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