1 /* 2 * Copyright (c) 1998, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "jfr/support/jfrThreadId.hpp" 29 #include "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "oops/markWord.hpp" 34 #include "oops/oop.inline.hpp" 35 #include "runtime/atomic.hpp" 36 #include "runtime/handles.inline.hpp" 37 #include "runtime/interfaceSupport.inline.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "runtime/objectMonitor.hpp" 40 #include "runtime/objectMonitor.inline.hpp" 41 #include "runtime/orderAccess.hpp" 42 #include "runtime/osThread.hpp" 43 #include "runtime/safepointMechanism.inline.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/stubRoutines.hpp" 46 #include "runtime/thread.inline.hpp" 47 #include "services/threadService.hpp" 48 #include "utilities/dtrace.hpp" 49 #include "utilities/macros.hpp" 50 #include "utilities/preserveException.hpp" 51 #if INCLUDE_JFR 52 #include "jfr/support/jfrFlush.hpp" 53 #endif 54 55 #ifdef DTRACE_ENABLED 56 57 // Only bother with this argument setup if dtrace is available 58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95 #else // ndef DTRACE_ENABLED 96 97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100 #endif // ndef DTRACE_ENABLED 101 102 // Tunables ... 103 // The knob* variables are effectively final. Once set they should 104 // never be modified hence. Consider using __read_mostly with GCC. 105 106 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 107 108 static int Knob_Bonus = 100; // spin success bonus 109 static int Knob_BonusB = 100; // spin success bonus 110 static int Knob_Penalty = 200; // spin failure penalty 111 static int Knob_Poverty = 1000; 112 static int Knob_FixedSpin = 0; 113 static int Knob_PreSpin = 10; // 20-100 likely better 114 115 DEBUG_ONLY(static volatile bool InitDone = false;) 116 117 // ----------------------------------------------------------------------------- 118 // Theory of operations -- Monitors lists, thread residency, etc: 119 // 120 // * A thread acquires ownership of a monitor by successfully 121 // CAS()ing the _owner field from null to non-null. 122 // 123 // * Invariant: A thread appears on at most one monitor list -- 124 // cxq, EntryList or WaitSet -- at any one time. 125 // 126 // * Contending threads "push" themselves onto the cxq with CAS 127 // and then spin/park. 128 // 129 // * After a contending thread eventually acquires the lock it must 130 // dequeue itself from either the EntryList or the cxq. 131 // 132 // * The exiting thread identifies and unparks an "heir presumptive" 133 // tentative successor thread on the EntryList. Critically, the 134 // exiting thread doesn't unlink the successor thread from the EntryList. 135 // After having been unparked, the wakee will recontend for ownership of 136 // the monitor. The successor (wakee) will either acquire the lock or 137 // re-park itself. 138 // 139 // Succession is provided for by a policy of competitive handoff. 140 // The exiting thread does _not_ grant or pass ownership to the 141 // successor thread. (This is also referred to as "handoff" succession"). 142 // Instead the exiting thread releases ownership and possibly wakes 143 // a successor, so the successor can (re)compete for ownership of the lock. 144 // If the EntryList is empty but the cxq is populated the exiting 145 // thread will drain the cxq into the EntryList. It does so by 146 // by detaching the cxq (installing null with CAS) and folding 147 // the threads from the cxq into the EntryList. The EntryList is 148 // doubly linked, while the cxq is singly linked because of the 149 // CAS-based "push" used to enqueue recently arrived threads (RATs). 150 // 151 // * Concurrency invariants: 152 // 153 // -- only the monitor owner may access or mutate the EntryList. 154 // The mutex property of the monitor itself protects the EntryList 155 // from concurrent interference. 156 // -- Only the monitor owner may detach the cxq. 157 // 158 // * The monitor entry list operations avoid locks, but strictly speaking 159 // they're not lock-free. Enter is lock-free, exit is not. 160 // For a description of 'Methods and apparatus providing non-blocking access 161 // to a resource,' see U.S. Pat. No. 7844973. 162 // 163 // * The cxq can have multiple concurrent "pushers" but only one concurrent 164 // detaching thread. This mechanism is immune from the ABA corruption. 165 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 166 // 167 // * Taken together, the cxq and the EntryList constitute or form a 168 // single logical queue of threads stalled trying to acquire the lock. 169 // We use two distinct lists to improve the odds of a constant-time 170 // dequeue operation after acquisition (in the ::enter() epilogue) and 171 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 172 // A key desideratum is to minimize queue & monitor metadata manipulation 173 // that occurs while holding the monitor lock -- that is, we want to 174 // minimize monitor lock holds times. Note that even a small amount of 175 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 176 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 177 // locks and monitor metadata. 178 // 179 // Cxq points to the set of Recently Arrived Threads attempting entry. 180 // Because we push threads onto _cxq with CAS, the RATs must take the form of 181 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 182 // the unlocking thread notices that EntryList is null but _cxq is != null. 183 // 184 // The EntryList is ordered by the prevailing queue discipline and 185 // can be organized in any convenient fashion, such as a doubly-linked list or 186 // a circular doubly-linked list. Critically, we want insert and delete operations 187 // to operate in constant-time. If we need a priority queue then something akin 188 // to Solaris' sleepq would work nicely. Viz., 189 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 190 // Queue discipline is enforced at ::exit() time, when the unlocking thread 191 // drains the cxq into the EntryList, and orders or reorders the threads on the 192 // EntryList accordingly. 193 // 194 // Barring "lock barging", this mechanism provides fair cyclic ordering, 195 // somewhat similar to an elevator-scan. 196 // 197 // * The monitor synchronization subsystem avoids the use of native 198 // synchronization primitives except for the narrow platform-specific 199 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 200 // the semantics of park-unpark. Put another way, this monitor implementation 201 // depends only on atomic operations and park-unpark. The monitor subsystem 202 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 203 // underlying OS manages the READY<->RUN transitions. 204 // 205 // * Waiting threads reside on the WaitSet list -- wait() puts 206 // the caller onto the WaitSet. 207 // 208 // * notify() or notifyAll() simply transfers threads from the WaitSet to 209 // either the EntryList or cxq. Subsequent exit() operations will 210 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 211 // it's likely the notifyee would simply impale itself on the lock held 212 // by the notifier. 213 // 214 // * An interesting alternative is to encode cxq as (List,LockByte) where 215 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 216 // variable, like _recursions, in the scheme. The threads or Events that form 217 // the list would have to be aligned in 256-byte addresses. A thread would 218 // try to acquire the lock or enqueue itself with CAS, but exiting threads 219 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 220 // Note that is is *not* word-tearing, but it does presume that full-word 221 // CAS operations are coherent with intermix with STB operations. That's true 222 // on most common processors. 223 // 224 // * See also http://blogs.sun.com/dave 225 226 227 void* ObjectMonitor::operator new (size_t size) throw() { 228 return AllocateHeap(size, mtInternal); 229 } 230 void* ObjectMonitor::operator new[] (size_t size) throw() { 231 return operator new (size); 232 } 233 void ObjectMonitor::operator delete(void* p) { 234 FreeHeap(p); 235 } 236 void ObjectMonitor::operator delete[] (void *p) { 237 operator delete(p); 238 } 239 240 // ----------------------------------------------------------------------------- 241 // Enter support 242 243 bool ObjectMonitor::enter(TRAPS) { 244 // The following code is ordered to check the most common cases first 245 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 246 Thread * const Self = THREAD; 247 248 void* cur = try_set_owner_from(NULL, Self); 249 if (cur == NULL) { 250 assert(_recursions == 0, "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 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 264 return true; 265 } 266 267 // We've encountered genuine contention. 268 assert(Self->_Stalled == 0, "invariant"); 269 Self->_Stalled = intptr_t(this); 270 271 // Try one round of spinning *before* enqueueing Self 272 // and before going through the awkward and expensive state 273 // transitions. The following spin is strictly optional ... 274 // Note that if we acquire the monitor from an initial spin 275 // we forgo posting JVMTI events and firing DTRACE probes. 276 if (TrySpin(Self) > 0) { 277 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 278 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 279 assert(((oop)object())->mark() == markWord::encode(this), 280 "object mark must match encoded this: mark=" INTPTR_FORMAT 281 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 282 markWord::encode(this).value()); 283 Self->_Stalled = 0; 284 return true; 285 } 286 287 assert(_owner != Self, "invariant"); 288 assert(_succ != Self, "invariant"); 289 assert(Self->is_Java_thread(), "invariant"); 290 JavaThread * jt = (JavaThread *) Self; 291 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 292 assert(jt->thread_state() != _thread_blocked, "invariant"); 293 assert(AsyncDeflateIdleMonitors || this->object() != NULL, "invariant"); 294 assert(AsyncDeflateIdleMonitors || contentions() >= 0, "must not be negative: contentions=%d", contentions()); 295 296 // Keep track of contention for JVM/TI and M&M queries. 297 Atomic::inc(&_contentions); 298 if (AsyncDeflateIdleMonitors && is_being_async_deflated()) { 299 // Async deflation is in progress and our contentions increment 300 // above lost the race to async deflation. Undo the work and 301 // force the caller to retry. 