq is prepend-only. 703 // The head of cxq is volatile but the interior is stable. 704 // In addition, Self.TState is stable. 705 706 assert(_owner == Self, "invariant"); 707 assert(object() != NULL, "invariant"); 708 // I'd like to write: 709 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 710 // but as we're at a safepoint that's not safe. 711 712 UnlinkAfterAcquire(Self, &node); 713 if (_succ == Self) _succ = NULL; 714 715 assert(_succ != Self, "invariant"); 716 if (_Responsible == Self) { 717 _Responsible = NULL; 718 OrderAccess::fence(); // Dekker pivot-point 719 720 // We may leave threads on cxq|EntryList without a designated 721 // "Responsible" thread. This is benign. When this thread subsequently 722 // exits the monitor it can "see" such preexisting "old" threads -- 723 // threads that arrived on the cxq|EntryList before the fence, above -- 724 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 725 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 726 // non-null and elect a new "Responsible" timer thread. 727 // 728 // This thread executes: 729 // ST Responsible=null; MEMBAR (in enter epilogue - here) 730 // LD cxq|EntryList (in subsequent exit) 731 // 732 // Entering threads in the slow/contended path execute: 733 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 734 // The (ST cxq; MEMBAR) is accomplished with CAS(). 735 // 736 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 737 // exit operation from floating above the ST Responsible=null. 738 } 739 740 // We've acquired ownership with CAS(). 741 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 742 // But since the CAS() this thread may have also stored into _succ, 743 // EntryList, cxq or Responsible. These meta-data updates must be 744 // visible __before this thread subsequently drops the lock. 745 // Consider what could occur if we didn't enforce this constraint -- 746 // STs to monitor meta-data and user-data could reorder with (become 747 // visible after) the ST in exit that drops ownership of the lock. 748 // Some other thread could then acquire the lock, but observe inconsistent 749 // or old monitor meta-data and heap data. That violates the JMM. 750 // To that end, the 1-0 exit() operation must have at least STST|LDST 751 // "release" barrier semantics. Specifically, there must be at least a 752 // STST|LDST barrier in exit() before the ST of null into _owner that drops 753 // the lock. The barrier ensures that changes to monitor meta-data and data 754 // protected by the lock will be visible before we release the lock, and 755 // therefore before some other thread (CPU) has a chance to acquire the lock. 756 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 757 // 758 // Critically, any prior STs to _succ or EntryList must be visible before 759 // the ST of null into _owner in the *subsequent* (following) corresponding 760 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 761 // execute a serializing instruction. 762 763 return; 764 } 765 766 // ReenterI() is a specialized inline form of the latter half of the 767 // contended slow-path from EnterI(). We use ReenterI() only for 768 // monitor reentry in wait(). 769 // 770 // In the future we should reconcile EnterI() and ReenterI(). 771 772 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 773 ADIM_guarantee(ref_count() > 0, "must be positive: ref_count=%d", ref_count()); 774 775 assert(Self != NULL, "invariant"); 776 assert(SelfNode != NULL, "invariant"); 777 assert(SelfNode->_thread == Self, "invariant"); 778 assert(_waiters > 0, "invariant"); 779 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 780 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 781 JavaThread * jt = (JavaThread *) Self; 782 783 int nWakeups = 0; 784 for (;;) { 785 ObjectWaiter::TStates v = SelfNode->TState; 786 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 787 assert(_owner != Self, "invariant"); 788 789 if (TryLock(Self) > 0) break; 790 if (TrySpin(Self) > 0) break; 791 792 if (AsyncDeflateIdleMonitors && 793 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 794 // The deflation protocol finished the first part (setting owner), 795 // but it failed the second part (making ref_count negative) and 796 // bailed. Or the ObjectMonitor was async deflated and reused. 797 // Acquired the monitor. 798 break; 799 } 800 801 // State transition wrappers around park() ... 802 // ReenterI() wisely defers state transitions until 803 // it's clear we must park the thread. 804 { 805 OSThreadContendState osts(Self->osthread()); 806 ThreadBlockInVM tbivm(jt); 807 808 // cleared by handle_special_suspend_equivalent_condition() 809 // or java_suspend_self() 810 jt->set_suspend_equivalent(); 811 Self->_ParkEvent->park(); 812 813 // were we externally suspended while we were waiting? 814 for (;;) { 815 if (!ExitSuspendEquivalent(jt)) break; 816 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 817 jt->java_suspend_self(); 818 jt->set_suspend_equivalent(); 819 } 820 } 821 822 // Try again, but just so we distinguish between futile wakeups and 823 // successful wakeups. The following test isn't algorithmically 824 // necessary, but it helps us maintain sensible statistics. 825 if (TryLock(Self) > 0) break; 826 827 // The lock is still contested. 828 // Keep a tally of the # of futile wakeups. 829 // Note that the counter is not protected by a lock or updated by atomics. 830 // That is by design - we trade "lossy" counters which are exposed to 831 // races during updates for a lower probe effect. 832 ++nWakeups; 833 834 // Assuming this is not a spurious wakeup we'll normally 835 // find that _succ == Self. 836 if (_succ == Self) _succ = NULL; 837 838 // Invariant: after clearing _succ a contending thread 839 // *must* retry _owner before parking. 840 OrderAccess::fence(); 841 842 // This PerfData object can be used in parallel with a safepoint. 843 // See the work around in PerfDataManager::destroy(). 844 OM_PERFDATA_OP(FutileWakeups, inc()); 845 } 846 847 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 848 // Normally we'll find Self on the EntryList. 849 // Unlinking from the EntryList is constant-time and atomic-free. 850 // From the perspective of the lock owner (this thread), the 851 // EntryList is stable and cxq is prepend-only. 852 // The head of cxq is volatile but the interior is stable. 853 // In addition, Self.TState is stable. 854 855 assert(_owner == Self, "invariant"); 856 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 857 UnlinkAfterAcquire(Self, SelfNode); 858 if (_succ == Self) _succ = NULL; 859 assert(_succ != Self, "invariant"); 860 SelfNode->TState = ObjectWaiter::TS_RUN; 861 OrderAccess::fence(); // see comments at the end of EnterI() 862 } 863 864 // By convention we unlink a contending thread from EntryList|cxq immediately 865 // after the thread acquires the lock in ::enter(). Equally, we could defer 866 // unlinking the thread until ::exit()-time. 867 868 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 869 assert(_owner == Self, "invariant"); 870 assert(SelfNode->_thread == Self, "invariant"); 871 872 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 873 // Normal case: remove Self from the DLL EntryList . 874 // This is a constant-time operation. 875 ObjectWaiter * nxt = SelfNode->_next; 876 ObjectWaiter * prv = SelfNode->_prev; 877 if (nxt != NULL) nxt->_prev = prv; 878 if (prv != NULL) prv->_next = nxt; 879 if (SelfNode == _EntryList) _EntryList = nxt; 880 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 881 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 882 } else { 883 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 884 // Inopportune interleaving -- Self is still on the cxq. 885 // This usually means the enqueue of self raced an exiting thread. 