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 "runtime/vframe_hp.hpp" 48 #include "services/threadService.hpp" 49 #include "utilities/dtrace.hpp" 50 #include "utilities/macros.hpp" 51 #include "utilities/preserveException.hpp" 52 #if INCLUDE_JFR 53 #include "jfr/support/jfrFlush.hpp" 54 #endif 55 56 #ifdef DTRACE_ENABLED 57 58 // Only bother with this argument setup if dtrace is available 59 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 60 61 62 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 63 char* bytes = NULL; \ 64 int len = 0; \ 65 jlong jtid = SharedRuntime::get_java_tid(thread); \ 66 Symbol* klassname = ((oop)obj)->klass()->name(); \ 67 if (klassname != NULL) { \ 68 bytes = (char*)klassname->bytes(); \ 69 len = klassname->utf8_length(); \ 70 } 71 72 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 73 { \ 74 if (DTraceMonitorProbes) { \ 75 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 76 HOTSPOT_MONITOR_WAIT(jtid, \ 77 (monitor), bytes, len, (millis)); \ 78 } \ 79 } 80 81 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 82 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 83 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 84 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 85 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 86 87 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 88 { \ 89 if (DTraceMonitorProbes) { \ 90 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 91 HOTSPOT_MONITOR_##probe(jtid, \ 92 (uintptr_t)(monitor), bytes, len); \ 93 } \ 94 } 95 96 #else // ndef DTRACE_ENABLED 97 98 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 99 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 100 101 #endif // ndef DTRACE_ENABLED 102 103 // Tunables ... 104 // The knob* variables are effectively final. Once set they should 105 // never be modified hence. Consider using __read_mostly with GCC. 106 107 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 108 109 static int Knob_Bonus = 100; // spin success bonus 110 static int Knob_BonusB = 100; // spin success bonus 111 static int Knob_Penalty = 200; // spin failure penalty 112 static int Knob_Poverty = 1000; 113 static int Knob_FixedSpin = 0; 114 static int Knob_PreSpin = 10; // 20-100 likely better 115 116 DEBUG_ONLY(static volatile bool InitDone = false;) 117 118 // ----------------------------------------------------------------------------- 119 // Theory of operations -- Monitors lists, thread residency, etc: 120 // 121 // * A thread acquires ownership of a monitor by successfully 122 // CAS()ing the _owner field from null to non-null. 123 // 124 // * Invariant: A thread appears on at most one monitor list -- 125 // cxq, EntryList or WaitSet -- at any one time. 126 // 127 // * Contending threads "push" themselves onto the cxq with CAS 128 // and then spin/park. 129 // 130 // * After a contending thread eventually acquires the lock it must 131 // dequeue itself from either the EntryList or the cxq. 132 // 133 // * The exiting thread identifies and unparks an "heir presumptive" 134 // tentative successor thread on the EntryList. Critically, the 135 // exiting thread doesn't unlink the successor thread from the EntryList. 136 // After having been unparked, the wakee will recontend for ownership of 137 // the monitor. The successor (wakee) will either acquire the lock or 138 // re-park itself. 139 // 140 // Succession is provided for by a policy of competitive handoff. 141 // The exiting thread does _not_ grant or pass ownership to the 142 // successor thread. (This is also referred to as "handoff" succession"). 143 // Instead the exiting thread releases ownership and possibly wakes 144 // a successor, so the successor can (re)compete for ownership of the lock. 145 // If the EntryList is empty but the cxq is populated the exiting 146 // thread will drain the cxq into the EntryList. It does so by 147 // by detaching the cxq (installing null with CAS) and folding 148 // the threads from the cxq into the EntryList. The EntryList is 149 // doubly linked, while the cxq is singly linked because of the 150 // CAS-based "push" used to enqueue recently arrived threads (RATs). 151 // 152 // * Concurrency invariants: 153 // 154 // -- only the monitor owner may access or mutate the EntryList. 155 // The mutex property of the monitor itself protects the EntryList 156 // from concurrent interference. 157 // -- Only the monitor owner may detach the cxq. 158 // 159 // * The monitor entry list operations avoid locks, but strictly speaking 160 // they're not lock-free. Enter is lock-free, exit is not. 161 // For a description of 'Methods and apparatus providing non-blocking access 162 // to a resource,' see U.S. Pat. No. 7844973. 163 // 164 // * The cxq can have multiple concurrent "pushers" but only one concurrent 165 // detaching thread. This mechanism is immune from the ABA corruption. 166 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 167 // 168 // * Taken together, the cxq and the EntryList constitute or form a 169 // single logical queue of threads stalled trying to acquire the lock. 170 // We use two distinct lists to improve the odds of a constant-time 171 // dequeue operation after acquisition (in the ::enter() epilogue) and 172 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 173 // A key desideratum is to minimize queue & monitor metadata manipulation 174 // that occurs while holding the monitor lock -- that is, we want to 175 // minimize monitor lock holds times. Note that even a small amount of 176 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 177 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 178 // locks and monitor metadata. 179 // 180 // Cxq points to the set of Recently Arrived Threads attempting entry. 181 // Because we push threads onto _cxq with CAS, the RATs must take the form of 182 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 183 // the unlocking thread notices that EntryList is null but _cxq is != null. 184 // 185 // The EntryList is ordered by the prevailing queue discipline and 186 // can be organized in any convenient fashion, such as a doubly-linked list or 187 // a circular doubly-linked list. Critically, we want insert and delete operations 188 // to operate in constant-time. If we need a priority queue then something akin 189 // to Solaris' sleepq would work nicely. Viz., 190 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 191 // Queue discipline is enforced at ::exit() time, when the unlocking thread 192 // drains the cxq into the EntryList, and orders or reorders the threads on the 193 // EntryList accordingly. 194 // 195 // Barring "lock barging", this mechanism provides fair cyclic ordering, 196 // somewhat similar to an elevator-scan. 197 // 198 // * The monitor synchronization subsystem avoids the use of native 199 // synchronization primitives except for the narrow platform-specific 200 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 201 // the semantics of park-unpark. Put another way, this monitor implementation 202 // depends only on atomic operations and park-unpark. The monitor subsystem 203 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 204 // underlying OS manages the READY<->RUN transitions. 205 // 206 // * Waiting threads reside on the WaitSet list -- wait() puts 207 // the caller onto the WaitSet. 208 // 209 // * notify() or notifyAll() simply transfers threads from the WaitSet to 210 // either the EntryList or cxq. Subsequent exit() operations will 211 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 212 // it's likely the notifyee would simply impale itself on the lock held 213 // by the notifier. 214 // 215 // * An interesting alternative is to encode cxq as (List,LockByte) where 216 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 217 // variable, like _recursions, in the scheme. The threads or Events that form 218 // the list would have to be aligned in 256-byte addresses. A thread would 219 // try to acquire the lock or enqueue itself with CAS, but exiting threads 220 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 221 // Note that is is *not* word-tearing, but it does presume that full-word 222 // CAS operations are coherent with intermix with STB operations. That's true 223 // on most common processors. 224 // 225 // * See also http://blogs.sun.com/dave 226 227 228 void* ObjectMonitor::operator new (size_t size) throw() { 229 return AllocateHeap(size, mtInternal); 230 } 231 void* ObjectMonitor::operator new[] (size_t size) throw() { 232 return operator new (size); 233 } 234 void ObjectMonitor::operator delete(void* p) { 235 FreeHeap(p); 236 } 237 void ObjectMonitor::operator delete[] (void *p) { 238 operator delete(p); 239 } 240 241 // ----------------------------------------------------------------------------- 242 // Enter support 243 244 bool ObjectMonitor::enter(TRAPS) { 245 // The following code is ordered to check the most common cases first 246 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 247 Thread * const Self = THREAD; 248 249 void* cur = try_set_owner_from(NULL, Self); 250 if (cur == NULL) { 251 assert(_recursions == 0, "invariant"); 252 return true; 253 } 254 255 if (cur == Self) { 256 // TODO-FIXME: check for integer overflow! BUGID 6557169. 257 _recursions++; 258 return true; 259 } 260 261 if (Self->is_lock_owned((address)cur)) { 262 assert(_recursions == 0, "internal state error"); 263 _recursions = 1; 264 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 265 return true; 266 } 267 268 // We've encountered genuine contention. 269 assert(Self->_Stalled == 0, "invariant"); 270 Self->_Stalled = intptr_t(this); 271 272 // Try one round of spinning *before* enqueueing Self 273 // and before going through the awkward and expensive state 274 // transitions. The following spin is strictly optional ... 275 // Note that if we acquire the monitor from an initial spin 276 // we forgo posting JVMTI events and firing DTRACE probes. 277 if (TrySpin(Self) > 0) { 278 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner)); 279 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 280 assert(((oop)object())->mark() == markWord::encode(this), 281 "object mark must match encoded this: mark=" INTPTR_FORMAT 282 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(), 283 markWord::encode(this).value()); 284 Self->_Stalled = 0; 285 return true; 286 } 287 288 assert(_owner != Self, "invariant"); 289 assert(_succ != Self, "invariant"); 290 assert(Self->is_Java_thread(), "invariant"); 291 JavaThread * jt = (JavaThread *) Self; 292 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 293 assert(jt->thread_state() != _thread_blocked, "invariant"); 294 assert(AsyncDeflateIdleMonitors || this->object() != NULL, "invariant"); 295 assert(AsyncDeflateIdleMonitors || contentions() >= 0, "must not be negative: contentions=%d", contentions()); 296 297 // Keep track of contention for JVM/TI and M&M queries. 