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