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 add_to_contentions(1); 298 if (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 add_to_contentions(-1); 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 add_to_contentions(-1); 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 // Separate loads in is_being_async_deflated(), which is almost always 443 // called before this function, from the load of dmw/header below. 444 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 445 // A non-multiple copy atomic (nMCA) machine needs a bigger 446 // hammer to separate the loads before and the load below. 447 OrderAccess::fence(); 448 } else { 449 OrderAccess::loadload(); 450 } 451 452 const oop l_object = (oop)object(); 453 if (l_object == NULL) { 454 // ObjectMonitor's object ref has already been cleared by async 455 // deflation so we're done here. 456 return; 457 } 458 ADIM_guarantee(l_object == obj, "object=" INTPTR_FORMAT " must equal obj=" 459 INTPTR_FORMAT, p2i(l_object), p2i(obj)); 460 461 markWord dmw = header(); 462 // The dmw has to be neutral (not NULL, not locked and not marked). 463 ADIM_guarantee(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 464 465 // Install displaced mark word if the object's header still points 466 // to this ObjectMonitor. More than one racing caller to this function 467 // can rarely reach this point, but only one can win. 468 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 469 if (res != markWord::encode(this)) { 470 // This should be rare so log at the Info level when it happens. 471 log_info(monitorinflation)("install_displaced_markword_in_object: " 472 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 473 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 474 dmw.value(), markWord::encode(this).value(), 475 res.value()); 476 } 477 478 // Note: It does not matter which thread restored the header/dmw 479 // into the object's header. The thread deflating the monitor just 480 // wanted the object's header restored and it is. The threads that 481 // detected a race with the deflation process also wanted the 482 // object's header restored before they retry their operation and 483 // because it is restored they will only retry once. 484 } 485 486 // Convert the fields used by is_busy() to a string that can be 487 // used for diagnostic output. 488 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 489 ss->print("is_busy: waiters=%d, ", _waiters); 490 if (!AsyncDeflateIdleMonitors) { 491 ss->print("contentions=%d, ", contentions()); 492 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 493 } else { 494 if (contentions() > 0) { 495 ss->print("contentions=%d, ", contentions()); 496 } else { 497 ss->print("contentions=0"); 498 } 499 if (_owner != DEFLATER_MARKER) { 500 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 501 } else { 502 // We report NULL instead of DEFLATER_MARKER here because is_busy() 503 // ignores DEFLATER_MARKER values. 504 ss->print("owner=" INTPTR_FORMAT, NULL); 505 } 506 } 507 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq), 508 p2i(_EntryList)); 509 return ss->base(); 510 } 511 512 #define MAX_RECHECK_INTERVAL 1000 513 514 void ObjectMonitor::EnterI(TRAPS) { 515 Thread * const Self = THREAD; 516 assert(Self->is_Java_thread(), "invariant"); 517 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 518 519 // Try the lock - TATAS 520 if (TryLock (Self) > 0) { 521 assert(_succ != Self, "invariant"); 522 assert(_owner == Self, "invariant"); 523 assert(_Responsible != Self, "invariant"); 524 return; 525 } 526 527 if (AsyncDeflateIdleMonitors && 528 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 529 // The deflation protocol finished the first part (setting owner), 530 // but it failed the second part (making contentions negative) and 531 // bailed. Acquired the monitor. 532 assert(_succ != Self, "invariant"); 533 assert(_Responsible != Self, "invariant"); 534 return; 535 } 536 537 assert(InitDone, "Unexpectedly not initialized"); 538 539 // We try one round of spinning *before* enqueueing Self. 540 // 541 // If the _owner is ready but OFFPROC we could use a YieldTo() 542 // operation to donate the remainder of this thread's quantum 543 // to the owner. This has subtle but beneficial affinity 544 // effects. 545 546 if (TrySpin(Self) > 0) { 547 assert(_owner == Self, "invariant"); 548 assert(_succ != Self, "invariant"); 549 assert(_Responsible != Self, "invariant"); 550 return; 551 } 552 553 // The Spin failed -- Enqueue and park the thread ... 554 assert(_succ != Self, "invariant"); 555 assert(_owner != Self, "invariant"); 556 assert(_Responsible != Self, "invariant"); 557 558 // Enqueue "Self" on ObjectMonitor's _cxq. 559 // 560 // Node acts as a proxy for Self. 561 // As an aside, if were to ever rewrite the synchronization code mostly 562 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 563 // Java objects. This would avoid awkward lifecycle and liveness issues, 564 // as well as eliminate a subset of ABA issues. 565 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 566 567 ObjectWaiter node(Self); 568 Self->_ParkEvent->reset(); 569 node._prev = (ObjectWaiter *) 0xBAD; 570 node.TState = ObjectWaiter::TS_CXQ; 571 572 // Push "Self" onto the front of the _cxq. 573 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 574 // Note that spinning tends to reduce the rate at which threads 575 // enqueue and dequeue on EntryList|cxq. 576 ObjectWaiter * nxt; 577 for (;;) { 578 node._next = nxt = _cxq; 579 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 580 581 // Interference - the CAS failed because _cxq changed. Just retry. 582 // As an optional optimization we retry the lock. 583 if (TryLock (Self) > 0) { 584 assert(_succ != Self, "invariant"); 585 assert(_owner == Self, "invariant"); 586 assert(_Responsible != Self, "invariant"); 587 return; 588 } 589 } 590 591 // Check for cxq|EntryList edge transition to non-null. This indicates 592 // the onset of contention. While contention persists exiting threads 593 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 594 // operations revert to the faster 1-0 mode. This enter operation may interleave 595 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 596 // arrange for one of the contending thread to use a timed park() operations 597 // to detect and recover from the race. (Stranding is form of progress failure 598 // where the monitor is unlocked but all the contending threads remain parked). 599 // That is, at least one of the contended threads will periodically poll _owner. 600 // One of the contending threads will become the designated "Responsible" thread. 601 // The Responsible thread uses a timed park instead of a normal indefinite park 602 // operation -- it periodically wakes and checks for and recovers from potential 603 // strandings admitted by 1-0 exit operations. We need at most one Responsible 604 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 605 // be responsible for a monitor. 606 // 607 // Currently, one of the contended threads takes on the added role of "Responsible". 608 // A viable alternative would be to use a dedicated "stranding checker" thread 609 // that periodically iterated over all the threads (or active monitors) and unparked 610 // successors where there was risk of stranding. This would help eliminate the 611 // timer scalability issues we see on some platforms as we'd only have one thread 612 // -- the checker -- parked on a timer. 613 614 if (nxt == NULL && _EntryList == NULL) { 615 // Try to assume the role of responsible thread for the monitor. 616 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 617 Atomic::replace_if_null(&_Responsible, Self); 618 } 619 620 // The lock might have been released while this thread was occupied queueing 621 // itself onto _cxq. To close the race and avoid "stranding" and 622 // progress-liveness failure we must resample-retry _owner before parking. 623 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 624 // In this case the ST-MEMBAR is accomplished with CAS(). 625 // 626 // TODO: Defer all thread state transitions until park-time. 627 // Since state transitions are heavy and inefficient we'd like 628 // to defer the state transitions until absolutely necessary, 629 // and in doing so avoid some transitions ... 630 631 int nWakeups = 0; 632 int recheckInterval = 1; 633 634 for (;;) { 635 636 if (TryLock(Self) > 0) break; 637 assert(_owner != Self, "invariant"); 638 639 // park self 640 if (_Responsible == Self) { 641 Self->_ParkEvent->park((jlong) recheckInterval); 642 // Increase the recheckInterval, but clamp the value. 