298 add_to_contentions(1); 299 if (is_being_async_deflated()) { 300 // Async deflation is in progress and our contentions increment 301 // above lost the race to async deflation. Undo the work and 302 // force the caller to retry. 303 const oop l_object = (oop)object(); 304 if (l_object != NULL) { 305 // Attempt to restore the header/dmw to the object's header so that 306 // we only retry once if the deflater thread happens to be slow. 307 install_displaced_markword_in_object(l_object); 308 } 309 Self->_Stalled = 0; 310 add_to_contentions(-1); 311 return false; 312 } 313 314 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);) 315 EventJavaMonitorEnter event; 316 if (event.should_commit()) { 317 event.set_monitorClass(((oop)this->object())->klass()); 318 event.set_address((uintptr_t)(this->object_addr())); 319 } 320 321 { // Change java thread status to indicate blocked on monitor enter. 322 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 323 324 Self->set_current_pending_monitor(this); 325 326 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 327 if (JvmtiExport::should_post_monitor_contended_enter()) { 328 JvmtiExport::post_monitor_contended_enter(jt, this); 329 330 // The current thread does not yet own the monitor and does not 331 // yet appear on any queues that would get it made the successor. 332 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 333 // handler cannot accidentally consume an unpark() meant for the 334 // ParkEvent associated with this ObjectMonitor. 335 } 336 337 OSThreadContendState osts(Self->osthread()); 338 ThreadBlockInVM tbivm(jt); 339 340 // TODO-FIXME: change the following for(;;) loop to straight-line code. 341 for (;;) { 342 jt->set_suspend_equivalent(); 343 // cleared by handle_special_suspend_equivalent_condition() 344 // or java_suspend_self() 345 346 EnterI(THREAD); 347 348 if (!ExitSuspendEquivalent(jt)) break; 349 350 // We have acquired the contended monitor, but while we were 351 // waiting another thread suspended us. We don't want to enter 352 // the monitor while suspended because that would surprise the 353 // thread that suspended us. 354 // 355 _recursions = 0; 356 _succ = NULL; 357 exit(false, Self); 358 359 jt->java_suspend_self(); 360 } 361 Self->set_current_pending_monitor(NULL); 362 363 // We cleared the pending monitor info since we've just gotten past 364 // the enter-check-for-suspend dance and we now own the monitor free 365 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 366 // destructor can go to a safepoint at the end of this block. If we 367 // do a thread dump during that safepoint, then this thread will show 368 // as having "-locked" the monitor, but the OS and java.lang.Thread 369 // states will still report that the thread is blocked trying to 370 // acquire it. 371 } 372 373 add_to_contentions(-1); 374 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions()); 375 Self->_Stalled = 0; 376 377 // Must either set _recursions = 0 or ASSERT _recursions == 0. 378 assert(_recursions == 0, "invariant"); 379 assert(_owner == Self, "invariant"); 380 assert(_succ != Self, "invariant"); 381 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 382 383 // The thread -- now the owner -- is back in vm mode. 384 // Report the glorious news via TI,DTrace and jvmstat. 385 // The probe effect is non-trivial. All the reportage occurs 386 // while we hold the monitor, increasing the length of the critical 387 // section. Amdahl's parallel speedup law comes vividly into play. 388 // 389 // Another option might be to aggregate the events (thread local or 390 // per-monitor aggregation) and defer reporting until a more opportune 391 // time -- such as next time some thread encounters contention but has 392 // yet to acquire the lock. While spinning that thread could 393 // spinning we could increment JVMStat counters, etc. 394 395 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 396 if (JvmtiExport::should_post_monitor_contended_entered()) { 397 JvmtiExport::post_monitor_contended_entered(jt, this); 398 399 // The current thread already owns the monitor and is not going to 400 // call park() for the remainder of the monitor enter protocol. So 401 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 402 // event handler consumed an unpark() issued by the thread that 403 // just exited the monitor. 404 } 405 if (event.should_commit()) { 406 event.set_previousOwner((uintptr_t)_previous_owner_tid); 407 event.commit(); 408 } 409 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 410 return true; 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 (try_set_owner_from(NULL, Self) == NULL) { 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 // && contentions <= 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"); 442 443 // Separate loads in is_being_async_deflated(), which is almost always 444 // called before this function, from the load of dmw/header below. 445 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 446 // A non-multiple copy atomic (nMCA) machine needs a bigger 447 // hammer to separate the loads before and the load below. 448 OrderAccess::fence(); 449 } else { 450 OrderAccess::loadload(); 451 } 452 453 const oop l_object = (oop)object(); 454 if (l_object == NULL) { 455 // ObjectMonitor's object ref has already been cleared by async 456 // deflation so we're done here. 457 return; 458 } 459 ADIM_guarantee(l_object == obj, "object=" INTPTR_FORMAT " must equal obj=" 460 INTPTR_FORMAT, p2i(l_object), p2i(obj)); 461 462 markWord dmw = header(); 463 // The dmw has to be neutral (not NULL, not locked and not marked). 464 ADIM_guarantee(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 465 466 // Install displaced mark word if the object's header still points 467 // to this ObjectMonitor. More than one racing caller to this function 468 // can rarely reach this point, but only one can win. 469 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 470 if (res != markWord::encode(this)) { 471 // This should be rare so log at the Info level when it happens. 472 log_info(monitorinflation)("install_displaced_markword_in_object: " 473 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 474 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 475 dmw.value(), markWord::encode(this).value(), 476 res.value()); 477 } 478 479 // Note: It does not matter which thread restored the header/dmw 480 // into the object's header. The thread deflating the monitor just 481 // wanted the object's header restored and it is. The threads that 482 // detected a race with the deflation process also wanted the 483 // object's header restored before they retry their operation and 484 // because it is restored they will only retry once. 485 } 486 487 // Convert the fields used by is_busy() to a string that can be 488 // used for diagnostic output. 489 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 490 ss->print("is_busy: waiters=%d, ", _waiters); 491 if (!AsyncDeflateIdleMonitors) { 492 ss->print("contentions=%d, ", contentions()); 493 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 494 } else { 495 if (contentions() > 0) { 496 ss->print("contentions=%d, ", contentions()); 497 } else { 498 ss->print("contentions=0"); 499 } 500 if (_owner != DEFLATER_MARKER) { 501 ss->print("owner=" INTPTR_FORMAT, p2i(_owner)); 502 } else { 503 // We report NULL instead of DEFLATER_MARKER here because is_busy() 504 // ignores DEFLATER_MARKER values. 505 ss->print("owner=" INTPTR_FORMAT, NULL); 506 } 507 } 508 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq), 509 p2i(_EntryList)); 510 return ss->base(); 511 } 512 513 #define MAX_RECHECK_INTERVAL 1000 514 515 void ObjectMonitor::EnterI(TRAPS) { 516 Thread * const Self = THREAD; 517 assert(Self->is_Java_thread(), "invariant"); 518 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 519 520 // Try the lock - TATAS 521 if (TryLock (Self) > 0) { 522 assert(_succ != Self, "invariant"); 523 assert(_owner == Self, "invariant"); 524 assert(_Responsible != Self, "invariant"); 525 return; 526 } 527 528 if (AsyncDeflateIdleMonitors && 529 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 530 // Cancelled the in-progress async deflation by changing owner from 531 // DEFLATER_MARKER to Self. As part of the contended enter protocol, 532 // contentions was incremented to a positive value before EnterI() 533 // was called and that prevents the deflater thread from winning the 534 // last part of the 2-part async deflation protocol. After EnterI() 535 // returns to enter(), contentions is decremented because the caller 536 // now owns the monitor. We bump contentions an extra time here to 537 // prevent the deflater thread from winning the last part of the 538 // 2-part async deflation protocol after the regular decrement 539 // occurs in enter(). The deflater thread will decrement contentions 540 // after it recognizes that the async deflation was cancelled. 541 add_to_contentions(1); 542 assert(_succ != Self, "invariant"); 543 assert(_Responsible != Self, "invariant"); 544 return; 545 } 546 547 assert(InitDone, "Unexpectedly not initialized"); 548 549 // We try one round of spinning *before* enqueueing Self. 550 // 551 // If the _owner is ready but OFFPROC we could use a YieldTo() 552 // operation to donate the remainder of this thread's quantum 553 // to the owner. This has subtle but beneficial affinity 554 // effects. 555 556 if (TrySpin(Self) > 0) { 557 assert(_owner == Self, "invariant"); 558 assert(_succ != Self, "invariant"); 559 assert(_Responsible != Self, "invariant"); 560 return; 561 } 562 563 // The Spin failed -- Enqueue and park the thread ... 564 assert(_succ != Self, "invariant"); 565 assert(_owner != Self, "invariant"); 566 assert(_Responsible != Self, "invariant"); 567 568 // Enqueue "Self" on ObjectMonitor's _cxq. 569 // 570 // Node acts as a proxy for Self. 571 // As an aside, if were to ever rewrite the synchronization code mostly 572 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 573 // Java objects. This would avoid awkward lifecycle and liveness issues, 574 // as well as eliminate a subset of ABA issues. 