886 // Normally we'll find Self near the front of the cxq, so 887 // dequeueing is typically fast. If needbe we can accelerate 888 // this with some MCS/CHL-like bidirectional list hints and advisory 889 // back-links so dequeueing from the interior will normally operate 890 // in constant-time. 891 // Dequeue Self from either the head (with CAS) or from the interior 892 // with a linear-time scan and normal non-atomic memory operations. 893 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 894 // and then unlink Self from EntryList. We have to drain eventually, 895 // so it might as well be now. 896 897 ObjectWaiter * v = _cxq; 898 assert(v != NULL, "invariant"); 899 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) { 900 // The CAS above can fail from interference IFF a "RAT" arrived. 901 // In that case Self must be in the interior and can no longer be 902 // at the head of cxq. 903 if (v == SelfNode) { 904 assert(_cxq != v, "invariant"); 905 v = _cxq; // CAS above failed - start scan at head of list 906 } 907 ObjectWaiter * p; 908 ObjectWaiter * q = NULL; 909 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 910 q = p; 911 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 912 } 913 assert(v != SelfNode, "invariant"); 914 assert(p == SelfNode, "Node not found on cxq"); 915 assert(p != _cxq, "invariant"); 916 assert(q != NULL, "invariant"); 917 assert(q->_next == p, "invariant"); 918 q->_next = p->_next; 919 } 920 } 921 922 #ifdef ASSERT 923 // Diagnostic hygiene ... 924 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 925 SelfNode->_next = (ObjectWaiter *) 0xBAD; 926 SelfNode->TState = ObjectWaiter::TS_RUN; 927 #endif 928 } 929 930 // ----------------------------------------------------------------------------- 931 // Exit support 932 // 933 // exit() 934 // ~~~~~~ 935 // Note that the collector can't reclaim the objectMonitor or deflate 936 // the object out from underneath the thread calling ::exit() as the 937 // thread calling ::exit() never transitions to a stable state. 938 // This inhibits GC, which in turn inhibits asynchronous (and 939 // inopportune) reclamation of "this". 940 // 941 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 942 // There's one exception to the claim above, however. EnterI() can call 943 // exit() to drop a lock if the acquirer has been externally suspended. 944 // In that case exit() is called with _thread_state as _thread_blocked, 945 // but the monitor's _contentions field is > 0, which inhibits reclamation. 946 // 947 // 1-0 exit 948 // ~~~~~~~~ 949 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 950 // the fast-path operators have been optimized so the common ::exit() 951 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 952 // The code emitted by fast_unlock() elides the usual MEMBAR. This 953 // greatly improves latency -- MEMBAR and CAS having considerable local 954 // latency on modern processors -- but at the cost of "stranding". Absent the 955 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 956 // ::enter() path, resulting in the entering thread being stranding 957 // and a progress-liveness failure. Stranding is extremely rare. 958 // We use timers (timed park operations) & periodic polling to detect 959 // and recover from stranding. Potentially stranded threads periodically 960 // wake up and poll the lock. See the usage of the _Responsible variable. 961 // 962 // The CAS() in enter provides for safety and exclusion, while the CAS or 963 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 964 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 965 // We detect and recover from stranding with timers. 966 // 967 // If a thread transiently strands it'll park until (a) another 968 // thread acquires the lock and then drops the lock, at which time the 969 // exiting thread will notice and unpark the stranded thread, or, (b) 970 // the timer expires. If the lock is high traffic then the stranding latency 971 // will be low due to (a). If the lock is low traffic then the odds of 972 // stranding are lower, although the worst-case stranding latency 973 // is longer. Critically, we don't want to put excessive load in the 974 // platform's timer subsystem. We want to minimize both the timer injection 975 // rate (timers created/sec) as well as the number of timers active at 976 // any one time. (more precisely, we want to minimize timer-seconds, which is 977 // the integral of the # of active timers at any instant over time). 978 // Both impinge on OS scalability. Given that, at most one thread parked on 979 // a monitor will use a timer. 980 // 981 // There is also the risk of a futile wake-up. If we drop the lock 982 // another thread can reacquire the lock immediately, and we can 983 // then wake a thread unnecessarily. This is benign, and we've 984 // structured the code so the windows are short and the frequency 985 // of such futile wakups is low. 986 987 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 988 Thread * const Self = THREAD; 989 if (THREAD != _owner) { 990 void* cur = _owner; 991 if (THREAD->is_lock_owned((address)cur)) { 992 assert(_recursions == 0, "invariant"); 993 set_owner_from_BasicLock(Self, cur); // Convert from BasicLock* to Thread*. 994 _recursions = 0; 995 } else { 996 // Apparent unbalanced locking ... 997 // Naively we'd like to throw IllegalMonitorStateException. 998 // As a practical matter we can neither allocate nor throw an 999 // exception as ::exit() can be called from leaf routines. 1000 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1001 // Upon deeper reflection, however, in a properly run JVM the only 1002 // way we should encounter this situation is in the presence of 1003 // unbalanced JNI locking. TODO: CheckJNICalls. 1004 // See also: CR4414101 1005 tty->print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 1006 " is exiting an ObjectMonitor it does not own.", 1007 p2i(THREAD)); 1008 tty->print_cr("The imbalance is possibly caused by JNI locking."); 1009 print_debug_style_on(tty); 1010 // Changing this from an assert() to ADIM_guarantee() may run 1011 // afoul of any test that is inducing non-balanced JNI locking. 1012 ADIM_guarantee(false, "Non-balanced monitor enter/exit!"); 1013 return; 1014 } 1015 } 1016 1017 if (_recursions != 0) { 1018 _recursions--; // this is simple recursive enter 1019 return; 1020 } 1021 1022 // Invariant: after setting Responsible=null an thread must execute 1023 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1024 _Responsible = NULL; 1025 1026 #if INCLUDE_JFR 1027 // get the owner's thread id for the MonitorEnter event 1028 // if it is enabled and the thread isn't suspended 1029 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1030 _previous_owner_tid = JFR_THREAD_ID(Self); 1031 } 1032 #endif 1033 1034 for (;;) { 1035 assert(THREAD == _owner, "invariant"); 1036 1037 // release semantics: prior loads and stores from within the critical section 1038 // must not float (reorder) past the following store that drops the lock. 1039 // On SPARC that requires MEMBAR #loadstore|#storestore. 1040 // But of course in TSO #loadstore|#storestore is not required. 1041 if (AsyncDeflateIdleMonitors) { 1042 set_owner_from(NULL, Self); 1043 } else { 1044 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock 1045 OrderAccess::storeload(); // See if we need to wake a successor 1046 } 1047 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1048 return; 1049 } 1050 // Other threads are blocked trying to acquire the lock. 1051 1052 // Normally the exiting thread is responsible for ensuring succession, 1053 // but if other successors are ready or other entering threads are spinning 1054 // then this thread can simply store NULL into _owner and exit without 1055 // waking a successor. The existence of spinners or ready successors 1056 // guarantees proper succession (liveness). Responsibility passes to the 1057 // ready or running successors. The exiting thread delegates the duty. 1058 // More precisely, if a successor already exists this thread is absolved 1059 // of the responsibility of waking (unparking) one. 1060 // 1061 // The _succ variable is critical to reducing futile wakeup frequency. 