643 recheckInterval *= 8; 644 if (recheckInterval > MAX_RECHECK_INTERVAL) { 645 recheckInterval = MAX_RECHECK_INTERVAL; 646 } 647 } else { 648 Self->_ParkEvent->park(); 649 } 650 651 if (TryLock(Self) > 0) break; 652 653 if (AsyncDeflateIdleMonitors && 654 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 655 // The deflation protocol finished the first part (setting owner), 656 // but it failed the second part (making contentions negative) and 657 // bailed. Acquired the monitor. 658 break; 659 } 660 661 // The lock is still contested. 662 // Keep a tally of the # of futile wakeups. 663 // Note that the counter is not protected by a lock or updated by atomics. 664 // That is by design - we trade "lossy" counters which are exposed to 665 // races during updates for a lower probe effect. 666 667 // This PerfData object can be used in parallel with a safepoint. 668 // See the work around in PerfDataManager::destroy(). 669 OM_PERFDATA_OP(FutileWakeups, inc()); 670 ++nWakeups; 671 672 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 673 // We can defer clearing _succ until after the spin completes 674 // TrySpin() must tolerate being called with _succ == Self. 675 // Try yet another round of adaptive spinning. 676 if (TrySpin(Self) > 0) break; 677 678 // We can find that we were unpark()ed and redesignated _succ while 679 // we were spinning. That's harmless. If we iterate and call park(), 680 // park() will consume the event and return immediately and we'll 681 // just spin again. This pattern can repeat, leaving _succ to simply 682 // spin on a CPU. 683 684 if (_succ == Self) _succ = NULL; 685 686 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 687 OrderAccess::fence(); 688 } 689 690 // Egress : 691 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 692 // Normally we'll find Self on the EntryList . 693 // From the perspective of the lock owner (this thread), the 694 // EntryList is stable and cxq is prepend-only. 695 // The head of cxq is volatile but the interior is stable. 696 // In addition, Self.TState is stable. 697 698 assert(_owner == Self, "invariant"); 699 assert(object() != NULL, "invariant"); 700 // I'd like to write: 701 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 702 // but as we're at a safepoint that's not safe. 703 704 UnlinkAfterAcquire(Self, &node); 705 if (_succ == Self) _succ = NULL; 706 707 assert(_succ != Self, "invariant"); 708 if (_Responsible == Self) { 709 _Responsible = NULL; 710 OrderAccess::fence(); // Dekker pivot-point 711 712 // We may leave threads on cxq|EntryList without a designated 713 // "Responsible" thread. This is benign. When this thread subsequently 714 // exits the monitor it can "see" such preexisting "old" threads -- 715 // threads that arrived on the cxq|EntryList before the fence, above -- 716 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 717 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 718 // non-null and elect a new "Responsible" timer thread. 719 // 720 // This thread executes: 721 // ST Responsible=null; MEMBAR (in enter epilogue - here) 722 // LD cxq|EntryList (in subsequent exit) 723 // 724 // Entering threads in the slow/contended path execute: 725 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 726 // The (ST cxq; MEMBAR) is accomplished with CAS(). 727 // 728 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 729 // exit operation from floating above the ST Responsible=null. 730 } 731 732 // We've acquired ownership with CAS(). 733 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 734 // But since the CAS() this thread may have also stored into _succ, 735 // EntryList, cxq or Responsible. These meta-data updates must be 736 // visible __before this thread subsequently drops the lock. 737 // Consider what could occur if we didn't enforce this constraint -- 738 // STs to monitor meta-data and user-data could reorder with (become 739 // visible after) the ST in exit that drops ownership of the lock. 740 // Some other thread could then acquire the lock, but observe inconsistent 741 // or old monitor meta-data and heap data. That violates the JMM. 742 // To that end, the 1-0 exit() operation must have at least STST|LDST 743 // "release" barrier semantics. Specifically, there must be at least a 744 // STST|LDST barrier in exit() before the ST of null into _owner that drops 745 // the lock. The barrier ensures that changes to monitor meta-data and data 746 // protected by the lock will be visible before we release the lock, and 747 // therefore before some other thread (CPU) has a chance to acquire the lock. 748 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 749 // 750 // Critically, any prior STs to _succ or EntryList must be visible before 751 // the ST of null into _owner in the *subsequent* (following) corresponding 752 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 753 // execute a serializing instruction. 754 755 return; 756 } 757 758 // ReenterI() is a specialized inline form of the latter half of the 759 // contended slow-path from EnterI(). We use ReenterI() only for 760 // monitor reentry in wait(). 761 // 762 // In the future we should reconcile EnterI() and ReenterI(). 763 764 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 765 assert(Self != NULL, "invariant"); 766 assert(SelfNode != NULL, "invariant"); 767 assert(SelfNode->_thread == Self, "invariant"); 768 assert(_waiters > 0, "invariant"); 769 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 770 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 771 JavaThread * jt = (JavaThread *) Self; 772 773 int nWakeups = 0; 774 for (;;) { 775 ObjectWaiter::TStates v = SelfNode->TState; 776 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 777 assert(_owner != Self, "invariant"); 778 779 if (TryLock(Self) > 0) break; 780 if (TrySpin(Self) > 0) break; 781 782 // State transition wrappers around park() ... 783 // ReenterI() wisely defers state transitions until 784 // it's clear we must park the thread. 785 { 786 OSThreadContendState osts(Self->osthread()); 787 ThreadBlockInVM tbivm(jt); 788 789 // cleared by handle_special_suspend_equivalent_condition() 790 // or java_suspend_self() 791 jt->set_suspend_equivalent(); 792 Self->_ParkEvent->park(); 793 794 // were we externally suspended while we were waiting? 795 for (;;) { 796 if (!ExitSuspendEquivalent(jt)) break; 797 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 798 jt->java_suspend_self(); 799 jt->set_suspend_equivalent(); 800 } 801 } 802 803 // Try again, but just so we distinguish between futile wakeups and 804 // successful wakeups. The following test isn't algorithmically 805 // necessary, but it helps us maintain sensible statistics. 806 if (TryLock(Self) > 0) break; 807 808 // The lock is still contested. 809 // Keep a tally of the # of futile wakeups. 810 // Note that the counter is not protected by a lock or updated by atomics. 811 // That is by design - we trade "lossy" counters which are exposed to 812 // races during updates for a lower probe effect. 813 ++nWakeups; 814 815 // Assuming this is not a spurious wakeup we'll normally 816 // find that _succ == Self. 817 if (_succ == Self) _succ = NULL; 818 819 // Invariant: after clearing _succ a contending thread 820 // *must* retry _owner before parking. 821 OrderAccess::fence(); 822 823 // This PerfData object can be used in parallel with a safepoint. 824 // See the work around in PerfDataManager::destroy(). 825 OM_PERFDATA_OP(FutileWakeups, inc()); 826 } 827 828 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 829 // Normally we'll find Self on the EntryList. 830 // Unlinking from the EntryList is constant-time and atomic-free. 831 // From the perspective of the lock owner (this thread), the 832 // EntryList is stable and cxq is prepend-only. 833 // The head of cxq is volatile but the interior is stable. 834 // In addition, Self.TState is stable. 835 836 assert(_owner == Self, "invariant"); 837 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 838 UnlinkAfterAcquire(Self, SelfNode); 839 if (_succ == Self) _succ = NULL; 840 assert(_succ != Self, "invariant"); 841 SelfNode->TState = ObjectWaiter::TS_RUN; 842 OrderAccess::fence(); // see comments at the end of EnterI() 843 } 844 845 // By convention we unlink a contending thread from EntryList|cxq immediately 846 // after the thread acquires the lock in ::enter(). Equally, we could defer 847 // unlinking the thread until ::exit()-time. 848 849 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 850 assert(_owner == Self, "invariant"); 851 assert(SelfNode->_thread == Self, "invariant"); 852 853 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 854 // Normal case: remove Self from the DLL EntryList . 