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 "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "oops/markWord.hpp" 34 #include "oops/oop.inline.hpp" 35 #include "runtime/atomic.hpp" 36 #include "runtime/handles.inline.hpp" 37 #include "runtime/interfaceSupport.inline.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "runtime/objectMonitor.hpp" 40 #include "runtime/objectMonitor.inline.hpp" 41 #include "runtime/orderAccess.hpp" 42 #include "runtime/osThread.hpp" 43 #include "runtime/safepointMechanism.inline.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/stubRoutines.hpp" 46 #include "runtime/thread.inline.hpp" 47 #include "services/threadService.hpp" 48 #include "utilities/dtrace.hpp" 49 #include "utilities/macros.hpp" 50 #include "utilities/preserveException.hpp" 51 #if INCLUDE_JFR 52 #include "jfr/support/jfrFlush.hpp" 53 #endif 54 55 #ifdef DTRACE_ENABLED 56 57 // Only bother with this argument setup if dtrace is available 58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95 #else // ndef DTRACE_ENABLED 96 97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100 #endif // ndef DTRACE_ENABLED 101 102 // Tunables ... 103 // The knob* variables are effectively final. Once set they should 104 // never be modified hence. Consider using __read_mostly with GCC. 105 106 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 107 108 static int Knob_Bonus = 100; // spin success bonus 109 static int Knob_BonusB = 100; // spin success bonus 110 static int Knob_Penalty = 200; // spin failure penalty 111 static int Knob_Poverty = 1000; 112 static int Knob_FixedSpin = 0; 113 static int Knob_PreSpin = 10; // 20-100 likely better 114 115 DEBUG_ONLY(static volatile bool InitDone = false;) 116 117 // ----------------------------------------------------------------------------- 118 // Theory of operations -- Monitors lists, thread residency, etc: 119 // 120 // * A thread acquires ownership of a monitor by successfully 121 // CAS()ing the _owner field from null to non-null. 122 // 123 // * Invariant: A thread appears on at most one monitor list -- 124 // cxq, EntryList or WaitSet -- at any one time. 125 // 126 // * Contending threads "push" themselves onto the cxq with CAS 127 // and then spin/park. 128 // 129 // * After a contending thread eventually acquires the lock it must 130 // dequeue itself from either the EntryList or the cxq. 131 // 132 // * The exiting thread identifies and unparks an "heir presumptive" 133 // tentative successor thread on the EntryList. Critically, the 134 // exiting thread doesn't unlink the successor thread from the EntryList. 135 // After having been unparked, the wakee will recontend for ownership of 136 // the monitor. The successor (wakee) will either acquire the lock or 137 // re-park itself. 138 // 139 // Succession is provided for by a policy of competitive handoff. 140 // The exiting thread does _not_ grant or pass ownership to the 141 // successor thread. (This is also referred to as "handoff" succession"). 142 // Instead the exiting thread releases ownership and possibly wakes 143 // a successor, so the successor can (re)compete for ownership of the lock. 144 // If the EntryList is empty but the cxq is populated the exiting 145 // thread will drain the cxq into the EntryList. It does so by 146 // by detaching the cxq (installing null with CAS) and folding 147 // the threads from the cxq into the EntryList. The EntryList is 148 // doubly linked, while the cxq is singly linked because of the 149 // CAS-based "push" used to enqueue recently arrived threads (RATs). 150 // 151 // * Concurrency invariants: 152 // 153 // -- only the monitor owner may access or mutate the EntryList. 154 // The mutex property of the monitor itself protects the EntryList 155 // from concurrent interference. 156 // -- Only the monitor owner may detach the cxq. 157 // 158 // * The monitor entry list operations avoid locks, but strictly speaking 159 // they're not lock-free. Enter is lock-free, exit is not. 160 // For a description of 'Methods and apparatus providing non-blocking access 161 // to a resource,' see U.S. Pat. No. 7844973. 162 // 163 // * The cxq can have multiple concurrent "pushers" but only one concurrent 164 // detaching thread. This mechanism is immune from the ABA corruption. 165 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 166 // 167 // * Taken together, the cxq and the EntryList constitute or form a 168 // single logical queue of threads stalled trying to acquire the lock. 169 // We use two distinct lists to improve the odds of a constant-time 170 // dequeue operation after acquisition (in the ::enter() epilogue) and 171 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 172 // A key desideratum is to minimize queue & monitor metadata manipulation 173 // that occurs while holding the monitor lock -- that is, we want to 174 // minimize monitor lock holds times. Note that even a small amount of 175 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 176 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 177 // locks and monitor metadata. 178 // 179 // Cxq points to the set of Recently Arrived Threads attempting entry. 180 // Because we push threads onto _cxq with CAS, the RATs must take the form of 181 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 182 // the unlocking thread notices that EntryList is null but _cxq is != null. 183 // 184 // The EntryList is ordered by the prevailing queue discipline and 185 // can be organized in any convenient fashion, such as a doubly-linked list or 186 // a circular doubly-linked list. Critically, we want insert and delete operations 187 // to operate in constant-time. If we need a priority queue then something akin 188 // to Solaris' sleepq would work nicely. Viz., 189 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 190 // Queue discipline is enforced at ::exit() time, when the unlocking thread 191 // drains the cxq into the EntryList, and orders or reorders the threads on the 192 // EntryList accordingly. 193 // 194 // Barring "lock barging", this mechanism provides fair cyclic ordering, 195 // somewhat similar to an elevator-scan. 196 // 197 // * The monitor synchronization subsystem avoids the use of native 198 // synchronization primitives except for the narrow platform-specific 199 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 200 // the semantics of park-unpark. Put another way, this monitor implementation 201 // depends only on atomic operations and park-unpark. The monitor subsystem 202 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 203 // underlying OS manages the READY<->RUN transitions. 204 // 205 // * Waiting threads reside on the WaitSet list -- wait() puts 206 // the caller onto the WaitSet. 207 // 208 // * notify() or notifyAll() simply transfers threads from the WaitSet to 209 // either the EntryList or cxq. Subsequent exit() operations will 210 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 211 // it's likely the notifyee would simply impale itself on the lock held 212 // by the notifier. 213 // 214 // * An interesting alternative is to encode cxq as (List,LockByte) where 215 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 216 // variable, like _recursions, in the scheme. The threads or Events that form 217 // the list would have to be aligned in 256-byte addresses. A thread would 218 // try to acquire the lock or enqueue itself with CAS, but exiting threads 219 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 220 // Note that is is *not* word-tearing, but it does presume that full-word 221 // CAS operations are coherent with intermix with STB operations. That's true 222 // on most common processors. 223 // 224 // * See also http://blogs.sun.com/dave 225 226 227 void* ObjectMonitor::operator new (size_t size) throw() { 228 return AllocateHeap(size, mtInternal); 229 } 230 void* ObjectMonitor::operator new[] (size_t size) throw() { 231 return operator new (size); 232 } 233 void ObjectMonitor::operator delete(void* p) { 234 FreeHeap(p); 235 } 236 void ObjectMonitor::operator delete[] (void *p) { 237 operator delete(p); 238 } 239 240 // ----------------------------------------------------------------------------- 241 // Enter support 242 243 void ObjectMonitor::enter(TRAPS) { 244 jint l_ref_count = ref_count(); 245 ADIM_guarantee(l_ref_count > 0, "must be positive: l_ref_count=%d, ref_count=%d", l_ref_count, ref_count()); 246 247 // The following code is ordered to check the most common cases first 248 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 249 Thread * const Self = THREAD; 250 251 void* cur = try_set_owner_from(Self, NULL); 252 if (cur == NULL) { 253 assert(_recursions == 0, "invariant"); 254 return; 255 } 256 257 if (cur == Self) { 258 // TODO-FIXME: check for integer overflow! BUGID 6557169. 259 _recursions++; 260 return; 261 } 262 263 if (Self->is_lock_owned((address)cur)) { 264 assert(_recursions == 0, "internal state error"); 265 _recursions = 1; 266 simply_set_owner_from_BasicLock(Self, cur); // Convert from BasicLock* to Thread*. 267 return; 268 } 269 270 if (AsyncDeflateIdleMonitors && 271 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 272 // The deflation protocol finished the first part (setting owner), 273 // but it failed the second part (making ref_count negative) and 274 // bailed. Or the ObjectMonitor was async deflated and reused. 275 // Acquired the monitor. 276 assert(_recursions == 0, "invariant"); 277 return; 278 } 279 280 // We've encountered genuine contention. 281 assert(Self->_Stalled == 0, "invariant"); 282 Self->_Stalled = intptr_t(this); 283 284 // Try one round of spinning *before* enqueueing Self 285 // and before going through the awkward and expensive state 286 // transitions. The following spin is strictly optional ... 287 // Note that if we acquire the monitor from an initial spin 288 // we forgo posting JVMTI events and firing DTRACE probes. 289 if (TrySpin(Self) > 0) { 290 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 291 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 292 assert(((oop)object())->mark() == markWord::encode(this), 293 "object mark must match encoded this: mark=" INTPTR_FORMAT 294 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 295 markWord::encode(this).value()); 296 Self->_Stalled = 0; 297 return; 298 } 299 300 assert(_owner != Self, "invariant"); 301 assert(_succ != Self, "invariant"); 302 assert(Self->is_Java_thread(), "invariant"); 303 JavaThread * jt = (JavaThread *) Self; 304 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 305 assert(jt->thread_state() != _thread_blocked, "invariant"); 306 assert(this->object() != NULL, "invariant"); 307 assert(_contentions >= 0, "must not be negative: contentions=%d", _contentions); 308 309 // Prevent deflation. See ObjectSynchronizer::deflate_monitor(), 310 // ObjectSynchronizer::deflate_monitor_using_JT() and is_busy(). 311 // Ensure the object <-> monitor relationship remains stable while 312 // there's contention. 313 Atomic::add(&_contentions, 1); 314 315 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 316 EventJavaMonitorEnter event; 317 if (event.should_commit()) { 318 event.set_monitorClass(((oop)this->object())->klass()); 319 event.set_address((uintptr_t)(this->object_addr())); 320 } 321 322 { // Change java thread status to indicate blocked on monitor enter. 