1062 // _succ identifies the "heir presumptive" thread that has been made 1063 // ready (unparked) but that has not yet run. We need only one such 1064 // successor thread to guarantee progress. 1065 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1066 // section 3.3 "Futile Wakeup Throttling" for details. 1067 // 1068 // Note that spinners in Enter() also set _succ non-null. 1069 // In the current implementation spinners opportunistically set 1070 // _succ so that exiting threads might avoid waking a successor. 1071 // Another less appealing alternative would be for the exiting thread 1072 // to drop the lock and then spin briefly to see if a spinner managed 1073 // to acquire the lock. If so, the exiting thread could exit 1074 // immediately without waking a successor, otherwise the exiting 1075 // thread would need to dequeue and wake a successor. 1076 // (Note that we'd need to make the post-drop spin short, but no 1077 // shorter than the worst-case round-trip cache-line migration time. 1078 // The dropped lock needs to become visible to the spinner, and then 1079 // the acquisition of the lock by the spinner must become visible to 1080 // the exiting thread). 1081 1082 // It appears that an heir-presumptive (successor) must be made ready. 1083 // Only the current lock owner can manipulate the EntryList or 1084 // drain _cxq, so we need to reacquire the lock. If we fail 1085 // to reacquire the lock the responsibility for ensuring succession 1086 // falls to the new owner. 1087 // 1088 if (try_set_owner_from(Self, NULL) != NULL) { 1089 return; 1090 } 1091 1092 guarantee(_owner == THREAD, "invariant"); 1093 1094 ObjectWaiter * w = NULL; 1095 1096 w = _EntryList; 1097 if (w != NULL) { 1098 // I'd like to write: guarantee (w->_thread != Self). 1099 // But in practice an exiting thread may find itself on the EntryList. 1100 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1101 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1102 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1103 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1104 // release the lock "O". T2 resumes immediately after the ST of null into 1105 // _owner, above. T2 notices that the EntryList is populated, so it 1106 // reacquires the lock and then finds itself on the EntryList. 1107 // Given all that, we have to tolerate the circumstance where "w" is 1108 // associated with Self. 1109 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1110 ExitEpilog(Self, w); 1111 return; 1112 } 1113 1114 // If we find that both _cxq and EntryList are null then just 1115 // re-run the exit protocol from the top. 1116 w = _cxq; 1117 if (w == NULL) continue; 1118 1119 // Drain _cxq into EntryList - bulk transfer. 1120 // First, detach _cxq. 1121 // The following loop is tantamount to: w = swap(&cxq, NULL) 1122 for (;;) { 1123 assert(w != NULL, "Invariant"); 1124 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w); 1125 if (u == w) break; 1126 w = u; 1127 } 1128 1129 assert(w != NULL, "invariant"); 1130 assert(_EntryList == NULL, "invariant"); 1131 1132 // Convert the LIFO SLL anchored by _cxq into a DLL. 1133 // The list reorganization step operates in O(LENGTH(w)) time. 1134 // It's critical that this step operate quickly as 1135 // "Self" still holds the outer-lock, restricting parallelism 1136 // and effectively lengthening the critical section. 1137 // Invariant: s chases t chases u. 1138 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1139 // we have faster access to the tail. 1140 1141 _EntryList = w; 1142 ObjectWaiter * q = NULL; 1143 ObjectWaiter * p; 1144 for (p = w; p != NULL; p = p->_next) { 1145 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1146 p->TState = ObjectWaiter::TS_ENTER; 1147 p->_prev = q; 1148 q = p; 1149 } 1150 1151 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1152 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1153 1154 // See if we can abdicate to a spinner instead of waking a thread. 1155 // A primary goal of the implementation is to reduce the 1156 // context-switch rate. 1157 if (_succ != NULL) continue; 1158 1159 w = _EntryList; 1160 if (w != NULL) { 1161 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1162 ExitEpilog(Self, w); 1163 return; 1164 } 1165 } 1166 } 1167 1168 // ExitSuspendEquivalent: 1169 // A faster alternate to handle_special_suspend_equivalent_condition() 1170 // 1171 // handle_special_suspend_equivalent_condition() unconditionally 1172 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1173 // operations have high latency. Note that in ::enter() we call HSSEC 1174 // while holding the monitor, so we effectively lengthen the critical sections. 1175 // 1176 // There are a number of possible solutions: 1177 // 1178 // A. To ameliorate the problem we might also defer state transitions 1179 // to as late as possible -- just prior to parking. 1180 // Given that, we'd call HSSEC after having returned from park(), 1181 // but before attempting to acquire the monitor. This is only a 1182 // partial solution. It avoids calling HSSEC while holding the 1183 // monitor (good), but it still increases successor reacquisition latency -- 1184 // the interval between unparking a successor and the time the successor 1185 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1186 // If we use this technique we can also avoid EnterI()-exit() loop 1187 // in ::enter() where we iteratively drop the lock and then attempt 1188 // to reacquire it after suspending. 1189 // 1190 // B. In the future we might fold all the suspend bits into a 1191 // composite per-thread suspend flag and then update it with CAS(). 1192 // Alternately, a Dekker-like mechanism with multiple variables 1193 // would suffice: 1194 // ST Self->_suspend_equivalent = false 1195 // MEMBAR 1196 // LD Self_>_suspend_flags 1197 1198 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1199 return jSelf->handle_special_suspend_equivalent_condition(); 1200 } 1201 1202 1203 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1204 assert(_owner == Self, "invariant"); 1205 1206 // Exit protocol: 1207 // 1. ST _succ = wakee 1208 // 2. membar #loadstore|#storestore; 1209 // 2. ST _owner = NULL 1210 // 3. unpark(wakee) 1211 1212 _succ = Wakee->_thread; 1213 ParkEvent * Trigger = Wakee->_event; 1214 1215 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1216 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1217 // out-of-scope (non-extant). 1218 Wakee = NULL; 1219 1220 // Drop the lock 1221 if (AsyncDeflateIdleMonitors) { 1222 set_owner_from(NULL, Self); 1223 } else { 1224 OrderAccess::release_store(&_owner, (void*)NULL); 1225 OrderAccess::fence(); // ST _owner vs LD in unpark() 1226 } 1227 1228 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1229 Trigger->unpark(); 1230 1231 // Maintain stats and report events to JVMTI 1232 OM_PERFDATA_OP(Parks, inc()); 1233 } 1234 1235 1236 // ----------------------------------------------------------------------------- 1237 // Class Loader deadlock handling. 1238 // 1239 // complete_exit exits a lock returning recursion count 1240 // complete_exit/reenter operate as a wait without waiting 1241 // complete_exit requires an inflated monitor 1242 // The _owner field is not always the Thread addr even with an 1243 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1244 // thread due to contention. 