855 // This is a constant-time operation. 856 ObjectWaiter * nxt = SelfNode->_next; 857 ObjectWaiter * prv = SelfNode->_prev; 858 if (nxt != NULL) nxt->_prev = prv; 859 if (prv != NULL) prv->_next = nxt; 860 if (SelfNode == _EntryList) _EntryList = nxt; 861 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 862 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 863 } else { 864 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 865 // Inopportune interleaving -- Self is still on the cxq. 866 // This usually means the enqueue of self raced an exiting thread. 867 // Normally we'll find Self near the front of the cxq, so 868 // dequeueing is typically fast. If needbe we can accelerate 869 // this with some MCS/CHL-like bidirectional list hints and advisory 870 // back-links so dequeueing from the interior will normally operate 871 // in constant-time. 872 // Dequeue Self from either the head (with CAS) or from the interior 873 // with a linear-time scan and normal non-atomic memory operations. 874 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 875 // and then unlink Self from EntryList. We have to drain eventually, 876 // so it might as well be now. 877 878 ObjectWaiter * v = _cxq; 879 assert(v != NULL, "invariant"); 880 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) { 881 // The CAS above can fail from interference IFF a "RAT" arrived. 882 // In that case Self must be in the interior and can no longer be 883 // at the head of cxq. 884 if (v == SelfNode) { 885 assert(_cxq != v, "invariant"); 886 v = _cxq; // CAS above failed - start scan at head of list 887 } 888 ObjectWaiter * p; 889 ObjectWaiter * q = NULL; 890 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 891 q = p; 892 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 893 } 894 assert(v != SelfNode, "invariant"); 895 assert(p == SelfNode, "Node not found on cxq"); 896 assert(p != _cxq, "invariant"); 897 assert(q != NULL, "invariant"); 898 assert(q->_next == p, "invariant"); 899 q->_next = p->_next; 900 } 901 } 902 903 #ifdef ASSERT 904 // Diagnostic hygiene ... 905 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 906 SelfNode->_next = (ObjectWaiter *) 0xBAD; 907 SelfNode->TState = ObjectWaiter::TS_RUN; 908 #endif 909 } 910 911 // ----------------------------------------------------------------------------- 912 // Exit support 913 // 914 // exit() 915 // ~~~~~~ 916 // Note that the collector can't reclaim the objectMonitor or deflate 917 // the object out from underneath the thread calling ::exit() as the 918 // thread calling ::exit() never transitions to a stable state. 919 // This inhibits GC, which in turn inhibits asynchronous (and 920 // inopportune) reclamation of "this". 921 // 922 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 923 // There's one exception to the claim above, however. EnterI() can call 924 // exit() to drop a lock if the acquirer has been externally suspended. 925 // In that case exit() is called with _thread_state == _thread_blocked, 926 // but the monitor's _contentions field is > 0, which inhibits reclamation. 927 // 928 // 1-0 exit 929 // ~~~~~~~~ 930 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 931 // the fast-path operators have been optimized so the common ::exit() 932 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 933 // The code emitted by fast_unlock() elides the usual MEMBAR. This 934 // greatly improves latency -- MEMBAR and CAS having considerable local 935 // latency on modern processors -- but at the cost of "stranding". Absent the 936 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 937 // ::enter() path, resulting in the entering thread being stranding 938 // and a progress-liveness failure. Stranding is extremely rare. 939 // We use timers (timed park operations) & periodic polling to detect 940 // and recover from stranding. Potentially stranded threads periodically 941 // wake up and poll the lock. See the usage of the _Responsible variable. 942 // 943 // The CAS() in enter provides for safety and exclusion, while the CAS or 944 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 945 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 946 // We detect and recover from stranding with timers. 947 // 948 // If a thread transiently strands it'll park until (a) another 949 // thread acquires the lock and then drops the lock, at which time the 950 // exiting thread will notice and unpark the stranded thread, or, (b) 951 // the timer expires. If the lock is high traffic then the stranding latency 952 // will be low due to (a). If the lock is low traffic then the odds of 953 // stranding are lower, although the worst-case stranding latency 954 // is longer. Critically, we don't want to put excessive load in the 955 // platform's timer subsystem. We want to minimize both the timer injection 956 // rate (timers created/sec) as well as the number of timers active at 957 // any one time. (more precisely, we want to minimize timer-seconds, which is 958 // the integral of the # of active timers at any instant over time). 959 // Both impinge on OS scalability. Given that, at most one thread parked on 960 // a monitor will use a timer. 961 // 962 // There is also the risk of a futile wake-up. If we drop the lock 963 // another thread can reacquire the lock immediately, and we can 964 // then wake a thread unnecessarily. This is benign, and we've 965 // structured the code so the windows are short and the frequency 966 // of such futile wakups is low. 967 968 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 969 Thread* const Self = THREAD; 970 void* cur = Atomic::load(&_owner); 971 if (THREAD != cur) { 972 if (THREAD->is_lock_owned((address)cur)) { 973 assert(_recursions == 0, "invariant"); 974 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 975 _recursions = 0; 976 } else { 977 // Apparent unbalanced locking ... 978 // Naively we'd like to throw IllegalMonitorStateException. 979 // As a practical matter we can neither allocate nor throw an 980 // exception as ::exit() can be called from leaf routines. 981 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 982 // Upon deeper reflection, however, in a properly run JVM the only 983 // way we should encounter this situation is in the presence of 984 // unbalanced JNI locking. TODO: CheckJNICalls. 985 // See also: CR4414101 986 #ifdef ASSERT 987 LogStreamHandle(Error, monitorinflation) lsh; 988 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 989 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 990 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 991 print_debug_style_on(&lsh); 992 #endif 993 assert(false, "Non-balanced monitor enter/exit!"); 994 return; 995 } 996 } 997 998 if (_recursions != 0) { 999 _recursions--; // this is simple recursive enter 1000 return; 1001 } 1002 1003 // Invariant: after setting Responsible=null an thread must execute 1004 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1005 _Responsible = NULL; 1006 1007 #if INCLUDE_JFR 1008 // get the owner's thread id for the MonitorEnter event 1009 // if it is enabled and the thread isn't suspended 1010 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1011 _previous_owner_tid = JFR_THREAD_ID(Self); 1012 } 1013 #endif 1014 1015 for (;;) { 1016 assert(THREAD == _owner, "invariant"); 1017 1018 // Drop the lock. 1019 // release semantics: prior loads and stores from within the critical section 1020 // must not float (reorder) past the following store that drops the lock. 1021 // Uses a storeload to separate release_store(owner) from the 1022 // successor check. The try_set_owner() below uses cmpxchg() so 1023 // we get the fence down there. 1024 release_clear_owner(Self); 1025 OrderAccess::storeload(); 1026 1027 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1028 return; 1029 } 1030 // Other threads are blocked trying to acquire the lock. 1031 1032 // Normally the exiting thread is responsible for ensuring succession, 1033 // but if other successors are ready or other entering threads are spinning 1034 // then this thread can simply store NULL into _owner and exit without 1035 // waking a successor. The existence of spinners or ready successors 1036 // guarantees proper succession (liveness). Responsibility passes to the 1037 // ready or running successors. The exiting thread delegates the duty. 1038 // More precisely, if a successor already exists this thread is absolved 1039 // of the responsibility of waking (unparking) one. 1040 // 1041 // The _succ variable is critical to reducing futile wakeup frequency. 