323 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 324 325 Self->set_current_pending_monitor(this); 326 327 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 328 if (JvmtiExport::should_post_monitor_contended_enter()) { 329 JvmtiExport::post_monitor_contended_enter(jt, this); 330 331 // The current thread does not yet own the monitor and does not 332 // yet appear on any queues that would get it made the successor. 333 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 334 // handler cannot accidentally consume an unpark() meant for the 335 // ParkEvent associated with this ObjectMonitor. 336 } 337 338 OSThreadContendState osts(Self->osthread()); 339 ThreadBlockInVM tbivm(jt); 340 341 // TODO-FIXME: change the following for(;;) loop to straight-line code. 342 for (;;) { 343 jt->set_suspend_equivalent(); 344 // cleared by handle_special_suspend_equivalent_condition() 345 // or java_suspend_self() 346 347 EnterI(THREAD); 348 349 if (!ExitSuspendEquivalent(jt)) break; 350 351 // We have acquired the contended monitor, but while we were 352 // waiting another thread suspended us. We don't want to enter 353 // the monitor while suspended because that would surprise the 354 // thread that suspended us. 355 // 356 _recursions = 0; 357 _succ = NULL; 358 exit(false, Self); 359 360 jt->java_suspend_self(); 361 } 362 Self->set_current_pending_monitor(NULL); 363 364 // We cleared the pending monitor info since we've just gotten past 365 // the enter-check-for-suspend dance and we now own the monitor free 366 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 367 // destructor can go to a safepoint at the end of this block. If we 368 // do a thread dump during that safepoint, then this thread will show 369 // as having "-locked" the monitor, but the OS and java.lang.Thread 370 // states will still report that the thread is blocked trying to 371 // acquire it. 372 } 373 374 Atomic::dec(&_contentions); 375 assert(_contentions >= 0, "must not be negative: contentions=%d", _contentions); 376 Self->_Stalled = 0; 377 378 // Must either set _recursions = 0 or ASSERT _recursions == 0. 379 assert(_recursions == 0, "invariant"); 380 assert(_owner == Self, "invariant"); 381 assert(_succ != Self, "invariant"); 382 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 383 384 // The thread -- now the owner -- is back in vm mode. 385 // Report the glorious news via TI,DTrace and jvmstat. 386 // The probe effect is non-trivial. All the reportage occurs 387 // while we hold the monitor, increasing the length of the critical 388 // section. Amdahl's parallel speedup law comes vividly into play. 389 // 390 // Another option might be to aggregate the events (thread local or 391 // per-monitor aggregation) and defer reporting until a more opportune 392 // time -- such as next time some thread encounters contention but has 393 // yet to acquire the lock. While spinning that thread could 394 // spinning we could increment JVMStat counters, etc. 395 396 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 397 if (JvmtiExport::should_post_monitor_contended_entered()) { 398 JvmtiExport::post_monitor_contended_entered(jt, this); 399 400 // The current thread already owns the monitor and is not going to 401 // call park() for the remainder of the monitor enter protocol. So 402 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 403 // event handler consumed an unpark() issued by the thread that 404 // just exited the monitor. 405 } 406 if (event.should_commit()) { 407 event.set_previousOwner((uintptr_t)_previous_owner_tid); 408 event.commit(); 409 } 410 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 411 } 412 413 // Caveat: TryLock() is not necessarily serializing if it returns failure. 414 // Callers must compensate as needed. 415 416 int ObjectMonitor::TryLock(Thread * Self) { 417 void * own = _owner; 418 if (own != NULL) return 0; 419 if (try_set_owner_from(Self, NULL) == NULL) { 420 assert(_recursions == 0, "invariant"); 421 return 1; 422 } 423 // The lock had been free momentarily, but we lost the race to the lock. 424 // Interference -- the CAS failed. 425 // We can either return -1 or retry. 426 // Retry doesn't make as much sense because the lock was just acquired. 427 return -1; 428 } 429 430 // Install the displaced mark word (dmw) of a deflating ObjectMonitor 431 // into the header of the object associated with the monitor. This 432 // idempotent method is called by a thread that is deflating a 433 // monitor and by other threads that have detected a race with the 434 // deflation process. 435 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) { 436 // This function must only be called when (owner == DEFLATER_MARKER 437 // && ref_count <= 0), but we can't guarantee that here because 438 // those values could change when the ObjectMonitor gets moved from 439 // the global free list to a per-thread free list. 440 441 guarantee(obj != NULL, "must be non-NULL"); 442 if (object() != obj) { 443 // ObjectMonitor's object ref no longer refers to the target object 444 // so the object's header has already been restored. 445 return; 446 } 447 448 markWord dmw = header(); 449 if (dmw.value() == 0) { 450 // ObjectMonitor's header/dmw has been cleared so the ObjectMonitor 451 // has been deflated and taken off the global free list. 452 return; 453 } 454 455 // A non-NULL dmw has to be either neutral (not locked and not marked) 456 // or is already participating in this restoration protocol. 457 assert(dmw.is_neutral() || (dmw.is_marked() && dmw.hash() == 0), 458 "failed precondition: dmw=" INTPTR_FORMAT, dmw.value()); 459 460 markWord marked_dmw = markWord::zero(); 461 if (!dmw.is_marked() && dmw.hash() == 0) { 462 // This dmw has not yet started the restoration protocol so we 463 // mark a copy of the dmw to begin the protocol. 464 // Note: A dmw with a hashcode does not take this code path. 465 marked_dmw = dmw.set_marked(); 466 467 // All of the callers to this function can be racing with each 468 // other trying to update the _header field. 469 dmw = (markWord) Atomic::cmpxchg(&_header, dmw, marked_dmw); 470 if (dmw.value() == 0) { 471 // ObjectMonitor's header/dmw has been cleared so the object's 472 // header has already been restored. 473 return; 474 } 475 // The _header field is now marked. The winner's 'dmw' variable 476 // contains the original, unmarked header/dmw value and any 477 // losers have a marked header/dmw value that will be cleaned 478 // up below. 479 } 480 481 if (dmw.is_marked()) { 482 // Clear the mark from the header/dmw copy in preparation for 483 // possible restoration from this thread. 484 assert(dmw.hash() == 0, "hashcode must be 0: dmw=" INTPTR_FORMAT, 485 dmw.value()); 486 dmw = dmw.set_unmarked(); 487 } 488 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 489 490 // Install displaced mark word if the object's header still points 491 // to this ObjectMonitor. All racing callers to this function will 492 // reach this point, but only one can win. 493 obj->cas_set_mark(dmw, markWord::encode(this)); 494 495 // Note: It does not matter which thread restored the header/dmw 496 // into the object's header. The thread deflating the monitor just 497 // wanted the object's header restored and it is. The threads that 498 // detected a race with the deflation process also wanted the 499 // object's header restored before they retry their operation and 500 // because it is restored they will only retry once. 501 } 502 503 // Convert the fields used by is_busy() to a string that can be 504 // used for diagnostic output. 505 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 506 ss->print("is_busy: contentions=%d, waiters=%d, ", _contentions, _waiters); 507 if (!AsyncDeflateIdleMonitors) { 508 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 509 } else if (_owner != DEFLATER_MARKER) { 510 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 511 } else { 512 // We report NULL instead of DEFLATER_MARKER here because is_busy() 513 // ignores DEFLATER_MARKER values. 514 ss->print("owner=" INTPTR_FORMAT, NULL); 515 } 516 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq), 517 p2i(_EntryList)); 518 return ss->base(); 519 } 520 521 #define MAX_RECHECK_INTERVAL 1000 522 523 void ObjectMonitor::EnterI(TRAPS) { 524 jint l_ref_count = ref_count(); 525 ADIM_guarantee(l_ref_count > 0, "must be positive: l_ref_count=%d, ref_count=%d", l_ref_count, ref_count()); 526 527 Thread * const Self = THREAD; 528 assert(Self->is_Java_thread(), "invariant"); 529 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 530 531 // Try the lock - TATAS 532 if (TryLock (Self) > 0) { 533 assert(_succ != Self, "invariant"); 534 assert(_owner == Self, "invariant"); 535 assert(_Responsible != Self, "invariant"); 536 return; 537 } 538 539 if (AsyncDeflateIdleMonitors && 540 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 541 // The deflation protocol finished the first part (setting owner), 542 // but it failed the second part (making ref_count negative) and 543 // bailed. Or the ObjectMonitor was async deflated and reused. 544 // Acquired the monitor. 545 assert(_succ != Self, "invariant"); 546 assert(_Responsible != Self, "invariant"); 547 return; 548 } 549 550 assert(InitDone, "Unexpectedly not initialized"); 551 552 // We try one round of spinning *before* enqueueing Self. 553 // 554 // If the _owner is ready but OFFPROC we could use a YieldTo() 555 // operation to donate the remainder of this thread's quantum 556 // to the owner. This has subtle but beneficial affinity 557 // effects. 558 559 if (TrySpin(Self) > 0) { 560 assert(_owner == Self, "invariant"); 561 assert(_succ != Self, "invariant"); 562 assert(_Responsible != Self, "invariant"); 563 return; 564 } 565 566 // The Spin failed -- Enqueue and park the thread ... 567 assert(_succ != Self, "invariant"); 568 assert(_owner != Self, "invariant"); 569 assert(_Responsible != Self, "invariant"); 570 571 // Enqueue "Self" on ObjectMonitor's _cxq. 572 // 573 // Node acts as a proxy for Self. 574 // As an aside, if were to ever rewrite the synchronization code mostly 575 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 576 // Java objects. This would avoid awkward lifecycle and liveness issues, 577 // as well as eliminate a subset of ABA issues. 578 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 579 580 ObjectWaiter node(Self); 581 Self->_ParkEvent->reset(); 582 node._prev = (ObjectWaiter *) 0xBAD; 583 node.TState = ObjectWaiter::TS_CXQ; 584 585 // Push "Self" onto the front of the _cxq. 586 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 587 // Note that spinning tends to reduce the rate at which threads 588 // enqueue and dequeue on EntryList|cxq. 589 ObjectWaiter * nxt; 590 for (;;) { 591 node._next = nxt = _cxq; 592 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 593 594 // Interference - the CAS failed because _cxq changed. Just retry. 595 // As an optional optimization we retry the lock. 596 if (TryLock (Self) > 0) { 597 assert(_succ != Self, "invariant"); 598 assert(_owner == Self, "invariant"); 599 assert(_Responsible != Self, "invariant"); 600 return; 601 } 602 } 603 604 // Check for cxq|EntryList edge transition to non-null. This indicates 605 // the onset of contention. While contention persists exiting threads 606 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 607 // operations revert to the faster 1-0 mode. This enter operation may interleave 608 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 609 // arrange for one of the contending thread to use a timed park() operations 610 // to detect and recover from the race. (Stranding is form of progress failure 611 // where the monitor is unlocked but all the contending threads remain parked). 