302 const oop l_object = (oop)object(); 303 if (l_object != NULL) { 304 // Attempt to restore the header/dmw to the object's header so that 305 // we only retry once if the deflater thread happens to be slow. 306 install_displaced_markword_in_object(l_object); 307 } 308 Self->_Stalled = 0; 309 Atomic::dec(&_contentions); 310 return false; 311 } 312 313 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 314 EventJavaMonitorEnter event; 315 if (event.should_commit()) { 316 event.set_monitorClass(((oop)this->object())->klass()); 317 event.set_address((uintptr_t)(this->object_addr())); 318 } 319 320 { // Change java thread status to indicate blocked on monitor enter. 321 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 322 323 Self->set_current_pending_monitor(this); 324 325 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 326 if (JvmtiExport::should_post_monitor_contended_enter()) { 327 JvmtiExport::post_monitor_contended_enter(jt, this); 328 329 // The current thread does not yet own the monitor and does not 330 // yet appear on any queues that would get it made the successor. 331 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 332 // handler cannot accidentally consume an unpark() meant for the 333 // ParkEvent associated with this ObjectMonitor. 334 } 335 336 OSThreadContendState osts(Self->osthread()); 337 ThreadBlockInVM tbivm(jt); 338 339 // TODO-FIXME: change the following for(;;) loop to straight-line code. 340 for (;;) { 341 jt->set_suspend_equivalent(); 342 // cleared by handle_special_suspend_equivalent_condition() 343 // or java_suspend_self() 344 345 EnterI(THREAD); 346 347 if (!ExitSuspendEquivalent(jt)) break; 348 349 // We have acquired the contended monitor, but while we were 350 // waiting another thread suspended us. We don't want to enter 351 // the monitor while suspended because that would surprise the 352 // thread that suspended us. 353 // 354 _recursions = 0; 355 _succ = NULL; 356 exit(false, Self); 357 358 jt->java_suspend_self(); 359 } 360 Self->set_current_pending_monitor(NULL); 361 362 // We cleared the pending monitor info since we've just gotten past 363 // the enter-check-for-suspend dance and we now own the monitor free 364 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 365 // destructor can go to a safepoint at the end of this block. If we 366 // do a thread dump during that safepoint, then this thread will show 367 // as having "-locked" the monitor, but the OS and java.lang.Thread 368 // states will still report that the thread is blocked trying to 369 // acquire it. 370 } 371 372 Atomic::dec(&_contentions); 373 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions()); 374 Self->_Stalled = 0; 375 376 // Must either set _recursions = 0 or ASSERT _recursions == 0. 377 assert(_recursions == 0, "invariant"); 378 assert(_owner == Self, "invariant"); 379 assert(_succ != Self, "invariant"); 380 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 381 382 // The thread -- now the owner -- is back in vm mode. 383 // Report the glorious news via TI,DTrace and jvmstat. 384 // The probe effect is non-trivial. All the reportage occurs 385 // while we hold the monitor, increasing the length of the critical 386 // section. Amdahl's parallel speedup law comes vividly into play. 387 // 388 // Another option might be to aggregate the events (thread local or 389 // per-monitor aggregation) and defer reporting until a more opportune 390 // time -- such as next time some thread encounters contention but has 391 // yet to acquire the lock. While spinning that thread could 392 // spinning we could increment JVMStat counters, etc. 393 394 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 395 if (JvmtiExport::should_post_monitor_contended_entered()) { 396 JvmtiExport::post_monitor_contended_entered(jt, this); 397 398 // The current thread already owns the monitor and is not going to 399 // call park() for the remainder of the monitor enter protocol. So 400 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 401 // event handler consumed an unpark() issued by the thread that 402 // just exited the monitor. 403 } 404 if (event.should_commit()) { 405 event.set_previousOwner((uintptr_t)_previous_owner_tid); 406 event.commit(); 407 } 408 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 409 return true; 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 (try_set_owner_from(NULL, Self) == NULL) { 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 // && contentions <= 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"); 441 442 const oop l_object = (oop)object(); 443 if (l_object == NULL) { 444 // ObjectMonitor's object ref has already been cleared by async 445 // deflation so we're done here. 446 return; 447 } 448 ADIM_guarantee(l_object == obj, "object=" INTPTR_FORMAT " must equal obj=" 449 INTPTR_FORMAT, p2i(l_object), p2i(obj)); 450 451 markWord dmw = header(); 452 // The dmw has to be neutral (not NULL, not locked and not marked). 453 ADIM_guarantee(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 454 455 // Install displaced mark word if the object's header still points 456 // to this ObjectMonitor. All racing callers to this function will 457 // reach this point, but only one can win. 458 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 459 if (res != markWord::encode(this)) { 460 // This should be rare so log at the Info level when it happens. 461 log_info(monitorinflation)("install_displaced_markword_in_object: " 462 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 463 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 464 dmw.value(), markWord::encode(this).value(), 465 res.value()); 466 } 467 468 // Note: It does not matter which thread restored the header/dmw 469 // into the object's header. The thread deflating the monitor just 470 // wanted the object's header restored and it is. The threads that 471 // detected a race with the deflation process also wanted the 472 // object's header restored before they retry their operation and 473 // because it is restored they will only retry once. 474 } 475 476 // Convert the fields used by is_busy() to a string that can be 477 // used for diagnostic output. 478 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 479 ss->print("is_busy: waiters=%d, ", _waiters); 480 if (!AsyncDeflateIdleMonitors) { 481 ss->print("contentions=%d, ", contentions()); 482 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 483 } else { 484 if (contentions() > 0) { 485 ss->print("contentions=%d, ", contentions()); 486 } else { 487 ss->print("contentions=0"); 488 } 489 if (_owner != DEFLATER_MARKER) { 490 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 491 } else { 492 // We report NULL instead of DEFLATER_MARKER here because is_busy() 493 // ignores DEFLATER_MARKER values. 494 ss->print("owner=" INTPTR_FORMAT, NULL); 495 } 496 } 497 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq), 498 p2i(_EntryList)); 499 return ss->base(); 500 } 501 502 #define MAX_RECHECK_INTERVAL 1000 503 504 void ObjectMonitor::EnterI(TRAPS) { 505 Thread * const Self = THREAD; 506 assert(Self->is_Java_thread(), "invariant"); 507 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 508 509 // Try the lock - TATAS 510 if (TryLock (Self) > 0) { 511 assert(_succ != Self, "invariant"); 512 assert(_owner == Self, "invariant"); 513 assert(_Responsible != Self, "invariant"); 514 return; 515 } 516 517 if (AsyncDeflateIdleMonitors && 518 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 519 // The deflation protocol finished the first part (setting owner), 520 // but it failed the second part (making contentions negative) and 521 // bailed. Acquired the monitor. 522 assert(_succ != Self, "invariant"); 523 assert(_Responsible != Self, "invariant"); 524 return; 525 } 526 527 assert(InitDone, "Unexpectedly not initialized"); 528 529 // We try one round of spinning *before* enqueueing Self. 530 // 531 // If the _owner is ready but OFFPROC we could use a YieldTo() 532 // operation to donate the remainder of this thread's quantum 533 // to the owner. This has subtle but beneficial affinity 534 // effects. 535 536 if (TrySpin(Self) > 0) { 537 assert(_owner == Self, "invariant"); 538 assert(_succ != Self, "invariant"); 539 assert(_Responsible != Self, "invariant"); 540 return; 541 } 542 543 // The Spin failed -- Enqueue and park the thread ... 544 assert(_succ != Self, "invariant"); 545 assert(_owner != Self, "invariant"); 546 assert(_Responsible != Self, "invariant"); 547 548 // Enqueue "Self" on ObjectMonitor's _cxq. 549 // 550 // Node acts as a proxy for Self. 551 // As an aside, if were to ever rewrite the synchronization code mostly 552 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 553 // Java objects. This would avoid awkward lifecycle and liveness issues, 554 // as well as eliminate a subset of ABA issues. 555 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 556 557 ObjectWaiter node(Self); 558 Self->_ParkEvent->reset(); 559 node._prev = (ObjectWaiter *) 0xBAD; 560 node.TState = ObjectWaiter::TS_CXQ; 561 562 // Push "Self" onto the front of the _cxq. 563 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 564 // Note that spinning tends to reduce the rate at which threads 565 // enqueue and dequeue on EntryList|cxq. 566 ObjectWaiter * nxt; 567 for (;;) { 568 node._next = nxt = _cxq; 569 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 570 571 // Interference - the CAS failed because _cxq changed. Just retry. 572 // As an optional optimization we retry the lock. 573 if (TryLock (Self) > 0) { 574 assert(_succ != Self, "invariant"); 575 assert(_owner == Self, "invariant"); 576 assert(_Responsible != Self, "invariant"); 577 return; 578 } 579 } 580 581 // Check for cxq|EntryList edge transition to non-null. This indicates 582 // the onset of contention. While contention persists exiting threads 583 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 584 // operations revert to the faster 1-0 mode. This enter operation may interleave 585 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 586 // arrange for one of the contending thread to use a timed park() operations 587 // to detect and recover from the race. (Stranding is form of progress failure 588 // where the monitor is unlocked but all the contending threads remain parked). 