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