1042 // _succ identifies the "heir presumptive" thread that has been made 1043 // ready (unparked) but that has not yet run. We need only one such 1044 // successor thread to guarantee progress. 1045 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1046 // section 3.3 "Futile Wakeup Throttling" for details. 1047 // 1048 // Note that spinners in Enter() also set _succ non-null. 1049 // In the current implementation spinners opportunistically set 1050 // _succ so that exiting threads might avoid waking a successor. 1051 // Another less appealing alternative would be for the exiting thread 1052 // to drop the lock and then spin briefly to see if a spinner managed 1053 // to acquire the lock. If so, the exiting thread could exit 1054 // immediately without waking a successor, otherwise the exiting 1055 // thread would need to dequeue and wake a successor. 1056 // (Note that we'd need to make the post-drop spin short, but no 1057 // shorter than the worst-case round-trip cache-line migration time. 1058 // The dropped lock needs to become visible to the spinner, and then 1059 // the acquisition of the lock by the spinner must become visible to 1060 // the exiting thread). 1061 1062 // It appears that an heir-presumptive (successor) must be made ready. 1063 // Only the current lock owner can manipulate the EntryList or 1064 // drain _cxq, so we need to reacquire the lock. If we fail 1065 // to reacquire the lock the responsibility for ensuring succession 1066 // falls to the new owner. 1067 // 1068 if (try_set_owner_from(NULL, Self) != NULL) { 1069 return; 1070 } 1071 1072 guarantee(_owner == THREAD, "invariant"); 1073 1074 ObjectWaiter * w = NULL; 1075 1076 w = _EntryList; 1077 if (w != NULL) { 1078 // I'd like to write: guarantee (w->_thread != Self). 1079 // But in practice an exiting thread may find itself on the EntryList. 1080 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1081 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1082 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1083 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1084 // release the lock "O". T2 resumes immediately after the ST of null into 1085 // _owner, above. T2 notices that the EntryList is populated, so it 1086 // reacquires the lock and then finds itself on the EntryList. 1087 // Given all that, we have to tolerate the circumstance where "w" is 1088 // associated with Self. 1089 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1090 ExitEpilog(Self, w); 1091 return; 1092 } 1093 1094 // If we find that both _cxq and EntryList are null then just 1095 // re-run the exit protocol from the top. 1096 w = _cxq; 1097 if (w == NULL) continue; 1098 1099 // Drain _cxq into EntryList - bulk transfer. 1100 // First, detach _cxq. 1101 // The following loop is tantamount to: w = swap(&cxq, NULL) 1102 for (;;) { 1103 assert(w != NULL, "Invariant"); 1104 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL); 1105 if (u == w) break; 1106 w = u; 1107 } 1108 1109 assert(w != NULL, "invariant"); 1110 assert(_EntryList == NULL, "invariant"); 1111 1112 // Convert the LIFO SLL anchored by _cxq into a DLL. 1113 // The list reorganization step operates in O(LENGTH(w)) time. 1114 // It's critical that this step operate quickly as 1115 // "Self" still holds the outer-lock, restricting parallelism 1116 // and effectively lengthening the critical section. 1117 // Invariant: s chases t chases u. 1118 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1119 // we have faster access to the tail. 1120 1121 _EntryList = w; 1122 ObjectWaiter * q = NULL; 1123 ObjectWaiter * p; 1124 for (p = w; p != NULL; p = p->_next) { 1125 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1126 p->TState = ObjectWaiter::TS_ENTER; 1127 p->_prev = q; 1128 q = p; 1129 } 1130 1131 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1132 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1133 1134 // See if we can abdicate to a spinner instead of waking a thread. 1135 // A primary goal of the implementation is to reduce the 1136 // context-switch rate. 1137 if (_succ != NULL) continue; 1138 1139 w = _EntryList; 1140 if (w != NULL) { 1141 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1142 ExitEpilog(Self, w); 1143 return; 1144 } 1145 } 1146 } 1147 1148 // ExitSuspendEquivalent: 1149 // A faster alternate to handle_special_suspend_equivalent_condition() 1150 // 1151 // handle_special_suspend_equivalent_condition() unconditionally 1152 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1153 // operations have high latency. Note that in ::enter() we call HSSEC 1154 // while holding the monitor, so we effectively lengthen the critical sections. 1155 // 1156 // There are a number of possible solutions: 1157 // 1158 // A. To ameliorate the problem we might also defer state transitions 1159 // to as late as possible -- just prior to parking. 1160 // Given that, we'd call HSSEC after having returned from park(), 1161 // but before attempting to acquire the monitor. This is only a 1162 // partial solution. It avoids calling HSSEC while holding the 1163 // monitor (good), but it still increases successor reacquisition latency -- 1164 // the interval between unparking a successor and the time the successor 1165 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1166 // If we use this technique we can also avoid EnterI()-exit() loop 1167 // in ::enter() where we iteratively drop the lock and then attempt 1168 // to reacquire it after suspending. 1169 // 1170 // B. In the future we might fold all the suspend bits into a 1171 // composite per-thread suspend flag and then update it with CAS(). 1172 // Alternately, a Dekker-like mechanism with multiple variables 1173 // would suffice: 1174 // ST Self->_suspend_equivalent = false 1175 // MEMBAR 1176 // LD Self_>_suspend_flags 1177 1178 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1179 return jSelf->handle_special_suspend_equivalent_condition(); 1180 } 1181 1182 1183 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1184 assert(_owner == Self, "invariant"); 1185 1186 // Exit protocol: 1187 // 1. ST _succ = wakee 1188 // 2. membar #loadstore|#storestore; 1189 // 2. ST _owner = NULL 1190 // 3. unpark(wakee) 1191 1192 _succ = Wakee->_thread; 1193 ParkEvent * Trigger = Wakee->_event; 1194 1195 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1196 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1197 // out-of-scope (non-extant). 1198 Wakee = NULL; 1199 1200 // Drop the lock. 1201 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1202 release_clear_owner(Self); 1203 OrderAccess::fence(); 1204 1205 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1206 Trigger->unpark(); 1207 1208 // Maintain stats and report events to JVMTI 1209 OM_PERFDATA_OP(Parks, inc()); 1210 } 1211 1212 1213 // ----------------------------------------------------------------------------- 1214 // Class Loader deadlock handling. 1215 // 1216 // complete_exit exits a lock returning recursion count 1217 // complete_exit/reenter operate as a wait without waiting 1218 // complete_exit requires an inflated monitor 1219 // The _owner field is not always the Thread addr even with an 1220 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1221 // thread due to contention. 1222 intx ObjectMonitor::complete_exit(TRAPS) { 1223 Thread * const Self = THREAD; 1224 assert(Self->is_Java_thread(), "Must be Java thread!"); 1225 JavaThread *jt = (JavaThread *)THREAD; 1226 1227 assert(InitDone, "Unexpectedly not initialized"); 1228 1229 void* cur = Atomic::load(&_owner); 1230 if (THREAD != cur) { 1231 if (THREAD->is_lock_owned((address)cur)) { 1232 assert(_recursions == 0, "internal state error"); 1233 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 1234 _recursions = 0; 1235 } 1236 } 1237 1238 guarantee(Self == _owner, "complete_exit not owner"); 1239 intx save = _recursions; // record the old recursion count 1240 _recursions = 0; // set the recursion level to be 0 1241 exit(true, Self); // exit the monitor 1242 guarantee(_owner != Self, "invariant"); 1243 return save; 1244 } 1245 1246 // reenter() enters a lock and sets recursion count 1247 // complete_exit/reenter operate as a wait without waiting 1248 bool ObjectMonitor::reenter(intx recursions, TRAPS) { 1249 Thread * const Self = THREAD; 1250 assert(Self->is_Java_thread(), "Must be Java thread!"); 1251 JavaThread *jt = (JavaThread *)THREAD; 1252 1253 guarantee(_owner != Self, "reenter already owner"); 1254 if (!enter(THREAD)) { 1255 return false; 1256 } 1257 // Entered the monitor. 