575 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 576 577 ObjectWaiter node(Self); 578 Self->_ParkEvent->reset(); 579 node._prev = (ObjectWaiter *) 0xBAD; 580 node.TState = ObjectWaiter::TS_CXQ; 581 582 // Push "Self" onto the front of the _cxq. 583 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 584 // Note that spinning tends to reduce the rate at which threads 585 // enqueue and dequeue on EntryList|cxq. 586 ObjectWaiter * nxt; 587 for (;;) { 588 node._next = nxt = _cxq; 589 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 590 591 // Interference - the CAS failed because _cxq changed. Just retry. 592 // As an optional optimization we retry the lock. 593 if (TryLock (Self) > 0) { 594 assert(_succ != Self, "invariant"); 595 assert(_owner == Self, "invariant"); 596 assert(_Responsible != Self, "invariant"); 597 return; 598 } 599 } 600 601 // Check for cxq|EntryList edge transition to non-null. This indicates 602 // the onset of contention. While contention persists exiting threads 603 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 604 // operations revert to the faster 1-0 mode. This enter operation may interleave 605 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 606 // arrange for one of the contending thread to use a timed park() operations 607 // to detect and recover from the race. (Stranding is form of progress failure 608 // where the monitor is unlocked but all the contending threads remain parked). 609 // That is, at least one of the contended threads will periodically poll _owner. 610 // One of the contending threads will become the designated "Responsible" thread. 611 // The Responsible thread uses a timed park instead of a normal indefinite park 612 // operation -- it periodically wakes and checks for and recovers from potential 613 // strandings admitted by 1-0 exit operations. We need at most one Responsible 614 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 615 // be responsible for a monitor. 616 // 617 // Currently, one of the contended threads takes on the added role of "Responsible". 618 // A viable alternative would be to use a dedicated "stranding checker" thread 619 // that periodically iterated over all the threads (or active monitors) and unparked 620 // successors where there was risk of stranding. This would help eliminate the 621 // timer scalability issues we see on some platforms as we'd only have one thread 622 // -- the checker -- parked on a timer. 623 624 if (nxt == NULL && _EntryList == NULL) { 625 // Try to assume the role of responsible thread for the monitor. 626 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 627 Atomic::replace_if_null(&_Responsible, Self); 628 } 629 630 // The lock might have been released while this thread was occupied queueing 631 // itself onto _cxq. To close the race and avoid "stranding" and 632 // progress-liveness failure we must resample-retry _owner before parking. 633 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 634 // In this case the ST-MEMBAR is accomplished with CAS(). 635 // 636 // TODO: Defer all thread state transitions until park-time. 637 // Since state transitions are heavy and inefficient we'd like 638 // to defer the state transitions until absolutely necessary, 639 // and in doing so avoid some transitions ... 640 641 int nWakeups = 0; 642 int recheckInterval = 1; 643 644 for (;;) { 645 646 if (TryLock(Self) > 0) break; 647 assert(_owner != Self, "invariant"); 648 649 // park self 650 if (_Responsible == Self) { 651 Self->_ParkEvent->park((jlong) recheckInterval); 652 // Increase the recheckInterval, but clamp the value. 653 recheckInterval *= 8; 654 if (recheckInterval > MAX_RECHECK_INTERVAL) { 655 recheckInterval = MAX_RECHECK_INTERVAL; 656 } 657 } else { 658 Self->_ParkEvent->park(); 659 } 660 661 if (TryLock(Self) > 0) break; 662 663 if (AsyncDeflateIdleMonitors && 664 try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) { 665 // Cancelled the in-progress async deflation by changing owner from 666 // DEFLATER_MARKER to Self. As part of the contended enter protocol, 667 // contentions was incremented to a positive value before EnterI() 668 // was called and that prevents the deflater thread from winning the 669 // last part of the 2-part async deflation protocol. After EnterI() 670 // returns to enter(), contentions is decremented because the caller 671 // now owns the monitor. We bump contentions an extra time here to 672 // prevent the deflater thread from winning the last part of the 673 // 2-part async deflation protocol after the regular decrement 674 // occurs in enter(). The deflater thread will decrement contentions 675 // after it recognizes that the async deflation was cancelled. 676 add_to_contentions(1); 677 break; 678 } 679 680 // The lock is still contested. 681 // Keep a tally of the # of futile wakeups. 682 // Note that the counter is not protected by a lock or updated by atomics. 683 // That is by design - we trade "lossy" counters which are exposed to 684 // races during updates for a lower probe effect. 685 686 // This PerfData object can be used in parallel with a safepoint. 687 // See the work around in PerfDataManager::destroy(). 688 OM_PERFDATA_OP(FutileWakeups, inc()); 689 ++nWakeups; 690 691 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 692 // We can defer clearing _succ until after the spin completes 693 // TrySpin() must tolerate being called with _succ == Self. 694 // Try yet another round of adaptive spinning. 695 if (TrySpin(Self) > 0) break; 696 697 // We can find that we were unpark()ed and redesignated _succ while 698 // we were spinning. That's harmless. If we iterate and call park(), 699 // park() will consume the event and return immediately and we'll 700 // just spin again. This pattern can repeat, leaving _succ to simply 701 // spin on a CPU. 702 703 if (_succ == Self) _succ = NULL; 704 705 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 706 OrderAccess::fence(); 707 } 708 709 // Egress : 710 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 711 // Normally we'll find Self on the EntryList . 712 // From the perspective of the lock owner (this thread), the 713 // EntryList is stable and cxq is prepend-only. 714 // The head of cxq is volatile but the interior is stable. 715 // In addition, Self.TState is stable. 716 717 assert(_owner == Self, "invariant"); 718 assert(object() != NULL, "invariant"); 719 // I'd like to write: 720 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ; 721 // but as we're at a safepoint that's not safe. 722 723 UnlinkAfterAcquire(Self, &node); 724 if (_succ == Self) _succ = NULL; 725 726 assert(_succ != Self, "invariant"); 727 if (_Responsible == Self) { 728 _Responsible = NULL; 729 OrderAccess::fence(); // Dekker pivot-point 730 731 // We may leave threads on cxq|EntryList without a designated 732 // "Responsible" thread. This is benign. When this thread subsequently 733 // exits the monitor it can "see" such preexisting "old" threads -- 734 // threads that arrived on the cxq|EntryList before the fence, above -- 735 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 736 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 737 // non-null and elect a new "Responsible" timer thread. 738 // 739 // This thread executes: 740 // ST Responsible=null; MEMBAR (in enter epilogue - here) 741 // LD cxq|EntryList (in subsequent exit) 742 // 743 // Entering threads in the slow/contended path execute: 744 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 745 // The (ST cxq; MEMBAR) is accomplished with CAS(). 746 // 747 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 748 // exit operation from floating above the ST Responsible=null. 749 } 750 751 // We've acquired ownership with CAS(). 752 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 753 // But since the CAS() this thread may have also stored into _succ, 754 // EntryList, cxq or Responsible. These meta-data updates must be 755 // visible __before this thread subsequently drops the lock. 756 // Consider what could occur if we didn't enforce this constraint -- 757 // STs to monitor meta-data and user-data could reorder with (become 758 // visible after) the ST in exit that drops ownership of the lock. 759 // Some other thread could then acquire the lock, but observe inconsistent 760 // or old monitor meta-data and heap data. That violates the JMM. 761 // To that end, the 1-0 exit() operation must have at least STST|LDST 762 // "release" barrier semantics. Specifically, there must be at least a 763 // STST|LDST barrier in exit() before the ST of null into _owner that drops 764 // the lock. The barrier ensures that changes to monitor meta-data and data 765 // protected by the lock will be visible before we release the lock, and 766 // therefore before some other thread (CPU) has a chance to acquire the lock. 767 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 768 // 769 // Critically, any prior STs to _succ or EntryList must be visible before 770 // the ST of null into _owner in the *subsequent* (following) corresponding 771 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 772 // execute a serializing instruction. 773 774 return; 775 } 776 777 // ReenterI() is a specialized inline form of the latter half of the 778 // contended slow-path from EnterI(). We use ReenterI() only for 779 // monitor reentry in wait(). 780 // 781 // In the future we should reconcile EnterI() and ReenterI(). 782 783 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 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() == markWord::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 // State transition wrappers around park() ... 802 // ReenterI() wisely defers state transitions until 803 // it's clear we must park the thread. 804 { 805 OSThreadContendState osts(Self->osthread()); 806 ThreadBlockInVM tbivm(jt); 807 808 // cleared by handle_special_suspend_equivalent_condition() 809 // or java_suspend_self() 810 jt->set_suspend_equivalent(); 811 Self->_ParkEvent->park(); 812 813 // were we externally suspended while we were waiting? 814 for (;;) { 815 if (!ExitSuspendEquivalent(jt)) break; 816 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 817 jt->java_suspend_self(); 818 jt->set_suspend_equivalent(); 819 } 820 } 821 822 // Try again, but just so we distinguish between futile wakeups and 823 // successful wakeups. The following test isn't algorithmically 824 // necessary, but it helps us maintain sensible statistics. 825 if (TryLock(Self) > 0) break; 826 827 // The lock is still contested. 828 // Keep a tally of the # of futile wakeups. 829 // Note that the counter is not protected by a lock or updated by atomics. 830 // That is by design - we trade "lossy" counters which are exposed to 831 // races during updates for a lower probe effect. 