1 /* 2 * Copyright (c) 1998, 2016, 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 "memory/resourceArea.hpp" 28 #include "oops/markOop.hpp" 29 #include "oops/oop.inline.hpp" 30 #include "runtime/atomic.hpp" 31 #include "runtime/handles.inline.hpp" 32 #include "runtime/interfaceSupport.hpp" 33 #include "runtime/mutexLocker.hpp" 34 #include "runtime/objectMonitor.hpp" 35 #include "runtime/objectMonitor.inline.hpp" 36 #include "runtime/orderAccess.inline.hpp" 37 #include "runtime/osThread.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #include "runtime/thread.inline.hpp" 40 #include "services/threadService.hpp" 41 #include "trace/tracing.hpp" 42 #include "trace/traceMacros.hpp" 43 #include "utilities/dtrace.hpp" 44 #include "utilities/macros.hpp" 45 #include "utilities/preserveException.hpp" 46 47 #ifdef DTRACE_ENABLED 48 49 // Only bother with this argument setup if dtrace is available 50 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 51 52 53 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 54 char* bytes = NULL; \ 55 int len = 0; \ 56 jlong jtid = SharedRuntime::get_java_tid(thread); \ 57 Symbol* klassname = ((oop)obj)->klass()->name(); \ 58 if (klassname != NULL) { \ 59 bytes = (char*)klassname->bytes(); \ 60 len = klassname->utf8_length(); \ 61 } 62 63 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 64 { \ 65 if (DTraceMonitorProbes) { \ 66 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 67 HOTSPOT_MONITOR_WAIT(jtid, \ 68 (monitor), bytes, len, (millis)); \ 69 } \ 70 } 71 72 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 73 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 74 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 75 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 76 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 77 78 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 79 { \ 80 if (DTraceMonitorProbes) { \ 81 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 82 HOTSPOT_MONITOR_##probe(jtid, \ 83 (uintptr_t)(monitor), bytes, len); \ 84 } \ 85 } 86 87 #else // ndef DTRACE_ENABLED 88 89 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 90 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 91 92 #endif // ndef DTRACE_ENABLED 93 94 // Tunables ... 95 // The knob* variables are effectively final. Once set they should 96 // never be modified hence. Consider using __read_mostly with GCC. 97 98 int ObjectMonitor::Knob_ExitRelease = 0; 99 int ObjectMonitor::Knob_Verbose = 0; 100 int ObjectMonitor::Knob_VerifyInUse = 0; 101 int ObjectMonitor::Knob_VerifyMatch = 0; 102 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 103 static int Knob_LogSpins = 0; // enable jvmstat tally for spins 104 static int Knob_HandOff = 0; 105 static int Knob_ReportSettings = 0; 106 107 static int Knob_SpinBase = 0; // Floor AKA SpinMin 108 static int Knob_SpinBackOff = 0; // spin-loop backoff 109 static int Knob_CASPenalty = -1; // Penalty for failed CAS 110 static int Knob_OXPenalty = -1; // Penalty for observed _owner change 111 static int Knob_SpinSetSucc = 1; // spinners set the _succ field 112 static int Knob_SpinEarly = 1; 113 static int Knob_SuccEnabled = 1; // futile wake throttling 114 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 115 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 116 static int Knob_Bonus = 100; // spin success bonus 117 static int Knob_BonusB = 100; // spin success bonus 118 static int Knob_Penalty = 200; // spin failure penalty 119 static int Knob_Poverty = 1000; 120 static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 121 static int Knob_FixedSpin = 0; 122 static int Knob_OState = 3; // Spinner checks thread state of _owner 123 static int Knob_UsePause = 1; 124 static int Knob_ExitPolicy = 0; 125 static int Knob_PreSpin = 10; // 20-100 likely better 126 static int Knob_ResetEvent = 0; 127 static int BackOffMask = 0; 128 129 static int Knob_FastHSSEC = 0; 130 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 131 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 132 static volatile int InitDone = 0; 133 134 // ----------------------------------------------------------------------------- 135 // Theory of operations -- Monitors lists, thread residency, etc: 136 // 137 // * A thread acquires ownership of a monitor by successfully 138 // CAS()ing the _owner field from null to non-null. 139 // 140 // * Invariant: A thread appears on at most one monitor list -- 141 // cxq, EntryList or WaitSet -- at any one time. 142 // 143 // * Contending threads "push" themselves onto the cxq with CAS 144 // and then spin/park. 145 // 146 // * After a contending thread eventually acquires the lock it must 147 // dequeue itself from either the EntryList or the cxq. 148 // 149 // * The exiting thread identifies and unparks an "heir presumptive" 150 // tentative successor thread on the EntryList. Critically, the 151 // exiting thread doesn't unlink the successor thread from the EntryList. 152 // After having been unparked, the wakee will recontend for ownership of 153 // the monitor. The successor (wakee) will either acquire the lock or 154 // re-park itself. 155 // 156 // Succession is provided for by a policy of competitive handoff. 157 // The exiting thread does _not_ grant or pass ownership to the 158 // successor thread. (This is also referred to as "handoff" succession"). 159 // Instead the exiting thread releases ownership and possibly wakes 160 // a successor, so the successor can (re)compete for ownership of the lock. 161 // If the EntryList is empty but the cxq is populated the exiting 162 // thread will drain the cxq into the EntryList. It does so by 163 // by detaching the cxq (installing null with CAS) and folding 164 // the threads from the cxq into the EntryList. The EntryList is 165 // doubly linked, while the cxq is singly linked because of the 166 // CAS-based "push" used to enqueue recently arrived threads (RATs). 167 // 168 // * Concurrency invariants: 169 // 170 // -- only the monitor owner may access or mutate the EntryList. 171 // The mutex property of the monitor itself protects the EntryList 172 // from concurrent interference. 173 // -- Only the monitor owner may detach the cxq. 174 // 175 // * The monitor entry list operations avoid locks, but strictly speaking 176 // they're not lock-free. Enter is lock-free, exit is not. 177 // For a description of 'Methods and apparatus providing non-blocking access 178 // to a resource,' see U.S. Pat. No. 7844973. 179 // 180 // * The cxq can have multiple concurrent "pushers" but only one concurrent 181 // detaching thread. This mechanism is immune from the ABA corruption. 182 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 183 // 184 // * Taken together, the cxq and the EntryList constitute or form a 185 // single logical queue of threads stalled trying to acquire the lock. 186 // We use two distinct lists to improve the odds of a constant-time 187 // dequeue operation after acquisition (in the ::enter() epilogue) and 188 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 189 // A key desideratum is to minimize queue & monitor metadata manipulation 190 // that occurs while holding the monitor lock -- that is, we want to 191 // minimize monitor lock holds times. Note that even a small amount of 192 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 193 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 194 // locks and monitor metadata. 195 // 196 // Cxq points to the set of Recently Arrived Threads attempting entry. 197 // Because we push threads onto _cxq with CAS, the RATs must take the form of 198 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 199 // the unlocking thread notices that EntryList is null but _cxq is != null. 200 // 201 // The EntryList is ordered by the prevailing queue discipline and 202 // can be organized in any convenient fashion, such as a doubly-linked list or 203 // a circular doubly-linked list. Critically, we want insert and delete operations 204 // to operate in constant-time. If we need a priority queue then something akin 205 // to Solaris' sleepq would work nicely. Viz., 206 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 207 // Queue discipline is enforced at ::exit() time, when the unlocking thread 208 // drains the cxq into the EntryList, and orders or reorders the threads on the 209 // EntryList accordingly. 210 // 211 // Barring "lock barging", this mechanism provides fair cyclic ordering, 212 // somewhat similar to an elevator-scan. 213 // 214 // * The monitor synchronization subsystem avoids the use of native 215 // synchronization primitives except for the narrow platform-specific 216 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 217 // the semantics of park-unpark. Put another way, this monitor implementation 218 // depends only on atomic operations and park-unpark. The monitor subsystem 219 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 220 // underlying OS manages the READY<->RUN transitions. 221 // 222 // * Waiting threads reside on the WaitSet list -- wait() puts 223 // the caller onto the WaitSet. 224 // 225 // * notify() or notifyAll() simply transfers threads from the WaitSet to 226 // either the EntryList or cxq. Subsequent exit() operations will 227 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 228 // it's likely the notifyee would simply impale itself on the lock held 229 // by the notifier. 230 // 231 // * An interesting alternative is to encode cxq as (List,LockByte) where 232 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 233 // variable, like _recursions, in the scheme. The threads or Events that form 234 // the list would have to be aligned in 256-byte addresses. A thread would 235 // try to acquire the lock or enqueue itself with CAS, but exiting threads 236 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 237 // Note that is is *not* word-tearing, but it does presume that full-word 238 // CAS operations are coherent with intermix with STB operations. That's true 239 // on most common processors. 240 // 241 // * See also http://blogs.sun.com/dave 242 243 #define DEFLATER_MARKER reinterpret_cast<void*>(-1) 244 245 // ----------------------------------------------------------------------------- 246 // Enter support 247 248 bool ObjectMonitor::enter(TRAPS) { 249 // The following code is ordered to check the most common cases first 250 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 251 Thread * const Self = THREAD; 252 253 void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL); 254 if (cur == NULL) { 255 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 256 assert(_recursions == 0, "invariant"); 257 assert(_owner == Self, "invariant"); 258 return true; 259 } 260 261 if (cur == Self) { 262 // TODO-FIXME: check for integer overflow! BUGID 6557169. 263 _recursions++; 264 return true; 265 } 266 267 if (Self->is_lock_owned ((address)cur)) { 268 assert(_recursions == 0, "internal state error"); 269 _recursions = 1; 270 // Commute owner from a thread-specific on-stack BasicLockObject address to 271 // a full-fledged "Thread *". 272 _owner = Self; 273 return true; 274 } 275 276 // We've encountered genuine contention. 277 assert(Self->_Stalled == 0, "invariant"); 278 Self->_Stalled = intptr_t(this); 279 280 // Try one round of spinning *before* enqueueing Self 281 // and before going through the awkward and expensive state 282 // transitions. The following spin is strictly optional ... 283 // Note that if we acquire the monitor from an initial spin 284 // we forgo posting JVMTI events and firing DTRACE probes. 285 if (Knob_SpinEarly && TrySpin (Self) > 0) { 286 assert(_owner == Self, "invariant"); 287 assert(_recursions == 0, "invariant"); 288 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 289 Self->_Stalled = 0; 290 return true; 291 } 292 293 assert(_owner != Self, "invariant"); 294 assert(_succ != Self, "invariant"); 295 assert(Self->is_Java_thread(), "invariant"); 296 JavaThread * jt = (JavaThread *) Self; 297 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 298 assert(jt->thread_state() != _thread_blocked, "invariant"); 299 assert(this->object() != NULL, "invariant"); 300 301 // Prevent deflation. See deflate_idle_monitors(), try_disable_monitor, and is_busy(). 302 // Ensure the object-monitor relationship remains stable while there's contention. 