1258 guarantee(_recursions == 0, "reenter recursion"); 1259 _recursions = recursions; 1260 return true; 1261 } 1262 1263 // Checks that the current THREAD owns this monitor and causes an 1264 // immediate return if it doesn't. We don't use the CHECK macro 1265 // because we want the IMSE to be the only exception that is thrown 1266 // from the call site when false is returned. Any other pending 1267 // exception is ignored. 1268 #define CHECK_OWNER() \ 1269 do { \ 1270 if (!check_owner(THREAD)) { \ 1271 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1272 return; \ 1273 } \ 1274 } while (false) 1275 1276 // Returns true if the specified thread owns the ObjectMonitor. 1277 // Otherwise returns false and throws IllegalMonitorStateException 1278 // (IMSE). If there is a pending exception and the specified thread 1279 // is not the owner, that exception will be replaced by the IMSE. 1280 bool ObjectMonitor::check_owner(Thread* THREAD) { 1281 void* cur = Atomic::load(&_owner); 1282 if (cur == THREAD) { 1283 return true; 1284 } 1285 if (THREAD->is_lock_owned((address)cur)) { 1286 set_owner_from_BasicLock(cur, THREAD); // Convert from BasicLock* to Thread*. 1287 _recursions = 0; 1288 return true; 1289 } 1290 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1291 "current thread is not owner", false); 1292 } 1293 1294 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1295 ObjectMonitor* monitor, 1296 jlong notifier_tid, 1297 jlong timeout, 1298 bool timedout) { 1299 assert(event != NULL, "invariant"); 1300 assert(monitor != NULL, "invariant"); 1301 event->set_monitorClass(((oop)monitor->object())->klass()); 1302 event->set_timeout(timeout); 1303 event->set_address((uintptr_t)monitor->object_addr()); 1304 event->set_notifier(notifier_tid); 1305 event->set_timedOut(timedout); 1306 event->commit(); 1307 } 1308 1309 // ----------------------------------------------------------------------------- 1310 // Wait/Notify/NotifyAll 1311 // 1312 // Note: a subset of changes to ObjectMonitor::wait() 1313 // will need to be replicated in complete_exit 1314 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1315 Thread * const Self = THREAD; 1316 assert(Self->is_Java_thread(), "Must be Java thread!"); 1317 JavaThread *jt = (JavaThread *)THREAD; 1318 1319 assert(InitDone, "Unexpectedly not initialized"); 1320 1321 CHECK_OWNER(); // Throws IMSE if not owner. 1322 1323 EventJavaMonitorWait event; 1324 1325 // check for a pending interrupt 1326 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1327 // post monitor waited event. Note that this is past-tense, we are done waiting. 1328 if (JvmtiExport::should_post_monitor_waited()) { 1329 // Note: 'false' parameter is passed here because the 1330 // wait was not timed out due to thread interrupt. 1331 JvmtiExport::post_monitor_waited(jt, this, false); 1332 1333 // In this short circuit of the monitor wait protocol, the 1334 // current thread never drops ownership of the monitor and 1335 // never gets added to the wait queue so the current thread 1336 // cannot be made the successor. This means that the 1337 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1338 // consume an unpark() meant for the ParkEvent associated with 1339 // this ObjectMonitor. 1340 } 1341 if (event.should_commit()) { 1342 post_monitor_wait_event(&event, this, 0, millis, false); 1343 } 1344 THROW(vmSymbols::java_lang_InterruptedException()); 1345 return; 1346 } 1347 1348 assert(Self->_Stalled == 0, "invariant"); 1349 Self->_Stalled = intptr_t(this); 1350 jt->set_current_waiting_monitor(this); 1351 1352 // create a node to be put into the queue 1353 // Critically, after we reset() the event but prior to park(), we must check 1354 // for a pending interrupt. 1355 ObjectWaiter node(Self); 1356 node.TState = ObjectWaiter::TS_WAIT; 1357 Self->_ParkEvent->reset(); 1358 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1359 1360 // Enter the waiting queue, which is a circular doubly linked list in this case 1361 // but it could be a priority queue or any data structure. 1362 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1363 // by the the owner of the monitor *except* in the case where park() 1364 // returns because of a timeout of interrupt. Contention is exceptionally rare 1365 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1366 1367 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1368 AddWaiter(&node); 1369 Thread::SpinRelease(&_WaitSetLock); 1370 1371 _Responsible = NULL; 1372 1373 intx save = _recursions; // record the old recursion count 1374 _waiters++; // increment the number of waiters 1375 _recursions = 0; // set the recursion level to be 1 1376 exit(true, Self); // exit the monitor 1377 guarantee(_owner != Self, "invariant"); 1378 1379 // The thread is on the WaitSet list - now park() it. 1380 // On MP systems it's conceivable that a brief spin before we park 1381 // could be profitable. 1382 // 1383 // TODO-FIXME: change the following logic to a loop of the form 1384 // while (!timeout && !interrupted && _notified == 0) park() 1385 1386 int ret = OS_OK; 1387 int WasNotified = 0; 1388 1389 // Need to check interrupt state whilst still _thread_in_vm 1390 bool interrupted = interruptible && jt->is_interrupted(false); 1391 1392 { // State transition wrappers 1393 OSThread* osthread = Self->osthread(); 1394 OSThreadWaitState osts(osthread, true); 1395 { 1396 ThreadBlockInVM tbivm(jt); 1397 // Thread is in thread_blocked state and oop access is unsafe. 1398 jt->set_suspend_equivalent(); 1399 1400 if (interrupted || HAS_PENDING_EXCEPTION) { 1401 // Intentionally empty 1402 } else if (node._notified == 0) { 1403 if (millis <= 0) { 1404 Self->_ParkEvent->park(); 1405 } else { 1406 ret = Self->_ParkEvent->park(millis); 1407 } 1408 } 1409 1410 // were we externally suspended while we were waiting? 1411 if (ExitSuspendEquivalent (jt)) { 1412 // TODO-FIXME: add -- if succ == Self then succ = null. 1413 jt->java_suspend_self(); 1414 } 1415 1416 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1417 1418 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1419 // from the WaitSet to the EntryList. 1420 // See if we need to remove Node from the WaitSet. 1421 // We use double-checked locking to avoid grabbing _WaitSetLock 1422 // if the thread is not on the wait queue. 1423 // 1424 // Note that we don't need a fence before the fetch of TState. 1425 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1426 // written by the is thread. (perhaps the fetch might even be satisfied 1427 // by a look-aside into the processor's own store buffer, although given 1428 // the length of the code path between the prior ST and this load that's 1429 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1430 // then we'll acquire the lock and then re-fetch a fresh TState value. 1431 // That is, we fail toward safety. 1432 1433 if (node.TState == ObjectWaiter::TS_WAIT) { 1434 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1435 if (node.TState == ObjectWaiter::TS_WAIT) { 1436 DequeueSpecificWaiter(&node); // unlink from WaitSet 1437 assert(node._notified == 0, "invariant"); 1438 node.TState = ObjectWaiter::TS_RUN; 1439 } 1440 Thread::SpinRelease(&_WaitSetLock); 1441 } 1442 1443 // The thread is now either on off-list (TS_RUN), 1444 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1445 // The Node's TState variable is stable from the perspective of this thread. 1446 // No other threads will asynchronously modify TState. 1447 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1448 OrderAccess::loadload(); 1449 if (_succ == Self) _succ = NULL; 1450 WasNotified = node._notified; 1451 1452 // Reentry phase -- reacquire the monitor. 1453 // re-enter contended monitor after object.wait(). 1454 // retain OBJECT_WAIT state until re-enter successfully completes 1455 // Thread state is thread_in_vm and oop access is again safe, 1456 // although the raw address of the object may have changed. 1457 // (Don't cache naked oops over safepoints, of course). 1458 1459 // post monitor waited event. Note that this is past-tense, we are done waiting. 1460 if (JvmtiExport::should_post_monitor_waited()) { 1461 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1462 1463 if (node._notified != 0 && _succ == Self) { 1464 // In this part of the monitor wait-notify-reenter protocol it 1465 // is possible (and normal) for another thread to do a fastpath 1466 // monitor enter-exit while this thread is still trying to get 1467 // to the reenter portion of the protocol. 