612 // That is, at least one of the contended threads will periodically poll _owner. 613 // One of the contending threads will become the designated "Responsible" thread. 614 // The Responsible thread uses a timed park instead of a normal indefinite park 615 // operation -- it periodically wakes and checks for and recovers from potential 616 // strandings admitted by 1-0 exit operations. We need at most one Responsible 617 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 618 // be responsible for a monitor. 619 // 620 // Currently, one of the contended threads takes on the added role of "Responsible". 621 // A viable alternative would be to use a dedicated "stranding checker" thread 622 // that periodically iterated over all the threads (or active monitors) and unparked 623 // successors where there was risk of stranding. This would help eliminate the 624 // timer scalability issues we see on some platforms as we'd only have one thread 625 // -- the checker -- parked on a timer. 626 627 if (nxt == NULL && _EntryList == NULL) { 628 // Try to assume the role of responsible thread for the monitor. 629 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 630 Atomic::replace_if_null(&_Responsible, Self); 631 } 632 633 // The lock might have been released while this thread was occupied queueing 634 // itself onto _cxq. To close the race and avoid "stranding" and 635 // progress-liveness failure we must resample-retry _owner before parking. 636 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 637 // In this case the ST-MEMBAR is accomplished with CAS(). 638 // 639 // TODO: Defer all thread state transitions until park-time. 640 // Since state transitions are heavy and inefficient we'd like 641 // to defer the state transitions until absolutely necessary, 642 // and in doing so avoid some transitions ... 643 644 int nWakeups = 0; 645 int recheckInterval = 1; 646 647 for (;;) { 648 649 if (TryLock(Self) > 0) break; 650 assert(_owner != Self, "invariant"); 651 652 // park self 653 if (_Responsible == Self) { 654 Self->_ParkEvent->park((jlong) recheckInterval); 655 // Increase the recheckInterval, but clamp the value. 656 recheckInterval *= 8; 657 if (recheckInterval > MAX_RECHECK_INTERVAL) { 658 recheckInterval = MAX_RECHECK_INTERVAL; 659 } 660 } else { 661 Self->_ParkEvent->park(); 662 } 663 664 if (TryLock(Self) > 0) break; 665 666 if (AsyncDeflateIdleMonitors && 667 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 668 // The deflation protocol finished the first part (setting owner), 669 // but it failed the second part (making ref_count negative) and 670 // bailed. Or the ObjectMonitor was async deflated and reused. 671 // Acquired the monitor. 672 break; 673 } 674 675 // The lock is still contested. 676 // Keep a tally of the # of futile wakeups. 677 // Note that the counter is not protected by a lock or updated by atomics. 678 // That is by design - we trade "lossy" counters which are exposed to 679 // races during updates for a lower probe effect. 680 681 // This PerfData object can be used in parallel with a safepoint. 682 // See the work around in PerfDataManager::destroy(). 683 OM_PERFDATA_OP(FutileWakeups, inc()); 684 ++nWakeups; 685 686 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 687 // We can defer clearing _succ until after the spin completes 688 // TrySpin() must tolerate being called with _succ == Self. 689 // Try yet another round of adaptive spinning. 690 if (TrySpin(Self) > 0) break; 691 692 // We can find that we were unpark()ed and redesignated _succ while 693 // we were spinning. That's harmless. If we iterate and call park(), 694 // park() will consume the event and return immediately and we'll 695 // just spin again. This pattern can repeat, leaving _succ to simply 696 // spin on a CPU. 697 698 if (_succ == Self) _succ = NULL; 699 700 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 701 OrderAccess::fence(); 702 } 703 704 // Egress : 705 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 706 // Normally we'll find Self on the EntryList . 707 // From the perspective of the lock owner (this thread), the 708 // EntryList is stable and cxq is prepend-only. 709 // The head of cxq is volatile but the interior is stable. 710 // In addition, Self.TState is stable. 711 712 assert(_owner == Self, "invariant"); 713 assert(object() != NULL, "invariant"); 714 // I'd like to write: 715 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 716 // but as we're at a safepoint that's not safe. 717 718 UnlinkAfterAcquire(Self, &node); 719 if (_succ == Self) _succ = NULL; 720 721 assert(_succ != Self, "invariant"); 722 if (_Responsible == Self) { 723 _Responsible = NULL; 724 OrderAccess::fence(); // Dekker pivot-point 725 726 // We may leave threads on cxq|EntryList without a designated 727 // "Responsible" thread. This is benign. When this thread subsequently 728 // exits the monitor it can "see" such preexisting "old" threads -- 729 // threads that arrived on the cxq|EntryList before the fence, above -- 730 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 731 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 732 // non-null and elect a new "Responsible" timer thread. 733 // 734 // This thread executes: 735 // ST Responsible=null; MEMBAR (in enter epilogue - here) 736 // LD cxq|EntryList (in subsequent exit) 737 // 738 // Entering threads in the slow/contended path execute: 739 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 740 // The (ST cxq; MEMBAR) is accomplished with CAS(). 741 // 742 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 743 // exit operation from floating above the ST Responsible=null. 744 } 745 746 // We've acquired ownership with CAS(). 747 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 748 // But since the CAS() this thread may have also stored into _succ, 749 // EntryList, cxq or Responsible. These meta-data updates must be 750 // visible __before this thread subsequently drops the lock. 751 // Consider what could occur if we didn't enforce this constraint -- 752 // STs to monitor meta-data and user-data could reorder with (become 753 // visible after) the ST in exit that drops ownership of the lock. 754 // Some other thread could then acquire the lock, but observe inconsistent 755 // or old monitor meta-data and heap data. That violates the JMM. 756 // To that end, the 1-0 exit() operation must have at least STST|LDST 757 // "release" barrier semantics. Specifically, there must be at least a 758 // STST|LDST barrier in exit() before the ST of null into _owner that drops 759 // the lock. The barrier ensures that changes to monitor meta-data and data 760 // protected by the lock will be visible before we release the lock, and 761 // therefore before some other thread (CPU) has a chance to acquire the lock. 762 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 763 // 764 // Critically, any prior STs to _succ or EntryList must be visible before 765 // the ST of null into _owner in the *subsequent* (following) corresponding 766 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 767 // execute a serializing instruction. 768 769 return; 770 } 771 772 // ReenterI() is a specialized inline form of the latter half of the 773 // contended slow-path from EnterI(). We use ReenterI() only for 774 // monitor reentry in wait(). 775 // 776 // In the future we should reconcile EnterI() and ReenterI(). 777 778 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 779 jint l_ref_count = ref_count(); 780 ADIM_guarantee(l_ref_count > 0, "must be positive: l_ref_count=%d, ref_count=%d", l_ref_count, ref_count()); 781 782 assert(Self != NULL, "invariant"); 783 assert(SelfNode != NULL, "invariant"); 784 assert(SelfNode->_thread == Self, "invariant"); 785 assert(_waiters > 0, "invariant"); 786 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 787 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 788 JavaThread * jt = (JavaThread *) Self; 789 790 int nWakeups = 0; 791 for (;;) { 792 ObjectWaiter::TStates v = SelfNode->TState; 793 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 794 assert(_owner != Self, "invariant"); 795 796 if (TryLock(Self) > 0) break; 797 if (TrySpin(Self) > 0) break; 798 799 if (AsyncDeflateIdleMonitors && 800 try_set_owner_from(Self, DEFLATER_MARKER) == DEFLATER_MARKER) { 801 // The deflation protocol finished the first part (setting owner), 802 // but it failed the second part (making ref_count negative) and 803 // bailed. Or the ObjectMonitor was async deflated and reused. 804 // Acquired the monitor. 805 break; 806 } 807 808 // State transition wrappers around park() ... 809 // ReenterI() wisely defers state transitions until 810 // it's clear we must park the thread. 811 { 812 OSThreadContendState osts(Self->osthread()); 813 ThreadBlockInVM tbivm(jt); 814 815 // cleared by handle_special_suspend_equivalent_condition() 816 // or java_suspend_self() 817 jt->set_suspend_equivalent(); 818 Self->_ParkEvent->park(); 819 820 // were we externally suspended while we were waiting? 821 for (;;) { 822 if (!ExitSuspendEquivalent(jt)) break; 823 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 824 jt->java_suspend_self(); 825 jt->set_suspend_equivalent(); 826 } 827 } 828 829 // Try again, but just so we distinguish between futile wakeups and 830 // successful wakeups. The following test isn't algorithmically 831 // necessary, but it helps us maintain sensible statistics. 832 if (TryLock(Self) > 0) break; 833 834 // The lock is still contested. 835 // Keep a tally of the # of futile wakeups. 836 // Note that the counter is not protected by a lock or updated by atomics. 837 // That is by design - we trade "lossy" counters which are exposed to 838 // races during updates for a lower probe effect. 839 ++nWakeups; 840 841 // Assuming this is not a spurious wakeup we'll normally 842 // find that _succ == Self. 843 if (_succ == Self) _succ = NULL; 844 845 // Invariant: after clearing _succ a contending thread 846 // *must* retry _owner before parking. 847 OrderAccess::fence(); 848 849 // This PerfData object can be used in parallel with a safepoint. 850 // See the work around in PerfDataManager::destroy(). 851 OM_PERFDATA_OP(FutileWakeups, inc()); 852 } 853 854 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 855 // Normally we'll find Self on the EntryList. 856 // Unlinking from the EntryList is constant-time and atomic-free. 857 // From the perspective of the lock owner (this thread), the 858 // EntryList is stable and cxq is prepend-only. 859 // The head of cxq is volatile but the interior is stable. 860 // In addition, Self.TState is stable. 861 862 assert(_owner == Self, "invariant"); 863 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 864 UnlinkAfterAcquire(Self, SelfNode); 865 if (_succ == Self) _succ = NULL; 866 assert(_succ != Self, "invariant"); 867 SelfNode->TState = ObjectWaiter::TS_RUN; 868 OrderAccess::fence(); // see comments at the end of EnterI() 869 } 870 871 // By convention we unlink a contending thread from EntryList|cxq immediately 872 // after the thread acquires the lock in ::enter(). Equally, we could defer 873 // unlinking the thread until ::exit()-time. 874 875 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 876 assert(_owner == Self, "invariant"); 877 assert(SelfNode->_thread == Self, "invariant"); 878 879 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 880 // Normal case: remove Self from the DLL EntryList . 