303 const jint count = Atomic::add(1, &_count); 304 if (count <= 0 && _owner == DEFLATER_MARKER) { 305 // Deflation in progress. 306 // Help deflater thread install the mark word (in case deflater thread is slow). 307 install_displaced_markword_in_object(); 308 Self->_Stalled = 0; 309 return false; // Caller should retry. Never mind about _count as this monitor has been deflated. 310 } 311 // The deflater thread will not deflate this monitor and the monitor is contended, continue. 312 313 EventJavaMonitorEnter event; 314 315 { // Change java thread status to indicate blocked on monitor enter. 316 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 317 318 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 319 if (JvmtiExport::should_post_monitor_contended_enter()) { 320 JvmtiExport::post_monitor_contended_enter(jt, this); 321 322 // The current thread does not yet own the monitor and does not 323 // yet appear on any queues that would get it made the successor. 324 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 325 // handler cannot accidentally consume an unpark() meant for the 326 // ParkEvent associated with this ObjectMonitor. 327 } 328 329 OSThreadContendState osts(Self->osthread()); 330 ThreadBlockInVM tbivm(jt); 331 332 Self->set_current_pending_monitor(this); 333 334 // TODO-FIXME: change the following for(;;) loop to straight-line code. 335 for (;;) { 336 jt->set_suspend_equivalent(); 337 // cleared by handle_special_suspend_equivalent_condition() 338 // or java_suspend_self() 339 340 EnterI(THREAD); 341 342 if (!ExitSuspendEquivalent(jt)) break; 343 344 // We have acquired the contended monitor, but while we were 345 // waiting another thread suspended us. We don't want to enter 346 // the monitor while suspended because that would surprise the 347 // thread that suspended us. 348 // 349 _recursions = 0; 350 _succ = NULL; 351 exit(false, Self); 352 353 jt->java_suspend_self(); 354 } 355 Self->set_current_pending_monitor(NULL); 356 357 // We cleared the pending monitor info since we've just gotten past 358 // the enter-check-for-suspend dance and we now own the monitor free 359 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 360 // destructor can go to a safepoint at the end of this block. If we 361 // do a thread dump during that safepoint, then this thread will show 362 // as having "-locked" the monitor, but the OS and java.lang.Thread 363 // states will still report that the thread is blocked trying to 364 // acquire it. 365 } 366 367 Atomic::dec(&_count); 368 Self->_Stalled = 0; 369 370 // Must either set _recursions = 0 or ASSERT _recursions == 0. 371 assert(_recursions == 0, "invariant"); 372 assert(_owner == Self, "invariant"); 373 assert(_succ != Self, "invariant"); 374 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 375 376 // The thread -- now the owner -- is back in vm mode. 377 // Report the glorious news via TI,DTrace and jvmstat. 378 // The probe effect is non-trivial. All the reportage occurs 379 // while we hold the monitor, increasing the length of the critical 380 // section. Amdahl's parallel speedup law comes vividly into play. 381 // 382 // Another option might be to aggregate the events (thread local or 383 // per-monitor aggregation) and defer reporting until a more opportune 384 // time -- such as next time some thread encounters contention but has 385 // yet to acquire the lock. While spinning that thread could 386 // spinning we could increment JVMStat counters, etc. 387 388 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 389 if (JvmtiExport::should_post_monitor_contended_entered()) { 390 JvmtiExport::post_monitor_contended_entered(jt, this); 391 392 // The current thread already owns the monitor and is not going to 393 // call park() for the remainder of the monitor enter protocol. So 394 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 395 // event handler consumed an unpark() issued by the thread that 396 // just exited the monitor. 397 } 398 399 if (event.should_commit()) { 400 event.set_monitorClass(((oop)this->object())->klass()); 401 event.set_previousOwner((TYPE_THREAD)_previous_owner_tid); 402 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 403 event.commit(); 404 } 405 406 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 407 return true; 408 } 409 410 411 // Caveat: TryLock() is not necessarily serializing if it returns failure. 412 // Callers must compensate as needed. 413 414 int ObjectMonitor::TryLock(Thread * Self) { 415 void * own = _owner; 416 if (own != NULL) return 0; 417 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 418 // Either guarantee _recursions == 0 or set _recursions = 0. 419 assert(_recursions == 0, "invariant"); 420 assert(_owner == Self, "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 // Try disabling this monitor. Returns true iff successful. 431 // The method will install DEFLATER_MARKER (-1) as the owner of the monitor, 432 // check _waiters == 0, and make _count negative if it is currently 0. The 433 // monitor is successfully disabled if _count is negative and this monitor is 434 // still owned by DEFLATER_MARKER. 435 // 436 // All threads trying to acquire the monitor must before parking themselves 437 // increment _count and check that _owner != DEFLATER_MARKER. If _owner == 438 // DEFLATER_MARKER and _count is positive, then a thread can still win the lock 439 // by atomically installing its thread pointer in _owner. If _count is 440 // negative and _owned == DEFLATER_MARKER, then the monitor has been 441 // successfully disabled and the acquiring threads should help install the 442 // displaced mark word back into the object and retry acquiring the lock. 443 // 444 // A thread wanting to wait on the monitor must increase _waiters while owning the monitor. 445 bool ObjectMonitor::try_disable_monitor() { 446 assert(Thread::current()->is_Java_thread(), "precondition"); 447 // We don't want to disable newly allocated monitors as it could result in a endless inflate/deflate cycle. 448 assert(is_old(), "precondition"); 449 450 // Set _owned to DEFLATER_MARKER if monitor is not owned by another thread. 451 // This forced contending thread through the slow path. 452 if (!is_busy() && Atomic::cmpxchg_ptr(DEFLATER_MARKER, &_owner, NULL) == NULL) { 453 // Another thread might still enter the monitor. 454 // Signal that other threads should retry if the owner is DEFLATER_MARKER by making _count negative. 455 if (_waiters == 0 && Atomic::cmpxchg(- max_jint, &_count, 0) == 0) { 456 // ABA problem with _count: 457 // Another thread might have acquired this monitor and finished using it. 458 // Check owner to see if that happened (no other thread installs DEFLATER_MARKER as owner). 459 if (_owner == DEFLATER_MARKER) { 460 // We successfully signalled to all threads entering that they should 461 // retry. 462 // Nobody acquired this monitor between installing DEFLATER_MARKER into 463 // _owner and now (such a thread would have changed _owner). If any 464 // thread is now waiting on the monitor, then _waiters must have been 465 // incremented as it was 0 before. _waiters is changed only when owning 466 // the monitor, but no other thread can have owned the monitor since we 467 // installed DEFLATER_MARKER, and thus _waiters must still be 0. 468 guarantee(_waiters == 0, "Not changed since the previous read"); 469 guarantee(_cxq == NULL, "All contending threads should retry"); 470 guarantee(_EntryList == NULL, "All contending threads should retry"); 471 // Install the old mark word if nobody else has already done it. 472 install_displaced_markword_in_object(); 473 set_allocation_state(Free); 474 // Leave this monitor locked to ensure acquiring threads take the slow-path and 475 // leave _count negative to make them retry. 476 return true; // Success, lock has been deflated. 477 } 478 // We do not own the monitor. Do not deflate. 479 Atomic::add(max_jint, &_count); 480 } 481 // Never mind. Another thread managed to acquire this monitor or there were 482 // threads waiting. The threads that saw (or will see) 0 <= _count and 483 // _owner == DEFLATER_MARKER will compete for ownership and one will 484 // eventually install itself as owner and subsequently run the exit protocol. 485 } 486 assert(0 <= _count, "Nobody else should be make _count negative"); 487 return false; 488 } 489 490 // Install the displaced mark word of a disabled monitor into the object 491 // associated with the monitor. 492 // This method is idempotent and is expected to be executed by both mutators 493 // wanting to acquire a monitor for an object, mutators wanting to install a 494 // hashcode in an object, and the thread deflating monitors. 495 void ObjectMonitor::install_displaced_markword_in_object() { 496 markOop dmw = header(); 497 assert(dmw->is_neutral() || dmw->hash() == 0 && dmw->is_marked(), "precondition"); 498 if (dmw->hash() == 0 && !dmw->is_marked()) { 499 // Another thread might update the displaced mark word by computing a hash 500 // code for the oject or running this method. 501 // Signal to other threads that the object mark word should be installed 502 // and read from the object by marking the displaced mark word. 503 markOop marked_dmw = dmw->set_marked(); 504 assert(marked_dmw->hash() == 0 && marked_dmw->is_marked(), "oops"); 505 dmw = (markOop) Atomic::cmpxchg_ptr(marked_dmw, &_header, dmw); 506 // If the CAS failed because another thread installed a hash value, then dmw 507 // will contain the hash and be unmarked. If the CAS failed because another thread 508 // marked the displaced mark word, then dmw->hash() is zero and dmw is marked. 509 } 510 if (dmw->is_marked()) { 511 assert(dmw->hash() == 0, "invariant"); 512 dmw = dmw->set_unmarked(); 513 } 514 oop const obj = (oop) object(); 515 // Install displaced mark word if object mark word still points to this monitor. 516 assert(dmw->is_neutral(), "Must not install non-neutral markword into object"); 517 obj->cas_set_mark(dmw, markOopDesc::encode(this)); 518 } 519 520 #define MAX_RECHECK_INTERVAL 1000 521 522 void ObjectMonitor::EnterI(TRAPS) { 523 Thread * const Self = THREAD; 524 assert(Self->is_Java_thread(), "invariant"); 525 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant"); 526 527 // Try the lock - TATAS 528 if (TryLock (Self) > 0) { 529 assert(_succ != Self, "invariant"); 530 assert(_owner == Self, "invariant"); 531 assert(_Responsible != Self, "invariant"); 532 return; 533 } 534 535 if (_owner == DEFLATER_MARKER) { 536 guarantee(0 < _count, "_owner == DEFLATER_MARKER && _count <= 0 should have been handled by the caller"); 537 // Deflater thread tried to lock this monitor, but it failed to make _count negative and gave up. 538 // Try to acquire monitor. 539 if (Atomic::cmpxchg_ptr(Self, &_owner, DEFLATER_MARKER) == DEFLATER_MARKER) { 540 assert(_succ != Self, "invariant"); 541 assert(_owner == Self, "invariant"); 542 assert(_Responsible != Self, "invariant"); 543 return; 544 } 545 } 546 547 DeferredInitialize(); 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_ptr(&node, &_cxq, nxt) == 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 ((SyncFlags & 16) == 0 && 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::cmpxchg_ptr(Self, &_Responsible, NULL); 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 TEVENT(Inflated enter - Contention); 642 int nWakeups = 0; 643 int recheckInterval = 1; 644 645 for (;;) { 646 647 if (TryLock(Self) > 0) break; 648 assert(_owner != Self, "invariant"); 649 650 if ((SyncFlags & 2) && _Responsible == NULL) { 651 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 652 } 653 654 // park self 655 if (_Responsible == Self || (SyncFlags & 1)) { 656 TEVENT(Inflated enter - park TIMED); 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 TEVENT(Inflated enter - park UNTIMED); 665 Self->_ParkEvent->park(); 666 } 667 668 if (TryLock(Self) > 0) break; 669 670 if (_owner == DEFLATER_MARKER) { 671 guarantee(0 < _count, "_owner == DEFLATER_MARKER && _count <= 0 should have been handled by the caller"); 672 // Deflater thread tried to lock this monitor, but it failed to make _count negative and gave up. 673 // Try to acquire monitor. 674 if (Atomic::cmpxchg_ptr(Self, &_owner, DEFLATER_MARKER) == DEFLATER_MARKER) { 675 break; 676 } 677 } 678 679 // The lock is still contested. 680 // Keep a tally of the # of futile wakeups. 