1245 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1246 Thread * const Self = THREAD; 1247 assert(Self->is_Java_thread(), "Must be Java thread!"); 1248 JavaThread *jt = (JavaThread *)THREAD; 1249 1250 assert(InitDone, "Unexpectedly not initialized"); 1251 1252 if (THREAD != _owner) { 1253 void* cur = _owner; 1254 if (THREAD->is_lock_owned((address)cur)) { 1255 assert(_recursions == 0, "internal state error"); 1256 set_owner_from_BasicLock(Self, cur); // Convert from BasicLock* to Thread*. 1257 _recursions = 0; 1258 } 1259 } 1260 1261 guarantee(Self == _owner, "complete_exit not owner"); 1262 intptr_t save = _recursions; // record the old recursion count 1263 _recursions = 0; // set the recursion level to be 0 1264 exit(true, Self); // exit the monitor 1265 guarantee(_owner != Self, "invariant"); 1266 return save; 1267 } 1268 1269 // reenter() enters a lock and sets recursion count 1270 // complete_exit/reenter operate as a wait without waiting 1271 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1272 Thread * const Self = THREAD; 1273 assert(Self->is_Java_thread(), "Must be Java thread!"); 1274 JavaThread *jt = (JavaThread *)THREAD; 1275 1276 guarantee(_owner != Self, "reenter already owner"); 1277 enter(THREAD); 1278 // Entered the monitor. 1279 guarantee(_recursions == 0, "reenter recursion"); 1280 _recursions = recursions; 1281 } 1282 1283 // Checks that the current THREAD owns this monitor and causes an 1284 // immediate return if it doesn't. We don't use the CHECK macro 1285 // because we want the IMSE to be the only exception that is thrown 1286 // from the call site when false is returned. Any other pending 1287 // exception is ignored. 1288 #define CHECK_OWNER() \ 1289 do { \ 1290 if (!check_owner(THREAD)) { \ 1291 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1292 return; \ 1293 } \ 1294 } while (false) 1295 1296 // Returns true if the specified thread owns the ObjectMonitor. 1297 // Otherwise returns false and throws IllegalMonitorStateException 1298 // (IMSE). If there is a pending exception and the specified thread 1299 // is not the owner, that exception will be replaced by the IMSE. 1300 bool ObjectMonitor::check_owner(Thread* THREAD) { 1301 if (_owner == THREAD) { 1302 return true; 1303 } 1304 void* cur = _owner; 1305 if (THREAD->is_lock_owned((address)cur)) { 1306 set_owner_from_BasicLock(THREAD, cur); // Convert from BasicLock* to Thread*. 1307 _recursions = 0; 1308 return true; 1309 } 1310 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1311 "current thread is not owner", false); 1312 } 1313 1314 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1315 ObjectMonitor* monitor, 1316 jlong notifier_tid, 1317 jlong timeout, 1318 bool timedout) { 1319 assert(event != NULL, "invariant"); 1320 assert(monitor != NULL, "invariant"); 1321 event->set_monitorClass(((oop)monitor->object())->klass()); 1322 event->set_timeout(timeout); 1323 event->set_address((uintptr_t)monitor->object_addr()); 1324 event->set_notifier(notifier_tid); 1325 event->set_timedOut(timedout); 1326 event->commit(); 1327 } 1328 1329 // ----------------------------------------------------------------------------- 1330 // Wait/Notify/NotifyAll 1331 // 1332 // Note: a subset of changes to ObjectMonitor::wait() 1333 // will need to be replicated in complete_exit 1334 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1335 Thread * const Self = THREAD; 1336 assert(Self->is_Java_thread(), "Must be Java thread!"); 1337 JavaThread *jt = (JavaThread *)THREAD; 1338 1339 assert(InitDone, "Unexpectedly not initialized"); 1340 1341 CHECK_OWNER(); // Throws IMSE if not owner. 1342 1343 EventJavaMonitorWait event; 1344 1345 // check for a pending interrupt 1346 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1347 // post monitor waited event. Note that this is past-tense, we are done waiting. 1348 if (JvmtiExport::should_post_monitor_waited()) { 1349 // Note: 'false' parameter is passed here because the 1350 // wait was not timed out due to thread interrupt. 1351 JvmtiExport::post_monitor_waited(jt, this, false); 1352 1353 // In this short circuit of the monitor wait protocol, the 1354 // current thread never drops ownership of the monitor and 1355 // never gets added to the wait queue so the current thread 1356 // cannot be made the successor. This means that the 1357 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1358 // consume an unpark() meant for the ParkEvent associated with 1359 // this ObjectMonitor. 1360 } 1361 if (event.should_commit()) { 1362 post_monitor_wait_event(&event, this, 0, millis, false); 1363 } 1364 THROW(vmSymbols::java_lang_InterruptedException()); 1365 return; 1366 } 1367 1368 assert(Self->_Stalled == 0, "invariant"); 1369 Self->_Stalled = intptr_t(this); 1370 jt->set_current_waiting_monitor(this); 1371 1372 // create a node to be put into the queue 1373 // Critically, after we reset() the event but prior to park(), we must check 1374 // for a pending interrupt. 1375 ObjectWaiter node(Self); 1376 node.TState = ObjectWaiter::TS_WAIT; 1377 Self->_ParkEvent->reset(); 1378 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1379 1380 // Enter the waiting queue, which is a circular doubly linked list in this case 1381 // but it could be a priority queue or any data structure. 1382 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1383 // by the the owner of the monitor *except* in the case where park() 1384 // returns because of a timeout of interrupt. Contention is exceptionally rare 1385 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1386 1387 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1388 AddWaiter(&node); 1389 Thread::SpinRelease(&_WaitSetLock); 1390 1391 _Responsible = NULL; 1392 1393 intptr_t save = _recursions; // record the old recursion count 1394 _waiters++; // increment the number of waiters 1395 _recursions = 0; // set the recursion level to be 1 1396 exit(true, Self); // exit the monitor 1397 guarantee(_owner != Self, "invariant"); 1398 1399 // The thread is on the WaitSet list - now park() it. 1400 // On MP systems it's conceivable that a brief spin before we park 1401 // could be profitable. 1402 // 1403 // TODO-FIXME: change the following logic to a loop of the form 1404 // while (!timeout && !interrupted && _notified == 0) park() 1405 1406 int ret = OS_OK; 1407 int WasNotified = 0; 1408 { // State transition wrappers 1409 OSThread* osthread = Self->osthread(); 1410 OSThreadWaitState osts(osthread, true); 1411 { 1412 ThreadBlockInVM tbivm(jt); 1413 // Thread is in thread_blocked state and oop access is unsafe. 1414 jt->set_suspend_equivalent(); 1415 1416 if (interruptible && (jt->is_interrupted(false) || HAS_PENDING_EXCEPTION)) { 1417 // Intentionally empty 1418 } else if (node._notified == 0) { 1419 if (millis <= 0) { 1420 Self->_ParkEvent->park(); 1421 } else { 1422 ret = Self->_ParkEvent->park(millis); 1423 } 1424 } 1425 1426 // were we externally suspended while we were waiting? 1427 if (ExitSuspendEquivalent (jt)) { 1428 // TODO-FIXME: add -- if succ == Self then succ = null. 1429 jt->java_suspend_self(); 1430 } 1431 1432 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1433 1434 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1435 // from the WaitSet to the EntryList. 1436 // See if we need to remove Node from the WaitSet. 1437 // We use double-checked locking to avoid grabbing _WaitSetLock 1438 // if the thread is not on the wait queue. 1439 // 1440 // Note that we don't need a fence before the fetch of TState. 1441 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1442 // written by the is thread. (perhaps the fetch might even be satisfied 1443 // by a look-aside into the processor's own store buffer, although given 1444 // the length of the code path between the prior ST and this load that's 1445 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1446 // then we'll acquire the lock and then re-fetch a fresh TState value. 