881 // This is a constant-time operation. 882 ObjectWaiter * nxt = SelfNode->_next; 883 ObjectWaiter * prv = SelfNode->_prev; 884 if (nxt != NULL) nxt->_prev = prv; 885 if (prv != NULL) prv->_next = nxt; 886 if (SelfNode == _EntryList) _EntryList = nxt; 887 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 888 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 889 } else { 890 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 891 // Inopportune interleaving -- Self is still on the cxq. 892 // This usually means the enqueue of self raced an exiting thread. 893 // Normally we'll find Self near the front of the cxq, so 894 // dequeueing is typically fast. If needbe we can accelerate 895 // this with some MCS/CHL-like bidirectional list hints and advisory 896 // back-links so dequeueing from the interior will normally operate 897 // in constant-time. 898 // Dequeue Self from either the head (with CAS) or from the interior 899 // with a linear-time scan and normal non-atomic memory operations. 900 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 901 // and then unlink Self from EntryList. We have to drain eventually, 902 // so it might as well be now. 903 904 ObjectWaiter * v = _cxq; 905 assert(v != NULL, "invariant"); 906 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) { 907 // The CAS above can fail from interference IFF a "RAT" arrived. 908 // In that case Self must be in the interior and can no longer be 909 // at the head of cxq. 910 if (v == SelfNode) { 911 assert(_cxq != v, "invariant"); 912 v = _cxq; // CAS above failed - start scan at head of list 913 } 914 ObjectWaiter * p; 915 ObjectWaiter * q = NULL; 916 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 917 q = p; 918 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 919 } 920 assert(v != SelfNode, "invariant"); 921 assert(p == SelfNode, "Node not found on cxq"); 922 assert(p != _cxq, "invariant"); 923 assert(q != NULL, "invariant"); 924 assert(q->_next == p, "invariant"); 925 q->_next = p->_next; 926 } 927 } 928 929 #ifdef ASSERT 930 // Diagnostic hygiene ... 931 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 932 SelfNode->_next = (ObjectWaiter *) 0xBAD; 933 SelfNode->TState = ObjectWaiter::TS_RUN; 934 #endif 935 } 936 937 // ----------------------------------------------------------------------------- 938 // Exit support 939 // 940 // exit() 941 // ~~~~~~ 942 // Note that the collector can't reclaim the objectMonitor or deflate 943 // the object out from underneath the thread calling ::exit() as the 944 // thread calling ::exit() never transitions to a stable state. 945 // This inhibits GC, which in turn inhibits asynchronous (and 946 // inopportune) reclamation of "this". 947 // 948 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 949 // There's one exception to the claim above, however. EnterI() can call 950 // exit() to drop a lock if the acquirer has been externally suspended. 951 // In that case exit() is called with _thread_state as _thread_blocked, 952 // but the monitor's _contentions field is > 0, which inhibits reclamation. 953 // 954 // 1-0 exit 955 // ~~~~~~~~ 956 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 957 // the fast-path operators have been optimized so the common ::exit() 958 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 959 // The code emitted by fast_unlock() elides the usual MEMBAR. This 960 // greatly improves latency -- MEMBAR and CAS having considerable local 961 // latency on modern processors -- but at the cost of "stranding". Absent the 962 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 963 // ::enter() path, resulting in the entering thread being stranding 964 // and a progress-liveness failure. Stranding is extremely rare. 965 // We use timers (timed park operations) & periodic polling to detect 966 // and recover from stranding. Potentially stranded threads periodically 967 // wake up and poll the lock. See the usage of the _Responsible variable. 968 // 969 // The CAS() in enter provides for safety and exclusion, while the CAS or 970 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 971 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 972 // We detect and recover from stranding with timers. 973 // 974 // If a thread transiently strands it'll park until (a) another 975 // thread acquires the lock and then drops the lock, at which time the 976 // exiting thread will notice and unpark the stranded thread, or, (b) 977 // the timer expires. If the lock is high traffic then the stranding latency 978 // will be low due to (a). If the lock is low traffic then the odds of 979 // stranding are lower, although the worst-case stranding latency 980 // is longer. Critically, we don't want to put excessive load in the 981 // platform's timer subsystem. We want to minimize both the timer injection 982 // rate (timers created/sec) as well as the number of timers active at 983 // any one time. (more precisely, we want to minimize timer-seconds, which is 984 // the integral of the # of active timers at any instant over time). 985 // Both impinge on OS scalability. Given that, at most one thread parked on 986 // a monitor will use a timer. 987 // 988 // There is also the risk of a futile wake-up. If we drop the lock 989 // another thread can reacquire the lock immediately, and we can 990 // then wake a thread unnecessarily. This is benign, and we've 991 // structured the code so the windows are short and the frequency 992 // of such futile wakups is low. 993 994 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 995 Thread * const Self = THREAD; 996 if (THREAD != _owner) { 997 void* cur = _owner; 998 if (THREAD->is_lock_owned((address)cur)) { 999 assert(_recursions == 0, "invariant"); 1000 simply_set_owner_from_BasicLock(Self, cur); // Convert from BasicLock* to Thread*. 1001 _recursions = 0; 1002 } else { 1003 // Apparent unbalanced locking ... 1004 // Naively we'd like to throw IllegalMonitorStateException. 1005 // As a practical matter we can neither allocate nor throw an 1006 // exception as ::exit() can be called from leaf routines. 1007 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1008 // Upon deeper reflection, however, in a properly run JVM the only 1009 // way we should encounter this situation is in the presence of 1010 // unbalanced JNI locking. TODO: CheckJNICalls. 1011 // See also: CR4414101 1012 //#ifdef ASSERT 1013 LogStreamHandle(Error, monitorinflation) lsh; 1014 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 1015 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 1016 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 1017 print_debug_style_on(&lsh); 1018 //#endif 1019 // Changing this from an assert() to ADIM_guarantee() may run 1020 // afoul of any test that is inducing non-balanced JNI locking. 1021 ADIM_guarantee(false, "Non-balanced monitor enter/exit!"); 1022 return; 1023 } 1024 } 1025 1026 if (_recursions != 0) { 1027 _recursions--; // this is simple recursive enter 1028 return; 1029 } 1030 1031 // Invariant: after setting Responsible=null an thread must execute 1032 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1033 _Responsible = NULL; 1034 1035 #if INCLUDE_JFR 1036 // get the owner's thread id for the MonitorEnter event 1037 // if it is enabled and the thread isn't suspended 1038 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1039 _previous_owner_tid = JFR_THREAD_ID(Self); 1040 } 1041 #endif 1042 1043 for (;;) { 1044 assert(THREAD == _owner, "invariant"); 1045 1046 // release semantics: prior loads and stores from within the critical section 1047 // must not float (reorder) past the following store that drops the lock. 1048 if (AsyncDeflateIdleMonitors) { 1049 set_owner_from(NULL, Self); 1050 } else { 1051 Atomic::release_store(&_owner, (void*)NULL); // drop the lock 1052 OrderAccess::storeload(); // See if we need to wake a successor 1053 } 1054 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1055 return; 1056 } 1057 // Other threads are blocked trying to acquire the lock. 1058 1059 // Normally the exiting thread is responsible for ensuring succession, 1060 // but if other successors are ready or other entering threads are spinning 1061 // then this thread can simply store NULL into _owner and exit without 1062 // waking a successor. The existence of spinners or ready successors 1063 // guarantees proper succession (liveness). Responsibility passes to the 1064 // ready or running successors. The exiting thread delegates the duty. 1065 // More precisely, if a successor already exists this thread is absolved 1066 // of the responsibility of waking (unparking) one. 1067 // 1068 // The _succ variable is critical to reducing futile wakeup frequency. 1069 // _succ identifies the "heir presumptive" thread that has been made 1070 // ready (unparked) but that has not yet run. We need only one such 1071 // successor thread to guarantee progress. 1072 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1073 // section 3.3 "Futile Wakeup Throttling" for details. 1074 // 1075 // Note that spinners in Enter() also set _succ non-null. 1076 // In the current implementation spinners opportunistically set 1077 // _succ so that exiting threads might avoid waking a successor. 1078 // Another less appealing alternative would be for the exiting thread 1079 // to drop the lock and then spin briefly to see if a spinner managed 1080 // to acquire the lock. If so, the exiting thread could exit 1081 // immediately without waking a successor, otherwise the exiting 1082 // thread would need to dequeue and wake a successor. 1083 // (Note that we'd need to make the post-drop spin short, but no 1084 // shorter than the worst-case round-trip cache-line migration time. 1085 // The dropped lock needs to become visible to the spinner, and then 1086 // the acquisition of the lock by the spinner must become visible to 1087 // the exiting thread). 1088 1089 // It appears that an heir-presumptive (successor) must be made ready. 1090 // Only the current lock owner can manipulate the EntryList or 1091 // drain _cxq, so we need to reacquire the lock. If we fail 1092 // to reacquire the lock the responsibility for ensuring succession 1093 // falls to the new owner. 1094 // 1095 if (try_set_owner_from(Self, NULL) != NULL) { 1096 return; 1097 } 1098 1099 guarantee(_owner == THREAD, "invariant"); 1100 1101 ObjectWaiter * w = NULL; 1102 1103 w = _EntryList; 1104 if (w != NULL) { 1105 // I'd like to write: guarantee (w->_thread != Self). 1106 // But in practice an exiting thread may find itself on the EntryList. 1107 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1108 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1109 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1110 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1111 // release the lock "O". T2 resumes immediately after the ST of null into 1112 // _owner, above. T2 notices that the EntryList is populated, so it 1113 // reacquires the lock and then finds itself on the EntryList. 