681 // Note that the counter is not protected by a lock or updated by atomics. 682 // That is by design - we trade "lossy" counters which are exposed to 683 // races during updates for a lower probe effect. 684 TEVENT(Inflated enter - Futile wakeup); 685 // This PerfData object can be used in parallel with a safepoint. 686 // See the work around in PerfDataManager::destroy(). 687 OM_PERFDATA_OP(FutileWakeups, inc()); 688 ++nWakeups; 689 690 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 691 // We can defer clearing _succ until after the spin completes 692 // TrySpin() must tolerate being called with _succ == Self. 693 // Try yet another round of adaptive spinning. 694 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 695 696 // We can find that we were unpark()ed and redesignated _succ while 697 // we were spinning. That's harmless. If we iterate and call park(), 698 // park() will consume the event and return immediately and we'll 699 // just spin again. This pattern can repeat, leaving _succ to simply 700 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 701 // Alternately, we can sample fired() here, and if set, forgo spinning 702 // in the next iteration. 703 704 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 705 Self->_ParkEvent->reset(); 706 OrderAccess::fence(); 707 } 708 if (_succ == Self) _succ = NULL; 709 710 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 711 OrderAccess::fence(); 712 } 713 714 // Egress : 715 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 716 // Normally we'll find Self on the EntryList . 717 // From the perspective of the lock owner (this thread), the 718 // EntryList is stable and cxq is prepend-only. 719 // The head of cxq is volatile but the interior is stable. 720 // In addition, Self.TState is stable. 721 722 assert(_owner == Self, "invariant"); 723 assert(object() != NULL, "invariant"); 724 // I'd like to write: 725 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 726 // but as we're at a safepoint that's not safe. 727 728 UnlinkAfterAcquire(Self, &node); 729 if (_succ == Self) _succ = NULL; 730 731 assert(_succ != Self, "invariant"); 732 if (_Responsible == Self) { 733 _Responsible = NULL; 734 OrderAccess::fence(); // Dekker pivot-point 735 736 // We may leave threads on cxq|EntryList without a designated 737 // "Responsible" thread. This is benign. When this thread subsequently 738 // exits the monitor it can "see" such preexisting "old" threads -- 739 // threads that arrived on the cxq|EntryList before the fence, above -- 740 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 741 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 742 // non-null and elect a new "Responsible" timer thread. 743 // 744 // This thread executes: 745 // ST Responsible=null; MEMBAR (in enter epilogue - here) 746 // LD cxq|EntryList (in subsequent exit) 747 // 748 // Entering threads in the slow/contended path execute: 749 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 750 // The (ST cxq; MEMBAR) is accomplished with CAS(). 751 // 752 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 753 // exit operation from floating above the ST Responsible=null. 754 } 755 756 // We've acquired ownership with CAS(). 757 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 758 // But since the CAS() this thread may have also stored into _succ, 759 // EntryList, cxq or Responsible. These meta-data updates must be 760 // visible __before this thread subsequently drops the lock. 761 // Consider what could occur if we didn't enforce this constraint -- 762 // STs to monitor meta-data and user-data could reorder with (become 763 // visible after) the ST in exit that drops ownership of the lock. 764 // Some other thread could then acquire the lock, but observe inconsistent 765 // or old monitor meta-data and heap data. That violates the JMM. 766 // To that end, the 1-0 exit() operation must have at least STST|LDST 767 // "release" barrier semantics. Specifically, there must be at least a 768 // STST|LDST barrier in exit() before the ST of null into _owner that drops 769 // the lock. The barrier ensures that changes to monitor meta-data and data 770 // protected by the lock will be visible before we release the lock, and 771 // therefore before some other thread (CPU) has a chance to acquire the lock. 772 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 773 // 774 // Critically, any prior STs to _succ or EntryList must be visible before 775 // the ST of null into _owner in the *subsequent* (following) corresponding 776 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 777 // execute a serializing instruction. 778 779 if (SyncFlags & 8) { 780 OrderAccess::fence(); 781 } 782 return; 783 } 784 785 // ReenterI() is a specialized inline form of the latter half of the 786 // contended slow-path from EnterI(). We use ReenterI() only for 787 // monitor reentry in wait(). 788 // 789 // In the future we should reconcile EnterI() and ReenterI(), adding 790 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the 791 // loop accordingly. 792 793 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) { 794 assert(Self != NULL, "invariant"); 795 assert(SelfNode != NULL, "invariant"); 796 assert(SelfNode->_thread == Self, "invariant"); 797 assert(_waiters > 0, "invariant"); 798 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 799 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 800 JavaThread * jt = (JavaThread *) Self; 801 802 int nWakeups = 0; 803 for (;;) { 804 ObjectWaiter::TStates v = SelfNode->TState; 805 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 806 assert(_owner != Self, "invariant"); 807 808 if (TryLock(Self) > 0) break; 809 if (TrySpin(Self) > 0) break; 810 811 if (_owner == DEFLATER_MARKER) { 812 guarantee(0 <= _count, "Impossible: _owner == DEFLATER_MARKER && _count < 0, monitor must not be owned by deflater thread here"); 813 if (Atomic::cmpxchg_ptr(Self, &_owner, DEFLATER_MARKER) == DEFLATER_MARKER) { 814 break; 815 } 816 } 817 818 TEVENT(Wait Reentry - parking); 819 820 // State transition wrappers around park() ... 821 // ReenterI() wisely defers state transitions until 822 // it's clear we must park the thread. 823 { 824 OSThreadContendState osts(Self->osthread()); 825 ThreadBlockInVM tbivm(jt); 826 827 // cleared by handle_special_suspend_equivalent_condition() 828 // or java_suspend_self() 829 jt->set_suspend_equivalent(); 830 if (SyncFlags & 1) { 831 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL); 832 } else { 833 Self->_ParkEvent->park(); 834 } 835 836 // were we externally suspended while we were waiting? 837 for (;;) { 838 if (!ExitSuspendEquivalent(jt)) break; 839 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 840 jt->java_suspend_self(); 841 jt->set_suspend_equivalent(); 842 } 843 } 844 845 // Try again, but just so we distinguish between futile wakeups and 846 // successful wakeups. The following test isn't algorithmically 847 // necessary, but it helps us maintain sensible statistics. 848 if (TryLock(Self) > 0) break; 849 850 // The lock is still contested. 851 // Keep a tally of the # of futile wakeups. 852 // Note that the counter is not protected by a lock or updated by atomics. 853 // That is by design - we trade "lossy" counters which are exposed to 854 // races during updates for a lower probe effect. 855 TEVENT(Wait Reentry - futile wakeup); 856 ++nWakeups; 857 858 // Assuming this is not a spurious wakeup we'll normally 859 // find that _succ == Self. 860 if (_succ == Self) _succ = NULL; 861 862 // Invariant: after clearing _succ a contending thread 863 // *must* retry _owner before parking. 864 OrderAccess::fence(); 865 866 // This PerfData object can be used in parallel with a safepoint. 867 // See the work around in PerfDataManager::destroy(). 868 OM_PERFDATA_OP(FutileWakeups, inc()); 869 } 870 871 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 872 // Normally we'll find Self on the EntryList. 873 // Unlinking from the EntryList is constant-time and atomic-free. 874 // From the perspective of the lock owner (this thread), the 875 // EntryList is stable and cxq is prepend-only. 876 // The head of cxq is volatile but the interior is stable. 877 // In addition, Self.TState is stable. 878 879 assert(_owner == Self, "invariant"); 880 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 881 UnlinkAfterAcquire(Self, SelfNode); 882 if (_succ == Self) _succ = NULL; 883 assert(_succ != Self, "invariant"); 884 SelfNode->TState = ObjectWaiter::TS_RUN; 885 OrderAccess::fence(); // see comments at the end of EnterI() 886 } 887 888 // By convention we unlink a contending thread from EntryList|cxq immediately 889 // after the thread acquires the lock in ::enter(). Equally, we could defer 890 // unlinking the thread until ::exit()-time. 891 892 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) { 893 assert(_owner == Self, "invariant"); 894 assert(SelfNode->_thread == Self, "invariant"); 895 896 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 897 // Normal case: remove Self from the DLL EntryList . 898 // This is a constant-time operation. 899 ObjectWaiter * nxt = SelfNode->_next; 900 ObjectWaiter * prv = SelfNode->_prev; 901 if (nxt != NULL) nxt->_prev = prv; 902 if (prv != NULL) prv->_next = nxt; 903 if (SelfNode == _EntryList) _EntryList = nxt; 904 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 905 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 906 TEVENT(Unlink from EntryList); 907 } else { 908 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 909 // Inopportune interleaving -- Self is still on the cxq. 910 // This usually means the enqueue of self raced an exiting thread. 911 // Normally we'll find Self near the front of the cxq, so 912 // dequeueing is typically fast. If needbe we can accelerate 913 // this with some MCS/CHL-like bidirectional list hints and advisory 914 // back-links so dequeueing from the interior will normally operate 915 // in constant-time. 916 // Dequeue Self from either the head (with CAS) or from the interior 917 // with a linear-time scan and normal non-atomic memory operations. 918 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 919 // and then unlink Self from EntryList. We have to drain eventually, 920 // so it might as well be now. 921 922 ObjectWaiter * v = _cxq; 923 assert(v != NULL, "invariant"); 924 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 925 // The CAS above can fail from interference IFF a "RAT" arrived. 926 // In that case Self must be in the interior and can no longer be 927 // at the head of cxq. 928 if (v == SelfNode) { 929 assert(_cxq != v, "invariant"); 930 v = _cxq; // CAS above failed - start scan at head of list 931 } 932 ObjectWaiter * p; 933 ObjectWaiter * q = NULL; 934 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 935 q = p; 936 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 937 } 938 assert(v != SelfNode, "invariant"); 939 assert(p == SelfNode, "Node not found on cxq"); 940 assert(p != _cxq, "invariant"); 941 assert(q != NULL, "invariant"); 942 assert(q->_next == p, "invariant"); 943 q->_next = p->_next; 944 } 945 TEVENT(Unlink from cxq); 946 } 947 948 #ifdef ASSERT 949 // Diagnostic hygiene ... 950 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 951 SelfNode->_next = (ObjectWaiter *) 0xBAD; 952 SelfNode->TState = ObjectWaiter::TS_RUN; 953 #endif 954 } 955 956 // ----------------------------------------------------------------------------- 957 // Exit support 958 // 959 // exit() 960 // ~~~~~~ 961 // Note that the collector can't reclaim the objectMonitor or deflate 962 // the object out from underneath the thread calling ::exit() as the 963 // thread calling ::exit() never transitions to a stable state. 964 // This inhibits GC, which in turn inhibits asynchronous (and 965 // inopportune) reclamation of "this". 966 // 967 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 968 // There's one exception to the claim above, however. EnterI() can call 969 // exit() to drop a lock if the acquirer has been externally suspended. 970 // In that case exit() is called with _thread_state as _thread_blocked, 971 // but the monitor's _count field is > 0, which inhibits reclamation. 