832 ++nWakeups; 833 834 // Assuming this is not a spurious wakeup we'll normally 835 // find that _succ == Self. 836 if (_succ == Self) _succ = NULL; 837 838 // Invariant: after clearing _succ a contending thread 839 // *must* retry _owner before parking. 840 OrderAccess::fence(); 841 842 // This PerfData object can be used in parallel with a safepoint. 843 // See the work around in PerfDataManager::destroy(). 844 OM_PERFDATA_OP(FutileWakeups, inc()); 845 } 846 847 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 848 // Normally we'll find Self on the EntryList. 849 // Unlinking from the EntryList is constant-time and atomic-free. 850 // From the perspective of the lock owner (this thread), the 851 // EntryList is stable and cxq is prepend-only. 852 // The head of cxq is volatile but the interior is stable. 853 // In addition, Self.TState is stable. 854 855 assert(_owner == Self, "invariant"); 856 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 857 UnlinkAfterAcquire(Self, SelfNode); 858 if (_succ == Self) _succ = NULL; 859 assert(_succ != Self, "invariant"); 860 SelfNode->TState = ObjectWaiter::TS_RUN; 861 OrderAccess::fence(); // see comments at the end of EnterI() 862 } 863 864 // By convention we unlink a contending thread from EntryList|cxq immediately 865 // after the thread acquires the lock in ::enter(). Equally, we could defer 866 // unlinking the thread until ::exit()-time. 867 868 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 869 assert(_owner == Self, "invariant"); 870 assert(SelfNode->_thread == Self, "invariant"); 871 872 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 873 // Normal case: remove Self from the DLL EntryList . 874 // This is a constant-time operation. 875 ObjectWaiter * nxt = SelfNode->_next; 876 ObjectWaiter * prv = SelfNode->_prev; 877 if (nxt != NULL) nxt->_prev = prv; 878 if (prv != NULL) prv->_next = nxt; 879 if (SelfNode == _EntryList) _EntryList = nxt; 880 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 881 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 882 } else { 883 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 884 // Inopportune interleaving -- Self is still on the cxq. 885 // This usually means the enqueue of self raced an exiting thread. 886 // Normally we'll find Self near the front of the cxq, so 887 // dequeueing is typically fast. If needbe we can accelerate 888 // this with some MCS/CHL-like bidirectional list hints and advisory 889 // back-links so dequeueing from the interior will normally operate 890 // in constant-time. 891 // Dequeue Self from either the head (with CAS) or from the interior 892 // with a linear-time scan and normal non-atomic memory operations. 893 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 894 // and then unlink Self from EntryList. We have to drain eventually, 895 // so it might as well be now. 896 897 ObjectWaiter * v = _cxq; 898 assert(v != NULL, "invariant"); 899 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) { 900 // The CAS above can fail from interference IFF a "RAT" arrived. 901 // In that case Self must be in the interior and can no longer be 902 // at the head of cxq. 903 if (v == SelfNode) { 904 assert(_cxq != v, "invariant"); 905 v = _cxq; // CAS above failed - start scan at head of list 906 } 907 ObjectWaiter * p; 908 ObjectWaiter * q = NULL; 909 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 910 q = p; 911 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 912 } 913 assert(v != SelfNode, "invariant"); 914 assert(p == SelfNode, "Node not found on cxq"); 915 assert(p != _cxq, "invariant"); 916 assert(q != NULL, "invariant"); 917 assert(q->_next == p, "invariant"); 918 q->_next = p->_next; 919 } 920 } 921 922 #ifdef ASSERT 923 // Diagnostic hygiene ... 924 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 925 SelfNode->_next = (ObjectWaiter *) 0xBAD; 926 SelfNode->TState = ObjectWaiter::TS_RUN; 927 #endif 928 } 929 930 // ----------------------------------------------------------------------------- 931 // Exit support 932 // 933 // exit() 934 // ~~~~~~ 935 // Note that the collector can't reclaim the objectMonitor or deflate 936 // the object out from underneath the thread calling ::exit() as the 937 // thread calling ::exit() never transitions to a stable state. 938 // This inhibits GC, which in turn inhibits asynchronous (and 939 // inopportune) reclamation of "this". 940 // 941 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 942 // There's one exception to the claim above, however. EnterI() can call 943 // exit() to drop a lock if the acquirer has been externally suspended. 944 // In that case exit() is called with _thread_state == _thread_blocked, 945 // but the monitor's _contentions field is > 0, which inhibits reclamation. 946 // 947 // 1-0 exit 948 // ~~~~~~~~ 949 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 950 // the fast-path operators have been optimized so the common ::exit() 951 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 952 // The code emitted by fast_unlock() elides the usual MEMBAR. This 953 // greatly improves latency -- MEMBAR and CAS having considerable local 954 // latency on modern processors -- but at the cost of "stranding". Absent the 955 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 956 // ::enter() path, resulting in the entering thread being stranding 957 // and a progress-liveness failure. Stranding is extremely rare. 958 // We use timers (timed park operations) & periodic polling to detect 959 // and recover from stranding. Potentially stranded threads periodically 960 // wake up and poll the lock. See the usage of the _Responsible variable. 961 // 962 // The CAS() in enter provides for safety and exclusion, while the CAS or 963 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 964 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 965 // We detect and recover from stranding with timers. 966 // 967 // If a thread transiently strands it'll park until (a) another 968 // thread acquires the lock and then drops the lock, at which time the 969 // exiting thread will notice and unpark the stranded thread, or, (b) 970 // the timer expires. If the lock is high traffic then the stranding latency 971 // will be low due to (a). If the lock is low traffic then the odds of 972 // stranding are lower, although the worst-case stranding latency 973 // is longer. Critically, we don't want to put excessive load in the 974 // platform's timer subsystem. We want to minimize both the timer injection 975 // rate (timers created/sec) as well as the number of timers active at 976 // any one time. (more precisely, we want to minimize timer-seconds, which is 977 // the integral of the # of active timers at any instant over time). 978 // Both impinge on OS scalability. Given that, at most one thread parked on 979 // a monitor will use a timer. 980 // 981 // There is also the risk of a futile wake-up. If we drop the lock 982 // another thread can reacquire the lock immediately, and we can 983 // then wake a thread unnecessarily. This is benign, and we've 984 // structured the code so the windows are short and the frequency 985 // of such futile wakups is low. 986 987 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 988 Thread* const Self = THREAD; 989 void* cur = Atomic::load(&_owner); 990 if (THREAD != cur) { 991 if (THREAD->is_lock_owned((address)cur)) { 992 assert(_recursions == 0, "invariant"); 993 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 994 _recursions = 0; 995 } else { 996 // Apparent unbalanced locking ... 997 // Naively we'd like to throw IllegalMonitorStateException. 998 // As a practical matter we can neither allocate nor throw an 999 // exception as ::exit() can be called from leaf routines. 1000 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1001 // Upon deeper reflection, however, in a properly run JVM the only 1002 // way we should encounter this situation is in the presence of 1003 // unbalanced JNI locking. TODO: CheckJNICalls. 1004 // See also: CR4414101 1005 #ifdef ASSERT 1006 LogStreamHandle(Error, monitorinflation) lsh; 1007 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 1008 " is exiting an ObjectMonitor it does not own.", p2i(THREAD)); 1009 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 1010 print_debug_style_on(&lsh); 1011 #endif 1012 assert(false, "Non-balanced monitor enter/exit!"); 1013 return; 1014 } 1015 } 1016 1017 if (_recursions != 0) { 1018 _recursions--; // this is simple recursive enter 1019 return; 1020 } 1021 1022 // Invariant: after setting Responsible=null an thread must execute 1023 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1024 _Responsible = NULL; 1025 1026 #if INCLUDE_JFR 1027 // get the owner's thread id for the MonitorEnter event 1028 // if it is enabled and the thread isn't suspended 1029 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1030 _previous_owner_tid = JFR_THREAD_ID(Self); 1031 } 1032 #endif 1033 1034 for (;;) { 1035 assert(THREAD == _owner, "invariant"); 1036 1037 // Drop the lock. 1038 // release semantics: prior loads and stores from within the critical section 1039 // must not float (reorder) past the following store that drops the lock. 1040 // Uses a storeload to separate release_store(owner) from the 1041 // successor check. The try_set_owner() below uses cmpxchg() so 1042 // we get the fence down there. 1043 release_clear_owner(Self); 1044 OrderAccess::storeload(); 1045 1046 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1047 return; 1048 } 1049 // Other threads are blocked trying to acquire the lock. 1050 1051 // Normally the exiting thread is responsible for ensuring succession, 1052 // but if other successors are ready or other entering threads are spinning 1053 // then this thread can simply store NULL into _owner and exit without 1054 // waking a successor. The existence of spinners or ready successors 1055 // guarantees proper succession (liveness). Responsibility passes to the 1056 // ready or running successors. The exiting thread delegates the duty. 1057 // More precisely, if a successor already exists this thread is absolved 1058 // of the responsibility of waking (unparking) one. 1059 // 1060 // The _succ variable is critical to reducing futile wakeup frequency. 1061 // _succ identifies the "heir presumptive" thread that has been made 1062 // ready (unparked) but that has not yet run. We need only one such 1063 // successor thread to guarantee progress. 1064 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1065 // section 3.3 "Futile Wakeup Throttling" for details. 1066 // 1067 // Note that spinners in Enter() also set _succ non-null. 1068 // In the current implementation spinners opportunistically set 1069 // _succ so that exiting threads might avoid waking a successor. 