1114 // Given all that, we have to tolerate the circumstance where "w" is 1115 // associated with Self. 1116 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1117 ExitEpilog(Self, w); 1118 return; 1119 } 1120 1121 // If we find that both _cxq and EntryList are null then just 1122 // re-run the exit protocol from the top. 1123 w = _cxq; 1124 if (w == NULL) continue; 1125 1126 // Drain _cxq into EntryList - bulk transfer. 1127 // First, detach _cxq. 1128 // The following loop is tantamount to: w = swap(&cxq, NULL) 1129 for (;;) { 1130 assert(w != NULL, "Invariant"); 1131 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL); 1132 if (u == w) break; 1133 w = u; 1134 } 1135 1136 assert(w != NULL, "invariant"); 1137 assert(_EntryList == NULL, "invariant"); 1138 1139 // Convert the LIFO SLL anchored by _cxq into a DLL. 1140 // The list reorganization step operates in O(LENGTH(w)) time. 1141 // It's critical that this step operate quickly as 1142 // "Self" still holds the outer-lock, restricting parallelism 1143 // and effectively lengthening the critical section. 1144 // Invariant: s chases t chases u. 1145 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1146 // we have faster access to the tail. 1147 1148 _EntryList = w; 1149 ObjectWaiter * q = NULL; 1150 ObjectWaiter * p; 1151 for (p = w; p != NULL; p = p->_next) { 1152 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1153 p->TState = ObjectWaiter::TS_ENTER; 1154 p->_prev = q; 1155 q = p; 1156 } 1157 1158 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1159 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1160 1161 // See if we can abdicate to a spinner instead of waking a thread. 1162 // A primary goal of the implementation is to reduce the 1163 // context-switch rate. 1164 if (_succ != NULL) continue; 1165 1166 w = _EntryList; 1167 if (w != NULL) { 1168 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1169 ExitEpilog(Self, w); 1170 return; 1171 } 1172 } 1173 } 1174 1175 // ExitSuspendEquivalent: 1176 // A faster alternate to handle_special_suspend_equivalent_condition() 1177 // 1178 // handle_special_suspend_equivalent_condition() unconditionally 1179 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1180 // operations have high latency. Note that in ::enter() we call HSSEC 1181 // while holding the monitor, so we effectively lengthen the critical sections. 1182 // 1183 // There are a number of possible solutions: 1184 // 1185 // A. To ameliorate the problem we might also defer state transitions 1186 // to as late as possible -- just prior to parking. 1187 // Given that, we'd call HSSEC after having returned from park(), 1188 // but before attempting to acquire the monitor. This is only a 1189 // partial solution. It avoids calling HSSEC while holding the 1190 // monitor (good), but it still increases successor reacquisition latency -- 1191 // the interval between unparking a successor and the time the successor 1192 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1193 // If we use this technique we can also avoid EnterI()-exit() loop 1194 // in ::enter() where we iteratively drop the lock and then attempt 1195 // to reacquire it after suspending. 1196 // 1197 // B. In the future we might fold all the suspend bits into a 1198 // composite per-thread suspend flag and then update it with CAS(). 1199 // Alternately, a Dekker-like mechanism with multiple variables 1200 // would suffice: 1201 // ST Self->_suspend_equivalent = false 1202 // MEMBAR 1203 // LD Self_>_suspend_flags 1204 1205 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1206 return jSelf->handle_special_suspend_equivalent_condition(); 1207 } 1208 1209 1210 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1211 assert(_owner == Self, "invariant"); 1212 1213 // Exit protocol: 1214 // 1. ST _succ = wakee 1215 // 2. membar #loadstore|#storestore; 1216 // 2. ST _owner = NULL 1217 // 3. unpark(wakee) 1218 1219 _succ = Wakee->_thread; 1220 ParkEvent * Trigger = Wakee->_event; 1221 1222 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1223 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1224 // out-of-scope (non-extant). 1225 Wakee = NULL; 1226 1227 // Drop the lock 1228 if (AsyncDeflateIdleMonitors) { 1229 set_owner_from(NULL, Self); 1230 } else { 1231 Atomic::release_store(&_owner, (void*)NULL); 1232 OrderAccess::fence(); // ST _owner vs LD in unpark() 1233 } 1234 1235 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1236 Trigger->unpark(); 1237 1238 // Maintain stats and report events to JVMTI 1239 OM_PERFDATA_OP(Parks, inc()); 1240 } 1241 1242 1243 // ----------------------------------------------------------------------------- 1244 // Class Loader deadlock handling. 1245 // 1246 // complete_exit exits a lock returning recursion count 1247 // complete_exit/reenter operate as a wait without waiting 1248 // complete_exit requires an inflated monitor 1249 // The _owner field is not always the Thread addr even with an 1250 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1251 // thread due to contention. 1252 intx ObjectMonitor::complete_exit(TRAPS) { 1253 Thread * const Self = THREAD; 1254 assert(Self->is_Java_thread(), "Must be Java thread!"); 1255 JavaThread *jt = (JavaThread *)THREAD; 1256 1257 assert(InitDone, "Unexpectedly not initialized"); 1258 1259 if (THREAD != _owner) { 1260 void* cur = _owner; 1261 if (THREAD->is_lock_owned((address)cur)) { 1262 assert(_recursions == 0, "internal state error"); 1263 simply_set_owner_from_BasicLock(Self, cur); // Convert from BasicLock* to Thread*. 1264 _recursions = 0; 1265 } 1266 } 1267 1268 guarantee(Self == _owner, "complete_exit not owner"); 1269 intx save = _recursions; // record the old recursion count 1270 _recursions = 0; // set the recursion level to be 0 1271 exit(true, Self); // exit the monitor 1272 guarantee(_owner != Self, "invariant"); 1273 return save; 1274 } 1275 1276 // reenter() enters a lock and sets recursion count 1277 // complete_exit/reenter operate as a wait without waiting 1278 void ObjectMonitor::reenter(intx recursions, TRAPS) { 1279 Thread * const Self = THREAD; 1280 assert(Self->is_Java_thread(), "Must be Java thread!"); 1281 JavaThread *jt = (JavaThread *)THREAD; 1282 1283 guarantee(_owner != Self, "reenter already owner"); 1284 enter(THREAD); 1285 // Entered the monitor. 1286 guarantee(_recursions == 0, "reenter recursion"); 1287 _recursions = recursions; 1288 } 1289 1290 // Checks that the current THREAD owns this monitor and causes an 1291 // immediate return if it doesn't. We don't use the CHECK macro 1292 // because we want the IMSE to be the only exception that is thrown 1293 // from the call site when false is returned. Any other pending 1294 // exception is ignored. 1295 #define CHECK_OWNER() \ 1296 do { \ 1297 if (!check_owner(THREAD)) { \ 1298 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1299 return; \ 1300 } \ 1301 } while (false) 1302 1303 // Returns true if the specified thread owns the ObjectMonitor. 1304 // Otherwise returns false and throws IllegalMonitorStateException 1305 // (IMSE). If there is a pending exception and the specified thread 1306 // is not the owner, that exception will be replaced by the IMSE. 1307 bool ObjectMonitor::check_owner(Thread* THREAD) { 1308 if (_owner == THREAD) { 1309 return true; 1310 } 1311 void* cur = _owner; 1312 if (THREAD->is_lock_owned((address)cur)) { 1313 simply_set_owner_from_BasicLock(THREAD, cur); // Convert from BasicLock* to Thread*. 1314 _recursions = 0; 1315 return true; 1316 } 1317 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1318 "current thread is not owner", false); 1319 } 1320 1321 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1322 ObjectMonitor* monitor, 1323 jlong notifier_tid, 1324 jlong timeout, 1325 bool timedout) { 1326 assert(event != NULL, "invariant"); 1327 assert(monitor != NULL, "invariant"); 1328 event->set_monitorClass(((oop)monitor->object())->klass()); 1329 event->set_timeout(timeout); 1330 event->set_address((uintptr_t)monitor->object_addr()); 1331 event->set_notifier(notifier_tid); 1332 event->set_timedOut(timedout); 1333 event->commit(); 1334 } 1335 1336 // ----------------------------------------------------------------------------- 1337 // Wait/Notify/NotifyAll 1338 // 1339 // Note: a subset of changes to ObjectMonitor::wait() 1340 // will need to be replicated in complete_exit 1341 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1342 Thread * const Self = THREAD; 1343 assert(Self->is_Java_thread(), "Must be Java thread!"); 1344 JavaThread *jt = (JavaThread *)THREAD; 1345 1346 assert(InitDone, "Unexpectedly not initialized"); 1347 1348 CHECK_OWNER(); // Throws IMSE if not owner. 1349 1350 EventJavaMonitorWait event; 1351 1352 // check for a pending interrupt 1353 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1354 // post monitor waited event. Note that this is past-tense, we are done waiting. 1355 if (JvmtiExport::should_post_monitor_waited()) { 1356 // Note: 'false' parameter is passed here because the 1357 // wait was not timed out due to thread interrupt. 1358 JvmtiExport::post_monitor_waited(jt, this, false); 1359 1360 // In this short circuit of the monitor wait protocol, the 1361 // current thread never drops ownership of the monitor and 1362 // never gets added to the wait queue so the current thread 1363 // cannot be made the successor. This means that the 1364 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1365 // consume an unpark() meant for the ParkEvent associated with 1366 // this ObjectMonitor. 1367 } 1368 if (event.should_commit()) { 1369 post_monitor_wait_event(&event, this, 0, millis, false); 1370 } 1371 THROW(vmSymbols::java_lang_InterruptedException()); 1372 return; 1373 } 1374 1375 assert(Self->_Stalled == 0, "invariant"); 1376 Self->_Stalled = intptr_t(this); 1377 jt->set_current_waiting_monitor(this); 1378 1379 // create a node to be put into the queue 1380 // Critically, after we reset() the event but prior to park(), we must check 1381 // for a pending interrupt. 1382 ObjectWaiter node(Self); 1383 node.TState = ObjectWaiter::TS_WAIT; 1384 Self->_ParkEvent->reset(); 1385 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1386 1387 // Enter the waiting queue, which is a circular doubly linked list in this case 1388 // but it could be a priority queue or any data structure. 1389 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1390 // by the the owner of the monitor *except* in the case where park() 1391 // returns because of a timeout of interrupt. Contention is exceptionally rare 1392 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1393 1394 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1395 AddWaiter(&node); 1396 Thread::SpinRelease(&_WaitSetLock); 1397 1398 _Responsible = NULL; 1399 1400 intx save = _recursions; // record the old recursion count 1401 _waiters++; // increment the number of waiters 1402 _recursions = 0; // set the recursion level to be 1 1403 exit(true, Self); // exit the monitor 1404 guarantee(_owner != Self, "invariant"); 1405 1406 // The thread is on the WaitSet list - now park() it. 1407 // On MP systems it's conceivable that a brief spin before we park 1408 // could be profitable. 1409 // 1410 // TODO-FIXME: change the following logic to a loop of the form 1411 // while (!timeout && !interrupted && _notified == 0) park() 1412 1413 int ret = OS_OK; 1414 int WasNotified = 0; 1415 1416 // Need to check interrupt state whilst still _thread_in_vm 1417 bool interrupted = interruptible && jt->is_interrupted(false); 1418 1419 { // State transition wrappers 1420 OSThread* osthread = Self->osthread(); 1421 OSThreadWaitState osts(osthread, true); 1422 { 1423 ThreadBlockInVM tbivm(jt); 1424 // Thread is in thread_blocked state and oop access is unsafe. 