1468 // 1469 // The ObjectMonitor was notified and the current thread is 1470 // the successor which also means that an unpark() has already 1471 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1472 // consume the unpark() that was done when the successor was 1473 // set because the same ParkEvent is shared between Java 1474 // monitors and JVM/TI RawMonitors (for now). 1475 // 1476 // We redo the unpark() to ensure forward progress, i.e., we 1477 // don't want all pending threads hanging (parked) with none 1478 // entering the unlocked monitor. 1479 node._event->unpark(); 1480 } 1481 } 1482 1483 if (event.should_commit()) { 1484 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1485 } 1486 1487 OrderAccess::fence(); 1488 1489 assert(Self->_Stalled != 0, "invariant"); 1490 Self->_Stalled = 0; 1491 1492 assert(_owner != Self, "invariant"); 1493 ObjectWaiter::TStates v = node.TState; 1494 if (v == ObjectWaiter::TS_RUN) { 1495 enter(Self); 1496 } else { 1497 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1498 ReenterI(Self, &node); 1499 node.wait_reenter_end(this); 1500 } 1501 1502 // Self has reacquired the lock. 1503 // Lifecycle - the node representing Self must not appear on any queues. 1504 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1505 // want residual elements associated with this thread left on any lists. 1506 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1507 assert(_owner == Self, "invariant"); 1508 assert(_succ != Self, "invariant"); 1509 } // OSThreadWaitState() 1510 1511 jt->set_current_waiting_monitor(NULL); 1512 1513 guarantee(_recursions == 0, "invariant"); 1514 _recursions = save; // restore the old recursion count 1515 _waiters--; // decrement the number of waiters 1516 1517 // Verify a few postconditions 1518 assert(_owner == Self, "invariant"); 1519 assert(_succ != Self, "invariant"); 1520 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1521 1522 // check if the notification happened 1523 if (!WasNotified) { 1524 // no, it could be timeout or Thread.interrupt() or both 1525 // check for interrupt event, otherwise it is timeout 1526 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1527 THROW(vmSymbols::java_lang_InterruptedException()); 1528 } 1529 } 1530 1531 // NOTE: Spurious wake up will be consider as timeout. 1532 // Monitor notify has precedence over thread interrupt. 1533 } 1534 1535 1536 // Consider: 1537 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1538 // then instead of transferring a thread from the WaitSet to the EntryList 1539 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1540 1541 void ObjectMonitor::INotify(Thread * Self) { 1542 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1543 ObjectWaiter * iterator = DequeueWaiter(); 1544 if (iterator != NULL) { 1545 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1546 guarantee(iterator->_notified == 0, "invariant"); 1547 // Disposition - what might we do with iterator ? 1548 // a. add it directly to the EntryList - either tail (policy == 1) 1549 // or head (policy == 0). 1550 // b. push it onto the front of the _cxq (policy == 2). 1551 // For now we use (b). 1552 1553 iterator->TState = ObjectWaiter::TS_ENTER; 1554 1555 iterator->_notified = 1; 1556 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1557 1558 ObjectWaiter * list = _EntryList; 1559 if (list != NULL) { 1560 assert(list->_prev == NULL, "invariant"); 1561 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1562 assert(list != iterator, "invariant"); 1563 } 1564 1565 // prepend to cxq 1566 if (list == NULL) { 1567 iterator->_next = iterator->_prev = NULL; 1568 _EntryList = iterator; 1569 } else { 1570 iterator->TState = ObjectWaiter::TS_CXQ; 1571 for (;;) { 1572 ObjectWaiter * front = _cxq; 1573 iterator->_next = front; 1574 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1575 break; 1576 } 1577 } 1578 } 1579 1580 // _WaitSetLock protects the wait queue, not the EntryList. We could 1581 // move the add-to-EntryList operation, above, outside the critical section 1582 // protected by _WaitSetLock. In practice that's not useful. With the 1583 // exception of wait() timeouts and interrupts the monitor owner 1584 // is the only thread that grabs _WaitSetLock. There's almost no contention 1585 // on _WaitSetLock so it's not profitable to reduce the length of the 1586 // critical section. 1587 1588 iterator->wait_reenter_begin(this); 1589 } 1590 Thread::SpinRelease(&_WaitSetLock); 1591 } 1592 1593 // Consider: a not-uncommon synchronization bug is to use notify() when 1594 // notifyAll() is more appropriate, potentially resulting in stranded 1595 // threads; this is one example of a lost wakeup. A useful diagnostic 1596 // option is to force all notify() operations to behave as notifyAll(). 1597 // 1598 // Note: We can also detect many such problems with a "minimum wait". 1599 // When the "minimum wait" is set to a small non-zero timeout value 1600 // and the program does not hang whereas it did absent "minimum wait", 1601 // that suggests a lost wakeup bug. 1602 1603 void ObjectMonitor::notify(TRAPS) { 1604 CHECK_OWNER(); // Throws IMSE if not owner. 1605 if (_WaitSet == NULL) { 1606 return; 1607 } 1608 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1609 INotify(THREAD); 1610 OM_PERFDATA_OP(Notifications, inc(1)); 1611 } 1612 1613 1614 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1615 // from the waitset to the EntryList. This could be done more efficiently with a 1616 // single bulk transfer but in practice it's not time-critical. Beware too, 1617 // that in prepend-mode we invert the order of the waiters. Let's say that the 1618 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1619 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1620 1621 void ObjectMonitor::notifyAll(TRAPS) { 1622 CHECK_OWNER(); // Throws IMSE if not owner. 1623 if (_WaitSet == NULL) { 1624 return; 1625 } 1626 1627 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1628 int tally = 0; 1629 while (_WaitSet != NULL) { 1630 tally++; 1631 INotify(THREAD); 1632 } 1633 1634 OM_PERFDATA_OP(Notifications, inc(tally)); 1635 } 1636 1637 // ----------------------------------------------------------------------------- 1638 // Adaptive Spinning Support 1639 // 1640 // Adaptive spin-then-block - rational spinning 1641 // 1642 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1643 // algorithm. On high order SMP systems it would be better to start with 1644 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1645 // a contending thread could enqueue itself on the cxq and then spin locally 1646 // on a thread-specific variable such as its ParkEvent._Event flag. 1647 // That's left as an exercise for the reader. Note that global spinning is 1648 // not problematic on Niagara, as the L2 cache serves the interconnect and 1649 // has both low latency and massive bandwidth. 1650 // 1651 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1652 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1653 // (duration) or we can fix the count at approximately the duration of 1654 // a context switch and vary the frequency. Of course we could also 1655 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1656 // For a description of 'Adaptive spin-then-block mutual exclusion in 1657 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1658 // 1659 // This implementation varies the duration "D", where D varies with 1660 // the success rate of recent spin attempts. (D is capped at approximately 1661 // length of a round-trip context switch). The success rate for recent 1662 // spin attempts is a good predictor of the success rate of future spin 1663 // attempts. The mechanism adapts automatically to varying critical 1664 // section length (lock modality), system load and degree of parallelism. 1665 // D is maintained per-monitor in _SpinDuration and is initialized 1666 // optimistically. Spin frequency is fixed at 100%. 1667 // 1668 // Note that _SpinDuration is volatile, but we update it without locks 1669 // or atomics. The code is designed so that _SpinDuration stays within 1670 // a reasonable range even in the presence of races. The arithmetic 1671 // operations on _SpinDuration are closed over the domain of legal values, 1672 // so at worst a race will install and older but still legal value. 1673 // At the very worst this introduces some apparent non-determinism. 1674 // We might spin when we shouldn't or vice-versa, but since the spin 1675 // count are relatively short, even in the worst case, the effect is harmless. 