1447 // That is, we fail toward safety. 1448 1449 if (node.TState == ObjectWaiter::TS_WAIT) { 1450 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1451 if (node.TState == ObjectWaiter::TS_WAIT) { 1452 DequeueSpecificWaiter(&node); // unlink from WaitSet 1453 assert(node._notified == 0, "invariant"); 1454 node.TState = ObjectWaiter::TS_RUN; 1455 } 1456 Thread::SpinRelease(&_WaitSetLock); 1457 } 1458 1459 // The thread is now either on off-list (TS_RUN), 1460 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1461 // The Node's TState variable is stable from the perspective of this thread. 1462 // No other threads will asynchronously modify TState. 1463 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1464 OrderAccess::loadload(); 1465 if (_succ == Self) _succ = NULL; 1466 WasNotified = node._notified; 1467 1468 // Reentry phase -- reacquire the monitor. 1469 // re-enter contended monitor after object.wait(). 1470 // retain OBJECT_WAIT state until re-enter successfully completes 1471 // Thread state is thread_in_vm and oop access is again safe, 1472 // although the raw address of the object may have changed. 1473 // (Don't cache naked oops over safepoints, of course). 1474 1475 // post monitor waited event. Note that this is past-tense, we are done waiting. 1476 if (JvmtiExport::should_post_monitor_waited()) { 1477 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1478 1479 if (node._notified != 0 && _succ == Self) { 1480 // In this part of the monitor wait-notify-reenter protocol it 1481 // is possible (and normal) for another thread to do a fastpath 1482 // monitor enter-exit while this thread is still trying to get 1483 // to the reenter portion of the protocol. 1484 // 1485 // The ObjectMonitor was notified and the current thread is 1486 // the successor which also means that an unpark() has already 1487 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1488 // consume the unpark() that was done when the successor was 1489 // set because the same ParkEvent is shared between Java 1490 // monitors and JVM/TI RawMonitors (for now). 1491 // 1492 // We redo the unpark() to ensure forward progress, i.e., we 1493 // don't want all pending threads hanging (parked) with none 1494 // entering the unlocked monitor. 1495 node._event->unpark(); 1496 } 1497 } 1498 1499 if (event.should_commit()) { 1500 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1501 } 1502 1503 OrderAccess::fence(); 1504 1505 assert(Self->_Stalled != 0, "invariant"); 1506 Self->_Stalled = 0; 1507 1508 assert(_owner != Self, "invariant"); 1509 ObjectWaiter::TStates v = node.TState; 1510 if (v == ObjectWaiter::TS_RUN) { 1511 enter(Self); 1512 } else { 1513 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1514 ReenterI(Self, &node); 1515 node.wait_reenter_end(this); 1516 } 1517 1518 // Self has reacquired the lock. 1519 // Lifecycle - the node representing Self must not appear on any queues. 1520 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1521 // want residual elements associated with this thread left on any lists. 1522 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1523 assert(_owner == Self, "invariant"); 1524 assert(_succ != Self, "invariant"); 1525 } // OSThreadWaitState() 1526 1527 jt->set_current_waiting_monitor(NULL); 1528 1529 guarantee(_recursions == 0, "invariant"); 1530 _recursions = save; // restore the old recursion count 1531 _waiters--; // decrement the number of waiters 1532 1533 // Verify a few postconditions 1534 assert(_owner == Self, "invariant"); 1535 assert(_succ != Self, "invariant"); 1536 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1537 1538 // check if the notification happened 1539 if (!WasNotified) { 1540 // no, it could be timeout or Thread.interrupt() or both 1541 // check for interrupt event, otherwise it is timeout 1542 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1543 THROW(vmSymbols::java_lang_InterruptedException()); 1544 } 1545 } 1546 1547 // NOTE: Spurious wake up will be consider as timeout. 1548 // Monitor notify has precedence over thread interrupt. 1549 } 1550 1551 1552 // Consider: 1553 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1554 // then instead of transferring a thread from the WaitSet to the EntryList 1555 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1556 1557 void ObjectMonitor::INotify(Thread * Self) { 1558 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1559 ObjectWaiter * iterator = DequeueWaiter(); 1560 if (iterator != NULL) { 1561 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1562 guarantee(iterator->_notified == 0, "invariant"); 1563 // Disposition - what might we do with iterator ? 1564 // a. add it directly to the EntryList - either tail (policy == 1) 1565 // or head (policy == 0). 1566 // b. push it onto the front of the _cxq (policy == 2). 1567 // For now we use (b). 1568 1569 iterator->TState = ObjectWaiter::TS_ENTER; 1570 1571 iterator->_notified = 1; 1572 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1573 1574 ObjectWaiter * list = _EntryList; 1575 if (list != NULL) { 1576 assert(list->_prev == NULL, "invariant"); 1577 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1578 assert(list != iterator, "invariant"); 1579 } 1580 1581 // prepend to cxq 1582 if (list == NULL) { 1583 iterator->_next = iterator->_prev = NULL; 1584 _EntryList = iterator; 1585 } else { 1586 iterator->TState = ObjectWaiter::TS_CXQ; 1587 for (;;) { 1588 ObjectWaiter * front = _cxq; 1589 iterator->_next = front; 1590 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) { 1591 break; 1592 } 1593 } 1594 } 1595 1596 // _WaitSetLock protects the wait queue, not the EntryList. We could 1597 // move the add-to-EntryList operation, above, outside the critical section 1598 // protected by _WaitSetLock. In practice that's not useful. With the 1599 // exception of wait() timeouts and interrupts the monitor owner 1600 // is the only thread that grabs _WaitSetLock. There's almost no contention 1601 // on _WaitSetLock so it's not profitable to reduce the length of the 1602 // critical section. 1603 1604 iterator->wait_reenter_begin(this); 1605 } 1606 Thread::SpinRelease(&_WaitSetLock); 1607 } 1608 1609 // Consider: a not-uncommon synchronization bug is to use notify() when 1610 // notifyAll() is more appropriate, potentially resulting in stranded 1611 // threads; this is one example of a lost wakeup. A useful diagnostic 1612 // option is to force all notify() operations to behave as notifyAll(). 1613 // 1614 // Note: We can also detect many such problems with a "minimum wait". 1615 // When the "minimum wait" is set to a small non-zero timeout value 1616 // and the program does not hang whereas it did absent "minimum wait", 1617 // that suggests a lost wakeup bug. 1618 1619 void ObjectMonitor::notify(TRAPS) { 1620 CHECK_OWNER(); // Throws IMSE if not owner. 1621 if (_WaitSet == NULL) { 1622 return; 1623 } 1624 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1625 INotify(THREAD); 1626 OM_PERFDATA_OP(Notifications, inc(1)); 1627 } 1628 1629 1630 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1631 // from the waitset to the EntryList. This could be done more efficiently with a 1632 // single bulk transfer but in practice it's not time-critical. Beware too, 1633 // that in prepend-mode we invert the order of the waiters. Let's say that the 1634 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1635 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1636 1637 void ObjectMonitor::notifyAll(TRAPS) { 1638 CHECK_OWNER(); // Throws IMSE if not owner. 