1425 jt->set_suspend_equivalent(); 1426 1427 if (interrupted || HAS_PENDING_EXCEPTION) { 1428 // Intentionally empty 1429 } else if (node._notified == 0) { 1430 if (millis <= 0) { 1431 Self->_ParkEvent->park(); 1432 } else { 1433 ret = Self->_ParkEvent->park(millis); 1434 } 1435 } 1436 1437 // were we externally suspended while we were waiting? 1438 if (ExitSuspendEquivalent (jt)) { 1439 // TODO-FIXME: add -- if succ == Self then succ = null. 1440 jt->java_suspend_self(); 1441 } 1442 1443 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1444 1445 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1446 // from the WaitSet to the EntryList. 1447 // See if we need to remove Node from the WaitSet. 1448 // We use double-checked locking to avoid grabbing _WaitSetLock 1449 // if the thread is not on the wait queue. 1450 // 1451 // Note that we don't need a fence before the fetch of TState. 1452 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1453 // written by the is thread. (perhaps the fetch might even be satisfied 1454 // by a look-aside into the processor's own store buffer, although given 1455 // the length of the code path between the prior ST and this load that's 1456 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1457 // then we'll acquire the lock and then re-fetch a fresh TState value. 1458 // That is, we fail toward safety. 1459 1460 if (node.TState == ObjectWaiter::TS_WAIT) { 1461 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1462 if (node.TState == ObjectWaiter::TS_WAIT) { 1463 DequeueSpecificWaiter(&node); // unlink from WaitSet 1464 assert(node._notified == 0, "invariant"); 1465 node.TState = ObjectWaiter::TS_RUN; 1466 } 1467 Thread::SpinRelease(&_WaitSetLock); 1468 } 1469 1470 // The thread is now either on off-list (TS_RUN), 1471 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1472 // The Node's TState variable is stable from the perspective of this thread. 1473 // No other threads will asynchronously modify TState. 1474 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1475 OrderAccess::loadload(); 1476 if (_succ == Self) _succ = NULL; 1477 WasNotified = node._notified; 1478 1479 // Reentry phase -- reacquire the monitor. 1480 // re-enter contended monitor after object.wait(). 1481 // retain OBJECT_WAIT state until re-enter successfully completes 1482 // Thread state is thread_in_vm and oop access is again safe, 1483 // although the raw address of the object may have changed. 1484 // (Don't cache naked oops over safepoints, of course). 1485 1486 // post monitor waited event. Note that this is past-tense, we are done waiting. 1487 if (JvmtiExport::should_post_monitor_waited()) { 1488 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1489 1490 if (node._notified != 0 && _succ == Self) { 1491 // In this part of the monitor wait-notify-reenter protocol it 1492 // is possible (and normal) for another thread to do a fastpath 1493 // monitor enter-exit while this thread is still trying to get 1494 // to the reenter portion of the protocol. 1495 // 1496 // The ObjectMonitor was notified and the current thread is 1497 // the successor which also means that an unpark() has already 1498 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1499 // consume the unpark() that was done when the successor was 1500 // set because the same ParkEvent is shared between Java 1501 // monitors and JVM/TI RawMonitors (for now). 1502 // 1503 // We redo the unpark() to ensure forward progress, i.e., we 1504 // don't want all pending threads hanging (parked) with none 1505 // entering the unlocked monitor. 1506 node._event->unpark(); 1507 } 1508 } 1509 1510 if (event.should_commit()) { 1511 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1512 } 1513 1514 OrderAccess::fence(); 1515 1516 assert(Self->_Stalled != 0, "invariant"); 1517 Self->_Stalled = 0; 1518 1519 assert(_owner != Self, "invariant"); 1520 ObjectWaiter::TStates v = node.TState; 1521 if (v == ObjectWaiter::TS_RUN) { 1522 enter(Self); 1523 } else { 1524 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1525 ReenterI(Self, &node); 1526 node.wait_reenter_end(this); 1527 } 1528 1529 // Self has reacquired the lock. 1530 // Lifecycle - the node representing Self must not appear on any queues. 1531 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1532 // want residual elements associated with this thread left on any lists. 1533 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1534 assert(_owner == Self, "invariant"); 1535 assert(_succ != Self, "invariant"); 1536 } // OSThreadWaitState() 1537 1538 jt->set_current_waiting_monitor(NULL); 1539 1540 guarantee(_recursions == 0, "invariant"); 1541 _recursions = save; // restore the old recursion count 1542 _waiters--; // decrement the number of waiters 1543 1544 // Verify a few postconditions 1545 assert(_owner == Self, "invariant"); 1546 assert(_succ != Self, "invariant"); 1547 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1548 1549 // check if the notification happened 1550 if (!WasNotified) { 1551 // no, it could be timeout or Thread.interrupt() or both 1552 // check for interrupt event, otherwise it is timeout 1553 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1554 THROW(vmSymbols::java_lang_InterruptedException()); 1555 } 1556 } 1557 1558 // NOTE: Spurious wake up will be consider as timeout. 1559 // Monitor notify has precedence over thread interrupt. 1560 } 1561 1562 1563 // Consider: 1564 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1565 // then instead of transferring a thread from the WaitSet to the EntryList 1566 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1567 1568 void ObjectMonitor::INotify(Thread * Self) { 1569 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1570 ObjectWaiter * iterator = DequeueWaiter(); 1571 if (iterator != NULL) { 1572 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1573 guarantee(iterator->_notified == 0, "invariant"); 1574 // Disposition - what might we do with iterator ? 1575 // a. add it directly to the EntryList - either tail (policy == 1) 1576 // or head (policy == 0). 1577 // b. push it onto the front of the _cxq (policy == 2). 1578 // For now we use (b). 1579 1580 iterator->TState = ObjectWaiter::TS_ENTER; 1581 1582 iterator->_notified = 1; 1583 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1584 1585 ObjectWaiter * list = _EntryList; 1586 if (list != NULL) { 1587 assert(list->_prev == NULL, "invariant"); 1588 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1589 assert(list != iterator, "invariant"); 1590 } 1591 1592 // prepend to cxq 1593 if (list == NULL) { 1594 iterator->_next = iterator->_prev = NULL; 1595 _EntryList = iterator; 1596 } else { 1597 iterator->TState = ObjectWaiter::TS_CXQ; 1598 for (;;) { 1599 ObjectWaiter * front = _cxq; 1600 iterator->_next = front; 1601 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1602 break; 1603 } 1604 } 1605 } 1606 1607 // _WaitSetLock protects the wait queue, not the EntryList. We could 1608 // move the add-to-EntryList operation, above, outside the critical section 1609 // protected by _WaitSetLock. In practice that's not useful. With the 1610 // exception of wait() timeouts and interrupts the monitor owner 1611 // is the only thread that grabs _WaitSetLock. There's almost no contention 1612 // on _WaitSetLock so it's not profitable to reduce the length of the 1613 // critical section. 1614 1615 iterator->wait_reenter_begin(this); 1616 } 1617 Thread::SpinRelease(&_WaitSetLock); 1618 } 1619 1620 // Consider: a not-uncommon synchronization bug is to use notify() when 1621 // notifyAll() is more appropriate, potentially resulting in stranded 1622 // threads; this is one example of a lost wakeup. A useful diagnostic 1623 // option is to force all notify() operations to behave as notifyAll(). 1624 // 1625 // Note: We can also detect many such problems with a "minimum wait". 1626 // When the "minimum wait" is set to a small non-zero timeout value 1627 // and the program does not hang whereas it did absent "minimum wait", 1628 // that suggests a lost wakeup bug. 1629 1630 void ObjectMonitor::notify(TRAPS) { 1631 CHECK_OWNER(); // Throws IMSE if not owner. 1632 if (_WaitSet == NULL) { 1633 return; 1634 } 1635 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1636 INotify(THREAD); 1637 OM_PERFDATA_OP(Notifications, inc(1)); 1638 } 1639 1640 1641 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1642 // from the waitset to the EntryList. This could be done more efficiently with a 1643 // single bulk transfer but in practice it's not time-critical. Beware too, 1644 // that in prepend-mode we invert the order of the waiters. Let's say that the 1645 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1646 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1647 1648 void ObjectMonitor::notifyAll(TRAPS) { 1649 CHECK_OWNER(); // Throws IMSE if not owner. 1650 if (_WaitSet == NULL) { 1651 return; 1652 } 1653 1654 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1655 int tally = 0; 1656 while (_WaitSet != NULL) { 1657 tally++; 1658 INotify(THREAD); 1659 } 1660 1661 OM_PERFDATA_OP(Notifications, inc(tally)); 1662 } 1663 1664 // ----------------------------------------------------------------------------- 1665 // Adaptive Spinning Support 1666 // 1667 // Adaptive spin-then-block - rational spinning 1668 // 1669 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1670 // algorithm. On high order SMP systems it would be better to start with 1671 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1672 // a contending thread could enqueue itself on the cxq and then spin locally 1673 // on a thread-specific variable such as its ParkEvent._Event flag. 1674 // That's left as an exercise for the reader. Note that global spinning is 1675 // not problematic on Niagara, as the L2 cache serves the interconnect and 1676 // has both low latency and massive bandwidth. 1677 // 1678 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1679 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1680 // (duration) or we can fix the count at approximately the duration of 1681 // a context switch and vary the frequency. Of course we could also 1682 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1683 // For a description of 'Adaptive spin-then-block mutual exclusion in 1684 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1685 // 1686 // This implementation varies the duration "D", where D varies with 1687 // the success rate of recent spin attempts. (D is capped at approximately 1688 // length of a round-trip context switch). The success rate for recent 1689 // spin attempts is a good predictor of the success rate of future spin 1690 // attempts. The mechanism adapts automatically to varying critical 1691 // section length (lock modality), system load and degree of parallelism. 1692 // D is maintained per-monitor in _SpinDuration and is initialized 1693 // optimistically. Spin frequency is fixed at 100%. 1694 // 1695 // Note that _SpinDuration is volatile, but we update it without locks 1696 // or atomics. The code is designed so that _SpinDuration stays within 1697 // a reasonable range even in the presence of races. The arithmetic 1698 // operations on _SpinDuration are closed over the domain of legal values, 1699 // so at worst a race will install and older but still legal value. 