1676 // 1677 // Care must be taken that a low "D" value does not become an 1678 // an absorbing state. Transient spinning failures -- when spinning 1679 // is overall profitable -- should not cause the system to converge 1680 // on low "D" values. We want spinning to be stable and predictable 1681 // and fairly responsive to change and at the same time we don't want 1682 // it to oscillate, become metastable, be "too" non-deterministic, 1683 // or converge on or enter undesirable stable absorbing states. 1684 // 1685 // We implement a feedback-based control system -- using past behavior 1686 // to predict future behavior. We face two issues: (a) if the 1687 // input signal is random then the spin predictor won't provide optimal 1688 // results, and (b) if the signal frequency is too high then the control 1689 // system, which has some natural response lag, will "chase" the signal. 1690 // (b) can arise from multimodal lock hold times. Transient preemption 1691 // can also result in apparent bimodal lock hold times. 1692 // Although sub-optimal, neither condition is particularly harmful, as 1693 // in the worst-case we'll spin when we shouldn't or vice-versa. 1694 // The maximum spin duration is rather short so the failure modes aren't bad. 1695 // To be conservative, I've tuned the gain in system to bias toward 1696 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1697 // "rings" or oscillates between spinning and not spinning. This happens 1698 // when spinning is just on the cusp of profitability, however, so the 1699 // situation is not dire. The state is benign -- there's no need to add 1700 // hysteresis control to damp the transition rate between spinning and 1701 // not spinning. 1702 1703 // Spinning: Fixed frequency (100%), vary duration 1704 int ObjectMonitor::TrySpin(Thread * Self) { 1705 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1706 int ctr = Knob_FixedSpin; 1707 if (ctr != 0) { 1708 while (--ctr >= 0) { 1709 if (TryLock(Self) > 0) return 1; 1710 SpinPause(); 1711 } 1712 return 0; 1713 } 1714 1715 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1716 if (TryLock(Self) > 0) { 1717 // Increase _SpinDuration ... 1718 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1719 // Raising _SpurDuration to the poverty line is key. 1720 int x = _SpinDuration; 1721 if (x < Knob_SpinLimit) { 1722 if (x < Knob_Poverty) x = Knob_Poverty; 1723 _SpinDuration = x + Knob_BonusB; 1724 } 1725 return 1; 1726 } 1727 SpinPause(); 1728 } 1729 1730 // Admission control - verify preconditions for spinning 1731 // 1732 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1733 // becoming an absorbing state. Put another way, we spin briefly to 1734 // sample, just in case the system load, parallelism, contention, or lock 1735 // modality changed. 1736 // 1737 // Consider the following alternative: 1738 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1739 // spin attempt. "Periodically" might mean after a tally of 1740 // the # of failed spin attempts (or iterations) reaches some threshold. 1741 // This takes us into the realm of 1-out-of-N spinning, where we 1742 // hold the duration constant but vary the frequency. 1743 1744 ctr = _SpinDuration; 1745 if (ctr <= 0) return 0; 1746 1747 if (NotRunnable(Self, (Thread *) _owner)) { 1748 return 0; 1749 } 1750 1751 // We're good to spin ... spin ingress. 1752 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1753 // when preparing to LD...CAS _owner, etc and the CAS is likely 1754 // to succeed. 1755 if (_succ == NULL) { 1756 _succ = Self; 1757 } 1758 Thread * prv = NULL; 1759 1760 // There are three ways to exit the following loop: 1761 // 1. A successful spin where this thread has acquired the lock. 1762 // 2. Spin failure with prejudice 1763 // 3. Spin failure without prejudice 1764 1765 while (--ctr >= 0) { 1766 1767 // Periodic polling -- Check for pending GC 1768 // Threads may spin while they're unsafe. 1769 // We don't want spinning threads to delay the JVM from reaching 1770 // a stop-the-world safepoint or to steal cycles from GC. 1771 // If we detect a pending safepoint we abort in order that 1772 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1773 // this thread, if safe, doesn't steal cycles from GC. 1774 // This is in keeping with the "no loitering in runtime" rule. 1775 // We periodically check to see if there's a safepoint pending. 1776 if ((ctr & 0xFF) == 0) { 1777 if (SafepointMechanism::should_block(Self)) { 1778 goto Abort; // abrupt spin egress 1779 } 1780 SpinPause(); 1781 } 1782 1783 // Probe _owner with TATAS 1784 // If this thread observes the monitor transition or flicker 1785 // from locked to unlocked to locked, then the odds that this 1786 // thread will acquire the lock in this spin attempt go down 1787 // considerably. The same argument applies if the CAS fails 1788 // or if we observe _owner change from one non-null value to 1789 // another non-null value. In such cases we might abort 1790 // the spin without prejudice or apply a "penalty" to the 1791 // spin count-down variable "ctr", reducing it by 100, say. 1792 1793 Thread * ox = (Thread *) _owner; 1794 if (ox == NULL) { 1795 ox = (Thread*)try_set_owner_from(NULL, Self); 1796 if (ox == NULL) { 1797 // The CAS succeeded -- this thread acquired ownership 1798 // Take care of some bookkeeping to exit spin state. 1799 if (_succ == Self) { 1800 _succ = NULL; 1801 } 1802 1803 // Increase _SpinDuration : 1804 // The spin was successful (profitable) so we tend toward 1805 // longer spin attempts in the future. 1806 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1807 // If we acquired the lock early in the spin cycle it 1808 // makes sense to increase _SpinDuration proportionally. 1809 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1810 int x = _SpinDuration; 1811 if (x < Knob_SpinLimit) { 1812 if (x < Knob_Poverty) x = Knob_Poverty; 1813 _SpinDuration = x + Knob_Bonus; 1814 } 1815 return 1; 1816 } 1817 1818 // The CAS failed ... we can take any of the following actions: 1819 // * penalize: ctr -= CASPenalty 1820 // * exit spin with prejudice -- goto Abort; 1821 // * exit spin without prejudice. 1822 // * Since CAS is high-latency, retry again immediately. 1823 prv = ox; 1824 goto Abort; 1825 } 1826 1827 // Did lock ownership change hands ? 1828 if (ox != prv && prv != NULL) { 1829 goto Abort; 1830 } 1831 prv = ox; 1832 1833 // Abort the spin if the owner is not executing. 1834 // The owner must be executing in order to drop the lock. 1835 // Spinning while the owner is OFFPROC is idiocy. 1836 // Consider: ctr -= RunnablePenalty ; 1837 if (NotRunnable(Self, ox)) { 1838 goto Abort; 1839 } 1840 if (_succ == NULL) { 1841 _succ = Self; 1842 } 1843 } 1844 1845 // Spin failed with prejudice -- reduce _SpinDuration. 1846 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1847 // AIMD is globally stable. 1848 { 1849 int x = _SpinDuration; 1850 if (x > 0) { 1851 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1852 // This is globally sample and tends to damp the response. 1853 x -= Knob_Penalty; 1854 if (x < 0) x = 0; 1855 _SpinDuration = x; 1856 } 1857 } 1858 1859 Abort: 1860 if (_succ == Self) { 1861 _succ = NULL; 1862 // Invariant: after setting succ=null a contending thread 1863 // must recheck-retry _owner before parking. This usually happens 1864 // in the normal usage of TrySpin(), but it's safest 1865 // to make TrySpin() as foolproof as possible. 1866 OrderAccess::fence(); 1867 if (TryLock(Self) > 0) return 1; 1868 } 1869 return 0; 1870 } 1871 1872 // NotRunnable() -- informed spinning 1873 // 1874 // Don't bother spinning if the owner is not eligible to drop the lock. 1875 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1876 // The thread must be runnable in order to drop the lock in timely fashion. 1877 // If the _owner is not runnable then spinning will not likely be 1878 // successful (profitable). 1879 // 1880 // Beware -- the thread referenced by _owner could have died 1881 // so a simply fetch from _owner->_thread_state might trap. 1882 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1883 // Because of the lifecycle issues, the _thread_state values 1884 // observed by NotRunnable() might be garbage. NotRunnable must 1885 // tolerate this and consider the observed _thread_state value 1886 // as advisory. 1887 // 1888 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1889 // a thread pointer. 1890 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1891 // with the LSB of _owner. Another option would be to probabilistically probe 1892 // the putative _owner->TypeTag value. 