972 // 973 // 1-0 exit 974 // ~~~~~~~~ 975 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 976 // the fast-path operators have been optimized so the common ::exit() 977 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 978 // The code emitted by fast_unlock() elides the usual MEMBAR. This 979 // greatly improves latency -- MEMBAR and CAS having considerable local 980 // latency on modern processors -- but at the cost of "stranding". Absent the 981 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 982 // ::enter() path, resulting in the entering thread being stranding 983 // and a progress-liveness failure. Stranding is extremely rare. 984 // We use timers (timed park operations) & periodic polling to detect 985 // and recover from stranding. Potentially stranded threads periodically 986 // wake up and poll the lock. See the usage of the _Responsible variable. 987 // 988 // The CAS() in enter provides for safety and exclusion, while the CAS or 989 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 990 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 991 // We detect and recover from stranding with timers. 992 // 993 // If a thread transiently strands it'll park until (a) another 994 // thread acquires the lock and then drops the lock, at which time the 995 // exiting thread will notice and unpark the stranded thread, or, (b) 996 // the timer expires. If the lock is high traffic then the stranding latency 997 // will be low due to (a). If the lock is low traffic then the odds of 998 // stranding are lower, although the worst-case stranding latency 999 // is longer. Critically, we don't want to put excessive load in the 1000 // platform's timer subsystem. We want to minimize both the timer injection 1001 // rate (timers created/sec) as well as the number of timers active at 1002 // any one time. (more precisely, we want to minimize timer-seconds, which is 1003 // the integral of the # of active timers at any instant over time). 1004 // Both impinge on OS scalability. Given that, at most one thread parked on 1005 // a monitor will use a timer. 1006 // 1007 // There is also the risk of a futile wake-up. If we drop the lock 1008 // another thread can reacquire the lock immediately, and we can 1009 // then wake a thread unnecessarily. This is benign, and we've 1010 // structured the code so the windows are short and the frequency 1011 // of such futile wakups is low. 1012 1013 void ObjectMonitor::exit(bool not_suspended, TRAPS) { 1014 Thread * const Self = THREAD; 1015 if (THREAD != _owner) { 1016 if (THREAD->is_lock_owned((address) _owner)) { 1017 // Transmute _owner from a BasicLock pointer to a Thread address. 1018 // We don't need to hold _mutex for this transition. 1019 // Non-null to Non-null is safe as long as all readers can 1020 // tolerate either flavor. 1021 assert(_recursions == 0, "invariant"); 1022 _owner = THREAD; 1023 _recursions = 0; 1024 } else { 1025 // Apparent unbalanced locking ... 1026 // Naively we'd like to throw IllegalMonitorStateException. 1027 // As a practical matter we can neither allocate nor throw an 1028 // exception as ::exit() can be called from leaf routines. 1029 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1030 // Upon deeper reflection, however, in a properly run JVM the only 1031 // way we should encounter this situation is in the presence of 1032 // unbalanced JNI locking. TODO: CheckJNICalls. 1033 // See also: CR4414101 1034 TEVENT(Exit - Throw IMSX); 1035 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking"); 1036 return; 1037 } 1038 } 1039 1040 if (_recursions != 0) { 1041 _recursions--; // this is simple recursive enter 1042 TEVENT(Inflated exit - recursive); 1043 return; 1044 } 1045 1046 // Invariant: after setting Responsible=null an thread must execute 1047 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1048 if ((SyncFlags & 4) == 0) { 1049 _Responsible = NULL; 1050 } 1051 1052 #if INCLUDE_TRACE 1053 // get the owner's thread id for the MonitorEnter event 1054 // if it is enabled and the thread isn't suspended 1055 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 1056 _previous_owner_tid = THREAD_TRACE_ID(Self); 1057 } 1058 #endif 1059 1060 for (;;) { 1061 assert(THREAD == _owner, "invariant"); 1062 1063 if (Knob_ExitPolicy == 0) { 1064 // release semantics: prior loads and stores from within the critical section 1065 // must not float (reorder) past the following store that drops the lock. 1066 // On SPARC that requires MEMBAR #loadstore|#storestore. 1067 // But of course in TSO #loadstore|#storestore is not required. 1068 // I'd like to write one of the following: 1069 // A. OrderAccess::release() ; _owner = NULL 1070 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 1071 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 1072 // store into a _dummy variable. That store is not needed, but can result 1073 // in massive wasteful coherency traffic on classic SMP systems. 1074 // Instead, I use release_store(), which is implemented as just a simple 1075 // ST on x64, x86 and SPARC. 1076 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1077 OrderAccess::storeload(); // See if we need to wake a successor 1078 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1079 TEVENT(Inflated exit - simple egress); 1080 return; 1081 } 1082 TEVENT(Inflated exit - complex egress); 1083 // Other threads are blocked trying to acquire the lock. 1084 1085 // Normally the exiting thread is responsible for ensuring succession, 1086 // but if other successors are ready or other entering threads are spinning 1087 // then this thread can simply store NULL into _owner and exit without 1088 // waking a successor. The existence of spinners or ready successors 1089 // guarantees proper succession (liveness). Responsibility passes to the 1090 // ready or running successors. The exiting thread delegates the duty. 1091 // More precisely, if a successor already exists this thread is absolved 1092 // of the responsibility of waking (unparking) one. 1093 // 1094 // The _succ variable is critical to reducing futile wakeup frequency. 1095 // _succ identifies the "heir presumptive" thread that has been made 1096 // ready (unparked) but that has not yet run. We need only one such 1097 // successor thread to guarantee progress. 1098 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1099 // section 3.3 "Futile Wakeup Throttling" for details. 1100 // 1101 // Note that spinners in Enter() also set _succ non-null. 1102 // In the current implementation spinners opportunistically set 1103 // _succ so that exiting threads might avoid waking a successor. 1104 // Another less appealing alternative would be for the exiting thread 1105 // to drop the lock and then spin briefly to see if a spinner managed 1106 // to acquire the lock. If so, the exiting thread could exit 1107 // immediately without waking a successor, otherwise the exiting 1108 // thread would need to dequeue and wake a successor. 1109 // (Note that we'd need to make the post-drop spin short, but no 1110 // shorter than the worst-case round-trip cache-line migration time. 1111 // The dropped lock needs to become visible to the spinner, and then 1112 // the acquisition of the lock by the spinner must become visible to 1113 // the exiting thread). 1114 1115 // It appears that an heir-presumptive (successor) must be made ready. 1116 // Only the current lock owner can manipulate the EntryList or 1117 // drain _cxq, so we need to reacquire the lock. If we fail 1118 // to reacquire the lock the responsibility for ensuring succession 1119 // falls to the new owner. 1120 // 1121 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1122 return; 1123 } 1124 TEVENT(Exit - Reacquired); 1125 } else { 1126 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1127 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1128 OrderAccess::storeload(); 1129 // Ratify the previously observed values. 1130 if (_cxq == NULL || _succ != NULL) { 1131 TEVENT(Inflated exit - simple egress); 1132 return; 1133 } 1134 1135 // inopportune interleaving -- the exiting thread (this thread) 1136 // in the fast-exit path raced an entering thread in the slow-enter 1137 // path. 1138 // We have two choices: 1139 // A. Try to reacquire the lock. 1140 // If the CAS() fails return immediately, otherwise 1141 // we either restart/rerun the exit operation, or simply 1142 // fall-through into the code below which wakes a successor. 1143 // B. If the elements forming the EntryList|cxq are TSM 1144 // we could simply unpark() the lead thread and return 1145 // without having set _succ. 1146 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1147 TEVENT(Inflated exit - reacquired succeeded); 1148 return; 1149 } 1150 TEVENT(Inflated exit - reacquired failed); 1151 } else { 1152 TEVENT(Inflated exit - complex egress); 1153 } 1154 } 1155 1156 guarantee(_owner == THREAD, "invariant"); 1157 1158 ObjectWaiter * w = NULL; 1159 int QMode = Knob_QMode; 1160 1161 if (QMode == 2 && _cxq != NULL) { 1162 // QMode == 2 : cxq has precedence over EntryList. 1163 // Try to directly wake a successor from the cxq. 1164 // If successful, the successor will need to unlink itself from cxq. 1165 w = _cxq; 1166 assert(w != NULL, "invariant"); 1167 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1168 ExitEpilog(Self, w); 1169 return; 1170 } 1171 1172 if (QMode == 3 && _cxq != NULL) { 1173 // Aggressively drain cxq into EntryList at the first opportunity. 1174 // This policy ensure that recently-run threads live at the head of EntryList. 1175 // Drain _cxq into EntryList - bulk transfer. 1176 // First, detach _cxq. 1177 // The following loop is tantamount to: w = swap(&cxq, NULL) 1178 w = _cxq; 1179 for (;;) { 1180 assert(w != NULL, "Invariant"); 1181 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1182 if (u == w) break; 1183 w = u; 1184 } 1185 assert(w != NULL, "invariant"); 1186 1187 ObjectWaiter * q = NULL; 1188 ObjectWaiter * p; 1189 for (p = w; p != NULL; p = p->_next) { 1190 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1191 p->TState = ObjectWaiter::TS_ENTER; 1192 p->_prev = q; 1193 q = p; 1194 } 1195 1196 // Append the RATs to the EntryList 1197 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1198 ObjectWaiter * Tail; 1199 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; 1200 Tail = Tail->_next) 1201 /* empty */; 1202 if (Tail == NULL) { 1203 _EntryList = w; 1204 } else { 1205 Tail->_next = w; 1206 w->_prev = Tail; 1207 } 1208 1209 // Fall thru into code that tries to wake a successor from EntryList 1210 } 1211 1212 if (QMode == 4 && _cxq != NULL) { 1213 // Aggressively drain cxq into EntryList at the first opportunity. 1214 // This policy ensure that recently-run threads live at the head of EntryList. 1215 1216 // Drain _cxq into EntryList - bulk transfer. 1217 // First, detach _cxq. 1218 // The following loop is tantamount to: w = swap(&cxq, NULL) 1219 w = _cxq; 1220 for (;;) { 1221 assert(w != NULL, "Invariant"); 1222 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1223 if (u == w) break; 1224 w = u; 1225 } 1226 assert(w != NULL, "invariant"); 1227 1228 ObjectWaiter * q = NULL; 1229 ObjectWaiter * p; 1230 for (p = w; p != NULL; p = p->_next) { 1231 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1232 p->TState = ObjectWaiter::TS_ENTER; 1233 p->_prev = q; 1234 q = p; 1235 } 1236 1237 // Prepend the RATs to the EntryList 1238 if (_EntryList != NULL) { 1239 q->_next = _EntryList; 1240 _EntryList->_prev = q; 1241 } 1242 _EntryList = w; 1243 1244 // Fall thru into code that tries to wake a successor from EntryList 1245 } 1246 1247 w = _EntryList; 1248 if (w != NULL) { 1249 // I'd like to write: guarantee (w->_thread != Self). 1250 // But in practice an exiting thread may find itself on the EntryList. 1251 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1252 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1253 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1254 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1255 // release the lock "O". T2 resumes immediately after the ST of null into 1256 // _owner, above. T2 notices that the EntryList is populated, so it 1257 // reacquires the lock and then finds itself on the EntryList. 1258 // Given all that, we have to tolerate the circumstance where "w" is 1259 // associated with Self. 1260 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1261 ExitEpilog(Self, w); 1262 return; 1263 } 1264 1265 // If we find that both _cxq and EntryList are null then just 1266 // re-run the exit protocol from the top. 1267 w = _cxq; 1268 if (w == NULL) continue; 1269 1270 // Drain _cxq into EntryList - bulk transfer. 