1070 // Another less appealing alternative would be for the exiting thread 1071 // to drop the lock and then spin briefly to see if a spinner managed 1072 // to acquire the lock. If so, the exiting thread could exit 1073 // immediately without waking a successor, otherwise the exiting 1074 // thread would need to dequeue and wake a successor. 1075 // (Note that we'd need to make the post-drop spin short, but no 1076 // shorter than the worst-case round-trip cache-line migration time. 1077 // The dropped lock needs to become visible to the spinner, and then 1078 // the acquisition of the lock by the spinner must become visible to 1079 // the exiting thread). 1080 1081 // It appears that an heir-presumptive (successor) must be made ready. 1082 // Only the current lock owner can manipulate the EntryList or 1083 // drain _cxq, so we need to reacquire the lock. If we fail 1084 // to reacquire the lock the responsibility for ensuring succession 1085 // falls to the new owner. 1086 // 1087 if (try_set_owner_from(NULL, Self) != NULL) { 1088 return; 1089 } 1090 1091 guarantee(_owner == THREAD, "invariant"); 1092 1093 ObjectWaiter * w = NULL; 1094 1095 w = _EntryList; 1096 if (w != NULL) { 1097 // I'd like to write: guarantee (w->_thread != Self). 1098 // But in practice an exiting thread may find itself on the EntryList. 1099 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1100 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1101 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1102 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1103 // release the lock "O". T2 resumes immediately after the ST of null into 1104 // _owner, above. T2 notices that the EntryList is populated, so it 1105 // reacquires the lock and then finds itself on the EntryList. 1106 // Given all that, we have to tolerate the circumstance where "w" is 1107 // associated with Self. 1108 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1109 ExitEpilog(Self, w); 1110 return; 1111 } 1112 1113 // If we find that both _cxq and EntryList are null then just 1114 // re-run the exit protocol from the top. 1115 w = _cxq; 1116 if (w == NULL) continue; 1117 1118 // Drain _cxq into EntryList - bulk transfer. 1119 // First, detach _cxq. 1120 // The following loop is tantamount to: w = swap(&cxq, NULL) 1121 for (;;) { 1122 assert(w != NULL, "Invariant"); 1123 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL); 1124 if (u == w) break; 1125 w = u; 1126 } 1127 1128 assert(w != NULL, "invariant"); 1129 assert(_EntryList == NULL, "invariant"); 1130 1131 // Convert the LIFO SLL anchored by _cxq into a DLL. 1132 // The list reorganization step operates in O(LENGTH(w)) time. 1133 // It's critical that this step operate quickly as 1134 // "Self" still holds the outer-lock, restricting parallelism 1135 // and effectively lengthening the critical section. 1136 // Invariant: s chases t chases u. 1137 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1138 // we have faster access to the tail. 1139 1140 _EntryList = w; 1141 ObjectWaiter * q = NULL; 1142 ObjectWaiter * p; 1143 for (p = w; p != NULL; p = p->_next) { 1144 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1145 p->TState = ObjectWaiter::TS_ENTER; 1146 p->_prev = q; 1147 q = p; 1148 } 1149 1150 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1151 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1152 1153 // See if we can abdicate to a spinner instead of waking a thread. 1154 // A primary goal of the implementation is to reduce the 1155 // context-switch rate. 1156 if (_succ != NULL) continue; 1157 1158 w = _EntryList; 1159 if (w != NULL) { 1160 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1161 ExitEpilog(Self, w); 1162 return; 1163 } 1164 } 1165 } 1166 1167 // ExitSuspendEquivalent: 1168 // A faster alternate to handle_special_suspend_equivalent_condition() 1169 // 1170 // handle_special_suspend_equivalent_condition() unconditionally 1171 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1172 // operations have high latency. Note that in ::enter() we call HSSEC 1173 // while holding the monitor, so we effectively lengthen the critical sections. 1174 // 1175 // There are a number of possible solutions: 1176 // 1177 // A. To ameliorate the problem we might also defer state transitions 1178 // to as late as possible -- just prior to parking. 1179 // Given that, we'd call HSSEC after having returned from park(), 1180 // but before attempting to acquire the monitor. This is only a 1181 // partial solution. It avoids calling HSSEC while holding the 1182 // monitor (good), but it still increases successor reacquisition latency -- 1183 // the interval between unparking a successor and the time the successor 1184 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1185 // If we use this technique we can also avoid EnterI()-exit() loop 1186 // in ::enter() where we iteratively drop the lock and then attempt 1187 // to reacquire it after suspending. 1188 // 1189 // B. In the future we might fold all the suspend bits into a 1190 // composite per-thread suspend flag and then update it with CAS(). 1191 // Alternately, a Dekker-like mechanism with multiple variables 1192 // would suffice: 1193 // ST Self->_suspend_equivalent = false 1194 // MEMBAR 1195 // LD Self_>_suspend_flags 1196 1197 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1198 return jSelf->handle_special_suspend_equivalent_condition(); 1199 } 1200 1201 1202 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1203 assert(_owner == Self, "invariant"); 1204 1205 // Exit protocol: 1206 // 1. ST _succ = wakee 1207 // 2. membar #loadstore|#storestore; 1208 // 2. ST _owner = NULL 1209 // 3. unpark(wakee) 1210 1211 _succ = Wakee->_thread; 1212 ParkEvent * Trigger = Wakee->_event; 1213 1214 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1215 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1216 // out-of-scope (non-extant). 1217 Wakee = NULL; 1218 1219 // Drop the lock. 1220 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1221 release_clear_owner(Self); 1222 OrderAccess::fence(); 1223 1224 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1225 Trigger->unpark(); 1226 1227 // Maintain stats and report events to JVMTI 1228 OM_PERFDATA_OP(Parks, inc()); 1229 } 1230 1231 1232 // ----------------------------------------------------------------------------- 1233 // Class Loader deadlock handling. 1234 // 1235 // complete_exit exits a lock returning recursion count 1236 // complete_exit/reenter operate as a wait without waiting 1237 // complete_exit requires an inflated monitor 1238 // The _owner field is not always the Thread addr even with an 1239 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1240 // thread due to contention. 1241 intx ObjectMonitor::complete_exit(TRAPS) { 1242 Thread * const Self = THREAD; 1243 assert(Self->is_Java_thread(), "Must be Java thread!"); 1244 JavaThread *jt = (JavaThread *)THREAD; 1245 1246 assert(InitDone, "Unexpectedly not initialized"); 1247 1248 void* cur = Atomic::load(&_owner); 1249 if (THREAD != cur) { 1250 if (THREAD->is_lock_owned((address)cur)) { 1251 assert(_recursions == 0, "internal state error"); 1252 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*. 1253 _recursions = 0; 1254 } 1255 } 1256 1257 guarantee(Self == _owner, "complete_exit not owner"); 1258 intx save = _recursions; // record the old recursion count 1259 _recursions = 0; // set the recursion level to be 0 1260 exit(true, Self); // exit the monitor 1261 guarantee(_owner != Self, "invariant"); 1262 return save; 1263 } 1264 1265 // reenter() enters a lock and sets recursion count 1266 // complete_exit/reenter operate as a wait without waiting 1267 bool ObjectMonitor::reenter(intx recursions, TRAPS) { 1268 Thread * const Self = THREAD; 1269 assert(Self->is_Java_thread(), "Must be Java thread!"); 1270 JavaThread *jt = (JavaThread *)THREAD; 1271 1272 guarantee(_owner != Self, "reenter already owner"); 1273 if (!enter(THREAD)) { 1274 return false; 1275 } 1276 // Entered the monitor. 1277 guarantee(_recursions == 0, "reenter recursion"); 1278 _recursions = recursions; 1279 return true; 1280 } 1281 1282 // Checks that the current THREAD owns this monitor and causes an 1283 // immediate return if it doesn't. We don't use the CHECK macro 1284 // because we want the IMSE to be the only exception that is thrown 1285 // from the call site when false is returned. Any other pending 1286 // exception is ignored. 1287 #define CHECK_OWNER() \ 1288 do { \ 1289 if (!check_owner(THREAD)) { \ 1290 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1291 return; \ 1292 } \ 1293 } while (false) 1294 1295 // Returns true if the specified thread owns the ObjectMonitor. 1296 // Otherwise returns false and throws IllegalMonitorStateException 1297 // (IMSE). If there is a pending exception and the specified thread 1298 // is not the owner, that exception will be replaced by the IMSE. 1299 bool ObjectMonitor::check_owner(Thread* THREAD) { 1300 void* cur = Atomic::load(&_owner); 1301 if (cur == THREAD) { 1302 return true; 1303 } 1304 if (THREAD->is_lock_owned((address)cur)) { 1305 set_owner_from_BasicLock(cur, THREAD); // Convert from BasicLock* to Thread*. 1306 _recursions = 0; 1307 return true; 1308 } 1309 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1310 "current thread is not owner", false); 1311 } 1312 1313 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1314 ObjectMonitor* monitor, 1315 jlong notifier_tid, 1316 jlong timeout, 1317 bool timedout) { 1318 assert(event != NULL, "invariant"); 1319 assert(monitor != NULL, "invariant"); 1320 event->set_monitorClass(((oop)monitor->object())->klass()); 1321 event->set_timeout(timeout); 1322 event->set_address((uintptr_t)monitor->object_addr()); 1323 event->set_notifier(notifier_tid); 1324 event->set_timedOut(timedout); 1325 event->commit(); 1326 } 1327 1328 // ----------------------------------------------------------------------------- 1329 // Wait/Notify/NotifyAll 1330 // 1331 // Note: a subset of changes to ObjectMonitor::wait() 1332 // will need to be replicated in complete_exit 1333 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1334 Thread * const Self = THREAD; 1335 assert(Self->is_Java_thread(), "Must be Java thread!"); 1336 JavaThread *jt = (JavaThread *)THREAD; 1337 1338 assert(InitDone, "Unexpectedly not initialized"); 1339 1340 CHECK_OWNER(); // Throws IMSE if not owner. 1341 1342 EventJavaMonitorWait event; 1343 1344 // check for a pending interrupt 1345 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1346 // post monitor waited event. Note that this is past-tense, we are done waiting. 