1639 if (_WaitSet == NULL) { 1640 return; 1641 } 1642 1643 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1644 int tally = 0; 1645 while (_WaitSet != NULL) { 1646 tally++; 1647 INotify(THREAD); 1648 } 1649 1650 OM_PERFDATA_OP(Notifications, inc(tally)); 1651 } 1652 1653 // ----------------------------------------------------------------------------- 1654 // Adaptive Spinning Support 1655 // 1656 // Adaptive spin-then-block - rational spinning 1657 // 1658 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1659 // algorithm. On high order SMP systems it would be better to start with 1660 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1661 // a contending thread could enqueue itself on the cxq and then spin locally 1662 // on a thread-specific variable such as its ParkEvent._Event flag. 1663 // That's left as an exercise for the reader. Note that global spinning is 1664 // not problematic on Niagara, as the L2 cache serves the interconnect and 1665 // has both low latency and massive bandwidth. 1666 // 1667 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1668 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1669 // (duration) or we can fix the count at approximately the duration of 1670 // a context switch and vary the frequency. Of course we could also 1671 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1672 // For a description of 'Adaptive spin-then-block mutual exclusion in 1673 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1674 // 1675 // This implementation varies the duration "D", where D varies with 1676 // the success rate of recent spin attempts. (D is capped at approximately 1677 // length of a round-trip context switch). The success rate for recent 1678 // spin attempts is a good predictor of the success rate of future spin 1679 // attempts. The mechanism adapts automatically to varying critical 1680 // section length (lock modality), system load and degree of parallelism. 1681 // D is maintained per-monitor in _SpinDuration and is initialized 1682 // optimistically. Spin frequency is fixed at 100%. 1683 // 1684 // Note that _SpinDuration is volatile, but we update it without locks 1685 // or atomics. The code is designed so that _SpinDuration stays within 1686 // a reasonable range even in the presence of races. The arithmetic 1687 // operations on _SpinDuration are closed over the domain of legal values, 1688 // so at worst a race will install and older but still legal value. 1689 // At the very worst this introduces some apparent non-determinism. 1690 // We might spin when we shouldn't or vice-versa, but since the spin 1691 // count are relatively short, even in the worst case, the effect is harmless. 1692 // 1693 // Care must be taken that a low "D" value does not become an 1694 // an absorbing state. Transient spinning failures -- when spinning 1695 // is overall profitable -- should not cause the system to converge 1696 // on low "D" values. We want spinning to be stable and predictable 1697 // and fairly responsive to change and at the same time we don't want 1698 // it to oscillate, become metastable, be "too" non-deterministic, 1699 // or converge on or enter undesirable stable absorbing states. 1700 // 1701 // We implement a feedback-based control system -- using past behavior 1702 // to predict future behavior. We face two issues: (a) if the 1703 // input signal is random then the spin predictor won't provide optimal 1704 // results, and (b) if the signal frequency is too high then the control 1705 // system, which has some natural response lag, will "chase" the signal. 1706 // (b) can arise from multimodal lock hold times. Transient preemption 1707 // can also result in apparent bimodal lock hold times. 1708 // Although sub-optimal, neither condition is particularly harmful, as 1709 // in the worst-case we'll spin when we shouldn't or vice-versa. 1710 // The maximum spin duration is rather short so the failure modes aren't bad. 1711 // To be conservative, I've tuned the gain in system to bias toward 1712 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1713 // "rings" or oscillates between spinning and not spinning. This happens 1714 // when spinning is just on the cusp of profitability, however, so the 1715 // situation is not dire. The state is benign -- there's no need to add 1716 // hysteresis control to damp the transition rate between spinning and 1717 // not spinning. 1718 1719 // Spinning: Fixed frequency (100%), vary duration 1720 int ObjectMonitor::TrySpin(Thread * Self) { 1721 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1722 int ctr = Knob_FixedSpin; 1723 if (ctr != 0) { 1724 while (--ctr >= 0) { 1725 if (TryLock(Self) > 0) return 1; 1726 SpinPause(); 1727 } 1728 return 0; 1729 } 1730 1731 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1732 if (TryLock(Self) > 0) { 1733 // Increase _SpinDuration ... 1734 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1735 // Raising _SpurDuration to the poverty line is key. 1736 int x = _SpinDuration; 1737 if (x < Knob_SpinLimit) { 1738 if (x < Knob_Poverty) x = Knob_Poverty; 1739 _SpinDuration = x + Knob_BonusB; 1740 } 1741 return 1; 1742 } 1743 SpinPause(); 1744 } 1745 1746 // Admission control - verify preconditions for spinning 1747 // 1748 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1749 // becoming an absorbing state. Put another way, we spin briefly to 1750 // sample, just in case the system load, parallelism, contention, or lock 1751 // modality changed. 1752 // 1753 // Consider the following alternative: 1754 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1755 // spin attempt. "Periodically" might mean after a tally of 1756 // the # of failed spin attempts (or iterations) reaches some threshold. 1757 // This takes us into the realm of 1-out-of-N spinning, where we 1758 // hold the duration constant but vary the frequency. 1759 1760 ctr = _SpinDuration; 1761 if (ctr <= 0) return 0; 1762 1763 if (NotRunnable(Self, (Thread *) _owner)) { 1764 return 0; 1765 } 1766 1767 // We're good to spin ... spin ingress. 1768 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1769 // when preparing to LD...CAS _owner, etc and the CAS is likely 1770 // to succeed. 1771 if (_succ == NULL) { 1772 _succ = Self; 1773 } 1774 Thread * prv = NULL; 1775 1776 // There are three ways to exit the following loop: 1777 // 1. A successful spin where this thread has acquired the lock. 1778 // 2. Spin failure with prejudice 1779 // 3. Spin failure without prejudice 1780 1781 while (--ctr >= 0) { 1782 1783 // Periodic polling -- Check for pending GC 1784 // Threads may spin while they're unsafe. 1785 // We don't want spinning threads to delay the JVM from reaching 1786 // a stop-the-world safepoint or to steal cycles from GC. 1787 // If we detect a pending safepoint we abort in order that 1788 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1789 // this thread, if safe, doesn't steal cycles from GC. 1790 // This is in keeping with the "no loitering in runtime" rule. 1791 // We periodically check to see if there's a safepoint pending. 1792 if ((ctr & 0xFF) == 0) { 1793 if (SafepointMechanism::should_block(Self)) { 1794 goto Abort; // abrupt spin egress 1795 } 1796 SpinPause(); 1797 } 1798 1799 // Probe _owner with TATAS 1800 // If this thread observes the monitor transition or flicker 1801 // from locked to unlocked to locked, then the odds that this 1802 // thread will acquire the lock in this spin attempt go down 1803 // considerably. The same argument applies if the CAS fails 1804 // or if we observe _owner change from one non-null value to 1805 // another non-null value. In such cases we might abort 1806 // the spin without prejudice or apply a "penalty" to the 1807 // spin count-down variable "ctr", reducing it by 100, say. 1808 1809 Thread * ox = (Thread *) _owner; 1810 if (ox == NULL) { 1811 ox = (Thread*)try_set_owner_from(Self, NULL); 1812 if (ox == NULL) { 1813 // The CAS succeeded -- this thread acquired ownership 1814 // Take care of some bookkeeping to exit spin state. 