1271 // First, detach _cxq. 1272 // The following loop is tantamount to: w = swap(&cxq, NULL) 1273 for (;;) { 1274 assert(w != NULL, "Invariant"); 1275 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1276 if (u == w) break; 1277 w = u; 1278 } 1279 TEVENT(Inflated exit - drain cxq into EntryList); 1280 1281 assert(w != NULL, "invariant"); 1282 assert(_EntryList == NULL, "invariant"); 1283 1284 // Convert the LIFO SLL anchored by _cxq into a DLL. 1285 // The list reorganization step operates in O(LENGTH(w)) time. 1286 // It's critical that this step operate quickly as 1287 // "Self" still holds the outer-lock, restricting parallelism 1288 // and effectively lengthening the critical section. 1289 // Invariant: s chases t chases u. 1290 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1291 // we have faster access to the tail. 1292 1293 if (QMode == 1) { 1294 // QMode == 1 : drain cxq to EntryList, reversing order 1295 // We also reverse the order of the list. 1296 ObjectWaiter * s = NULL; 1297 ObjectWaiter * t = w; 1298 ObjectWaiter * u = NULL; 1299 while (t != NULL) { 1300 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1301 t->TState = ObjectWaiter::TS_ENTER; 1302 u = t->_next; 1303 t->_prev = u; 1304 t->_next = s; 1305 s = t; 1306 t = u; 1307 } 1308 _EntryList = s; 1309 assert(s != NULL, "invariant"); 1310 } else { 1311 // QMode == 0 or QMode == 2 1312 _EntryList = w; 1313 ObjectWaiter * q = NULL; 1314 ObjectWaiter * p; 1315 for (p = w; p != NULL; p = p->_next) { 1316 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1317 p->TState = ObjectWaiter::TS_ENTER; 1318 p->_prev = q; 1319 q = p; 1320 } 1321 } 1322 1323 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1324 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1325 1326 // See if we can abdicate to a spinner instead of waking a thread. 1327 // A primary goal of the implementation is to reduce the 1328 // context-switch rate. 1329 if (_succ != NULL) continue; 1330 1331 w = _EntryList; 1332 if (w != NULL) { 1333 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1334 ExitEpilog(Self, w); 1335 return; 1336 } 1337 } 1338 } 1339 1340 // ExitSuspendEquivalent: 1341 // A faster alternate to handle_special_suspend_equivalent_condition() 1342 // 1343 // handle_special_suspend_equivalent_condition() unconditionally 1344 // acquires the SR_lock. On some platforms uncontended MutexLocker() 1345 // operations have high latency. Note that in ::enter() we call HSSEC 1346 // while holding the monitor, so we effectively lengthen the critical sections. 1347 // 1348 // There are a number of possible solutions: 1349 // 1350 // A. To ameliorate the problem we might also defer state transitions 1351 // to as late as possible -- just prior to parking. 1352 // Given that, we'd call HSSEC after having returned from park(), 1353 // but before attempting to acquire the monitor. This is only a 1354 // partial solution. It avoids calling HSSEC while holding the 1355 // monitor (good), but it still increases successor reacquisition latency -- 1356 // the interval between unparking a successor and the time the successor 1357 // resumes and retries the lock. See ReenterI(), which defers state transitions. 1358 // If we use this technique we can also avoid EnterI()-exit() loop 1359 // in ::enter() where we iteratively drop the lock and then attempt 1360 // to reacquire it after suspending. 1361 // 1362 // B. In the future we might fold all the suspend bits into a 1363 // composite per-thread suspend flag and then update it with CAS(). 1364 // Alternately, a Dekker-like mechanism with multiple variables 1365 // would suffice: 1366 // ST Self->_suspend_equivalent = false 1367 // MEMBAR 1368 // LD Self_>_suspend_flags 1369 // 1370 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem 1371 // with a more efficient implementation, the need to use "FastHSSEC" has 1372 // decreased. - Dave 1373 1374 1375 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) { 1376 const int Mode = Knob_FastHSSEC; 1377 if (Mode && !jSelf->is_external_suspend()) { 1378 assert(jSelf->is_suspend_equivalent(), "invariant"); 1379 jSelf->clear_suspend_equivalent(); 1380 if (2 == Mode) OrderAccess::storeload(); 1381 if (!jSelf->is_external_suspend()) return false; 1382 // We raced a suspension -- fall thru into the slow path 1383 TEVENT(ExitSuspendEquivalent - raced); 1384 jSelf->set_suspend_equivalent(); 1385 } 1386 return jSelf->handle_special_suspend_equivalent_condition(); 1387 } 1388 1389 1390 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) { 1391 assert(_owner == Self, "invariant"); 1392 1393 // Exit protocol: 1394 // 1. ST _succ = wakee 1395 // 2. membar #loadstore|#storestore; 1396 // 2. ST _owner = NULL 1397 // 3. unpark(wakee) 1398 1399 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1400 ParkEvent * Trigger = Wakee->_event; 1401 1402 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1403 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1404 // out-of-scope (non-extant). 1405 Wakee = NULL; 1406 1407 // Drop the lock 1408 OrderAccess::release_store_ptr(&_owner, NULL); 1409 OrderAccess::fence(); // ST _owner vs LD in unpark() 1410 1411 if (SafepointSynchronize::do_call_back()) { 1412 TEVENT(unpark before SAFEPOINT); 1413 } 1414 1415 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1416 Trigger->unpark(); 1417 1418 // Maintain stats and report events to JVMTI 1419 OM_PERFDATA_OP(Parks, inc()); 1420 } 1421 1422 1423 // ----------------------------------------------------------------------------- 1424 // Class Loader deadlock handling. 1425 // 1426 // complete_exit exits a lock returning recursion count 1427 // complete_exit/reenter operate as a wait without waiting 1428 // complete_exit requires an inflated monitor 1429 // The _owner field is not always the Thread addr even with an 1430 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1431 // thread due to contention. 1432 intptr_t ObjectMonitor::complete_exit(TRAPS) { 1433 Thread * const Self = THREAD; 1434 assert(Self->is_Java_thread(), "Must be Java thread!"); 1435 JavaThread *jt = (JavaThread *)THREAD; 1436 1437 DeferredInitialize(); 1438 1439 if (THREAD != _owner) { 1440 if (THREAD->is_lock_owned ((address)_owner)) { 1441 assert(_recursions == 0, "internal state error"); 1442 _owner = THREAD; // Convert from basiclock addr to Thread addr 1443 _recursions = 0; 1444 } 1445 } 1446 1447 guarantee(Self == _owner, "complete_exit not owner"); 1448 intptr_t save = _recursions; // record the old recursion count 1449 _recursions = 0; // set the recursion level to be 0 1450 exit(true, Self); // exit the monitor 1451 guarantee(_owner != Self, "invariant"); 1452 return save; 1453 } 1454 1455 // reenter() enters a lock and sets recursion count 1456 // complete_exit/reenter operate as a wait without waiting 1457 bool ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1458 Thread * const Self = THREAD; 1459 assert(Self->is_Java_thread(), "Must be Java thread!"); 1460 JavaThread *jt = (JavaThread *)THREAD; 1461 1462 guarantee(_owner != Self, "reenter already owner"); 1463 if (!enter(THREAD)) { return false; } // enter the monitor 1464 guarantee(_recursions == 0, "reenter recursion"); 1465 _recursions = recursions; 1466 return true; 1467 } 1468 1469 1470 // ----------------------------------------------------------------------------- 1471 // A macro is used below because there may already be a pending 1472 // exception which should not abort the execution of the routines 1473 // which use this (which is why we don't put this into check_slow and 1474 // call it with a CHECK argument). 1475 1476 #define CHECK_OWNER() \ 1477 do { \ 1478 if (THREAD != _owner) { \ 1479 if (THREAD->is_lock_owned((address) _owner)) { \ 1480 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1481 _recursions = 0; \ 1482 } else { \ 1483 TEVENT(Throw IMSX); \ 1484 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1485 } \ 1486 } \ 1487 } while (false) 1488 1489 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1490 // TODO-FIXME: remove check_slow() -- it's likely dead. 1491 1492 void ObjectMonitor::check_slow(TRAPS) { 1493 TEVENT(check_slow - throw IMSX); 1494 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1495 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1496 } 1497 1498 static int Adjust(volatile int * adr, int dx) { 1499 int v; 1500 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */; 1501 return v; 1502 } 1503 1504 // helper method for posting a monitor wait event 1505 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1506 jlong notifier_tid, 1507 jlong timeout, 1508 bool timedout) { 1509 assert(event != NULL, "invariant"); 1510 event->set_monitorClass(((oop)this->object())->klass()); 1511 event->set_timeout(timeout); 1512 event->set_address((TYPE_ADDRESS)this->object_addr()); 1513 event->set_notifier(notifier_tid); 1514 event->set_timedOut(timedout); 1515 event->commit(); 1516 } 1517 1518 // ----------------------------------------------------------------------------- 1519 // Wait/Notify/NotifyAll 1520 // 1521 // Note: a subset of changes to ObjectMonitor::wait() 1522 // will need to be replicated in complete_exit 1523 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1524 Thread * const Self = THREAD; 1525 assert(Self->is_Java_thread(), "Must be Java thread!"); 1526 JavaThread *jt = (JavaThread *)THREAD; 1527 1528 DeferredInitialize(); 1529 1530 // Throw IMSX or IEX. 1531 CHECK_OWNER(); 1532 1533 EventJavaMonitorWait event; 1534 1535 // check for a pending interrupt 1536 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1537 // post monitor waited event. Note that this is past-tense, we are done waiting. 1538 if (JvmtiExport::should_post_monitor_waited()) { 1539 // Note: 'false' parameter is passed here because the 1540 // wait was not timed out due to thread interrupt. 1541 JvmtiExport::post_monitor_waited(jt, this, false); 1542 1543 // In this short circuit of the monitor wait protocol, the 1544 // current thread never drops ownership of the monitor and 1545 // never gets added to the wait queue so the current thread 1546 // cannot be made the successor. This means that the 1547 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1548 // consume an unpark() meant for the ParkEvent associated with 1549 // this ObjectMonitor. 1550 } 1551 if (event.should_commit()) { 1552 post_monitor_wait_event(&event, 0, millis, false); 1553 } 1554 TEVENT(Wait - Throw IEX); 1555 THROW(vmSymbols::java_lang_InterruptedException()); 1556 return; 1557 } 1558 1559 TEVENT(Wait); 1560 1561 assert(Self->_Stalled == 0, "invariant"); 1562 Self->_Stalled = intptr_t(this); 1563 jt->set_current_waiting_monitor(this); 1564 1565 // create a node to be put into the queue 1566 // Critically, after we reset() the event but prior to park(), we must check 1567 // for a pending interrupt. 1568 ObjectWaiter node(Self); 1569 node.TState = ObjectWaiter::TS_WAIT; 1570 Self->_ParkEvent->reset(); 1571 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1572 1573 // Enter the waiting queue, which is a circular doubly linked list in this case 1574 // but it could be a priority queue or any data structure. 1575 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1576 // by the the owner of the monitor *except* in the case where park() 1577 // returns because of a timeout of interrupt. Contention is exceptionally rare 1578 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1579 1580 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1581 AddWaiter(&node); 1582 Thread::SpinRelease(&_WaitSetLock); 1583 1584 if ((SyncFlags & 4) == 0) { 1585 _Responsible = NULL; 1586 } 1587 intptr_t save = _recursions; // record the old recursion count 1588 _waiters++; // increment the number of waiters 1589 _recursions = 0; // set the recursion level to be 1 1590 exit(true, Self); // exit the monitor 1591 guarantee(_owner != Self, "invariant"); 1592 1593 // The thread is on the WaitSet list - now park() it. 1594 // On MP systems it's conceivable that a brief spin before we park 1595 // could be profitable. 