589 // That is, at least one of the contended threads will periodically poll _owner. 590 // One of the contending threads will become the designated "Responsible" thread. 591 // The Responsible thread uses a timed park instead of a normal indefinite park 592 // operation -- it periodically wakes and checks for and recovers from potential 593 // strandings admitted by 1-0 exit operations. We need at most one Responsible 594 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 595 // be responsible for a monitor. 596 // 597 // Currently, one of the contended threads takes on the added role of "Responsible". 598 // A viable alternative would be to use a dedicated "stranding checker" thread 599 // that periodically iterated over all the threads (or active monitors) and unparked 600 // successors where there was risk of stranding. This would help eliminate the 601 // timer scalability issues we see on some platforms as we'd only have one thread 602 // -- the checker -- parked on a timer. 603 604 if (nxt == NULL && _EntryList == NULL) { 605 // Try to assume the role of responsible thread for the monitor. 606 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 607 Atomic::replace_if_null(&_Responsible, Self); 608 } 609 610 // The lock might have been released while this thread was occupied queueing 611 // itself onto _cxq. To close the race and avoid "stranding" and 612 // progress-liveness failure we must resample-retry _owner before parking. 613 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 614 // In this case the ST-MEMBAR is accomplished with CAS(). 615 // 616 // TODO: Defer all thread state transitions until park-time. 617 // Since state transitions are heavy and inefficient we'd like 618 // to defer the state transitions until absolutely necessary, 619 // and in doing so avoid some transitions ... 620 621 int nWakeups = 0; 622 int recheckInterval = 1; 623 624 for (;;) { 625 626 if (TryLock(Self) > 0) break; 627 assert(_owner != Self, "invariant"); 628 629 // park self 630 if (_Responsible == Self) { 631 Self->_ParkEvent->park((jlong) recheckInterval); 632 // Increase the recheckInterval, but clamp the value. 633 recheckInterval *= 8; 634 if (recheckInterval > MAX_RECHECK_INTERVAL) { 635 recheckInterval = MAX_RECHECK_INTERVAL; 636 } 637 } else { 638 Self->_ParkEvent->park(); 639 } 640 641 if (TryLock(Self) > 0) break; 642 643 if (AsyncDeflateIdleMonitors && 644 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 645 // The deflation protocol finished the first part (setting owner), 646 // but it failed the second part (making contentions negative) and 647 // bailed. Acquired the monitor. 648 break; 649 } 650 651 // The lock is still contested. 652 // Keep a tally of the # of futile wakeups. 653 // Note that the counter is not protected by a lock or updated by atomics. 654 // That is by design - we trade "lossy" counters which are exposed to 655 // races during updates for a lower probe effect. 656 657 // This PerfData object can be used in parallel with a safepoint. 658 // See the work around in PerfDataManager::destroy(). 659 OM_PERFDATA_OP(FutileWakeups, inc()); 660 ++nWakeups; 661 662 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 663 // We can defer clearing _succ until after the spin completes 664 // TrySpin() must tolerate being called with _succ == Self. 665 // Try yet another round of adaptive spinning. 666 if (TrySpin(Self) > 0) break; 667 668 // We can find that we were unpark()ed and redesignated _succ while 669 // we were spinning. That's harmless. If we iterate and call park(), 670 // park() will consume the event and return immediately and we'll 671 // just spin again. This pattern can repeat, leaving _succ to simply 672 // spin on a CPU. 673 674 if (_succ == Self) _succ = NULL; 675 676 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 677 OrderAccess::fence(); 678 } 679 680 // Egress : 681 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 682 // Normally we'll find Self on the EntryList . 683 // From the perspective of the lock owner (this thread), the 684 // EntryList is stable and cxq is prepend-only. 685 // The head of cxq is volatile but the interior is stable. 686 // In addition, Self.TState is stable. 687 688 assert(_owner == Self, "invariant"); 689 assert(object() != NULL, "invariant"); 690 // I'd like to write: 691 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 692 // but as we're at a safepoint that's not safe. 693 694 UnlinkAfterAcquire(Self, &node); 695 if (_succ == Self) _succ = NULL; 696 697 assert(_succ != Self, "invariant"); 698 if (_Responsible == Self) { 699 _Responsible = NULL; 700 OrderAccess::fence(); // Dekker pivot-point 701 702 // We may leave threads on cxq|EntryList without a designated 703 // "Responsible" thread. This is benign. When this thread subsequently 704 // exits the monitor it can "see" such preexisting "old" threads -- 705 // threads that arrived on the cxq|EntryList before the fence, above -- 706 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 707 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 708 // non-null and elect a new "Responsible" timer thread. 709 // 710 // This thread executes: 711 // ST Responsible=null; MEMBAR (in enter epilogue - here) 712 // LD cxq|EntryList (in subsequent exit) 713 // 714 // Entering threads in the slow/contended path execute: 715 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 716 // The (ST cxq; MEMBAR) is accomplished with CAS(). 717 // 718 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 719 // exit operation from floating above the ST Responsible=null. 720 } 721 722 // We've acquired ownership with CAS(). 723 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 724 // But since the CAS() this thread may have also stored into _succ, 725 // EntryList, cxq or Responsible. These meta-data updates must be 726 // visible __before this thread subsequently drops the lock. 727 // Consider what could occur if we didn't enforce this constraint -- 728 // STs to monitor meta-data and user-data could reorder with (become 729 // visible after) the ST in exit that drops ownership of the lock. 730 // Some other thread could then acquire the lock, but observe inconsistent 731 // or old monitor meta-data and heap data. That violates the JMM. 732 // To that end, the 1-0 exit() operation must have at least STST|LDST 733 // "release" barrier semantics. Specifically, there must be at least a 734 // STST|LDST barrier in exit() before the ST of null into _owner that drops 735 // the lock. The barrier ensures that changes to monitor meta-data and data 736 // protected by the lock will be visible before we release the lock, and 737 // therefore before some other thread (CPU) has a chance to acquire the lock. 738 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 739 // 740 // Critically, any prior STs to _succ or EntryList must be visible before 741 // the ST of null into _owner in the *subsequent* (following) corresponding 742 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 743 // execute a serializing instruction. 744 745 return; 746 } 747 748 // ReenterI() is a specialized inline form of the latter half of the 749 // contended slow-path from EnterI(). We use ReenterI() only for 750 // monitor reentry in wait(). 751 // 752 // In the future we should reconcile EnterI() and ReenterI(). 753 754 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 755 assert(Self != NULL, "invariant"); 756 assert(SelfNode != NULL, "invariant"); 757 assert(SelfNode->_thread == Self, "invariant"); 758 assert(_waiters > 0, "invariant"); 759 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 760 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 761 JavaThread * jt = (JavaThread *) Self; 762 763 int nWakeups = 0; 764 for (;;) { 765 ObjectWaiter::TStates v = SelfNode->TState; 766 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 767 assert(_owner != Self, "invariant"); 768 769 if (TryLock(Self) > 0) break; 770 if (TrySpin(Self) > 0) break; 771 772 // State transition wrappers around park() ... 773 // ReenterI() wisely defers state transitions until 774 // it's clear we must park the thread. 775 { 776 OSThreadContendState osts(Self->osthread()); 777 ThreadBlockInVM tbivm(jt); 778 779 // cleared by handle_special_suspend_equivalent_condition() 780 // or java_suspend_self() 781 jt->set_suspend_equivalent(); 782 Self->_ParkEvent->park(); 783 784 // were we externally suspended while we were waiting? 785 for (;;) { 786 if (!ExitSuspendEquivalent(jt)) break; 787 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 788 jt->java_suspend_self(); 789 jt->set_suspend_equivalent(); 790 } 791 } 792 793 // Try again, but just so we distinguish between futile wakeups and 794 // successful wakeups. The following test isn't algorithmically 795 // necessary, but it helps us maintain sensible statistics. 796 if (TryLock(Self) > 0) break; 797 798 // The lock is still contested. 799 // Keep a tally of the # of futile wakeups. 800 // Note that the counter is not protected by a lock or updated by atomics. 801 // That is by design - we trade "lossy" counters which are exposed to 802 // races during updates for a lower probe effect. 803 ++nWakeups; 804 805 // Assuming this is not a spurious wakeup we'll normally 806 // find that _succ == Self. 807 if (_succ == Self) _succ = NULL; 808 809 // Invariant: after clearing _succ a contending thread 810 // *must* retry _owner before parking. 811 OrderAccess::fence(); 812 813 // This PerfData object can be used in parallel with a safepoint. 814 // See the work around in PerfDataManager::destroy(). 815 OM_PERFDATA_OP(FutileWakeups, inc()); 816 } 817 818 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 819 // Normally we'll find Self on the EntryList. 820 // Unlinking from the EntryList is constant-time and atomic-free. 821 // From the perspective of the lock owner (this thread), the 822 // EntryList is stable and cxq is prepend-only. 823 // The head of cxq is volatile but the interior is stable. 824 // In addition, Self.TState is stable. 