1893 // 1894 // Checking _thread_state isn't perfect. Even if the thread is 1895 // in_java it might be blocked on a page-fault or have been preempted 1896 // and sitting on a ready/dispatch queue. 1897 // 1898 // The return value from NotRunnable() is *advisory* -- the 1899 // result is based on sampling and is not necessarily coherent. 1900 // The caller must tolerate false-negative and false-positive errors. 1901 // Spinning, in general, is probabilistic anyway. 1902 1903 1904 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1905 // Check ox->TypeTag == 2BAD. 1906 if (ox == NULL) return 0; 1907 1908 // Avoid transitive spinning ... 1909 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1910 // Immediately after T1 acquires L it's possible that T2, also 1911 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1912 // This occurs transiently after T1 acquired L but before 1913 // T1 managed to clear T1.Stalled. T2 does not need to abort 1914 // its spin in this circumstance. 1915 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1916 1917 if (BlockedOn == 1) return 1; 1918 if (BlockedOn != 0) { 1919 return BlockedOn != intptr_t(this) && _owner == ox; 1920 } 1921 1922 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1923 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1924 // consider also: jst != _thread_in_Java -- but that's overspecific. 1925 return jst == _thread_blocked || jst == _thread_in_native; 1926 } 1927 1928 1929 // ----------------------------------------------------------------------------- 1930 // WaitSet management ... 1931 1932 ObjectWaiter::ObjectWaiter(Thread* thread) { 1933 _next = NULL; 1934 _prev = NULL; 1935 _notified = 0; 1936 _notifier_tid = 0; 1937 TState = TS_RUN; 1938 _thread = thread; 1939 _event = thread->_ParkEvent; 1940 _active = false; 1941 assert(_event != NULL, "invariant"); 1942 } 1943 1944 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1945 JavaThread *jt = (JavaThread *)this->_thread; 1946 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1947 } 1948 1949 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1950 JavaThread *jt = (JavaThread *)this->_thread; 1951 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1952 } 1953 1954 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1955 assert(node != NULL, "should not add NULL node"); 1956 assert(node->_prev == NULL, "node already in list"); 1957 assert(node->_next == NULL, "node already in list"); 1958 // put node at end of queue (circular doubly linked list) 1959 if (_WaitSet == NULL) { 1960 _WaitSet = node; 1961 node->_prev = node; 1962 node->_next = node; 1963 } else { 1964 ObjectWaiter* head = _WaitSet; 1965 ObjectWaiter* tail = head->_prev; 1966 assert(tail->_next == head, "invariant check"); 1967 tail->_next = node; 1968 head->_prev = node; 1969 node->_next = head; 1970 node->_prev = tail; 1971 } 1972 } 1973 1974 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1975 // dequeue the very first waiter 1976 ObjectWaiter* waiter = _WaitSet; 1977 if (waiter) { 1978 DequeueSpecificWaiter(waiter); 1979 } 1980 return waiter; 1981 } 1982 1983 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1984 assert(node != NULL, "should not dequeue NULL node"); 1985 assert(node->_prev != NULL, "node already removed from list"); 1986 assert(node->_next != NULL, "node already removed from list"); 1987 // when the waiter has woken up because of interrupt, 1988 // timeout or other spurious wake-up, dequeue the 1989 // waiter from waiting list 1990 ObjectWaiter* next = node->_next; 1991 if (next == node) { 1992 assert(node->_prev == node, "invariant check"); 1993 _WaitSet = NULL; 1994 } else { 1995 ObjectWaiter* prev = node->_prev; 1996 assert(prev->_next == node, "invariant check"); 1997 assert(next->_prev == node, "invariant check"); 1998 next->_prev = prev; 1999 prev->_next = next; 2000 if (_WaitSet == node) { 2001 _WaitSet = next; 2002 } 2003 } 2004 node->_next = NULL; 2005 node->_prev = NULL; 2006 } 2007 2008 // ----------------------------------------------------------------------------- 2009 // PerfData support 2010 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2011 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2012 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2013 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2014 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2015 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2016 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2017 2018 // One-shot global initialization for the sync subsystem. 2019 // We could also defer initialization and initialize on-demand 2020 // the first time we call ObjectSynchronizer::inflate(). 2021 // Initialization would be protected - like so many things - by 2022 // the MonitorCache_lock. 2023 2024 void ObjectMonitor::Initialize() { 2025 assert(!InitDone, "invariant"); 2026 2027 if (!os::is_MP()) { 2028 Knob_SpinLimit = 0; 2029 Knob_PreSpin = 0; 2030 Knob_FixedSpin = -1; 2031 } 2032 2033 if (UsePerfData) { 2034 EXCEPTION_MARK; 2035 #define NEWPERFCOUNTER(n) \ 2036 { \ 2037 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2038 CHECK); \ 2039 } 2040 #define NEWPERFVARIABLE(n) \ 2041 { \ 2042 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2043 CHECK); \ 2044 } 2045 NEWPERFCOUNTER(_sync_Inflations); 2046 NEWPERFCOUNTER(_sync_Deflations); 2047 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2048 NEWPERFCOUNTER(_sync_FutileWakeups); 2049 NEWPERFCOUNTER(_sync_Parks); 2050 NEWPERFCOUNTER(_sync_Notifications); 2051 NEWPERFVARIABLE(_sync_MonExtant); 2052 #undef NEWPERFCOUNTER 2053 #undef NEWPERFVARIABLE 2054 } 2055 2056 DEBUG_ONLY(InitDone = true;) 2057 } 2058 2059 void ObjectMonitor::print_on(outputStream* st) const { 2060 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2061 st->print("{contentions=0x%08x,waiters=0x%08x" 2062 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2063 contentions(), waiters(), recursions(), 2064 p2i(owner())); 2065 } 2066 void ObjectMonitor::print() const { print_on(tty); } 2067 2068 #ifdef ASSERT 2069 // Print the ObjectMonitor like a debugger would: 2070 // 2071 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2072 // _header = 0x0000000000000001 2073 // _object = 0x000000070ff45fd0 2074 // _allocation_state = Old 2075 // _pad_buf0 = { 2076 // [0] = '\0' 2077 // ... 2078 // [43] = '\0' 2079 // } 2080 // _owner = 0x0000000000000000 2081 // _previous_owner_tid = 0 2082 // _pad_buf1 = { 2083 // [0] = '\0' 2084 // ... 2085 // [47] = '\0' 2086 // } 2087 // _next_om = 0x0000000000000000 2088 // _recursions = 0 2089 // _EntryList = 0x0000000000000000 2090 // _cxq = 0x0000000000000000 2091 // _succ = 0x0000000000000000 2092 // _Responsible = 0x0000000000000000 2093 // _Spinner = 0 2094 // _SpinDuration = 5000 2095 // _contentions = 0 2096 // _WaitSet = 0x0000700009756248 2097 // _waiters = 1 2098 // _WaitSetLock = 0 2099 // } 2100 // 2101 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2102 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2103 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2104 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 2105 st->print(" _allocation_state = "); 2106 if (is_free()) { 2107 st->print("Free"); 2108 } else if (is_old()) { 2109 st->print("Old"); 2110 } else if (is_new()) { 2111 st->print("New"); 2112 } else { 2113 st->print("unknown=%d", _allocation_state); 2114 } 2115 st->cr(); 2116 st->print_cr(" _pad_buf0 = {"); 2117 st->print_cr(" [0] = '\\0'"); 2118 st->print_cr(" ..."); 2119 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2120 st->print_cr(" }"); 2121 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 2122 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 2123 st->print_cr(" _pad_buf1 = {"); 2124 st->print_cr(" [0] = '\\0'"); 2125 st->print_cr(" ..."); 2126 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2127 st->print_cr(" }"); 2128 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2129 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2130 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2131 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2132 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2133 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2134 st->print_cr(" _Spinner = %d", _Spinner); 2135 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2136 st->print_cr(" _contentions = %d", contentions()); 2137 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2138 st->print_cr(" _waiters = %d", _waiters); 2139 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2140 st->print_cr("}"); 2141 } 2142 #endif