1596 // 1597 // TODO-FIXME: change the following logic to a loop of the form 1598 // while (!timeout && !interrupted && _notified == 0) park() 1599 1600 int ret = OS_OK; 1601 int WasNotified = 0; 1602 { // State transition wrappers 1603 OSThread* osthread = Self->osthread(); 1604 OSThreadWaitState osts(osthread, true); 1605 { 1606 ThreadBlockInVM tbivm(jt); 1607 // Thread is in thread_blocked state and oop access is unsafe. 1608 jt->set_suspend_equivalent(); 1609 1610 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1611 // Intentionally empty 1612 } else if (node._notified == 0) { 1613 if (millis <= 0) { 1614 Self->_ParkEvent->park(); 1615 } else { 1616 ret = Self->_ParkEvent->park(millis); 1617 } 1618 } 1619 1620 // were we externally suspended while we were waiting? 1621 if (ExitSuspendEquivalent (jt)) { 1622 // TODO-FIXME: add -- if succ == Self then succ = null. 1623 jt->java_suspend_self(); 1624 } 1625 1626 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1627 1628 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1629 // from the WaitSet to the EntryList. 1630 // See if we need to remove Node from the WaitSet. 1631 // We use double-checked locking to avoid grabbing _WaitSetLock 1632 // if the thread is not on the wait queue. 1633 // 1634 // Note that we don't need a fence before the fetch of TState. 1635 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1636 // written by the is thread. (perhaps the fetch might even be satisfied 1637 // by a look-aside into the processor's own store buffer, although given 1638 // the length of the code path between the prior ST and this load that's 1639 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1640 // then we'll acquire the lock and then re-fetch a fresh TState value. 1641 // That is, we fail toward safety. 1642 1643 if (node.TState == ObjectWaiter::TS_WAIT) { 1644 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1645 if (node.TState == ObjectWaiter::TS_WAIT) { 1646 DequeueSpecificWaiter(&node); // unlink from WaitSet 1647 assert(node._notified == 0, "invariant"); 1648 node.TState = ObjectWaiter::TS_RUN; 1649 } 1650 Thread::SpinRelease(&_WaitSetLock); 1651 } 1652 1653 // The thread is now either on off-list (TS_RUN), 1654 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1655 // The Node's TState variable is stable from the perspective of this thread. 1656 // No other threads will asynchronously modify TState. 1657 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1658 OrderAccess::loadload(); 1659 if (_succ == Self) _succ = NULL; 1660 WasNotified = node._notified; 1661 1662 // Reentry phase -- reacquire the monitor. 1663 // re-enter contended monitor after object.wait(). 1664 // retain OBJECT_WAIT state until re-enter successfully completes 1665 // Thread state is thread_in_vm and oop access is again safe, 1666 // although the raw address of the object may have changed. 1667 // (Don't cache naked oops over safepoints, of course). 1668 1669 // post monitor waited event. Note that this is past-tense, we are done waiting. 1670 if (JvmtiExport::should_post_monitor_waited()) { 1671 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1672 1673 if (node._notified != 0 && _succ == Self) { 1674 // In this part of the monitor wait-notify-reenter protocol it 1675 // is possible (and normal) for another thread to do a fastpath 1676 // monitor enter-exit while this thread is still trying to get 1677 // to the reenter portion of the protocol. 1678 // 1679 // The ObjectMonitor was notified and the current thread is 1680 // the successor which also means that an unpark() has already 1681 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1682 // consume the unpark() that was done when the successor was 1683 // set because the same ParkEvent is shared between Java 1684 // monitors and JVM/TI RawMonitors (for now). 1685 // 1686 // We redo the unpark() to ensure forward progress, i.e., we 1687 // don't want all pending threads hanging (parked) with none 1688 // entering the unlocked monitor. 1689 node._event->unpark(); 1690 } 1691 } 1692 1693 if (event.should_commit()) { 1694 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1695 } 1696 1697 OrderAccess::fence(); 1698 1699 assert(Self->_Stalled != 0, "invariant"); 1700 Self->_Stalled = 0; 1701 1702 assert(_owner != Self, "invariant"); 1703 ObjectWaiter::TStates v = node.TState; 1704 if (v == ObjectWaiter::TS_RUN) { 1705 DEBUG_ONLY(const bool success = ) enter(Self); 1706 assert(success, "enter signalled that we should retry, but monitor should not be deflated as waiters > 0"); 1707 } else { 1708 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1709 ReenterI(Self, &node); 1710 node.wait_reenter_end(this); 1711 } 1712 1713 // Self has reacquired the lock. 1714 // Lifecycle - the node representing Self must not appear on any queues. 1715 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1716 // want residual elements associated with this thread left on any lists. 1717 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1718 assert(_owner == Self, "invariant"); 1719 assert(_succ != Self, "invariant"); 1720 } // OSThreadWaitState() 1721 1722 jt->set_current_waiting_monitor(NULL); 1723 1724 guarantee(_recursions == 0, "invariant"); 1725 _recursions = save; // restore the old recursion count 1726 _waiters--; // decrement the number of waiters 1727 1728 // Verify a few postconditions 1729 assert(_owner == Self, "invariant"); 1730 assert(_succ != Self, "invariant"); 1731 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1732 1733 if (SyncFlags & 32) { 1734 OrderAccess::fence(); 1735 } 1736 1737 // check if the notification happened 1738 if (!WasNotified) { 1739 // no, it could be timeout or Thread.interrupt() or both 1740 // check for interrupt event, otherwise it is timeout 1741 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1742 TEVENT(Wait - throw IEX from epilog); 1743 THROW(vmSymbols::java_lang_InterruptedException()); 1744 } 1745 } 1746 1747 // NOTE: Spurious wake up will be consider as timeout. 1748 // Monitor notify has precedence over thread interrupt. 1749 } 1750 1751 1752 // Consider: 1753 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1754 // then instead of transferring a thread from the WaitSet to the EntryList 1755 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1756 1757 void ObjectMonitor::INotify(Thread * Self) { 1758 const int policy = Knob_MoveNotifyee; 1759 1760 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1761 ObjectWaiter * iterator = DequeueWaiter(); 1762 if (iterator != NULL) { 1763 TEVENT(Notify1 - Transfer); 1764 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1765 guarantee(iterator->_notified == 0, "invariant"); 1766 // Disposition - what might we do with iterator ? 1767 // a. add it directly to the EntryList - either tail (policy == 1) 1768 // or head (policy == 0). 1769 // b. push it onto the front of the _cxq (policy == 2). 1770 // For now we use (b). 1771 if (policy != 4) { 1772 iterator->TState = ObjectWaiter::TS_ENTER; 1773 } 1774 iterator->_notified = 1; 1775 iterator->_notifier_tid = THREAD_TRACE_ID(Self); 1776 1777 ObjectWaiter * list = _EntryList; 1778 if (list != NULL) { 1779 assert(list->_prev == NULL, "invariant"); 1780 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1781 assert(list != iterator, "invariant"); 1782 } 1783 1784 if (policy == 0) { // prepend to EntryList 1785 if (list == NULL) { 1786 iterator->_next = iterator->_prev = NULL; 1787 _EntryList = iterator; 1788 } else { 1789 list->_prev = iterator; 1790 iterator->_next = list; 1791 iterator->_prev = NULL; 1792 _EntryList = iterator; 1793 } 1794 } else if (policy == 1) { // append to EntryList 1795 if (list == NULL) { 1796 iterator->_next = iterator->_prev = NULL; 1797 _EntryList = iterator; 1798 } else { 1799 // CONSIDER: finding the tail currently requires a linear-time walk of 1800 // the EntryList. We can make tail access constant-time by converting to 1801 // a CDLL instead of using our current DLL. 1802 ObjectWaiter * tail; 1803 for (tail = list; tail->_next != NULL; tail = tail->_next) /* empty */; 1804 assert(tail != NULL && tail->_next == NULL, "invariant"); 1805 tail->_next = iterator; 1806 iterator->_prev = tail; 1807 iterator->_next = NULL; 1808 } 1809 } else if (policy == 2) { // prepend to cxq 1810 if (list == NULL) { 1811 iterator->_next = iterator->_prev = NULL; 1812 _EntryList = iterator; 1813 } else { 1814 iterator->TState = ObjectWaiter::TS_CXQ; 1815 for (;;) { 1816 ObjectWaiter * front = _cxq; 1817 iterator->_next = front; 1818 if (Atomic::cmpxchg_ptr(iterator, &_cxq, front) == front) { 1819 break; 1820 } 1821 } 1822 } 1823 } else if (policy == 3) { // append to cxq 1824 iterator->TState = ObjectWaiter::TS_CXQ; 1825 for (;;) { 1826 ObjectWaiter * tail = _cxq; 1827 if (tail == NULL) { 1828 iterator->_next = NULL; 1829 if (Atomic::cmpxchg_ptr(iterator, &_cxq, NULL) == NULL) { 1830 break; 1831 } 1832 } else { 1833 while (tail->_next != NULL) tail = tail->_next; 1834 tail->_next = iterator; 1835 iterator->_prev = tail; 1836 iterator->_next = NULL; 1837 break; 1838 } 1839 } 1840 } else { 1841 ParkEvent * ev = iterator->_event; 1842 iterator->TState = ObjectWaiter::TS_RUN; 1843 OrderAccess::fence(); 1844 ev->unpark(); 1845 } 1846 1847 // _WaitSetLock protects the wait queue, not the EntryList. We could 1848 // move the add-to-EntryList operation, above, outside the critical section 1849 // protected by _WaitSetLock. In practice that's not useful. With the 1850 // exception of wait() timeouts and interrupts the monitor owner 1851 // is the only thread that grabs _WaitSetLock. There's almost no contention 1852 // on _WaitSetLock so it's not profitable to reduce the length of the 1853 // critical section. 1854 1855 if (policy < 4) { 1856 iterator->wait_reenter_begin(this); 1857 } 1858 } 1859 Thread::SpinRelease(&_WaitSetLock); 1860 } 1861 1862 // Consider: a not-uncommon synchronization bug is to use notify() when 1863 // notifyAll() is more appropriate, potentially resulting in stranded 1864 // threads; this is one example of a lost wakeup. A useful diagnostic 1865 // option is to force all notify() operations to behave as notifyAll(). 1866 // 1867 // Note: We can also detect many such problems with a "minimum wait". 1868 // When the "minimum wait" is set to a small non-zero timeout value 1869 // and the program does not hang whereas it did absent "minimum wait", 1870 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses 1871 // a "minimum wait" for all park() operations; see the recheckInterval 1872 // variable and MAX_RECHECK_INTERVAL. 1873 1874 void ObjectMonitor::notify(TRAPS) { 1875 CHECK_OWNER(); 1876 if (_WaitSet == NULL) { 1877 TEVENT(Empty-Notify); 1878 return; 1879 } 1880 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1881 INotify(THREAD); 1882 OM_PERFDATA_OP(Notifications, inc(1)); 1883 } 1884 1885 1886 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1887 // from the waitset to the EntryList. This could be done more efficiently with a 1888 // single bulk transfer but in practice it's not time-critical. Beware too, 1889 // that in prepend-mode we invert the order of the waiters. Let's say that the 1890 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1891 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1892 1893 void ObjectMonitor::notifyAll(TRAPS) { 1894 CHECK_OWNER(); 1895 if (_WaitSet == NULL) { 1896 TEVENT(Empty-NotifyAll); 1897 return; 1898 } 1899 1900 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1901 int tally = 0; 1902 while (_WaitSet != NULL) { 1903 tally++; 1904 INotify(THREAD); 1905 } 1906 1907 OM_PERFDATA_OP(Notifications, inc(tally)); 1908 } 1909 1910 // ----------------------------------------------------------------------------- 1911 // Adaptive Spinning Support 1912 // 1913 // Adaptive spin-then-block - rational spinning 1914 // 1915 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1916 // algorithm. On high order SMP systems it would be better to start with 1917 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1918 // a contending thread could enqueue itself on the cxq and then spin locally 1919 // on a thread-specific variable such as its ParkEvent._Event flag. 1920 // That's left as an exercise for the reader. Note that global spinning is 1921 // not problematic on Niagara, as the L2 cache serves the interconnect and 1922 // has both low latency and massive bandwidth. 