1347 if (JvmtiExport::should_post_monitor_waited()) { 1348 // Note: 'false' parameter is passed here because the 1349 // wait was not timed out due to thread interrupt. 1350 JvmtiExport::post_monitor_waited(jt, this, false); 1351 1352 // In this short circuit of the monitor wait protocol, the 1353 // current thread never drops ownership of the monitor and 1354 // never gets added to the wait queue so the current thread 1355 // cannot be made the successor. This means that the 1356 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1357 // consume an unpark() meant for the ParkEvent associated with 1358 // this ObjectMonitor. 1359 } 1360 if (event.should_commit()) { 1361 post_monitor_wait_event(&event, this, 0, millis, false); 1362 } 1363 THROW(vmSymbols::java_lang_InterruptedException()); 1364 return; 1365 } 1366 1367 assert(Self->_Stalled == 0, "invariant"); 1368 Self->_Stalled = intptr_t(this); 1369 jt->set_current_waiting_monitor(this); 1370 1371 // create a node to be put into the queue 1372 // Critically, after we reset() the event but prior to park(), we must check 1373 // for a pending interrupt. 1374 ObjectWaiter node(Self); 1375 node.TState = ObjectWaiter::TS_WAIT; 1376 Self->_ParkEvent->reset(); 1377 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1378 1379 // Enter the waiting queue, which is a circular doubly linked list in this case 1380 // but it could be a priority queue or any data structure. 1381 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1382 // by the the owner of the monitor *except* in the case where park() 1383 // returns because of a timeout of interrupt. Contention is exceptionally rare 1384 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1385 1386 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1387 AddWaiter(&node); 1388 Thread::SpinRelease(&_WaitSetLock); 1389 1390 _Responsible = NULL; 1391 1392 intx save = _recursions; // record the old recursion count 1393 _waiters++; // increment the number of waiters 1394 _recursions = 0; // set the recursion level to be 1 1395 exit(true, Self); // exit the monitor 1396 guarantee(_owner != Self, "invariant"); 1397 1398 // The thread is on the WaitSet list - now park() it. 1399 // On MP systems it's conceivable that a brief spin before we park 1400 // could be profitable. 1401 // 1402 // TODO-FIXME: change the following logic to a loop of the form 1403 // while (!timeout && !interrupted && _notified == 0) park() 1404 1405 int ret = OS_OK; 1406 int WasNotified = 0; 1407 1408 // Need to check interrupt state whilst still _thread_in_vm 1409 bool interrupted = interruptible && jt->is_interrupted(false); 1410 1411 { // State transition wrappers 1412 OSThread* osthread = Self->osthread(); 1413 OSThreadWaitState osts(osthread, true); 1414 { 1415 ThreadBlockInVM tbivm(jt); 1416 // Thread is in thread_blocked state and oop access is unsafe. 1417 jt->set_suspend_equivalent(); 1418 1419 if (interrupted || HAS_PENDING_EXCEPTION) { 1420 // Intentionally empty 1421 } else if (node._notified == 0) { 1422 if (millis <= 0) { 1423 Self->_ParkEvent->park(); 1424 } else { 1425 ret = Self->_ParkEvent->park(millis); 1426 } 1427 } 1428 1429 // were we externally suspended while we were waiting? 1430 if (ExitSuspendEquivalent (jt)) { 1431 // TODO-FIXME: add -- if succ == Self then succ = null. 1432 jt->java_suspend_self(); 1433 } 1434 1435 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1436 1437 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1438 // from the WaitSet to the EntryList. 1439 // See if we need to remove Node from the WaitSet. 1440 // We use double-checked locking to avoid grabbing _WaitSetLock 1441 // if the thread is not on the wait queue. 1442 // 1443 // Note that we don't need a fence before the fetch of TState. 1444 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1445 // written by the is thread. (perhaps the fetch might even be satisfied 1446 // by a look-aside into the processor's own store buffer, although given 1447 // the length of the code path between the prior ST and this load that's 1448 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1449 // then we'll acquire the lock and then re-fetch a fresh TState value. 1450 // That is, we fail toward safety. 1451 1452 if (node.TState == ObjectWaiter::TS_WAIT) { 1453 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1454 if (node.TState == ObjectWaiter::TS_WAIT) { 1455 DequeueSpecificWaiter(&node); // unlink from WaitSet 1456 assert(node._notified == 0, "invariant"); 1457 node.TState = ObjectWaiter::TS_RUN; 1458 } 1459 Thread::SpinRelease(&_WaitSetLock); 1460 } 1461 1462 // The thread is now either on off-list (TS_RUN), 1463 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1464 // The Node's TState variable is stable from the perspective of this thread. 1465 // No other threads will asynchronously modify TState. 1466 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1467 OrderAccess::loadload(); 1468 if (_succ == Self) _succ = NULL; 1469 WasNotified = node._notified; 1470 1471 // Reentry phase -- reacquire the monitor. 1472 // re-enter contended monitor after object.wait(). 1473 // retain OBJECT_WAIT state until re-enter successfully completes 1474 // Thread state is thread_in_vm and oop access is again safe, 1475 // although the raw address of the object may have changed. 1476 // (Don't cache naked oops over safepoints, of course). 1477 1478 // post monitor waited event. Note that this is past-tense, we are done waiting. 1479 if (JvmtiExport::should_post_monitor_waited()) { 1480 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1481 1482 if (node._notified != 0 && _succ == Self) { 1483 // In this part of the monitor wait-notify-reenter protocol it 1484 // is possible (and normal) for another thread to do a fastpath 1485 // monitor enter-exit while this thread is still trying to get 1486 // to the reenter portion of the protocol. 1487 // 1488 // The ObjectMonitor was notified and the current thread is 1489 // the successor which also means that an unpark() has already 1490 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1491 // consume the unpark() that was done when the successor was 1492 // set because the same ParkEvent is shared between Java 1493 // monitors and JVM/TI RawMonitors (for now). 1494 // 1495 // We redo the unpark() to ensure forward progress, i.e., we 1496 // don't want all pending threads hanging (parked) with none 1497 // entering the unlocked monitor. 1498 node._event->unpark(); 1499 } 1500 } 1501 1502 if (event.should_commit()) { 1503 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1504 } 1505 1506 OrderAccess::fence(); 1507 1508 assert(Self->_Stalled != 0, "invariant"); 1509 Self->_Stalled = 0; 1510 1511 assert(_owner != Self, "invariant"); 1512 ObjectWaiter::TStates v = node.TState; 1513 if (v == ObjectWaiter::TS_RUN) { 1514 enter(Self); 1515 } else { 1516 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1517 ReenterI(Self, &node); 1518 node.wait_reenter_end(this); 1519 } 1520 1521 // Self has reacquired the lock. 1522 // Lifecycle - the node representing Self must not appear on any queues. 1523 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1524 // want residual elements associated with this thread left on any lists. 1525 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1526 assert(_owner == Self, "invariant"); 1527 assert(_succ != Self, "invariant"); 1528 } // OSThreadWaitState() 1529 1530 jt->set_current_waiting_monitor(NULL); 1531 1532 guarantee(_recursions == 0, "invariant"); 1533 _recursions = save // restore the old recursion count 1534 + JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(jt); // increased by the deferred relock count 1535 _waiters--; // decrement the number of waiters 1536 1537 // Verify a few postconditions 1538 assert(_owner == Self, "invariant"); 1539 assert(_succ != Self, "invariant"); 1540 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant"); 1541 1542 // check if the notification happened 1543 if (!WasNotified) { 1544 // no, it could be timeout or Thread.interrupt() or both 1545 // check for interrupt event, otherwise it is timeout 1546 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1547 THROW(vmSymbols::java_lang_InterruptedException()); 1548 } 1549 } 1550 1551 // NOTE: Spurious wake up will be consider as timeout. 1552 // Monitor notify has precedence over thread interrupt. 1553 } 1554 1555 1556 // Consider: 1557 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1558 // then instead of transferring a thread from the WaitSet to the EntryList 1559 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1560 1561 void ObjectMonitor::INotify(Thread * Self) { 1562 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1563 ObjectWaiter * iterator = DequeueWaiter(); 1564 if (iterator != NULL) { 1565 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1566 guarantee(iterator->_notified == 0, "invariant"); 1567 // Disposition - what might we do with iterator ? 1568 // a. add it directly to the EntryList - either tail (policy == 1) 1569 // or head (policy == 0). 1570 // b. push it onto the front of the _cxq (policy == 2). 1571 // For now we use (b). 1572 1573 iterator->TState = ObjectWaiter::TS_ENTER; 1574 1575 iterator->_notified = 1; 1576 iterator->_notifier_tid = JFR_THREAD_ID(Self); 1577 1578 ObjectWaiter * list = _EntryList; 1579 if (list != NULL) { 1580 assert(list->_prev == NULL, "invariant"); 1581 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1582 assert(list != iterator, "invariant"); 1583 } 1584 1585 // prepend to cxq 1586 if (list == NULL) { 1587 iterator->_next = iterator->_prev = NULL; 1588 _EntryList = iterator; 1589 } else { 1590 iterator->TState = ObjectWaiter::TS_CXQ; 1591 for (;;) { 1592 ObjectWaiter * front = _cxq; 1593 iterator->_next = front; 1594 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1595 break; 1596 } 1597 } 1598 } 1599 1600 // _WaitSetLock protects the wait queue, not the EntryList. We could 1601 // move the add-to-EntryList operation, above, outside the critical section 1602 // protected by _WaitSetLock. In practice that's not useful. With the 1603 // exception of wait() timeouts and interrupts the monitor owner 1604 // is the only thread that grabs _WaitSetLock. There's almost no contention 1605 // on _WaitSetLock so it's not profitable to reduce the length of the 1606 // critical section. 