825 826 assert(_owner == Self, "invariant"); 827 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 828 UnlinkAfterAcquire(Self, SelfNode); 829 if (_succ == Self) _succ = NULL; 830 assert(_succ != Self, "invariant"); 831 SelfNode->TState = ObjectWaiter::TS_RUN; 832 OrderAccess::fence(); // see comments at the end of EnterI() 833 } 834 835 // By convention we unlink a contending thread from EntryList|cxq immediately 836 // after the thread acquires the lock in ::enter(). Equally, we could defer 837 // unlinking the thread until ::exit()-time. 838 839 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 840 assert(_owner == Self, "invariant"); 841 assert(SelfNode->_thread == Self, "invariant"); 842 843 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 844 // Normal case: remove Self from the DLL EntryList . 845 // This is a constant-time operation. 846 ObjectWaiter * nxt = SelfNode->_next; 847 ObjectWaiter * prv = SelfNode->_prev; 848 if (nxt != NULL) nxt->_prev = prv; 849 if (prv != NULL) prv->_next = nxt; 850 if (SelfNode == _EntryList) _EntryList = nxt; 851 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 852 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 853 } else { 854 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 855 // Inopportune interleaving -- Self is still on the cxq. 856 // This usually means the enqueue of self raced an exiting thread. 857 // Normally we'll find Self near the front of the cxq, so 858 // dequeueing is typically fast. If needbe we can accelerate 859 // this with some MCS/CHL-like bidirectional list hints and advisory 860 // back-links so dequeueing from the interior will normally operate 861 // in constant-time. 862 // Dequeue Self from either the head (with CAS) or from the interior 863 // with a linear-time scan and normal non-atomic memory operations. 864 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 865 // and then unlink Self from EntryList. We have to drain eventually, 866 // so it might as well be now. 867 868 ObjectWaiter * v = _cxq; 869 assert(v != NULL, "invariant"); 870 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) { 871 // The CAS above can fail from interference IFF a "RAT" arrived. 872 // In that case Self must be in the interior and can no longer be 873 // at the head of cxq. 874 if (v == SelfNode) { 875 assert(_cxq != v, "invariant"); 876 v = _cxq; // CAS above failed - start scan at head of list 877 } 878 ObjectWaiter * p; 879 ObjectWaiter * q = NULL; 880 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 881 q = p; 882 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 883 } 884 assert(v != SelfNode, "invariant"); 885 assert(p == SelfNode, "Node not found on cxq"); 886 assert(p != _cxq, "invariant"); 887 assert(q != NULL, "invariant"); 888 assert(q->_next == p, "invariant"); 889 q->_next = p->_next; 890 } 891 } 892 893 #ifdef ASSERT 894 // Diagnostic hygiene ... 895 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 896 SelfNode->_next = (ObjectWaiter *) 0xBAD; 897 SelfNode->TState = ObjectWaiter::TS_RUN; 898 #endif 899 } 900 901 // ----------------------------------------------------------------------------- 902 // Exit support 903 // 904 // exit() 905 // ~~~~~~ 906 // Note that the collector can't reclaim the objectMonitor or deflate 907 // the object out from underneath the thread calling ::exit() as the 908 // thread calling ::exit() never transitions to a stable state. 909 // This inhibits GC, which in turn inhibits asynchronous (and 910 // inopportune) reclamation of "this". 911 // 912 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 913 // There's one exception to the claim above, however. EnterI() can call 914 // exit() to drop a lock if the acquirer has been externally suspended. 915 // In that case exit() is called with _thread_state == _thread_blocked, 916 // but the monitor's _contentions field is > 0, which inhibits reclamation. 917 // 918 // 1-0 exit 919 // ~~~~~~~~ 920 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 921 // the fast-path operators have been optimized so the common ::exit() 922 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 923 // The code emitted by fast_unlock() elides the usual MEMBAR. This 924 // greatly improves latency -- MEMBAR and CAS having considerable local 925 // latency on modern processors -- but at the cost of "stranding". Absent the 926 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 927 // ::enter() path, resulting in the entering thread being stranding 928 // and a progress-liveness failure. Stranding is extremely rare. 929 // We use timers (timed park operations) & periodic polling to detect 930 // and recover from stranding. Potentially stranded threads periodically 931 // wake up and poll the lock. See the usage of the _Responsible variable. 932 // 933 // The CAS() in enter provides for safety and exclusion, while the CAS or 934 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 935 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 936 // We detect and recover from stranding with timers. 937 // 938 // If a thread transiently strands it'll park until (a) another 939 // thread acquires the lock and then drops the lock, at which time the 940 // exiting thread will notice and unpark the stranded thread, or, (b) 941 // the timer expires. If the lock is high traffic then the stranding latency 942 // will be low due to (a). If the lock is low traffic then the odds of 943 // stranding are lower, although the worst-case stranding latency 944 // is longer. Critically, we don't want to put excessive load in the 945 // platform's timer subsystem. We want to minimize both the timer injection 946 // rate (timers created/sec) as well as the number of timers active at 947 // any one time. (more precisely, we want to minimize timer-seconds, which is 948 // the integral of the # of active timers at any instant over time). 949 // Both impinge on OS scalability. Given that, at most one thread parked on 950 // a monitor will use a timer. 951 // 952 // There is also the risk of a futile wake-up. If we drop the lock 953 // another thread can reacquire the lock immediately, and we can 954 // then wake a thread unnecessarily. This is benign, and we've 955 // structured the code so the windows are short and the frequency 956 // of such futile wakups is low. 957 958 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 959 Thread* const Self = THREAD; 960 void* cur = Atomic::load(&_owner); 961 if (THREAD != cur) { 962 if (THREAD->is_lock_owned((address)cur)) { 963 assert(_recursions == 0, "invariant"); 964 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 965 _recursions = 0; 966 } else { 967 // Apparent unbalanced locking ... 968 // Naively we'd like to throw IllegalMonitorStateException. 969 // As a practical matter we can neither allocate nor throw an 970 // exception as ::exit() can be called from leaf routines. 971 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 972 // Upon deeper reflection, however, in a properly run JVM the only 973 // way we should encounter this situation is in the presence of 974 // unbalanced JNI locking. TODO: CheckJNICalls. 975 // See also: CR4414101 976 #ifdef ASSERT 977 LogStreamHandle(Error, monitorinflation) lsh; 978 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 979 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 980 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 981 print_debug_style_on(&lsh); 982 #endif 983 assert(false, "Non-balanced monitor enter/exit!"); 984 return; 985 } 986 } 987 988 if (_recursions != 0) { 989 _recursions--; // this is simple recursive enter 990 return; 991 } 992 993 // Invariant: after setting Responsible=null an thread must execute 994 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 995 _Responsible = NULL; 996 997 #if INCLUDE_JFR 998 // get the owner's thread id for the MonitorEnter event 999 // if it is enabled and the thread isn't suspended 1000 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1001 _previous_owner_tid = JFR_THREAD_ID(Self); 1002 } 1003 #endif 1004 1005 for (;;) { 1006 assert(THREAD == _owner, "invariant"); 1007 1008 // Drop the lock. 1009 // release semantics: prior loads and stores from within the critical section 1010 // must not float (reorder) past the following store that drops the lock. 1011 // Uses a storeload to separate release_store(owner) from the 1012 // successor check. The try_set_owner() below uses cmpxchg() so 1013 // we get the fence down there. 1014 release_clear_owner(Self); 1015 OrderAccess::storeload(); 1016 1017 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1018 return; 1019 } 1020 // Other threads are blocked trying to acquire the lock. 1021 1022 // Normally the exiting thread is responsible for ensuring succession, 1023 // but if other successors are ready or other entering threads are spinning 1024 // then this thread can simply store NULL into _owner and exit without 1025 // waking a successor. The existence of spinners or ready successors 1026 // guarantees proper succession (liveness). Responsibility passes to the 1027 // ready or running successors. The exiting thread delegates the duty. 1028 // More precisely, if a successor already exists this thread is absolved 1029 // of the responsibility of waking (unparking) one. 1030 // 1031 // The _succ variable is critical to reducing futile wakeup frequency. 1032 // _succ identifies the "heir presumptive" thread that has been made 1033 // ready (unparked) but that has not yet run. We need only one such 1034 // successor thread to guarantee progress. 1035 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1036 // section 3.3 "Futile Wakeup Throttling" for details. 1037 // 1038 // Note that spinners in Enter() also set _succ non-null. 1039 // In the current implementation spinners opportunistically set 1040 // _succ so that exiting threads might avoid waking a successor. 1041 // Another less appealing alternative would be for the exiting thread 1042 // to drop the lock and then spin briefly to see if a spinner managed 1043 // to acquire the lock. If so, the exiting thread could exit 1044 // immediately without waking a successor, otherwise the exiting 1045 // thread would need to dequeue and wake a successor. 1046 // (Note that we'd need to make the post-drop spin short, but no 1047 // shorter than the worst-case round-trip cache-line migration time. 1048 // The dropped lock needs to become visible to the spinner, and then 1049 // the acquisition of the lock by the spinner must become visible to 1050 // the exiting thread). 1051 1052 // It appears that an heir-presumptive (successor) must be made ready. 