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