1923 // 1924 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1925 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1926 // (duration) or we can fix the count at approximately the duration of 1927 // a context switch and vary the frequency. Of course we could also 1928 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1929 // For a description of 'Adaptive spin-then-block mutual exclusion in 1930 // multi-threaded processing,' see U.S. Pat. No. 8046758. 1931 // 1932 // This implementation varies the duration "D", where D varies with 1933 // the success rate of recent spin attempts. (D is capped at approximately 1934 // length of a round-trip context switch). The success rate for recent 1935 // spin attempts is a good predictor of the success rate of future spin 1936 // attempts. The mechanism adapts automatically to varying critical 1937 // section length (lock modality), system load and degree of parallelism. 1938 // D is maintained per-monitor in _SpinDuration and is initialized 1939 // optimistically. Spin frequency is fixed at 100%. 1940 // 1941 // Note that _SpinDuration is volatile, but we update it without locks 1942 // or atomics. The code is designed so that _SpinDuration stays within 1943 // a reasonable range even in the presence of races. The arithmetic 1944 // operations on _SpinDuration are closed over the domain of legal values, 1945 // so at worst a race will install and older but still legal value. 1946 // At the very worst this introduces some apparent non-determinism. 1947 // We might spin when we shouldn't or vice-versa, but since the spin 1948 // count are relatively short, even in the worst case, the effect is harmless. 1949 // 1950 // Care must be taken that a low "D" value does not become an 1951 // an absorbing state. Transient spinning failures -- when spinning 1952 // is overall profitable -- should not cause the system to converge 1953 // on low "D" values. We want spinning to be stable and predictable 1954 // and fairly responsive to change and at the same time we don't want 1955 // it to oscillate, become metastable, be "too" non-deterministic, 1956 // or converge on or enter undesirable stable absorbing states. 1957 // 1958 // We implement a feedback-based control system -- using past behavior 1959 // to predict future behavior. We face two issues: (a) if the 1960 // input signal is random then the spin predictor won't provide optimal 1961 // results, and (b) if the signal frequency is too high then the control 1962 // system, which has some natural response lag, will "chase" the signal. 1963 // (b) can arise from multimodal lock hold times. Transient preemption 1964 // can also result in apparent bimodal lock hold times. 1965 // Although sub-optimal, neither condition is particularly harmful, as 1966 // in the worst-case we'll spin when we shouldn't or vice-versa. 1967 // The maximum spin duration is rather short so the failure modes aren't bad. 1968 // To be conservative, I've tuned the gain in system to bias toward 1969 // _not spinning. Relatedly, the system can sometimes enter a mode where it 1970 // "rings" or oscillates between spinning and not spinning. This happens 1971 // when spinning is just on the cusp of profitability, however, so the 1972 // situation is not dire. The state is benign -- there's no need to add 1973 // hysteresis control to damp the transition rate between spinning and 1974 // not spinning. 1975 1976 // Spinning: Fixed frequency (100%), vary duration 1977 int ObjectMonitor::TrySpin(Thread * Self) { 1978 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1979 int ctr = Knob_FixedSpin; 1980 if (ctr != 0) { 1981 while (--ctr >= 0) { 1982 if (TryLock(Self) > 0) return 1; 1983 SpinPause(); 1984 } 1985 return 0; 1986 } 1987 1988 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1989 if (TryLock(Self) > 0) { 1990 // Increase _SpinDuration ... 1991 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1992 // Raising _SpurDuration to the poverty line is key. 1993 int x = _SpinDuration; 1994 if (x < Knob_SpinLimit) { 1995 if (x < Knob_Poverty) x = Knob_Poverty; 1996 _SpinDuration = x + Knob_BonusB; 1997 } 1998 return 1; 1999 } 2000 SpinPause(); 2001 } 2002 2003 // Admission control - verify preconditions for spinning 2004 // 2005 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2006 // becoming an absorbing state. Put another way, we spin briefly to 2007 // sample, just in case the system load, parallelism, contention, or lock 2008 // modality changed. 2009 // 2010 // Consider the following alternative: 2011 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2012 // spin attempt. "Periodically" might mean after a tally of 2013 // the # of failed spin attempts (or iterations) reaches some threshold. 2014 // This takes us into the realm of 1-out-of-N spinning, where we 2015 // hold the duration constant but vary the frequency. 2016 2017 ctr = _SpinDuration; 2018 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 2019 if (ctr <= 0) return 0; 2020 2021 if (Knob_SuccRestrict && _succ != NULL) return 0; 2022 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2023 TEVENT(Spin abort - notrunnable [TOP]); 2024 return 0; 2025 } 2026 2027 int MaxSpin = Knob_MaxSpinners; 2028 if (MaxSpin >= 0) { 2029 if (_Spinner > MaxSpin) { 2030 TEVENT(Spin abort -- too many spinners); 2031 return 0; 2032 } 2033 // Slightly racy, but benign ... 2034 Adjust(&_Spinner, 1); 2035 } 2036 2037 // We're good to spin ... spin ingress. 2038 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2039 // when preparing to LD...CAS _owner, etc and the CAS is likely 2040 // to succeed. 2041 int hits = 0; 2042 int msk = 0; 2043 int caspty = Knob_CASPenalty; 2044 int oxpty = Knob_OXPenalty; 2045 int sss = Knob_SpinSetSucc; 2046 if (sss && _succ == NULL) _succ = Self; 2047 Thread * prv = NULL; 2048 2049 // There are three ways to exit the following loop: 2050 // 1. A successful spin where this thread has acquired the lock. 2051 // 2. Spin failure with prejudice 2052 // 3. Spin failure without prejudice 2053 2054 while (--ctr >= 0) { 2055 2056 // Periodic polling -- Check for pending GC 2057 // Threads may spin while they're unsafe. 2058 // We don't want spinning threads to delay the JVM from reaching 2059 // a stop-the-world safepoint or to steal cycles from GC. 2060 // If we detect a pending safepoint we abort in order that 2061 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2062 // this thread, if safe, doesn't steal cycles from GC. 2063 // This is in keeping with the "no loitering in runtime" rule. 2064 // We periodically check to see if there's a safepoint pending. 2065 if ((ctr & 0xFF) == 0) { 2066 if (SafepointSynchronize::do_call_back()) { 2067 TEVENT(Spin: safepoint); 2068 goto Abort; // abrupt spin egress 2069 } 2070 if (Knob_UsePause & 1) SpinPause(); 2071 } 2072 2073 if (Knob_UsePause & 2) SpinPause(); 2074 2075 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2076 // This is useful on classic SMP systems, but is of less utility on 2077 // N1-style CMT platforms. 2078 // 2079 // Trade-off: lock acquisition latency vs coherency bandwidth. 2080 // Lock hold times are typically short. A histogram 2081 // of successful spin attempts shows that we usually acquire 2082 // the lock early in the spin. That suggests we want to 2083 // sample _owner frequently in the early phase of the spin, 2084 // but then back-off and sample less frequently as the spin 2085 // progresses. The back-off makes a good citizen on SMP big 2086 // SMP systems. Oversampling _owner can consume excessive 2087 // coherency bandwidth. Relatedly, if we _oversample _owner we 2088 // can inadvertently interfere with the the ST m->owner=null. 2089 // executed by the lock owner. 2090 if (ctr & msk) continue; 2091 ++hits; 2092 if ((hits & 0xF) == 0) { 2093 // The 0xF, above, corresponds to the exponent. 2094 // Consider: (msk+1)|msk 2095 msk = ((msk << 2)|3) & BackOffMask; 2096 } 2097 2098 // Probe _owner with TATAS 2099 // If this thread observes the monitor transition or flicker 2100 // from locked to unlocked to locked, then the odds that this 2101 // thread will acquire the lock in this spin attempt go down 2102 // considerably. The same argument applies if the CAS fails 2103 // or if we observe _owner change from one non-null value to 2104 // another non-null value. In such cases we might abort 2105 // the spin without prejudice or apply a "penalty" to the 2106 // spin count-down variable "ctr", reducing it by 100, say. 2107 2108 Thread * ox = (Thread *) _owner; 2109 if (ox == NULL) { 2110 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 2111 if (ox == NULL) { 2112 // The CAS succeeded -- this thread acquired ownership 2113 // Take care of some bookkeeping to exit spin state. 2114 if (sss && _succ == Self) { 2115 _succ = NULL; 2116 } 2117 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2118 2119 // Increase _SpinDuration : 2120 // The spin was successful (profitable) so we tend toward 2121 // longer spin attempts in the future. 2122 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2123 // If we acquired the lock early in the spin cycle it 2124 // makes sense to increase _SpinDuration proportionally. 2125 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2126 int x = _SpinDuration; 2127 if (x < Knob_SpinLimit) { 2128 if (x < Knob_Poverty) x = Knob_Poverty; 2129 _SpinDuration = x + Knob_Bonus; 2130 } 2131 return 1; 2132 } 2133 2134 // The CAS failed ... we can take any of the following actions: 2135 // * penalize: ctr -= Knob_CASPenalty 2136 // * exit spin with prejudice -- goto Abort; 2137 // * exit spin without prejudice. 2138 // * Since CAS is high-latency, retry again immediately. 2139 prv = ox; 2140 TEVENT(Spin: cas failed); 2141 if (caspty == -2) break; 2142 if (caspty == -1) goto Abort; 2143 ctr -= caspty; 2144 continue; 2145 } 2146 2147 // Did lock ownership change hands ? 2148 if (ox != prv && prv != NULL) { 2149 TEVENT(spin: Owner changed) 2150 if (oxpty == -2) break; 2151 if (oxpty == -1) goto Abort; 2152 ctr -= oxpty; 2153 } 2154 prv = ox; 2155 2156 // Abort the spin if the owner is not executing. 2157 // The owner must be executing in order to drop the lock. 2158 // Spinning while the owner is OFFPROC is idiocy. 2159 // Consider: ctr -= RunnablePenalty ; 2160 if (Knob_OState && NotRunnable (Self, ox)) { 2161 TEVENT(Spin abort - notrunnable); 2162 goto Abort; 2163 } 2164 if (sss && _succ == NULL) _succ = Self; 2165 } 2166 2167 // Spin failed with prejudice -- reduce _SpinDuration. 2168 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2169 // AIMD is globally stable. 2170 TEVENT(Spin failure); 2171 { 2172 int x = _SpinDuration; 2173 if (x > 0) { 2174 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2175 // This is globally sample and tends to damp the response. 2176 x -= Knob_Penalty; 2177 if (x < 0) x = 0; 2178 _SpinDuration = x; 2179 } 2180 } 2181 2182 Abort: 2183 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2184 if (sss && _succ == Self) { 2185 _succ = NULL; 2186 // Invariant: after setting succ=null a contending thread 2187 // must recheck-retry _owner before parking. This usually happens 2188 // in the normal usage of TrySpin(), but it's safest 2189 // to make TrySpin() as foolproof as possible. 2190 OrderAccess::fence(); 2191 if (TryLock(Self) > 0) return 1; 2192 } 2193 return 0; 2194 } 2195 2196 // NotRunnable() -- informed spinning 2197 // 2198 // Don't bother spinning if the owner is not eligible to drop the lock. 2199 // Peek at the owner's schedctl.sc_state and Thread._thread_values and 2200 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2201 // The thread must be runnable in order to drop the lock in timely fashion. 2202 // If the _owner is not runnable then spinning will not likely be 2203 // successful (profitable). 2204 // 2205 // Beware -- the thread referenced by _owner could have died 2206 // so a simply fetch from _owner->_thread_state might trap. 2207 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2208 // Because of the lifecycle issues the schedctl and _thread_state values 2209 // observed by NotRunnable() might be garbage. NotRunnable must 2210 // tolerate this and consider the observed _thread_state value 2211 // as advisory. 2212 // 2213 // Beware too, that _owner is sometimes a BasicLock address and sometimes 2214 // a thread pointer. 