1607 1608 iterator->wait_reenter_begin(this); 1609 } 1610 Thread::SpinRelease(&_WaitSetLock); 1611 } 1612 1613 // Consider: a not-uncommon synchronization bug is to use notify() when 1614 // notifyAll() is more appropriate, potentially resulting in stranded 1615 // threads; this is one example of a lost wakeup. A useful diagnostic 1616 // option is to force all notify() operations to behave as notifyAll(). 1617 // 1618 // Note: We can also detect many such problems with a "minimum wait". 1619 // When the "minimum wait" is set to a small non-zero timeout value 1620 // and the program does not hang whereas it did absent "minimum wait", 1621 // that suggests a lost wakeup bug. 1622 1623 void ObjectMonitor::notify(TRAPS) { 1624 CHECK_OWNER(); // Throws IMSE if not owner. 1625 if (_WaitSet == NULL) { 1626 return; 1627 } 1628 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1629 INotify(THREAD); 1630 OM_PERFDATA_OP(Notifications, inc(1)); 1631 } 1632 1633 1634 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1635 // from the waitset to the EntryList. This could be done more efficiently with a 1636 // single bulk transfer but in practice it's not time-critical. Beware too, 1637 // that in prepend-mode we invert the order of the waiters. Let's say that the 1638 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1639 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1640 1641 void ObjectMonitor::notifyAll(TRAPS) { 1642 CHECK_OWNER(); // Throws IMSE if not owner. 1643 if (_WaitSet == NULL) { 1644 return; 1645 } 1646 1647 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1648 int tally = 0; 1649 while (_WaitSet != NULL) { 1650 tally++; 1651 INotify(THREAD); 1652 } 1653 1654 OM_PERFDATA_OP(Notifications, inc(tally)); 1655 } 1656 1657 // ----------------------------------------------------------------------------- 1658 // Adaptive Spinning Support 1659 // 1660 // Adaptive spin-then-block - rational spinning 1661 // 1662 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1663 // algorithm. On high order SMP systems it would be better to start with 1664 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1665 // a contending thread could enqueue itself on the cxq and then spin locally 1666 // on a thread-specific variable such as its ParkEvent._Event flag. 1667 // That's left as an exercise for the reader. Note that global spinning is 1668 // not problematic on Niagara, as the L2 cache serves the interconnect and 1669 // has both low latency and massive bandwidth. 1670 // 1671 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1672 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1673 // (duration) or we can fix the count at approximately the duration of 1674 // a context switch and vary the frequency. Of course we could also 1675 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1676 // For a description of 'Adaptive spin-then-block mutual exclusion in 1677 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1678 // 1679 // This implementation varies the duration "D", where D varies with 1680 // the success rate of recent spin attempts. (D is capped at approximately 1681 // length of a round-trip context switch). The success rate for recent 1682 // spin attempts is a good predictor of the success rate of future spin 1683 // attempts. The mechanism adapts automatically to varying critical 1684 // section length (lock modality), system load and degree of parallelism. 1685 // D is maintained per-monitor in _SpinDuration and is initialized 1686 // optimistically. Spin frequency is fixed at 100%. 1687 // 1688 // Note that _SpinDuration is volatile, but we update it without locks 1689 // or atomics. The code is designed so that _SpinDuration stays within 1690 // a reasonable range even in the presence of races. The arithmetic 1691 // operations on _SpinDuration are closed over the domain of legal values, 1692 // so at worst a race will install and older but still legal value. 1693 // At the very worst this introduces some apparent non-determinism. 1694 // We might spin when we shouldn't or vice-versa, but since the spin 1695 // count are relatively short, even in the worst case, the effect is harmless. 1696 // 1697 // Care must be taken that a low "D" value does not become an 1698 // an absorbing state. Transient spinning failures -- when spinning 1699 // is overall profitable -- should not cause the system to converge 1700 // on low "D" values. We want spinning to be stable and predictable 1701 // and fairly responsive to change and at the same time we don't want 1702 // it to oscillate, become metastable, be "too" non-deterministic, 1703 // or converge on or enter undesirable stable absorbing states. 1704 // 1705 // We implement a feedback-based control system -- using past behavior 1706 // to predict future behavior. We face two issues: (a) if the 1707 // input signal is random then the spin predictor won't provide optimal 1708 // results, and (b) if the signal frequency is too high then the control 1709 // system, which has some natural response lag, will "chase" the signal. 1710 // (b) can arise from multimodal lock hold times. Transient preemption 1711 // can also result in apparent bimodal lock hold times. 1712 // Although sub-optimal, neither condition is particularly harmful, as 1713 // in the worst-case we'll spin when we shouldn't or vice-versa. 1714 // The maximum spin duration is rather short so the failure modes aren't bad. 1715 // To be conservative, I've tuned the gain in system to bias toward 1716 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1717 // "rings" or oscillates between spinning and not spinning. This happens 1718 // when spinning is just on the cusp of profitability, however, so the 1719 // situation is not dire. The state is benign -- there's no need to add 1720 // hysteresis control to damp the transition rate between spinning and 1721 // not spinning. 1722 1723 // Spinning: Fixed frequency (100%), vary duration 1724 int ObjectMonitor::TrySpin(Thread * Self) { 1725 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1726 int ctr = Knob_FixedSpin; 1727 if (ctr != 0) { 1728 while (--ctr >= 0) { 1729 if (TryLock(Self) > 0) return 1; 1730 SpinPause(); 1731 } 1732 return 0; 1733 } 1734 1735 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1736 if (TryLock(Self) > 0) { 1737 // Increase _SpinDuration ... 1738 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1739 // Raising _SpurDuration to the poverty line is key. 1740 int x = _SpinDuration; 1741 if (x < Knob_SpinLimit) { 1742 if (x < Knob_Poverty) x = Knob_Poverty; 1743 _SpinDuration = x + Knob_BonusB; 1744 } 1745 return 1; 1746 } 1747 SpinPause(); 1748 } 1749 1750 // Admission control - verify preconditions for spinning 1751 // 1752 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1753 // becoming an absorbing state. Put another way, we spin briefly to 1754 // sample, just in case the system load, parallelism, contention, or lock 1755 // modality changed. 1756 // 1757 // Consider the following alternative: 1758 // Periodically set _SpinDuration = _SpinLimit and try a long/full 1759 // spin attempt. "Periodically" might mean after a tally of 1760 // the # of failed spin attempts (or iterations) reaches some threshold. 1761 // This takes us into the realm of 1-out-of-N spinning, where we 1762 // hold the duration constant but vary the frequency. 1763 1764 ctr = _SpinDuration; 1765 if (ctr <= 0) return 0; 1766 1767 if (NotRunnable(Self, (Thread *) _owner)) { 1768 return 0; 1769 } 1770 1771 // We're good to spin ... spin ingress. 1772 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 1773 // when preparing to LD...CAS _owner, etc and the CAS is likely 1774 // to succeed. 1775 if (_succ == NULL) { 1776 _succ = Self; 1777 } 1778 Thread * prv = NULL; 1779 1780 // There are three ways to exit the following loop: 1781 // 1. A successful spin where this thread has acquired the lock. 1782 // 2. Spin failure with prejudice 1783 // 3. Spin failure without prejudice 1784 1785 while (--ctr >= 0) { 1786 1787 // Periodic polling -- Check for pending GC 1788 // Threads may spin while they're unsafe. 1789 // We don't want spinning threads to delay the JVM from reaching 1790 // a stop-the-world safepoint or to steal cycles from GC. 1791 // If we detect a pending safepoint we abort in order that 1792 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 1793 // this thread, if safe, doesn't steal cycles from GC. 1794 // This is in keeping with the "no loitering in runtime" rule. 1795 // We periodically check to see if there's a safepoint pending. 1796 if ((ctr & 0xFF) == 0) { 1797 if (SafepointMechanism::should_block(Self)) { 1798 goto Abort; // abrupt spin egress 1799 } 1800 SpinPause(); 1801 } 1802 1803 // Probe _owner with TATAS 1804 // If this thread observes the monitor transition or flicker 1805 // from locked to unlocked to locked, then the odds that this 1806 // thread will acquire the lock in this spin attempt go down 1807 // considerably. The same argument applies if the CAS fails 1808 // or if we observe _owner change from one non-null value to 1809 // another non-null value. In such cases we might abort 1810 // the spin without prejudice or apply a "penalty" to the 1811 // spin count-down variable "ctr", reducing it by 100, say. 1812 1813 Thread * ox = (Thread *) _owner; 1814 if (ox == NULL) { 1815 ox = (Thread*)try_set_owner_from(NULL, Self); 1816 if (ox == NULL) { 1817 // The CAS succeeded -- this thread acquired ownership 1818 // Take care of some bookkeeping to exit spin state. 1819 if (_succ == Self) { 1820 _succ = NULL; 1821 } 1822 1823 // Increase _SpinDuration : 1824 // The spin was successful (profitable) so we tend toward 1825 // longer spin attempts in the future. 1826 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 1827 // If we acquired the lock early in the spin cycle it 1828 // makes sense to increase _SpinDuration proportionally. 1829 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1830 int x = _SpinDuration; 1831 if (x < Knob_SpinLimit) { 1832 if (x < Knob_Poverty) x = Knob_Poverty; 1833 _SpinDuration = x + Knob_Bonus; 1834 } 1835 return 1; 1836 } 1837 1838 // The CAS failed ... we can take any of the following actions: 1839 // * penalize: ctr -= CASPenalty 1840 // * exit spin with prejudice -- goto Abort; 1841 // * exit spin without prejudice. 1842 // * Since CAS is high-latency, retry again immediately. 1843 prv = ox; 1844 goto Abort; 1845 } 1846 1847 // Did lock ownership change hands ? 1848 if (ox != prv && prv != NULL) { 1849 goto Abort; 1850 } 1851 prv = ox; 1852 1853 // Abort the spin if the owner is not executing. 1854 // The owner must be executing in order to drop the lock. 1855 // Spinning while the owner is OFFPROC is idiocy. 