1815 if (_succ == Self) { 1816 _succ = NULL; 1817 } 1818 1819 // Increase _SpinDuration : 1820 // The spin was successful (profitable) so we tend toward 1821 // longer spin attempts in the future. 1822 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1823 // If we acquired the lock early in the spin cycle it 1824 // makes sense to increase _SpinDuration proportionally. 1825 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1826 int x = _SpinDuration; 1827 if (x < Knob_SpinLimit) { 1828 if (x < Knob_Poverty) x = Knob_Poverty; 1829 _SpinDuration = x + Knob_Bonus; 1830 } 1831 return 1; 1832 } 1833 1834 // The CAS failed ... we can take any of the following actions: 1835 // * penalize: ctr -= CASPenalty 1836 // * exit spin with prejudice -- goto Abort; 1837 // * exit spin without prejudice. 1838 // * Since CAS is high-latency, retry again immediately. 1839 prv = ox; 1840 goto Abort; 1841 } 1842 1843 // Did lock ownership change hands ? 1844 if (ox != prv && prv != NULL) { 1845 goto Abort; 1846 } 1847 prv = ox; 1848 1849 // Abort the spin if the owner is not executing. 1850 // The owner must be executing in order to drop the lock. 1851 // Spinning while the owner is OFFPROC is idiocy. 1852 // Consider: ctr -= RunnablePenalty ; 1853 if (NotRunnable(Self, ox)) { 1854 goto Abort; 1855 } 1856 if (_succ == NULL) { 1857 _succ = Self; 1858 } 1859 } 1860 1861 // Spin failed with prejudice -- reduce _SpinDuration. 1862 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1863 // AIMD is globally stable. 1864 { 1865 int x = _SpinDuration; 1866 if (x > 0) { 1867 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1868 // This is globally sample and tends to damp the response. 1869 x -= Knob_Penalty; 1870 if (x < 0) x = 0; 1871 _SpinDuration = x; 1872 } 1873 } 1874 1875 Abort: 1876 if (_succ == Self) { 1877 _succ = NULL; 1878 // Invariant: after setting succ=null a contending thread 1879 // must recheck-retry _owner before parking. This usually happens 1880 // in the normal usage of TrySpin(), but it's safest 1881 // to make TrySpin() as foolproof as possible. 1882 OrderAccess::fence(); 1883 if (TryLock(Self) > 0) return 1; 1884 } 1885 return 0; 1886 } 1887 1888 // NotRunnable() -- informed spinning 1889 // 1890 // Don't bother spinning if the owner is not eligible to drop the lock. 1891 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1892 // The thread must be runnable in order to drop the lock in timely fashion. 1893 // If the _owner is not runnable then spinning will not likely be 1894 // successful (profitable). 1895 // 1896 // Beware -- the thread referenced by _owner could have died 1897 // so a simply fetch from _owner->_thread_state might trap. 1898 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1899 // Because of the lifecycle issues, the _thread_state values 1900 // observed by NotRunnable() might be garbage. NotRunnable must 1901 // tolerate this and consider the observed _thread_state value 1902 // as advisory. 1903 // 1904 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1905 // a thread pointer. 1906 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1907 // with the LSB of _owner. Another option would be to probabilistically probe 1908 // the putative _owner->TypeTag value. 1909 // 1910 // Checking _thread_state isn't perfect. Even if the thread is 1911 // in_java it might be blocked on a page-fault or have been preempted 1912 // and sitting on a ready/dispatch queue. 1913 // 1914 // The return value from NotRunnable() is *advisory* -- the 1915 // result is based on sampling and is not necessarily coherent. 1916 // The caller must tolerate false-negative and false-positive errors. 1917 // Spinning, in general, is probabilistic anyway. 1918 1919 1920 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1921 // Check ox->TypeTag == 2BAD. 1922 if (ox == NULL) return 0; 1923 1924 // Avoid transitive spinning ... 1925 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1926 // Immediately after T1 acquires L it's possible that T2, also 1927 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1928 // This occurs transiently after T1 acquired L but before 1929 // T1 managed to clear T1.Stalled. T2 does not need to abort 1930 // its spin in this circumstance. 1931 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1932 1933 if (BlockedOn == 1) return 1; 1934 if (BlockedOn != 0) { 1935 return BlockedOn != intptr_t(this) && _owner == ox; 1936 } 1937 1938 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1939 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1940 // consider also: jst != _thread_in_Java -- but that's overspecific. 1941 return jst == _thread_blocked || jst == _thread_in_native; 1942 } 1943 1944 1945 // ----------------------------------------------------------------------------- 1946 // WaitSet management ... 1947 1948 ObjectWaiter::ObjectWaiter(Thread* thread) { 1949 _next = NULL; 1950 _prev = NULL; 1951 _notified = 0; 1952 _notifier_tid = 0; 1953 TState = TS_RUN; 1954 _thread = thread; 1955 _event = thread->_ParkEvent; 1956 _active = false; 1957 assert(_event != NULL, "invariant"); 1958 } 1959 1960 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1961 JavaThread *jt = (JavaThread *)this->_thread; 1962 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1963 } 1964 1965 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1966 JavaThread *jt = (JavaThread *)this->_thread; 1967 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1968 } 1969 1970 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1971 assert(node != NULL, "should not add NULL node"); 1972 assert(node->_prev == NULL, "node already in list"); 1973 assert(node->_next == NULL, "node already in list"); 1974 // put node at end of queue (circular doubly linked list) 1975 if (_WaitSet == NULL) { 1976 _WaitSet = node; 1977 node->_prev = node; 1978 node->_next = node; 1979 } else { 1980 ObjectWaiter* head = _WaitSet; 1981 ObjectWaiter* tail = head->_prev; 1982 assert(tail->_next == head, "invariant check"); 1983 tail->_next = node; 1984 head->_prev = node; 1985 node->_next = head; 1986 node->_prev = tail; 1987 } 1988 } 1989 1990 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1991 // dequeue the very first waiter 1992 ObjectWaiter* waiter = _WaitSet; 1993 if (waiter) { 1994 DequeueSpecificWaiter(waiter); 1995 } 1996 return waiter; 1997 } 1998 1999 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2000 assert(node != NULL, "should not dequeue NULL node"); 2001 assert(node->_prev != NULL, "node already removed from list"); 2002 assert(node->_next != NULL, "node already removed from list"); 2003 // when the waiter has woken up because of interrupt, 2004 // timeout or other spurious wake-up, dequeue the 2005 // waiter from waiting list 2006 ObjectWaiter* next = node->_next; 2007 if (next == node) { 2008 assert(node->_prev == node, "invariant check"); 2009 _WaitSet = NULL; 2010 } else { 2011 ObjectWaiter* prev = node->_prev; 2012 assert(prev->_next == node, "invariant check"); 2013 assert(next->_prev == node, "invariant check"); 2014 next->_prev = prev; 2015 prev->_next = next; 2016 if (_WaitSet == node) { 2017 _WaitSet = next; 2018 } 2019 } 2020 node->_next = NULL; 2021 node->_prev = NULL; 2022 } 2023 2024 // ----------------------------------------------------------------------------- 2025 // PerfData support 2026 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2027 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2028 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2029 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2030 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2031 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2032 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2033 2034 // One-shot global initialization for the sync subsystem. 