1053 // Only the current lock owner can manipulate the EntryList or 1054 // drain _cxq, so we need to reacquire the lock. If we fail 1055 // to reacquire the lock the responsibility for ensuring succession 1056 // falls to the new owner. 1057 // 1058 if (try_set_owner_from(NULL, Self) != NULL) { 1059 return; 1060 } 1061 1062 guarantee(_owner == THREAD, "invariant"); 1063 1064 ObjectWaiter * w = NULL; 1065 1066 w = _EntryList; 1067 if (w != NULL) { 1068 // I'd like to write: guarantee (w->_thread != Self). 1069 // But in practice an exiting thread may find itself on the EntryList. 1070 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1071 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1072 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1073 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1074 // release the lock "O". T2 resumes immediately after the ST of null into 1075 // _owner, above. T2 notices that the EntryList is populated, so it 1076 // reacquires the lock and then finds itself on the EntryList. 1077 // Given all that, we have to tolerate the circumstance where "w" is 1078 // associated with Self. 1079 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1080 ExitEpilog(Self, w); 1081 return; 1082 } 1083 1084 // If we find that both _cxq and EntryList are null then just 1085 // re-run the exit protocol from the top. 1086 w = _cxq; 1087 if (w == NULL) continue; 1088 1089 // Drain _cxq into EntryList - bulk transfer. 1090 // First, detach _cxq. 1091 // The following loop is tantamount to: w = swap(&cxq, NULL) 1092 for (;;) { 1093 assert(w != NULL, "Invariant"); 1094 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL); 1095 if (u == w) break; 1096 w = u; 1097 } 1098 1099 assert(w != NULL, "invariant"); 1100 assert(_EntryList == NULL, "invariant"); 1101 1102 // Convert the LIFO SLL anchored by _cxq into a DLL. 1103 // The list reorganization step operates in O(LENGTH(w)) time. 1104 // It's critical that this step operate quickly as 1105 // "Self" still holds the outer-lock, restricting parallelism 1106 // and effectively lengthening the critical section. 1107 // Invariant: s chases t chases u. 1108 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1109 // we have faster access to the tail. 1110 1111 _EntryList = w; 1112 ObjectWaiter * q = NULL; 1113 ObjectWaiter * p; 1114 for (p = w; p != NULL; p = p->_next) { 1115 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1116 p->TState = ObjectWaiter::TS_ENTER; 1117 p->_prev = q; 1118 q = p; 1119 } 1120 1121 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1122 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1123 1124 // See if we can abdicate to a spinner instead of waking a thread. 1125 // A primary goal of the implementation is to reduce the 1126 // context-switch rate. 1127 if (_succ != NULL) continue; 1128 1129 w = _EntryList; 1130 if (w != NULL) { 1131 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1132 ExitEpilog(Self, w); 1133 return; 1134 } 1135 } 1136 } 1137 1138 // ExitSuspendEquivalent: 1139 // A faster alternate to handle_special_suspend_equivalent_condition() 1140 // 1141 // handle_special_suspend_equivalent_condition() unconditionally 1142 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1143 // operations have high latency. Note that in ::enter() we call HSSEC 1144 // while holding the monitor, so we effectively lengthen the critical sections. 1145 // 1146 // There are a number of possible solutions: 1147 // 1148 // A. To ameliorate the problem we might also defer state transitions 1149 // to as late as possible -- just prior to parking. 1150 // Given that, we'd call HSSEC after having returned from park(), 1151 // but before attempting to acquire the monitor. This is only a 1152 // partial solution. It avoids calling HSSEC while holding the 1153 // monitor (good), but it still increases successor reacquisition latency -- 1154 // the interval between unparking a successor and the time the successor 1155 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1156 // If we use this technique we can also avoid EnterI()-exit() loop 1157 // in ::enter() where we iteratively drop the lock and then attempt 1158 // to reacquire it after suspending. 1159 // 1160 // B. In the future we might fold all the suspend bits into a 1161 // composite per-thread suspend flag and then update it with CAS(). 1162 // Alternately, a Dekker-like mechanism with multiple variables 1163 // would suffice: 1164 // ST Self->_suspend_equivalent = false 1165 // MEMBAR 1166 // LD Self_>_suspend_flags 1167 1168 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1169 return jSelf->handle_special_suspend_equivalent_condition(); 1170 } 1171 1172 1173 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1174 assert(_owner == Self, "invariant"); 1175 1176 // Exit protocol: 1177 // 1. ST _succ = wakee 1178 // 2. membar #loadstore|#storestore; 1179 // 2. ST _owner = NULL 1180 // 3. unpark(wakee) 1181 1182 _succ = Wakee->_thread; 1183 ParkEvent * Trigger = Wakee->_event; 1184 1185 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1186 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1187 // out-of-scope (non-extant). 1188 Wakee = NULL; 1189 1190 // Drop the lock. 1191 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1192 release_clear_owner(Self); 1193 OrderAccess::fence(); 1194 1195 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1196 Trigger->unpark(); 1197 1198 // Maintain stats and report events to JVMTI 1199 OM_PERFDATA_OP(Parks, inc()); 1200 } 1201 1202 1203 // ----------------------------------------------------------------------------- 1204 // Class Loader deadlock handling. 1205 // 1206 // complete_exit exits a lock returning recursion count 1207 // complete_exit/reenter operate as a wait without waiting 1208 // complete_exit requires an inflated monitor 1209 // The _owner field is not always the Thread addr even with an 1210 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1211 // thread due to contention. 1212 intx ObjectMonitor::complete_exit(TRAPS) { 1213 Thread * const Self = THREAD; 1214 assert(Self->is_Java_thread(), "Must be Java thread!"); 1215 JavaThread *jt = (JavaThread *)THREAD; 1216 1217 assert(InitDone, "Unexpectedly not initialized"); 1218 1219 void* cur = Atomic::load(&_owner); 1220 if (THREAD != cur) { 1221 if (THREAD->is_lock_owned((address)cur)) { 1222 assert(_recursions == 0, "internal state error"); 1223 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 1224 _recursions = 0; 1225 } 1226 } 1227 1228 guarantee(Self == _owner, "complete_exit not owner"); 1229 intx save = _recursions; // record the old recursion count 1230 _recursions = 0; // set the recursion level to be 0 1231 exit(true, Self); // exit the monitor 1232 guarantee(_owner != Self, "invariant"); 1233 return save; 1234 } 1235 1236 // reenter() enters a lock and sets recursion count 1237 // complete_exit/reenter operate as a wait without waiting 1238 bool ObjectMonitor::reenter(intx recursions, TRAPS) { 1239 Thread * const Self = THREAD; 1240 assert(Self->is_Java_thread(), "Must be Java thread!"); 1241 JavaThread *jt = (JavaThread *)THREAD; 1242 1243 guarantee(_owner != Self, "reenter already owner"); 1244 if (!enter(THREAD)) { 1245 return false; 1246 } 1247 // Entered the monitor. 1248 guarantee(_recursions == 0, "reenter recursion"); 1249 _recursions = recursions; 1250 return true; 1251 } 1252 1253 // Checks that the current THREAD owns this monitor and causes an 1254 // immediate return if it doesn't. We don't use the CHECK macro 1255 // because we want the IMSE to be the only exception that is thrown 1256 // from the call site when false is returned. Any other pending 1257 // exception is ignored. 1258 #define CHECK_OWNER() \ 1259 do { \ 1260 if (!check_owner(THREAD)) { \ 1261 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1262 return; \ 1263 } \ 1264 } while (false) 1265 1266 // Returns true if the specified thread owns the ObjectMonitor. 1267 // Otherwise returns false and throws IllegalMonitorStateException 1268 // (IMSE). If there is a pending exception and the specified thread 1269 // is not the owner, that exception will be replaced by the IMSE. 1270 bool ObjectMonitor::check_owner(Thread* THREAD) { 1271 void* cur = Atomic::load(&_owner); 1272 if (cur == THREAD) { 1273 return true; 1274 } 1275 if (THREAD->is_lock_owned((address)cur)) { 1276 set_owner_from_BasicLock(cur, THREAD); // Convert from BasicLock* to Thread*. 1277 _recursions = 0; 1278 return true; 1279 } 1280 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1281 "current thread is not owner", false); 1282 } 1283 1284 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1285 ObjectMonitor* monitor, 1286 jlong notifier_tid, 1287 jlong timeout, 1288 bool timedout) { 1289 assert(event != NULL, "invariant"); 1290 assert(monitor != NULL, "invariant"); 1291 event->set_monitorClass(((oop)monitor->object())->klass()); 1292 event->set_timeout(timeout); 1293 event->set_address((uintptr_t)monitor->object_addr()); 1294 event->set_notifier(notifier_tid); 1295 event->set_timedOut(timedout); 1296 event->commit(); 1297 } 1298 1299 // ----------------------------------------------------------------------------- 1300 // Wait/Notify/NotifyAll 1301 // 1302 // Note: a subset of changes to ObjectMonitor::wait() 1303 // will need to be replicated in complete_exit 1304 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1305 Thread * const Self = THREAD; 1306 assert(Self->is_Java_thread(), "Must be Java thread!"); 1307 JavaThread *jt = (JavaThread *)THREAD; 1308 1309 assert(InitDone, "Unexpectedly not initialized"); 1310 1311 CHECK_OWNER(); // Throws IMSE if not owner. 1312 1313 EventJavaMonitorWait event; 1314 1315 // check for a pending interrupt 1316 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1317 // post monitor waited event. Note that this is past-tense, we are done waiting. 1318 if (JvmtiExport::should_post_monitor_waited()) { 1319 // Note: 'false' parameter is passed here because the 1320 // wait was not timed out due to thread interrupt. 1321 JvmtiExport::post_monitor_waited(jt, this, false); 1322 1323 // In this short circuit of the monitor wait protocol, the 1324 // current thread never drops ownership of the monitor and 1325 // never gets added to the wait queue so the current thread 1326 // cannot be made the successor. This means that the 1327 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1328 // consume an unpark() meant for the ParkEvent associated with 1329 // this ObjectMonitor. 