2215 // Alternately, we might tag the type (thread pointer vs basiclock pointer) 2216 // with the LSB of _owner. Another option would be to probablistically probe 2217 // the putative _owner->TypeTag value. 2218 // 2219 // Checking _thread_state isn't perfect. Even if the thread is 2220 // in_java it might be blocked on a page-fault or have been preempted 2221 // and sitting on a ready/dispatch queue. _thread state in conjunction 2222 // with schedctl.sc_state gives us a good picture of what the 2223 // thread is doing, however. 2224 // 2225 // TODO: check schedctl.sc_state. 2226 // We'll need to use SafeFetch32() to read from the schedctl block. 2227 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2228 // 2229 // The return value from NotRunnable() is *advisory* -- the 2230 // result is based on sampling and is not necessarily coherent. 2231 // The caller must tolerate false-negative and false-positive errors. 2232 // Spinning, in general, is probabilistic anyway. 2233 2234 2235 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) { 2236 // Check ox->TypeTag == 2BAD. 2237 if (ox == NULL) return 0; 2238 2239 // Avoid transitive spinning ... 2240 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2241 // Immediately after T1 acquires L it's possible that T2, also 2242 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2243 // This occurs transiently after T1 acquired L but before 2244 // T1 managed to clear T1.Stalled. T2 does not need to abort 2245 // its spin in this circumstance. 2246 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2247 2248 if (BlockedOn == 1) return 1; 2249 if (BlockedOn != 0) { 2250 return BlockedOn != intptr_t(this) && _owner == ox; 2251 } 2252 2253 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2254 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2255 // consider also: jst != _thread_in_Java -- but that's overspecific. 2256 return jst == _thread_blocked || jst == _thread_in_native; 2257 } 2258 2259 2260 // ----------------------------------------------------------------------------- 2261 // WaitSet management ... 2262 2263 ObjectWaiter::ObjectWaiter(Thread* thread) { 2264 _next = NULL; 2265 _prev = NULL; 2266 _notified = 0; 2267 TState = TS_RUN; 2268 _thread = thread; 2269 _event = thread->_ParkEvent; 2270 _active = false; 2271 assert(_event != NULL, "invariant"); 2272 } 2273 2274 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2275 JavaThread *jt = (JavaThread *)this->_thread; 2276 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2277 } 2278 2279 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2280 JavaThread *jt = (JavaThread *)this->_thread; 2281 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2282 } 2283 2284 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2285 assert(node != NULL, "should not add NULL node"); 2286 assert(node->_prev == NULL, "node already in list"); 2287 assert(node->_next == NULL, "node already in list"); 2288 // put node at end of queue (circular doubly linked list) 2289 if (_WaitSet == NULL) { 2290 _WaitSet = node; 2291 node->_prev = node; 2292 node->_next = node; 2293 } else { 2294 ObjectWaiter* head = _WaitSet; 2295 ObjectWaiter* tail = head->_prev; 2296 assert(tail->_next == head, "invariant check"); 2297 tail->_next = node; 2298 head->_prev = node; 2299 node->_next = head; 2300 node->_prev = tail; 2301 } 2302 } 2303 2304 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2305 // dequeue the very first waiter 2306 ObjectWaiter* waiter = _WaitSet; 2307 if (waiter) { 2308 DequeueSpecificWaiter(waiter); 2309 } 2310 return waiter; 2311 } 2312 2313 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2314 assert(node != NULL, "should not dequeue NULL node"); 2315 assert(node->_prev != NULL, "node already removed from list"); 2316 assert(node->_next != NULL, "node already removed from list"); 2317 // when the waiter has woken up because of interrupt, 2318 // timeout or other spurious wake-up, dequeue the 2319 // waiter from waiting list 2320 ObjectWaiter* next = node->_next; 2321 if (next == node) { 2322 assert(node->_prev == node, "invariant check"); 2323 _WaitSet = NULL; 2324 } else { 2325 ObjectWaiter* prev = node->_prev; 2326 assert(prev->_next == node, "invariant check"); 2327 assert(next->_prev == node, "invariant check"); 2328 next->_prev = prev; 2329 prev->_next = next; 2330 if (_WaitSet == node) { 2331 _WaitSet = next; 2332 } 2333 } 2334 node->_next = NULL; 2335 node->_prev = NULL; 2336 } 2337 2338 // ----------------------------------------------------------------------------- 2339 // PerfData support 2340 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2341 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2342 PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2343 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2344 PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2345 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2346 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2347 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2348 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2349 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2350 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2351 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2352 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2353 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2354 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2355 PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2356 PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2357 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2358 2359 // One-shot global initialization for the sync subsystem. 2360 // We could also defer initialization and initialize on-demand 2361 // the first time we call inflate(). Initialization would 2362 // be protected - like so many things - by the MonitorCache_lock. 2363 2364 void ObjectMonitor::Initialize() { 2365 static int InitializationCompleted = 0; 2366 assert(InitializationCompleted == 0, "invariant"); 2367 InitializationCompleted = 1; 2368 if (UsePerfData) { 2369 EXCEPTION_MARK; 2370 #define NEWPERFCOUNTER(n) \ 2371 { \ 2372 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2373 CHECK); \ 2374 } 2375 #define NEWPERFVARIABLE(n) \ 2376 { \ 2377 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2378 CHECK); \ 2379 } 2380 NEWPERFCOUNTER(_sync_Inflations); 2381 NEWPERFCOUNTER(_sync_Deflations); 2382 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2383 NEWPERFCOUNTER(_sync_FutileWakeups); 2384 NEWPERFCOUNTER(_sync_Parks); 2385 NEWPERFCOUNTER(_sync_EmptyNotifications); 2386 NEWPERFCOUNTER(_sync_Notifications); 2387 NEWPERFCOUNTER(_sync_SlowEnter); 2388 NEWPERFCOUNTER(_sync_SlowExit); 2389 NEWPERFCOUNTER(_sync_SlowNotify); 2390 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2391 NEWPERFCOUNTER(_sync_FailedSpins); 2392 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2393 NEWPERFCOUNTER(_sync_PrivateA); 2394 NEWPERFCOUNTER(_sync_PrivateB); 2395 NEWPERFCOUNTER(_sync_MonInCirculation); 2396 NEWPERFCOUNTER(_sync_MonScavenged); 2397 NEWPERFVARIABLE(_sync_MonExtant); 2398 #undef NEWPERFCOUNTER 2399 #undef NEWPERFVARIABLE 2400 } 2401 } 2402 2403 static char * kvGet(char * kvList, const char * Key) { 2404 if (kvList == NULL) return NULL; 2405 size_t n = strlen(Key); 2406 char * Search; 2407 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2408 if (strncmp (Search, Key, n) == 0) { 2409 if (Search[n] == '=') return Search + n + 1; 2410 if (Search[n] == 0) return(char *) "1"; 2411 } 2412 } 2413 return NULL; 2414 } 2415 2416 static int kvGetInt(char * kvList, const char * Key, int Default) { 2417 char * v = kvGet(kvList, Key); 2418 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2419 if (Knob_ReportSettings && v != NULL) { 2420 tty->print_cr("INFO: SyncKnob: %s %d(%d)", Key, rslt, Default) ; 2421 tty->flush(); 2422 } 2423 return rslt; 2424 } 2425 2426 void ObjectMonitor::DeferredInitialize() { 2427 if (InitDone > 0) return; 2428 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2429 while (InitDone != 1) /* empty */; 2430 return; 2431 } 2432 2433 // One-shot global initialization ... 2434 // The initialization is idempotent, so we don't need locks. 2435 // In the future consider doing this via os::init_2(). 2436 // SyncKnobs consist of <Key>=<Value> pairs in the style 2437 // of environment variables. Start by converting ':' to NUL. 2438 2439 if (SyncKnobs == NULL) SyncKnobs = ""; 2440 2441 size_t sz = strlen(SyncKnobs); 2442 char * knobs = (char *) malloc(sz + 2); 2443 if (knobs == NULL) { 2444 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2445 guarantee(0, "invariant"); 2446 } 2447 strcpy(knobs, SyncKnobs); 2448 knobs[sz+1] = 0; 2449 for (char * p = knobs; *p; p++) { 2450 if (*p == ':') *p = 0; 2451 } 2452 2453 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); } 2454 SETKNOB(ReportSettings); 2455 SETKNOB(ExitRelease); 2456 SETKNOB(Verbose); 2457 SETKNOB(VerifyInUse); 2458 SETKNOB(VerifyMatch); 2459 SETKNOB(FixedSpin); 2460 SETKNOB(SpinLimit); 2461 SETKNOB(SpinBase); 2462 SETKNOB(SpinBackOff); 2463 SETKNOB(CASPenalty); 2464 SETKNOB(OXPenalty); 2465 SETKNOB(LogSpins); 2466 SETKNOB(SpinSetSucc); 2467 SETKNOB(SuccEnabled); 2468 SETKNOB(SuccRestrict); 2469 SETKNOB(Penalty); 2470 SETKNOB(Bonus); 2471 SETKNOB(BonusB); 2472 SETKNOB(Poverty); 2473 SETKNOB(SpinAfterFutile); 2474 SETKNOB(UsePause); 2475 SETKNOB(SpinEarly); 2476 SETKNOB(OState); 2477 SETKNOB(MaxSpinners); 2478 SETKNOB(PreSpin); 2479 SETKNOB(ExitPolicy); 2480 SETKNOB(QMode); 2481 SETKNOB(ResetEvent); 2482 SETKNOB(MoveNotifyee); 2483 SETKNOB(FastHSSEC); 2484 #undef SETKNOB 2485 2486 if (Knob_Verbose) { 2487 sanity_checks(); 2488 } 2489 2490 if (os::is_MP()) { 2491 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2492 if (Knob_ReportSettings) { 2493 tty->print_cr("INFO: BackOffMask=0x%X", BackOffMask); 2494 } 2495 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2496 } else { 2497 Knob_SpinLimit = 0; 2498 Knob_SpinBase = 0; 2499 Knob_PreSpin = 0; 2500 Knob_FixedSpin = -1; 2501 } 2502 2503 if (Knob_LogSpins == 0) { 2504 ObjectMonitor::_sync_FailedSpins = NULL; 2505 } 2506 2507 free(knobs); 2508 OrderAccess::fence(); 2509 InitDone = 1; 2510 } 2511 2512 void ObjectMonitor::sanity_checks() { 2513 int error_cnt = 0; 2514 int warning_cnt = 0; 2515 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); 2516 2517 if (verbose) { 2518 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, 2519 sizeof(ObjectMonitor)); 2520 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT, 2521 sizeof(PaddedEnd<ObjectMonitor>)); 2522 } 2523 2524 uint cache_line_size = VM_Version::L1_data_cache_line_size(); 2525 if (verbose) { 2526 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); 2527 } 2528 2529 ObjectMonitor dummy; 2530 u_char *addr_begin = (u_char*)&dummy; 2531 u_char *addr_header = (u_char*)&dummy._header; 2532 u_char *addr_owner = (u_char*)&dummy._owner; 2533 2534 uint offset_header = (uint)(addr_header - addr_begin); 2535 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); 2536 2537 uint offset_owner = (uint)(addr_owner - addr_begin); 2538 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); 2539 2540 if ((uint)(addr_header - addr_begin) != 0) { 2541 tty->print_cr("ERROR: offset(_header) must be zero (0)."); 2542 error_cnt++; 2543 } 2544 2545 if (cache_line_size != 0) { 2546 // We were able to determine the L1 data cache line size so 2547 // do some cache line specific sanity checks 2548 2549 if ((offset_owner - offset_header) < cache_line_size) { 2550 tty->print_cr("WARNING: the _header and _owner fields are closer " 2551 "than a cache line which permits false sharing."); 2552 warning_cnt++; 2553 } 2554 2555 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) { 2556 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a " 2557 "multiple of a cache line which permits false sharing."); 2558 warning_cnt++; 2559 } 2560 } 2561 2562 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, 2563 &warning_cnt); 2564 2565 if (verbose || error_cnt != 0 || warning_cnt != 0) { 2566 tty->print_cr("INFO: error_cnt=%d", error_cnt); 2567 tty->print_cr("INFO: warning_cnt=%d", warning_cnt); 2568 } 2569 2570 guarantee(error_cnt == 0, 2571 "Fatal error(s) found in ObjectMonitor::sanity_checks()"); 2572 } 2573 2574 #ifndef PRODUCT 2575 void ObjectMonitor_test() { 2576 ObjectMonitor::sanity_checks(); 2577 } 2578 #endif