2035 // We could also defer initialization and initialize on-demand 2036 // the first time we call ObjectSynchronizer::inflate(). 2037 // Initialization would be protected - like so many things - by 2038 // the MonitorCache_lock. 2039 2040 void ObjectMonitor::Initialize() { 2041 assert(!InitDone, "invariant"); 2042 2043 if (!os::is_MP()) { 2044 Knob_SpinLimit = 0; 2045 Knob_PreSpin = 0; 2046 Knob_FixedSpin = -1; 2047 } 2048 2049 if (UsePerfData) { 2050 EXCEPTION_MARK; 2051 #define NEWPERFCOUNTER(n) \ 2052 { \ 2053 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2054 CHECK); \ 2055 } 2056 #define NEWPERFVARIABLE(n) \ 2057 { \ 2058 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2059 CHECK); \ 2060 } 2061 NEWPERFCOUNTER(_sync_Inflations); 2062 NEWPERFCOUNTER(_sync_Deflations); 2063 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2064 NEWPERFCOUNTER(_sync_FutileWakeups); 2065 NEWPERFCOUNTER(_sync_Parks); 2066 NEWPERFCOUNTER(_sync_Notifications); 2067 NEWPERFVARIABLE(_sync_MonExtant); 2068 #undef NEWPERFCOUNTER 2069 #undef NEWPERFVARIABLE 2070 } 2071 2072 DEBUG_ONLY(InitDone = true;) 2073 } 2074 2075 // For internal use by ObjectSynchronizer::monitors_iterate(). 2076 ObjectMonitorHandle::ObjectMonitorHandle(ObjectMonitor * om_ptr) { 2077 om_ptr->inc_ref_count(); 2078 _om_ptr = om_ptr; 2079 } 2080 2081 ObjectMonitorHandle::~ObjectMonitorHandle() { 2082 if (_om_ptr != NULL) { 2083 _om_ptr->dec_ref_count(); 2084 _om_ptr = NULL; 2085 } 2086 } 2087 2088 // Save the ObjectMonitor* associated with the specified markWord and 2089 // increment the ref_count. This function should only be called if 2090 // the caller has verified mark.has_monitor() == true. The object 2091 // parameter is needed to verify that ObjectMonitor* has not been 2092 // deflated and reused for another object. 2093 // 2094 // This function returns true if the ObjectMonitor* has been safely 2095 // saved. This function returns false if we have lost a race with 2096 // async deflation; the caller should retry as appropriate. 2097 // 2098 bool ObjectMonitorHandle::save_om_ptr(oop object, markWord mark) { 2099 guarantee(mark.has_monitor(), "sanity check: mark=" INTPTR_FORMAT, 2100 mark.value()); 2101 2102 ObjectMonitor * om_ptr = mark.monitor(); 2103 om_ptr->inc_ref_count(); 2104 2105 if (AsyncDeflateIdleMonitors) { 2106 // Race here if monitor is not owned! The above ref_count bump 2107 // will cause subsequent async deflation to skip it. However, 2108 // previous or concurrent async deflation is a race. 2109 if (om_ptr->owner_is_DEFLATER_MARKER() && om_ptr->ref_count() <= 0) { 2110 // Async deflation is in progress and our ref_count increment 2111 // above lost the race to async deflation. Attempt to restore 2112 // the header/dmw to the object's header so that we only retry 2113 // once if the deflater thread happens to be slow. 2114 om_ptr->install_displaced_markword_in_object(object); 2115 om_ptr->dec_ref_count(); 2116 return false; 2117 } 2118 if (om_ptr->ref_count() <= 0) { 2119 // Async deflation is in the process of bailing out, but has not 2120 // yet restored the ref_count field so we return false to force 2121 // a retry. We want a positive ref_count value for a true return. 2122 om_ptr->dec_ref_count(); 2123 return false; 2124 } 2125 // The ObjectMonitor could have been deflated and reused for 2126 // another object before we bumped the ref_count so make sure 2127 // our object still refers to this ObjectMonitor. 2128 const markWord tmp = object->mark(); 2129 if (!tmp.has_monitor() || tmp.monitor() != om_ptr) { 2130 // Async deflation and reuse won the race so we have to retry. 2131 // Skip object header restoration since that's already done. 2132 om_ptr->dec_ref_count(); 2133 return false; 2134 } 2135 } 2136 2137 ADIM_guarantee(_om_ptr == NULL, "sanity check: _om_ptr=" INTPTR_FORMAT, 2138 p2i(_om_ptr)); 2139 _om_ptr = om_ptr; 2140 return true; 2141 } 2142 2143 // For internal use by ObjectSynchronizer::inflate(). 2144 void ObjectMonitorHandle::set_om_ptr(ObjectMonitor * om_ptr) { 2145 if (_om_ptr == NULL) { 2146 ADIM_guarantee(om_ptr != NULL, "cannot clear an unset om_ptr"); 2147 om_ptr->inc_ref_count(); 2148 _om_ptr = om_ptr; 2149 } else { 2150 ADIM_guarantee(om_ptr == NULL, "can only clear a set om_ptr"); 2151 _om_ptr->dec_ref_count(); 2152 _om_ptr = NULL; 2153 } 2154 } 2155 2156 void ObjectMonitor::print_on(outputStream* st) const { 2157 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2158 st->print("{contentions=0x%08x,waiters=0x%08x" 2159 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2160 contentions(), waiters(), recursions(), 2161 p2i(owner())); 2162 } 2163 void ObjectMonitor::print() const { print_on(tty); } 2164 2165 // Print the ObjectMonitor like a debugger would: 2166 // 2167 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2168 // _header = (_value = 1) 2169 // _object = 0x000000070ff45fd0 2170 // _allocation_state = Old 2171 // _pad_buf0 = { 2172 // [0] = '\0' 2173 // ... 2174 // [43] = '\0' 2175 // } 2176 // _owner = 0x0000000000000000 2177 // _previous_owner_tid = 0 2178 // _pad_buf1 = { 2179 // [0] = '\0' 2180 // ... 2181 // [47] = '\0' 2182 // } 2183 // _ref_count = 1 2184 // _pad_buf2 = { 2185 // [0] = '\0' 2186 // ... 2187 // [59] = '\0' 2188 // } 2189 // _next_om = 0x0000000000000000 2190 // _recursions = 0 2191 // _EntryList = 0x0000000000000000 2192 // _cxq = 0x0000000000000000 2193 // _succ = 0x0000000000000000 2194 // _Responsible = 0x0000000000000000 2195 // _Spinner = 0 2196 // _SpinDuration = 5000 2197 // _contentions = 0 2198 // _WaitSet = 0x0000700009756248 2199 // _waiters = 1 2200 // _WaitSetLock = 0 2201 // } 2202 // 2203 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2204 st->print_cr("(ObjectMonitor *) " INTPTR_FORMAT " = {", p2i(this)); 2205 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2206 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 2207 st->print(" _allocation_state = "); 2208 if (is_free()) { 2209 st->print("Free"); 2210 } else if (is_old()) { 2211 st->print("Old"); 2212 } else if (is_new()) { 2213 st->print("New"); 2214 } else { 2215 st->print("unknown=%d", _allocation_state); 2216 } 2217 st->cr(); 2218 st->print_cr(" _pad_buf0 = {"); 2219 st->print_cr(" [0] = '\\0'"); 2220 st->print_cr(" ..."); 2221 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2222 st->print_cr(" }"); 2223 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 2224 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 2225 st->print_cr(" _pad_buf1 = {"); 2226 st->print_cr(" [0] = '\\0'"); 2227 st->print_cr(" ..."); 2228 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2229 st->print_cr(" }"); 2230 st->print_cr(" _ref_count = %d", ref_count()); 2231 st->print_cr(" _pad_buf2 = {"); 2232 st->print_cr(" [0] = '\\0'"); 2233 st->print_cr(" ..."); 2234 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2235 st->print_cr(" }"); 2236 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(_next_om)); 2237 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2238 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2239 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2240 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2241 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2242 st->print_cr(" _Spinner = %d", _Spinner); 2243 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2244 st->print_cr(" _contentions = %d", _contentions); 2245 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2246 st->print_cr(" _waiters = %d", _waiters); 2247 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2248 st->print_cr("}"); 2249 }