1330 } 1331 if (event.should_commit()) { 1332 post_monitor_wait_event(&event, this, 0, millis, false); 1333 } 1334 THROW(vmSymbols::java_lang_InterruptedException()); 1335 return; 1336 } 1337 1338 assert(Self->_Stalled == 0, "invariant"); 1339 Self->_Stalled = intptr_t(this); 1340 jt->set_current_waiting_monitor(this); 1341 1342 // create a node to be put into the queue 1343 // Critically, after we reset() the event but prior to park(), we must check 1344 // for a pending interrupt. 1345 ObjectWaiter node(Self); 1346 node.TState = ObjectWaiter::TS_WAIT; 1347 Self->_ParkEvent->reset(); 1348 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1349 1350 // Enter the waiting queue, which is a circular doubly linked list in this case 1351 // but it could be a priority queue or any data structure. 1352 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1353 // by the the owner of the monitor *except* in the case where park() 1354 // returns because of a timeout of interrupt. Contention is exceptionally rare 1355 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1356 1357 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1358 AddWaiter(&node); 1359 Thread::SpinRelease(&_WaitSetLock); 1360 1361 _Responsible = NULL; 1362 1363 intx save = _recursions; // record the old recursion count 1364 _waiters++; // increment the number of waiters 1365 _recursions = 0; // set the recursion level to be 1 1366 exit(true, Self); // exit the monitor 1367 guarantee(_owner != Self, "invariant"); 1368 1369 // The thread is on the WaitSet list - now park() it. 1370 // On MP systems it's conceivable that a brief spin before we park 1371 // could be profitable. 1372 // 1373 // TODO-FIXME: change the following logic to a loop of the form 1374 // while (!timeout && !interrupted && _notified == 0) park() 1375 1376 int ret = OS_OK; 1377 int WasNotified = 0; 1378 1379 // Need to check interrupt state whilst still _thread_in_vm 1380 bool interrupted = interruptible && jt->is_interrupted(false); 1381 1382 { // State transition wrappers 1383 OSThread* osthread = Self->osthread(); 1384 OSThreadWaitState osts(osthread, true); 1385 { 1386 ThreadBlockInVM tbivm(jt); 1387 // Thread is in thread_blocked state and oop access is unsafe. 1388 jt->set_suspend_equivalent(); 1389 1390 if (interrupted || HAS_PENDING_EXCEPTION) { 1391 // Intentionally empty 1392 } else if (node._notified == 0) { 1393 if (millis <= 0) { 1394 Self->_ParkEvent->park(); 1395 } else { 1396 ret = Self->_ParkEvent->park(millis); 1397 } 1398 } 1399 1400 // were we externally suspended while we were waiting? 1401 if (ExitSuspendEquivalent (jt)) { 1402 // TODO-FIXME: add -- if succ == Self then succ = null. 1403 jt->java_suspend_self(); 1404 } 1405 1406 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1407 1408 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1409 // from the WaitSet to the EntryList. 1410 // See if we need to remove Node from the WaitSet. 1411 // We use double-checked locking to avoid grabbing _WaitSetLock 1412 // if the thread is not on the wait queue. 1413 // 1414 // Note that we don't need a fence before the fetch of TState. 1415 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1416 // written by the is thread. (perhaps the fetch might even be satisfied 1417 // by a look-aside into the processor's own store buffer, although given 1418 // the length of the code path between the prior ST and this load that's 1419 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1420 // then we'll acquire the lock and then re-fetch a fresh TState value. 1421 // That is, we fail toward safety. 1422 1423 if (node.TState == ObjectWaiter::TS_WAIT) { 1424 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1425 if (node.TState == ObjectWaiter::TS_WAIT) { 1426 DequeueSpecificWaiter(&node); // unlink from WaitSet 1427 assert(node._notified == 0, "invariant"); 1428 node.TState = ObjectWaiter::TS_RUN; 1429 } 1430 Thread::SpinRelease(&_WaitSetLock); 1431 } 1432 1433 // The thread is now either on off-list (TS_RUN), 1434 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1435 // The Node's TState variable is stable from the perspective of this thread. 1436 // No other threads will asynchronously modify TState. 1437 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1438 OrderAccess::loadload(); 1439 if (_succ == Self) _succ = NULL; 1440 WasNotified = node._notified; 1441 1442 // Reentry phase -- reacquire the monitor. 1443 // re-enter contended monitor after object.wait(). 1444 // retain OBJECT_WAIT state until re-enter successfully completes 1445 // Thread state is thread_in_vm and oop access is again safe, 1446 // although the raw address of the object may have changed. 1447 // (Don't cache naked oops over safepoints, of course). 1448 1449 // post monitor waited event. Note that this is past-tense, we are done waiting. 1450 if (JvmtiExport::should_post_monitor_waited()) { 1451 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1452 1453 if (node._notified != 0 && _succ == Self) { 1454 // In this part of the monitor wait-notify-reenter protocol it 1455 // is possible (and normal) for another thread to do a fastpath 1456 // monitor enter-exit while this thread is still trying to get 1457 // to the reenter portion of the protocol. 1458 // 1459 // The ObjectMonitor was notified and the current thread is 1460 // the successor which also means that an unpark() has already 1461 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1462 // consume the unpark() that was done when the successor was 1463 // set because the same ParkEvent is shared between Java 1464 // monitors and JVM/TI RawMonitors (for now). 1465 // 1466 // We redo the unpark() to ensure forward progress, i.e., we 1467 // don't want all pending threads hanging (parked) with none 1468 // entering the unlocked monitor. 1469 node._event->unpark(); 1470 } 1471 } 1472 1473 if (event.should_commit()) { 1474 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1475 } 1476 1477 OrderAccess::fence(); 1478 1479 assert(Self->_Stalled != 0, "invariant"); 1480 Self->_Stalled = 0; 1481 1482 assert(_owner != Self, "invariant"); 1483 ObjectWaiter::TStates v = node.TState; 1484 if (v == ObjectWaiter::TS_RUN) { 1485 enter(Self); 1486 } else { 1487 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1488 ReenterI(Self, &node); 1489 node.wait_reenter_end(this); 1490 } 1491 1492 // Self has reacquired the lock. 1493 // Lifecycle - the node representing Self must not appear on any queues. 1494 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1495 // want residual elements associated with this thread left on any lists. 1496 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1497 assert(_owner == Self, "invariant"); 1498 assert(_succ != Self, "invariant"); 1499 } // OSThreadWaitState() 1500 1501 jt->set_current_waiting_monitor(NULL); 1502 1503 guarantee(_recursions == 0, "invariant"); 1504 _recursions = save; // restore the old recursion count 1505 _waiters--; // decrement the number of waiters 1506 1507 // Verify a few postconditions 1508 assert(_owner == Self, "invariant"); 1509 assert(_succ != Self, "invariant"); 1510 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1511 1512 // check if the notification happened 1513 if (!WasNotified) { 1514 // no, it could be timeout or Thread.interrupt() or both 1515 // check for interrupt event, otherwise it is timeout 1516 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1517 THROW(vmSymbols::java_lang_InterruptedException()); 1518 } 1519 } 1520 1521 // NOTE: Spurious wake up will be consider as timeout. 1522 // Monitor notify has precedence over thread interrupt. 1523 } 1524 1525 1526 // Consider: 1527 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1528 // then instead of transferring a thread from the WaitSet to the EntryList 1529 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1530 1531 void ObjectMonitor::INotify(Thread * Self) { 1532 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1533 ObjectWaiter * iterator = DequeueWaiter(); 1534 if (iterator != NULL) { 1535 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1536 guarantee(iterator->_notified == 0, "invariant"); 1537 // Disposition - what might we do with iterator ? 1538 // a. add it directly to the EntryList - either tail (policy == 1) 1539 // or head (policy == 0). 1540 // b. push it onto the front of the _cxq (policy == 2). 1541 // For now we use (b). 1542 1543 iterator->TState = ObjectWaiter::TS_ENTER; 1544 1545 iterator->_notified = 1; 1546 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1547 1548 ObjectWaiter * list = _EntryList; 1549 if (list != NULL) { 1550 assert(list->_prev == NULL, "invariant"); 1551 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1552 assert(list != iterator, "invariant"); 1553 } 1554 1555 // prepend to cxq 1556 if (list == NULL) { 1557 iterator->_next = iterator->_prev = NULL; 1558 _EntryList = iterator; 1559 } else { 1560 iterator->TState = ObjectWaiter::TS_CXQ; 1561 for (;;) { 1562 ObjectWaiter * front = _cxq; 1563 iterator->_next = front; 1564 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1565 break; 1566 } 1567 } 1568 } 1569 1570 // _WaitSetLock protects the wait queue, not the EntryList. We could 1571 // move the add-to-EntryList operation, above, outside the critical section 1572 // protected by _WaitSetLock. In practice that's not useful. With the 1573 // exception of wait() timeouts and interrupts the monitor owner 1574 // is the only thread that grabs _WaitSetLock. There's almost no contention 1575 // on _WaitSetLock so it's not profitable to reduce the length of the 1576 // critical section. 1577 1578 iterator->wait_reenter_begin(this); 1579 } 1580 Thread::SpinRelease(&_WaitSetLock); 1581 } 1582 1583 // Consider: a not-uncommon synchronization bug is to use notify() when 1584 // notifyAll() is more appropriate, potentially resulting in stranded 1585 // threads; this is one example of a lost wakeup. A useful diagnostic 1586 // option is to force all notify() operations to behave as notifyAll(). 1587 // 1588 // Note: We can also detect many such problems with a "minimum wait". 1589 // When the "minimum wait" is set to a small non-zero timeout value 1590 // and the program does not hang whereas it did absent "minimum wait", 1591 // that suggests a lost wakeup bug. 1592 1593 void ObjectMonitor::notify(TRAPS) { 1594 CHECK_OWNER(); // Throws IMSE if not owner. 