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