1700 // At the very worst this introduces some apparent non-determinism. 1701 // We might spin when we shouldn't or vice-versa, but since the spin 1702 // count are relatively short, even in the worst case, the effect is harmless. 1703 // 1704 // Care must be taken that a low "D" value does not become an 1705 // an absorbing state. Transient spinning failures -- when spinning 1706 // is overall profitable -- should not cause the system to converge 1707 // on low "D" values. We want spinning to be stable and predictable 1708 // and fairly responsive to change and at the same time we don't want 1709 // it to oscillate, become metastable, be "too" non-deterministic, 1710 // or converge on or enter undesirable stable absorbing states. 1711 // 1712 // We implement a feedback-based control system -- using past behavior 1713 // to predict future behavior. We face two issues: (a) if the 1714 // input signal is random then the spin predictor won't provide optimal 1715 // results, and (b) if the signal frequency is too high then the control 1716 // system, which has some natural response lag, will "chase" the signal. 1717 // (b) can arise from multimodal lock hold times. Transient preemption 1718 // can also result in apparent bimodal lock hold times. 1719 // Although sub-optimal, neither condition is particularly harmful, as 1720 // in the worst-case we'll spin when we shouldn't or vice-versa. 1721 // The maximum spin duration is rather short so the failure modes aren't bad. 1722 // To be conservative, I've tuned the gain in system to bias toward 1723 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1724 // "rings" or oscillates between spinning and not spinning. This happens 1725 // when spinning is just on the cusp of profitability, however, so the 1726 // situation is not dire. The state is benign -- there's no need to add 1727 // hysteresis control to damp the transition rate between spinning and 1728 // not spinning. 1729 1730 // Spinning: Fixed frequency (100%), vary duration 1731 int ObjectMonitor::TrySpin(Thread * Self) { 1732 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1733 int ctr = Knob_FixedSpin; 1734 if (ctr != 0) { 1735 while (--ctr >= 0) { 1736 if (TryLock(Self) > 0) return 1; 1737 SpinPause(); 1738 } 1739 return 0; 1740 } 1741 1742 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1743 if (TryLock(Self) > 0) { 1744 // Increase _SpinDuration ... 1745 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1746 // Raising _SpurDuration to the poverty line is key. 1747 int x = _SpinDuration; 1748 if (x < Knob_SpinLimit) { 1749 if (x < Knob_Poverty) x = Knob_Poverty; 1750 _SpinDuration = x + Knob_BonusB; 1751 } 1752 return 1; 1753 } 1754 SpinPause(); 1755 } 1756 1757 // Admission control - verify preconditions for spinning 1758 // 1759 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1760 // becoming an absorbing state. Put another way, we spin briefly to 1761 // sample, just in case the system load, parallelism, contention, or lock 1762 // modality changed. 1763 // 1764 // Consider the following alternative: 1765 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1766 // spin attempt. "Periodically" might mean after a tally of 1767 // the # of failed spin attempts (or iterations) reaches some threshold. 1768 // This takes us into the realm of 1-out-of-N spinning, where we 1769 // hold the duration constant but vary the frequency. 1770 1771 ctr = _SpinDuration; 1772 if (ctr <= 0) return 0; 1773 1774 if (NotRunnable(Self, (Thread *) _owner)) { 1775 return 0; 1776 } 1777 1778 // We're good to spin ... spin ingress. 1779 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1780 // when preparing to LD...CAS _owner, etc and the CAS is likely 1781 // to succeed. 1782 if (_succ == NULL) { 1783 _succ = Self; 1784 } 1785 Thread * prv = NULL; 1786 1787 // There are three ways to exit the following loop: 1788 // 1. A successful spin where this thread has acquired the lock. 1789 // 2. Spin failure with prejudice 1790 // 3. Spin failure without prejudice 1791 1792 while (--ctr >= 0) { 1793 1794 // Periodic polling -- Check for pending GC 1795 // Threads may spin while they're unsafe. 1796 // We don't want spinning threads to delay the JVM from reaching 1797 // a stop-the-world safepoint or to steal cycles from GC. 1798 // If we detect a pending safepoint we abort in order that 1799 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1800 // this thread, if safe, doesn't steal cycles from GC. 1801 // This is in keeping with the "no loitering in runtime" rule. 1802 // We periodically check to see if there's a safepoint pending. 1803 if ((ctr & 0xFF) == 0) { 1804 if (SafepointMechanism::should_block(Self)) { 1805 goto Abort; // abrupt spin egress 1806 } 1807 SpinPause(); 1808 } 1809 1810 // Probe _owner with TATAS 1811 // If this thread observes the monitor transition or flicker 1812 // from locked to unlocked to locked, then the odds that this 1813 // thread will acquire the lock in this spin attempt go down 1814 // considerably. The same argument applies if the CAS fails 1815 // or if we observe _owner change from one non-null value to 1816 // another non-null value. In such cases we might abort 1817 // the spin without prejudice or apply a "penalty" to the 1818 // spin count-down variable "ctr", reducing it by 100, say. 1819 1820 Thread * ox = (Thread *) _owner; 1821 if (ox == NULL) { 1822 ox = (Thread*)try_set_owner_from(Self, NULL); 1823 if (ox == NULL) { 1824 // The CAS succeeded -- this thread acquired ownership 1825 // Take care of some bookkeeping to exit spin state. 1826 if (_succ == Self) { 1827 _succ = NULL; 1828 } 1829 1830 // Increase _SpinDuration : 1831 // The spin was successful (profitable) so we tend toward 1832 // longer spin attempts in the future. 1833 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1834 // If we acquired the lock early in the spin cycle it 1835 // makes sense to increase _SpinDuration proportionally. 1836 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1837 int x = _SpinDuration; 1838 if (x < Knob_SpinLimit) { 1839 if (x < Knob_Poverty) x = Knob_Poverty; 1840 _SpinDuration = x + Knob_Bonus; 1841 } 1842 return 1; 1843 } 1844 1845 // The CAS failed ... we can take any of the following actions: 1846 // * penalize: ctr -= CASPenalty 1847 // * exit spin with prejudice -- goto Abort; 1848 // * exit spin without prejudice. 1849 // * Since CAS is high-latency, retry again immediately. 1850 prv = ox; 1851 goto Abort; 1852 } 1853 1854 // Did lock ownership change hands ? 1855 if (ox != prv && prv != NULL) { 1856 goto Abort; 1857 } 1858 prv = ox; 1859 1860 // Abort the spin if the owner is not executing. 1861 // The owner must be executing in order to drop the lock. 1862 // Spinning while the owner is OFFPROC is idiocy. 1863 // Consider: ctr -= RunnablePenalty ; 1864 if (NotRunnable(Self, ox)) { 1865 goto Abort; 1866 } 1867 if (_succ == NULL) { 1868 _succ = Self; 1869 } 1870 } 1871 1872 // Spin failed with prejudice -- reduce _SpinDuration. 1873 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1874 // AIMD is globally stable. 1875 { 1876 int x = _SpinDuration; 1877 if (x > 0) { 1878 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1879 // This is globally sample and tends to damp the response. 1880 x -= Knob_Penalty; 1881 if (x < 0) x = 0; 1882 _SpinDuration = x; 1883 } 1884 } 1885 1886 Abort: 1887 if (_succ == Self) { 1888 _succ = NULL; 1889 // Invariant: after setting succ=null a contending thread 1890 // must recheck-retry _owner before parking. This usually happens 1891 // in the normal usage of TrySpin(), but it's safest 1892 // to make TrySpin() as foolproof as possible. 1893 OrderAccess::fence(); 1894 if (TryLock(Self) > 0) return 1; 1895 } 1896 return 0; 1897 } 1898 1899 // NotRunnable() -- informed spinning 1900 // 1901 // Don't bother spinning if the owner is not eligible to drop the lock. 1902 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1903 // The thread must be runnable in order to drop the lock in timely fashion. 1904 // If the _owner is not runnable then spinning will not likely be 1905 // successful (profitable). 1906 // 1907 // Beware -- the thread referenced by _owner could have died 1908 // so a simply fetch from _owner->_thread_state might trap. 1909 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1910 // Because of the lifecycle issues, the _thread_state values 1911 // observed by NotRunnable() might be garbage. NotRunnable must 1912 // tolerate this and consider the observed _thread_state value 1913 // as advisory. 1914 // 1915 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1916 // a thread pointer. 1917 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1918 // with the LSB of _owner. Another option would be to probabilistically probe 1919 // the putative _owner->TypeTag value. 1920 // 1921 // Checking _thread_state isn't perfect. Even if the thread is 1922 // in_java it might be blocked on a page-fault or have been preempted 1923 // and sitting on a ready/dispatch queue. 1924 // 1925 // The return value from NotRunnable() is *advisory* -- the 1926 // result is based on sampling and is not necessarily coherent. 1927 // The caller must tolerate false-negative and false-positive errors. 1928 // Spinning, in general, is probabilistic anyway. 1929 1930 1931 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1932 // Check ox->TypeTag == 2BAD. 1933 if (ox == NULL) return 0; 1934 1935 // Avoid transitive spinning ... 1936 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1937 // Immediately after T1 acquires L it's possible that T2, also 1938 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1939 // This occurs transiently after T1 acquired L but before 1940 // T1 managed to clear T1.Stalled. T2 does not need to abort 1941 // its spin in this circumstance. 1942 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1943 1944 if (BlockedOn == 1) return 1; 1945 if (BlockedOn != 0) { 1946 return BlockedOn != intptr_t(this) && _owner == ox; 1947 } 1948 1949 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1950 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1951 // consider also: jst != _thread_in_Java -- but that's overspecific. 1952 return jst == _thread_blocked || jst == _thread_in_native; 1953 } 1954 1955 1956 // ----------------------------------------------------------------------------- 1957 // WaitSet management ... 