1595 if (_WaitSet == NULL) { 1596 return; 1597 } 1598 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1599 INotify(THREAD); 1600 OM_PERFDATA_OP(Notifications, inc(1)); 1601 } 1602 1603 1604 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1605 // from the waitset to the EntryList. This could be done more efficiently with a 1606 // single bulk transfer but in practice it's not time-critical. Beware too, 1607 // that in prepend-mode we invert the order of the waiters. Let's say that the 1608 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1609 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1610 1611 void ObjectMonitor::notifyAll(TRAPS) { 1612 CHECK_OWNER(); // Throws IMSE if not owner. 1613 if (_WaitSet == NULL) { 1614 return; 1615 } 1616 1617 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1618 int tally = 0; 1619 while (_WaitSet != NULL) { 1620 tally++; 1621 INotify(THREAD); 1622 } 1623 1624 OM_PERFDATA_OP(Notifications, inc(tally)); 1625 } 1626 1627 // ----------------------------------------------------------------------------- 1628 // Adaptive Spinning Support 1629 // 1630 // Adaptive spin-then-block - rational spinning 1631 // 1632 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1633 // algorithm. On high order SMP systems it would be better to start with 1634 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1635 // a contending thread could enqueue itself on the cxq and then spin locally 1636 // on a thread-specific variable such as its ParkEvent._Event flag. 1637 // That's left as an exercise for the reader. Note that global spinning is 1638 // not problematic on Niagara, as the L2 cache serves the interconnect and 1639 // has both low latency and massive bandwidth. 1640 // 1641 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1642 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1643 // (duration) or we can fix the count at approximately the duration of 1644 // a context switch and vary the frequency. Of course we could also 1645 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1646 // For a description of 'Adaptive spin-then-block mutual exclusion in 1647 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1648 // 1649 // This implementation varies the duration "D", where D varies with 1650 // the success rate of recent spin attempts. (D is capped at approximately 1651 // length of a round-trip context switch). The success rate for recent 1652 // spin attempts is a good predictor of the success rate of future spin 1653 // attempts. The mechanism adapts automatically to varying critical 1654 // section length (lock modality), system load and degree of parallelism. 1655 // D is maintained per-monitor in _SpinDuration and is initialized 1656 // optimistically. Spin frequency is fixed at 100%. 1657 // 1658 // Note that _SpinDuration is volatile, but we update it without locks 1659 // or atomics. The code is designed so that _SpinDuration stays within 1660 // a reasonable range even in the presence of races. The arithmetic 1661 // operations on _SpinDuration are closed over the domain of legal values, 1662 // so at worst a race will install and older but still legal value. 1663 // At the very worst this introduces some apparent non-determinism. 1664 // We might spin when we shouldn't or vice-versa, but since the spin 1665 // count are relatively short, even in the worst case, the effect is harmless. 1666 // 1667 // Care must be taken that a low "D" value does not become an 1668 // an absorbing state. Transient spinning failures -- when spinning 1669 // is overall profitable -- should not cause the system to converge 1670 // on low "D" values. We want spinning to be stable and predictable 1671 // and fairly responsive to change and at the same time we don't want 1672 // it to oscillate, become metastable, be "too" non-deterministic, 1673 // or converge on or enter undesirable stable absorbing states. 1674 // 1675 // We implement a feedback-based control system -- using past behavior 1676 // to predict future behavior. We face two issues: (a) if the 1677 // input signal is random then the spin predictor won't provide optimal 1678 // results, and (b) if the signal frequency is too high then the control 1679 // system, which has some natural response lag, will "chase" the signal. 1680 // (b) can arise from multimodal lock hold times. Transient preemption 1681 // can also result in apparent bimodal lock hold times. 1682 // Although sub-optimal, neither condition is particularly harmful, as 1683 // in the worst-case we'll spin when we shouldn't or vice-versa. 1684 // The maximum spin duration is rather short so the failure modes aren't bad. 1685 // To be conservative, I've tuned the gain in system to bias toward 1686 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1687 // "rings" or oscillates between spinning and not spinning. This happens 1688 // when spinning is just on the cusp of profitability, however, so the 1689 // situation is not dire. The state is benign -- there's no need to add 1690 // hysteresis control to damp the transition rate between spinning and 1691 // not spinning. 1692 1693 // Spinning: Fixed frequency (100%), vary duration 1694 int ObjectMonitor::TrySpin(Thread * Self) { 1695 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1696 int ctr = Knob_FixedSpin; 1697 if (ctr != 0) { 1698 while (--ctr >= 0) { 1699 if (TryLock(Self) > 0) return 1; 1700 SpinPause(); 1701 } 1702 return 0; 1703 } 1704 1705 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1706 if (TryLock(Self) > 0) { 1707 // Increase _SpinDuration ... 1708 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1709 // Raising _SpurDuration to the poverty line is key. 1710 int x = _SpinDuration; 1711 if (x < Knob_SpinLimit) { 1712 if (x < Knob_Poverty) x = Knob_Poverty; 1713 _SpinDuration = x + Knob_BonusB; 1714 } 1715 return 1; 1716 } 1717 SpinPause(); 1718 } 1719 1720 // Admission control - verify preconditions for spinning 1721 // 1722 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1723 // becoming an absorbing state. Put another way, we spin briefly to 1724 // sample, just in case the system load, parallelism, contention, or lock 1725 // modality changed. 1726 // 1727 // Consider the following alternative: 1728 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1729 // spin attempt. "Periodically" might mean after a tally of 1730 // the # of failed spin attempts (or iterations) reaches some threshold. 1731 // This takes us into the realm of 1-out-of-N spinning, where we 1732 // hold the duration constant but vary the frequency. 1733 1734 ctr = _SpinDuration; 1735 if (ctr <= 0) return 0; 1736 1737 if (NotRunnable(Self, (Thread *) _owner)) { 1738 return 0; 1739 } 1740 1741 // We're good to spin ... spin ingress. 1742 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1743 // when preparing to LD...CAS _owner, etc and the CAS is likely 1744 // to succeed. 1745 if (_succ == NULL) { 1746 _succ = Self; 1747 } 1748 Thread * prv = NULL; 1749 1750 // There are three ways to exit the following loop: 1751 // 1. A successful spin where this thread has acquired the lock. 1752 // 2. Spin failure with prejudice 1753 // 3. Spin failure without prejudice 1754 1755 while (--ctr >= 0) { 1756 1757 // Periodic polling -- Check for pending GC 1758 // Threads may spin while they're unsafe. 1759 // We don't want spinning threads to delay the JVM from reaching 1760 // a stop-the-world safepoint or to steal cycles from GC. 1761 // If we detect a pending safepoint we abort in order that 1762 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1763 // this thread, if safe, doesn't steal cycles from GC. 1764 // This is in keeping with the "no loitering in runtime" rule. 1765 // We periodically check to see if there's a safepoint pending. 1766 if ((ctr & 0xFF) == 0) { 1767 if (SafepointMechanism::should_block(Self)) { 1768 goto Abort; // abrupt spin egress 1769 } 1770 SpinPause(); 1771 } 1772 1773 // Probe _owner with TATAS 1774 // If this thread observes the monitor transition or flicker 1775 // from locked to unlocked to locked, then the odds that this 1776 // thread will acquire the lock in this spin attempt go down 1777 // considerably. The same argument applies if the CAS fails 1778 // or if we observe _owner change from one non-null value to 1779 // another non-null value. In such cases we might abort 1780 // the spin without prejudice or apply a "penalty" to the 1781 // spin count-down variable "ctr", reducing it by 100, say. 1782 1783 Thread * ox = (Thread *) _owner; 1784 if (ox == NULL) { 1785 ox = (Thread*)try_set_owner_from(NULL, Self); 1786 if (ox == NULL) { 1787 // The CAS succeeded -- this thread acquired ownership 1788 // Take care of some bookkeeping to exit spin state. 1789 if (_succ == Self) { 1790 _succ = NULL; 1791 } 1792 1793 // Increase _SpinDuration : 1794 // The spin was successful (profitable) so we tend toward 1795 // longer spin attempts in the future. 1796 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1797 // If we acquired the lock early in the spin cycle it 1798 // makes sense to increase _SpinDuration proportionally. 1799 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1800 int x = _SpinDuration; 1801 if (x < Knob_SpinLimit) { 1802 if (x < Knob_Poverty) x = Knob_Poverty; 1803 _SpinDuration = x + Knob_Bonus; 1804 } 1805 return 1; 1806 } 1807 1808 // The CAS failed ... we can take any of the following actions: 1809 // * penalize: ctr -= CASPenalty 1810 // * exit spin with prejudice -- goto Abort; 1811 // * exit spin without prejudice. 1812 // * Since CAS is high-latency, retry again immediately. 1813 prv = ox; 1814 goto Abort; 1815 } 1816 1817 // Did lock ownership change hands ? 1818 if (ox != prv && prv != NULL) { 1819 goto Abort; 1820 } 1821 prv = ox; 1822 1823 // Abort the spin if the owner is not executing. 1824 // The owner must be executing in order to drop the lock. 1825 // Spinning while the owner is OFFPROC is idiocy. 1826 // Consider: ctr -= RunnablePenalty ; 1827 if (NotRunnable(Self, ox)) { 1828 goto Abort; 1829 } 1830 if (_succ == NULL) { 1831 _succ = Self; 1832 } 1833 } 1834 1835 // Spin failed with prejudice -- reduce _SpinDuration. 1836 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1837 // AIMD is globally stable. 