1856 // Consider: ctr -= RunnablePenalty ; 1857 if (NotRunnable(Self, ox)) { 1858 goto Abort; 1859 } 1860 if (_succ == NULL) { 1861 _succ = Self; 1862 } 1863 } 1864 1865 // Spin failed with prejudice -- reduce _SpinDuration. 1866 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 1867 // AIMD is globally stable. 1868 { 1869 int x = _SpinDuration; 1870 if (x > 0) { 1871 // Consider an AIMD scheme like: x -= (x >> 3) + 100 1872 // This is globally sample and tends to damp the response. 1873 x -= Knob_Penalty; 1874 if (x < 0) x = 0; 1875 _SpinDuration = x; 1876 } 1877 } 1878 1879 Abort: 1880 if (_succ == Self) { 1881 _succ = NULL; 1882 // Invariant: after setting succ=null a contending thread 1883 // must recheck-retry _owner before parking. This usually happens 1884 // in the normal usage of TrySpin(), but it's safest 1885 // to make TrySpin() as foolproof as possible. 1886 OrderAccess::fence(); 1887 if (TryLock(Self) > 0) return 1; 1888 } 1889 return 0; 1890 } 1891 1892 // NotRunnable() -- informed spinning 1893 // 1894 // Don't bother spinning if the owner is not eligible to drop the lock. 1895 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm. 1896 // The thread must be runnable in order to drop the lock in timely fashion. 1897 // If the _owner is not runnable then spinning will not likely be 1898 // successful (profitable). 1899 // 1900 // Beware -- the thread referenced by _owner could have died 1901 // so a simply fetch from _owner->_thread_state might trap. 1902 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 1903 // Because of the lifecycle issues, the _thread_state values 1904 // observed by NotRunnable() might be garbage. NotRunnable must 1905 // tolerate this and consider the observed _thread_state value 1906 // as advisory. 1907 // 1908 // Beware too, that _owner is sometimes a BasicLock address and sometimes 1909 // a thread pointer. 1910 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 1911 // with the LSB of _owner. Another option would be to probabilistically probe 1912 // the putative _owner->TypeTag value. 1913 // 1914 // Checking _thread_state isn't perfect. Even if the thread is 1915 // in_java it might be blocked on a page-fault or have been preempted 1916 // and sitting on a ready/dispatch queue. 1917 // 1918 // The return value from NotRunnable() is *advisory* -- the 1919 // result is based on sampling and is not necessarily coherent. 1920 // The caller must tolerate false-negative and false-positive errors. 1921 // Spinning, in general, is probabilistic anyway. 1922 1923 1924 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 1925 // Check ox->TypeTag == 2BAD. 1926 if (ox == NULL) return 0; 1927 1928 // Avoid transitive spinning ... 1929 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 1930 // Immediately after T1 acquires L it's possible that T2, also 1931 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 1932 // This occurs transiently after T1 acquired L but before 1933 // T1 managed to clear T1.Stalled. T2 does not need to abort 1934 // its spin in this circumstance. 1935 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 1936 1937 if (BlockedOn == 1) return 1; 1938 if (BlockedOn != 0) { 1939 return BlockedOn != intptr_t(this) && _owner == ox; 1940 } 1941 1942 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 1943 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 1944 // consider also: jst != _thread_in_Java -- but that's overspecific. 1945 return jst == _thread_blocked || jst == _thread_in_native; 1946 } 1947 1948 1949 // ----------------------------------------------------------------------------- 1950 // WaitSet management ... 1951 1952 ObjectWaiter::ObjectWaiter(Thread* thread) { 1953 _next = NULL; 1954 _prev = NULL; 1955 _notified = 0; 1956 _notifier_tid = 0; 1957 TState = TS_RUN; 1958 _thread = thread; 1959 _event = thread->_ParkEvent; 1960 _active = false; 1961 assert(_event != NULL, "invariant"); 1962 } 1963 1964 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 1965 JavaThread *jt = (JavaThread *)this->_thread; 1966 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 1967 } 1968 1969 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 1970 JavaThread *jt = (JavaThread *)this->_thread; 1971 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 1972 } 1973 1974 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1975 assert(node != NULL, "should not add NULL node"); 1976 assert(node->_prev == NULL, "node already in list"); 1977 assert(node->_next == NULL, "node already in list"); 1978 // put node at end of queue (circular doubly linked list) 1979 if (_WaitSet == NULL) { 1980 _WaitSet = node; 1981 node->_prev = node; 1982 node->_next = node; 1983 } else { 1984 ObjectWaiter* head = _WaitSet; 1985 ObjectWaiter* tail = head->_prev; 1986 assert(tail->_next == head, "invariant check"); 1987 tail->_next = node; 1988 head->_prev = node; 1989 node->_next = head; 1990 node->_prev = tail; 1991 } 1992 } 1993 1994 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1995 // dequeue the very first waiter 1996 ObjectWaiter* waiter = _WaitSet; 1997 if (waiter) { 1998 DequeueSpecificWaiter(waiter); 1999 } 2000 return waiter; 2001 } 2002 2003 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2004 assert(node != NULL, "should not dequeue NULL node"); 2005 assert(node->_prev != NULL, "node already removed from list"); 2006 assert(node->_next != NULL, "node already removed from list"); 2007 // when the waiter has woken up because of interrupt, 2008 // timeout or other spurious wake-up, dequeue the 2009 // waiter from waiting list 2010 ObjectWaiter* next = node->_next; 2011 if (next == node) { 2012 assert(node->_prev == node, "invariant check"); 2013 _WaitSet = NULL; 2014 } else { 2015 ObjectWaiter* prev = node->_prev; 2016 assert(prev->_next == node, "invariant check"); 2017 assert(next->_prev == node, "invariant check"); 2018 next->_prev = prev; 2019 prev->_next = next; 2020 if (_WaitSet == node) { 2021 _WaitSet = next; 2022 } 2023 } 2024 node->_next = NULL; 2025 node->_prev = NULL; 2026 } 2027 2028 // ----------------------------------------------------------------------------- 2029 // PerfData support 2030 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2031 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2032 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2033 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2034 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2035 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2036 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2037 2038 // One-shot global initialization for the sync subsystem. 2039 // We could also defer initialization and initialize on-demand 2040 // the first time we call ObjectSynchronizer::inflate(). 2041 // Initialization would be protected - like so many things - by 2042 // the MonitorCache_lock. 2043 2044 void ObjectMonitor::Initialize() { 2045 assert(!InitDone, "invariant"); 2046 2047 if (!os::is_MP()) { 2048 Knob_SpinLimit = 0; 2049 Knob_PreSpin = 0; 2050 Knob_FixedSpin = -1; 2051 } 2052 2053 if (UsePerfData) { 2054 EXCEPTION_MARK; 2055 #define NEWPERFCOUNTER(n) \ 2056 { \ 2057 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2058 CHECK); \ 2059 } 2060 #define NEWPERFVARIABLE(n) \ 2061 { \ 2062 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2063 CHECK); \ 2064 } 2065 NEWPERFCOUNTER(_sync_Inflations); 2066 NEWPERFCOUNTER(_sync_Deflations); 2067 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2068 NEWPERFCOUNTER(_sync_FutileWakeups); 2069 NEWPERFCOUNTER(_sync_Parks); 2070 NEWPERFCOUNTER(_sync_Notifications); 2071 NEWPERFVARIABLE(_sync_MonExtant); 2072 #undef NEWPERFCOUNTER 2073 #undef NEWPERFVARIABLE 2074 } 2075 2076 DEBUG_ONLY(InitDone = true;) 2077 } 2078 2079 void ObjectMonitor::print_on(outputStream* st) const { 2080 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2081 st->print("{contentions=0x%08x,waiters=0x%08x" 2082 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2083 contentions(), waiters(), recursions(), 2084 p2i(owner())); 2085 } 2086 void ObjectMonitor::print() const { print_on(tty); } 2087 2088 #ifdef ASSERT 2089 // Print the ObjectMonitor like a debugger would: 2090 // 2091 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2092 // _header = 0x0000000000000001 2093 // _object = 0x000000070ff45fd0 2094 // _allocation_state = Old 2095 // _pad_buf0 = { 2096 // [0] = '\0' 2097 // ... 2098 // [43] = '\0' 2099 // } 2100 // _owner = 0x0000000000000000 2101 // _previous_owner_tid = 0 2102 // _pad_buf1 = { 2103 // [0] = '\0' 2104 // ... 2105 // [47] = '\0' 2106 // } 2107 // _next_om = 0x0000000000000000 2108 // _recursions = 0 2109 // _EntryList = 0x0000000000000000 2110 // _cxq = 0x0000000000000000 2111 // _succ = 0x0000000000000000 2112 // _Responsible = 0x0000000000000000 2113 // _Spinner = 0 2114 // _SpinDuration = 5000 2115 // _contentions = 0 2116 // _WaitSet = 0x0000700009756248 2117 // _waiters = 1 2118 // _WaitSetLock = 0 2119 // } 2120 // 2121 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2122 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2123 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2124 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object)); 2125 st->print(" _allocation_state = "); 2126 if (is_free()) { 2127 st->print("Free"); 2128 } else if (is_old()) { 2129 st->print("Old"); 2130 } else if (is_new()) { 2131 st->print("New"); 2132 } else { 2133 st->print("unknown=%d", _allocation_state); 2134 } 2135 st->cr(); 2136 st->print_cr(" _pad_buf0 = {"); 2137 st->print_cr(" [0] = '\\0'"); 2138 st->print_cr(" ..."); 2139 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2140 st->print_cr(" }"); 2141 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner)); 2142 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid); 2143 st->print_cr(" _pad_buf1 = {"); 2144 st->print_cr(" [0] = '\\0'"); 2145 st->print_cr(" ..."); 2146 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2147 st->print_cr(" }"); 2148 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2149 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2150 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2151 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2152 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2153 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2154 st->print_cr(" _Spinner = %d", _Spinner); 2155 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2156 st->print_cr(" _contentions = %d", contentions()); 2157 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2158 st->print_cr(" _waiters = %d", _waiters); 2159 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2160 st->print_cr("}"); 2161 } 2162 #endif