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