1838 { 1839 int x = _SpinDuration; 1840 if (x > 0) { 1841 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1842 // This is globally sample and tends to damp the response. 1843 x -= Knob_Penalty; 1844 if (x < 0) x = 0; 1845 _SpinDuration = x; 1846 } 1847 } 1848 1849 Abort: 1850 if (_succ == Self) { 1851 _succ = NULL; 1852 // Invariant: after setting succ=null a contending thread 1853 // must recheck-retry _owner before parking. This usually happens 1854 // in the normal usage of TrySpin(), but it's safest 1855 // to make TrySpin() as foolproof as possible. 1856 OrderAccess::fence(); 1857 if (TryLock(Self) > 0) return 1; 1858 } 1859 return 0; 1860 } 1861 1862 // NotRunnable() -- informed spinning 1863 // 1864 // Don't bother spinning if the owner is not eligible to drop the lock. 1865 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1866 // The thread must be runnable in order to drop the lock in timely fashion. 1867 // If the _owner is not runnable then spinning will not likely be 1868 // successful (profitable). 1869 // 1870 // Beware -- the thread referenced by _owner could have died 1871 // so a simply fetch from _owner->_thread_state might trap. 1872 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1873 // Because of the lifecycle issues, the _thread_state values 1874 // observed by NotRunnable() might be garbage. NotRunnable must 1875 // tolerate this and consider the observed _thread_state value 1876 // as advisory. 1877 // 1878 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1879 // a thread pointer. 1880 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1881 // with the LSB of _owner. Another option would be to probabilistically probe 1882 // the putative _owner->TypeTag value. 1883 // 1884 // Checking _thread_state isn't perfect. Even if the thread is 1885 // in_java it might be blocked on a page-fault or have been preempted 1886 // and sitting on a ready/dispatch queue. 1887 // 1888 // The return value from NotRunnable() is *advisory* -- the 1889 // result is based on sampling and is not necessarily coherent. 1890 // The caller must tolerate false-negative and false-positive errors. 1891 // Spinning, in general, is probabilistic anyway. 1892 1893 1894 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1895 // Check ox->TypeTag == 2BAD. 1896 if (ox == NULL) return 0; 1897 1898 // Avoid transitive spinning ... 1899 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1900 // Immediately after T1 acquires L it's possible that T2, also 1901 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1902 // This occurs transiently after T1 acquired L but before 1903 // T1 managed to clear T1.Stalled. T2 does not need to abort 1904 // its spin in this circumstance. 1905 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1906 1907 if (BlockedOn == 1) return 1; 1908 if (BlockedOn != 0) { 1909 return BlockedOn != intptr_t(this) && _owner == ox; 1910 } 1911 1912 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1913 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1914 // consider also: jst != _thread_in_Java -- but that's overspecific. 1915 return jst == _thread_blocked || jst == _thread_in_native; 1916 } 1917 1918 1919 // ----------------------------------------------------------------------------- 1920 // WaitSet management ... 1921 1922 ObjectWaiter::ObjectWaiter(Thread* thread) { 1923 _next = NULL; 1924 _prev = NULL; 1925 _notified = 0; 1926 _notifier_tid = 0; 1927 TState = TS_RUN; 1928 _thread = thread; 1929 _event = thread->_ParkEvent; 1930 _active = false; 1931 assert(_event != NULL, "invariant"); 1932 } 1933 1934 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1935 JavaThread *jt = (JavaThread *)this->_thread; 1936 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1937 } 1938 1939 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1940 JavaThread *jt = (JavaThread *)this->_thread; 1941 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1942 } 1943 1944 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1945 assert(node != NULL, "should not add NULL node"); 1946 assert(node->_prev == NULL, "node already in list"); 1947 assert(node->_next == NULL, "node already in list"); 1948 // put node at end of queue (circular doubly linked list) 1949 if (_WaitSet == NULL) { 1950 _WaitSet = node; 1951 node->_prev = node; 1952 node->_next = node; 1953 } else { 1954 ObjectWaiter* head = _WaitSet; 1955 ObjectWaiter* tail = head->_prev; 1956 assert(tail->_next == head, "invariant check"); 1957 tail->_next = node; 1958 head->_prev = node; 1959 node->_next = head; 1960 node->_prev = tail; 1961 } 1962 } 1963 1964 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1965 // dequeue the very first waiter 1966 ObjectWaiter* waiter = _WaitSet; 1967 if (waiter) { 1968 DequeueSpecificWaiter(waiter); 1969 } 1970 return waiter; 1971 } 1972 1973 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1974 assert(node != NULL, "should not dequeue NULL node"); 1975 assert(node->_prev != NULL, "node already removed from list"); 1976 assert(node->_next != NULL, "node already removed from list"); 1977 // when the waiter has woken up because of interrupt, 1978 // timeout or other spurious wake-up, dequeue the 1979 // waiter from waiting list 1980 ObjectWaiter* next = node->_next; 1981 if (next == node) { 1982 assert(node->_prev == node, "invariant check"); 1983 _WaitSet = NULL; 1984 } else { 1985 ObjectWaiter* prev = node->_prev; 1986 assert(prev->_next == node, "invariant check"); 1987 assert(next->_prev == node, "invariant check"); 1988 next->_prev = prev; 1989 prev->_next = next; 1990 if (_WaitSet == node) { 1991 _WaitSet = next; 1992 } 1993 } 1994 node->_next = NULL; 1995 node->_prev = NULL; 1996 } 1997 1998 // ----------------------------------------------------------------------------- 1999 // PerfData support 2000 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2001 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2002 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2003 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2004 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2005 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2006 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2007 2008 // One-shot global initialization for the sync subsystem. 2009 // We could also defer initialization and initialize on-demand 2010 // the first time we call ObjectSynchronizer::inflate(). 2011 // Initialization would be protected - like so many things - by 2012 // the MonitorCache_lock. 2013 2014 void ObjectMonitor::Initialize() { 2015 assert(!InitDone, "invariant"); 2016 2017 if (!os::is_MP()) { 2018 Knob_SpinLimit = 0; 2019 Knob_PreSpin = 0; 2020 Knob_FixedSpin = -1; 2021 } 2022 2023 if (UsePerfData) { 2024 EXCEPTION_MARK; 2025 #define NEWPERFCOUNTER(n) \ 2026 { \ 2027 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2028 CHECK); \ 2029 } 2030 #define NEWPERFVARIABLE(n) \ 2031 { \ 2032 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2033 CHECK); \ 2034 } 2035 NEWPERFCOUNTER(_sync_Inflations); 2036 NEWPERFCOUNTER(_sync_Deflations); 2037 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2038 NEWPERFCOUNTER(_sync_FutileWakeups); 2039 NEWPERFCOUNTER(_sync_Parks); 2040 NEWPERFCOUNTER(_sync_Notifications); 2041 NEWPERFVARIABLE(_sync_MonExtant); 2042 #undef NEWPERFCOUNTER 2043 #undef NEWPERFVARIABLE 2044 } 2045 2046 DEBUG_ONLY(InitDone = true;) 2047 } 2048 2049 void ObjectMonitor::print_on(outputStream* st) const { 2050 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2051 st->print("{contentions=0x%08x,waiters=0x%08x" 2052 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2053 contentions(), waiters(), recursions(), 2054 p2i(owner())); 2055 } 2056 void ObjectMonitor::print() const { print_on(tty); } 2057 2058 #ifdef ASSERT 2059 // Print the ObjectMonitor like a debugger would: 2060 // 2061 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2062 // _header = 0x0000000000000001 2063 // _object = 0x000000070ff45fd0 2064 // _allocation_state = Old 2065 // _pad_buf0 = { 2066 // [0] = '\0' 2067 // ... 2068 // [43] = '\0' 2069 // } 2070 // _owner = 0x0000000000000000 2071 // _previous_owner_tid = 0 2072 // _pad_buf1 = { 2073 // [0] = '\0' 2074 // ... 2075 // [47] = '\0' 2076 // } 2077 // _next_om = 0x0000000000000000 2078 // _recursions = 0 2079 // _EntryList = 0x0000000000000000 2080 // _cxq = 0x0000000000000000 2081 // _succ = 0x0000000000000000 2082 // _Responsible = 0x0000000000000000 2083 // _Spinner = 0 2084 // _SpinDuration = 5000 2085 // _contentions = 0 2086 // _WaitSet = 0x0000700009756248 2087 // _waiters = 1 2088 // _WaitSetLock = 0 2089 // } 2090 // 2091 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2092 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2093 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2094 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 2095 st->print(" _allocation_state = "); 2096 if (is_free()) { 2097 st->print("Free"); 2098 } else if (is_old()) { 2099 st->print("Old"); 2100 } else if (is_new()) { 2101 st->print("New"); 2102 } else { 2103 st->print("unknown=%d", _allocation_state); 2104 } 2105 st->cr(); 2106 st->print_cr(" _pad_buf0 = {"); 2107 st->print_cr(" [0] = '\\0'"); 2108 st->print_cr(" ..."); 2109 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2110 st->print_cr(" }"); 2111 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 2112 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 2113 st->print_cr(" _pad_buf1 = {"); 2114 st->print_cr(" [0] = '\\0'"); 2115 st->print_cr(" ..."); 2116 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2117 st->print_cr(" }"); 2118 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2119 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2120 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2121 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2122 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2123 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2124 st->print_cr(" _Spinner = %d", _Spinner); 2125 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2126 st->print_cr(" _contentions = %d", contentions()); 2127 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2128 st->print_cr(" _waiters = %d", _waiters); 2129 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2130 st->print_cr("}"); 2131 } 2132 #endif