1958 1959 ObjectWaiter::ObjectWaiter(Thread* thread) { 1960 _next = NULL; 1961 _prev = NULL; 1962 _notified = 0; 1963 _notifier_tid = 0; 1964 TState = TS_RUN; 1965 _thread = thread; 1966 _event = thread->_ParkEvent; 1967 _active = false; 1968 assert(_event != NULL, "invariant"); 1969 } 1970 1971 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1972 JavaThread *jt = (JavaThread *)this->_thread; 1973 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1974 } 1975 1976 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1977 JavaThread *jt = (JavaThread *)this->_thread; 1978 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1979 } 1980 1981 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1982 assert(node != NULL, "should not add NULL node"); 1983 assert(node->_prev == NULL, "node already in list"); 1984 assert(node->_next == NULL, "node already in list"); 1985 // put node at end of queue (circular doubly linked list) 1986 if (_WaitSet == NULL) { 1987 _WaitSet = node; 1988 node->_prev = node; 1989 node->_next = node; 1990 } else { 1991 ObjectWaiter* head = _WaitSet; 1992 ObjectWaiter* tail = head->_prev; 1993 assert(tail->_next == head, "invariant check"); 1994 tail->_next = node; 1995 head->_prev = node; 1996 node->_next = head; 1997 node->_prev = tail; 1998 } 1999 } 2000 2001 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2002 // dequeue the very first waiter 2003 ObjectWaiter* waiter = _WaitSet; 2004 if (waiter) { 2005 DequeueSpecificWaiter(waiter); 2006 } 2007 return waiter; 2008 } 2009 2010 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2011 assert(node != NULL, "should not dequeue NULL node"); 2012 assert(node->_prev != NULL, "node already removed from list"); 2013 assert(node->_next != NULL, "node already removed from list"); 2014 // when the waiter has woken up because of interrupt, 2015 // timeout or other spurious wake-up, dequeue the 2016 // waiter from waiting list 2017 ObjectWaiter* next = node->_next; 2018 if (next == node) { 2019 assert(node->_prev == node, "invariant check"); 2020 _WaitSet = NULL; 2021 } else { 2022 ObjectWaiter* prev = node->_prev; 2023 assert(prev->_next == node, "invariant check"); 2024 assert(next->_prev == node, "invariant check"); 2025 next->_prev = prev; 2026 prev->_next = next; 2027 if (_WaitSet == node) { 2028 _WaitSet = next; 2029 } 2030 } 2031 node->_next = NULL; 2032 node->_prev = NULL; 2033 } 2034 2035 // ----------------------------------------------------------------------------- 2036 // PerfData support 2037 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2038 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2039 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2040 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2041 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2042 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2043 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2044 2045 // One-shot global initialization for the sync subsystem. 2046 // We could also defer initialization and initialize on-demand 2047 // the first time we call ObjectSynchronizer::inflate(). 2048 // Initialization would be protected - like so many things - by 2049 // the MonitorCache_lock. 2050 2051 void ObjectMonitor::Initialize() { 2052 assert(!InitDone, "invariant"); 2053 2054 if (!os::is_MP()) { 2055 Knob_SpinLimit = 0; 2056 Knob_PreSpin = 0; 2057 Knob_FixedSpin = -1; 2058 } 2059 2060 if (UsePerfData) { 2061 EXCEPTION_MARK; 2062 #define NEWPERFCOUNTER(n) \ 2063 { \ 2064 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2065 CHECK); \ 2066 } 2067 #define NEWPERFVARIABLE(n) \ 2068 { \ 2069 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2070 CHECK); \ 2071 } 2072 NEWPERFCOUNTER(_sync_Inflations); 2073 NEWPERFCOUNTER(_sync_Deflations); 2074 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2075 NEWPERFCOUNTER(_sync_FutileWakeups); 2076 NEWPERFCOUNTER(_sync_Parks); 2077 NEWPERFCOUNTER(_sync_Notifications); 2078 NEWPERFVARIABLE(_sync_MonExtant); 2079 #undef NEWPERFCOUNTER 2080 #undef NEWPERFVARIABLE 2081 } 2082 2083 DEBUG_ONLY(InitDone = true;) 2084 } 2085 2086 ObjectMonitorHandle::~ObjectMonitorHandle() { 2087 if (_om_ptr != NULL) { 2088 _om_ptr->dec_ref_count(); 2089 _om_ptr = NULL; 2090 } 2091 } 2092 2093 // Save the ObjectMonitor* associated with the specified markWord and 2094 // increment the ref_count. This function should only be called if 2095 // the caller has verified mark.has_monitor() == true. The object 2096 // parameter is needed to verify that ObjectMonitor* has not been 2097 // deflated and reused for another object. 2098 // 2099 // This function returns true if the ObjectMonitor* has been safely 2100 // saved. This function returns false if we have lost a race with 2101 // async deflation; the caller should retry as appropriate. 2102 // 2103 bool ObjectMonitorHandle::save_om_ptr(oop object, markWord mark) { 2104 guarantee(mark.has_monitor(), "sanity check: mark=" INTPTR_FORMAT, 2105 mark.value()); 2106 2107 ObjectMonitor* om_ptr = mark.monitor(); 2108 om_ptr->inc_ref_count(); 2109 2110 if (AsyncDeflateIdleMonitors) { 2111 // Race here if monitor is not owned! The above ref_count bump 2112 // will cause subsequent async deflation to skip it. However, 2113 // previous or concurrent async deflation is a race. 2114 if (om_ptr->owner_is_DEFLATER_MARKER() && om_ptr->ref_count() <= 0) { 2115 // Async deflation is in progress and our ref_count increment 2116 // above lost the race to async deflation. Attempt to restore 2117 // the header/dmw to the object's header so that we only retry 2118 // once if the deflater thread happens to be slow. 2119 om_ptr->install_displaced_markword_in_object(object); 2120 om_ptr->dec_ref_count(); 2121 return false; 2122 } 2123 if (om_ptr->ref_count() <= 0) { 2124 // Async deflation is in the process of bailing out, but has not 2125 // yet restored the ref_count field so we return false to force 2126 // a retry. We want a positive ref_count value for a true return. 2127 om_ptr->dec_ref_count(); 2128 return false; 2129 } 2130 // The ObjectMonitor could have been deflated and reused for 2131 // another object before we bumped the ref_count so make sure 2132 // our object still refers to this ObjectMonitor. 2133 const markWord tmp = object->mark(); 2134 if (!tmp.has_monitor() || tmp.monitor() != om_ptr) { 2135 // Async deflation and reuse won the race so we have to retry. 2136 // Skip object header restoration since that's already done. 2137 om_ptr->dec_ref_count(); 2138 return false; 2139 } 2140 } 2141 2142 ADIM_guarantee(_om_ptr == NULL, "sanity check: _om_ptr=" INTPTR_FORMAT, 2143 p2i(_om_ptr)); 2144 _om_ptr = om_ptr; 2145 return true; 2146 } 2147 2148 // For internal use by ObjectSynchronizer::inflate(). 2149 // This function is only used when we don't have to worry about async 2150 // deflation of the specified ObjectMonitor*. 2151 void ObjectMonitorHandle::set_om_ptr(ObjectMonitor* om_ptr) { 2152 if (_om_ptr == NULL) { 2153 ADIM_guarantee(om_ptr != NULL, "cannot clear an unset om_ptr"); 2154 om_ptr->inc_ref_count(); 2155 _om_ptr = om_ptr; 2156 } else { 2157 ADIM_guarantee(om_ptr == NULL, "can only clear a set om_ptr"); 2158 _om_ptr->dec_ref_count(); 2159 _om_ptr = NULL; 2160 } 2161 } 2162 2163 // Save the specified ObjectMonitor* if it is safe, i.e., not being 2164 // async deflated. 2165 // 2166 // This function returns true if the ObjectMonitor* has been safely 2167 // saved. This function returns false if the specified ObjectMonitor* 2168 // is NULL or if we have lost a race with async deflation; the caller 2169 // can retry as appropriate. 2170 bool ObjectMonitorHandle::set_om_ptr_if_safe(ObjectMonitor* om_ptr) { 2171 if (om_ptr == NULL) { 2172 return false; // Nothing to save if input is NULL 2173 } 2174 2175 om_ptr->inc_ref_count(); 2176 2177 if (AsyncDeflateIdleMonitors) { 2178 if (om_ptr->owner_is_DEFLATER_MARKER() && om_ptr->ref_count() <= 0) { 2179 // Async deflation is in progress and our ref_count increment 2180 // above lost the race to async deflation. 2181 om_ptr->dec_ref_count(); 2182 return false; 2183 } 2184 if (om_ptr->ref_count() <= 0) { 2185 // Async deflation is in the process of bailing out, but has not 2186 // yet restored the ref_count field so we return false to force 2187 // a retry. We want a positive ref_count value for a true return. 2188 om_ptr->dec_ref_count(); 2189 return false; 2190 } 2191 // Unlike save_om_ptr(), we don't have context to determine if 2192 // the ObjectMonitor has been deflated and reused for another 2193 // object. 2194 } 2195 2196 ADIM_guarantee(_om_ptr == NULL, "sanity check: _om_ptr=" INTPTR_FORMAT, 2197 p2i(_om_ptr)); 2198 _om_ptr = om_ptr; 2199 return true; 2200 } 2201 2202 // Unset the _om_ptr field and decrement the ref_count field. 2203 void ObjectMonitorHandle::unset_om_ptr() { 2204 ADIM_guarantee(_om_ptr != NULL, "_om_ptr must not be NULL"); 2205 _om_ptr->dec_ref_count(); 2206 _om_ptr = NULL; 2207 } 2208 2209 void ObjectMonitor::print_on(outputStream* st) const { 2210 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2211 st->print("{contentions=0x%08x,waiters=0x%08x" 2212 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2213 contentions(), waiters(), recursions(), 2214 p2i(owner())); 2215 } 2216 void ObjectMonitor::print() const { print_on(tty); } 2217 2218 //#ifdef ASSERT 2219 // Print the ObjectMonitor like a debugger would: 2220 // 2221 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2222 // _header = 0x0000000000000001 2223 // _object = 0x000000070ff45fd0 2224 // _allocation_state = Old 2225 // _pad_buf0 = { 2226 // [0] = '\0' 2227 // ... 2228 // [43] = '\0' 2229 // } 2230 // _owner = 0x0000000000000000 2231 // _previous_owner_tid = 0 2232 // _pad_buf1 = { 2233 // [0] = '\0' 2234 // ... 2235 // [47] = '\0' 2236 // } 2237 // _ref_count = 1 2238 // _pad_buf2 = { 2239 // [0] = '\0' 2240 // ... 2241 // [47] = '\0' 2242 // } 2243 // _next_om = 0x0000000000000000 2244 // _recursions = 0 2245 // _EntryList = 0x0000000000000000 2246 // _cxq = 0x0000000000000000 2247 // _succ = 0x0000000000000000 2248 // _Responsible = 0x0000000000000000 2249 // _Spinner = 0 2250 // _SpinDuration = 5000 2251 // _contentions = 0 2252 // _WaitSet = 0x0000700009756248 2253 // _waiters = 1 2254 // _WaitSetLock = 0 2255 // } 2256 // 2257 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2258 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2259 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2260 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 2261 st->print(" _allocation_state = "); 2262 if (is_free()) { 2263 st->print("Free"); 2264 } else if (is_old()) { 2265 st->print("Old"); 2266 } else if (is_new()) { 2267 st->print("New"); 2268 } else { 2269 st->print("unknown=%d", _allocation_state); 2270 } 2271 st->cr(); 2272 st->print_cr(" _pad_buf0 = {"); 2273 st->print_cr(" [0] = '\\0'"); 2274 st->print_cr(" ..."); 2275 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2276 st->print_cr(" }"); 2277 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 2278 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 2279 st->print_cr(" _pad_buf1 = {"); 2280 st->print_cr(" [0] = '\\0'"); 2281 st->print_cr(" ..."); 2282 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2283 st->print_cr(" }"); 2284 st->print_cr(" _ref_count = %d", ref_count()); 2285 st->print_cr(" _pad_buf2 = {"); 2286 st->print_cr(" [0] = '\\0'"); 2287 st->print_cr(" ..."); 2288 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2289 st->print_cr(" }"); 2290 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(_next_om)); 2291 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2292 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2293 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2294 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2295 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2296 st->print_cr(" _Spinner = %d", _Spinner); 2297 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2298 st->print_cr(" _contentions = %d", _contentions); 2299 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2300 st->print_cr(" _waiters = %d", _waiters); 2301 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2302 st->print_cr("}"); 2303 } 2304 //#endif