synchronizer.cpp revision 579:0fbdb4381b99
1/* 2 * Copyright 1998-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25# include "incls/_precompiled.incl" 26# include "incls/_synchronizer.cpp.incl" 27 28#if defined(__GNUC__) && !defined(IA64) 29 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 30 #define ATTR __attribute__((noinline)) 31#else 32 #define ATTR 33#endif 34 35// Native markword accessors for synchronization and hashCode(). 36// 37// The "core" versions of monitor enter and exit reside in this file. 38// The interpreter and compilers contain specialized transliterated 39// variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 40// for instance. If you make changes here, make sure to modify the 41// interpreter, and both C1 and C2 fast-path inline locking code emission. 42// 43// TODO: merge the objectMonitor and synchronizer classes. 44// 45// ----------------------------------------------------------------------------- 46 47#ifdef DTRACE_ENABLED 48 49// Only bother with this argument setup if dtrace is available 50// TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 51 52HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait, 53 jlong, uintptr_t, char*, int, long); 54HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited, 55 jlong, uintptr_t, char*, int); 56HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, 57 jlong, uintptr_t, char*, int); 58HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, 59 jlong, uintptr_t, char*, int); 60HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, 61 jlong, uintptr_t, char*, int); 62HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, 63 jlong, uintptr_t, char*, int); 64HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, 65 jlong, uintptr_t, char*, int); 66 67#define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \ 68 char* bytes = NULL; \ 69 int len = 0; \ 70 jlong jtid = SharedRuntime::get_java_tid(thread); \ 71 symbolOop klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \ 72 if (klassname != NULL) { \ 73 bytes = (char*)klassname->bytes(); \ 74 len = klassname->utf8_length(); \ 75 } 76 77#define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ 78 { \ 79 if (DTraceMonitorProbes) { \ 80 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 81 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 82 (monitor), bytes, len, (millis)); \ 83 } \ 84 } 85 86#define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ 90 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95#else // ndef DTRACE_ENABLED 96 97#define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;} 98#define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;} 99 100#endif // ndef DTRACE_ENABLED 101 102// ObjectWaiter serves as a "proxy" or surrogate thread. 103// TODO-FIXME: Eliminate ObjectWaiter and use the thread-specific 104// ParkEvent instead. Beware, however, that the JVMTI code 105// knows about ObjectWaiters, so we'll have to reconcile that code. 106// See next_waiter(), first_waiter(), etc. 107 108class ObjectWaiter : public StackObj { 109 public: 110 enum TStates { TS_UNDEF, TS_READY, TS_RUN, TS_WAIT, TS_ENTER, TS_CXQ } ; 111 enum Sorted { PREPEND, APPEND, SORTED } ; 112 ObjectWaiter * volatile _next; 113 ObjectWaiter * volatile _prev; 114 Thread* _thread; 115 ParkEvent * _event; 116 volatile int _notified ; 117 volatile TStates TState ; 118 Sorted _Sorted ; // List placement disposition 119 bool _active ; // Contention monitoring is enabled 120 public: 121 ObjectWaiter(Thread* thread) { 122 _next = NULL; 123 _prev = NULL; 124 _notified = 0; 125 TState = TS_RUN ; 126 _thread = thread; 127 _event = thread->_ParkEvent ; 128 _active = false; 129 assert (_event != NULL, "invariant") ; 130 } 131 132 void wait_reenter_begin(ObjectMonitor *mon) { 133 JavaThread *jt = (JavaThread *)this->_thread; 134 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 135 } 136 137 void wait_reenter_end(ObjectMonitor *mon) { 138 JavaThread *jt = (JavaThread *)this->_thread; 139 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 140 } 141}; 142 143enum ManifestConstants { 144 ClearResponsibleAtSTW = 0, 145 MaximumRecheckInterval = 1000 146} ; 147 148 149#undef TEVENT 150#define TEVENT(nom) {if (SyncVerbose) FEVENT(nom); } 151 152#define FEVENT(nom) { static volatile int ctr = 0 ; int v = ++ctr ; if ((v & (v-1)) == 0) { ::printf (#nom " : %d \n", v); ::fflush(stdout); }} 153 154#undef TEVENT 155#define TEVENT(nom) {;} 156 157// Performance concern: 158// OrderAccess::storestore() calls release() which STs 0 into the global volatile 159// OrderAccess::Dummy variable. This store is unnecessary for correctness. 160// Many threads STing into a common location causes considerable cache migration 161// or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore() 162// until it's repaired. In some cases OrderAccess::fence() -- which incurs local 163// latency on the executing processor -- is a better choice as it scales on SMP 164// systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a 165// discussion of coherency costs. Note that all our current reference platforms 166// provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC. 167// 168// As a general policy we use "volatile" to control compiler-based reordering 169// and explicit fences (barriers) to control for architectural reordering performed 170// by the CPU(s) or platform. 171 172static int MBFence (int x) { OrderAccess::fence(); return x; } 173 174struct SharedGlobals { 175 // These are highly shared mostly-read variables. 176 // To avoid false-sharing they need to be the sole occupants of a $ line. 177 double padPrefix [8]; 178 volatile int stwRandom ; 179 volatile int stwCycle ; 180 181 // Hot RW variables -- Sequester to avoid false-sharing 182 double padSuffix [16]; 183 volatile int hcSequence ; 184 double padFinal [8] ; 185} ; 186 187static SharedGlobals GVars ; 188 189 190// Tunables ... 191// The knob* variables are effectively final. Once set they should 192// never be modified hence. Consider using __read_mostly with GCC. 193 194static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins 195static int Knob_HandOff = 0 ; 196static int Knob_Verbose = 0 ; 197static int Knob_ReportSettings = 0 ; 198 199static int Knob_SpinLimit = 5000 ; // derived by an external tool - 200static int Knob_SpinBase = 0 ; // Floor AKA SpinMin 201static int Knob_SpinBackOff = 0 ; // spin-loop backoff 202static int Knob_CASPenalty = -1 ; // Penalty for failed CAS 203static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change 204static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field 205static int Knob_SpinEarly = 1 ; 206static int Knob_SuccEnabled = 1 ; // futile wake throttling 207static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one 208static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs 209static int Knob_Bonus = 100 ; // spin success bonus 210static int Knob_BonusB = 100 ; // spin success bonus 211static int Knob_Penalty = 200 ; // spin failure penalty 212static int Knob_Poverty = 1000 ; 213static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() 214static int Knob_FixedSpin = 0 ; 215static int Knob_OState = 3 ; // Spinner checks thread state of _owner 216static int Knob_UsePause = 1 ; 217static int Knob_ExitPolicy = 0 ; 218static int Knob_PreSpin = 10 ; // 20-100 likely better 219static int Knob_ResetEvent = 0 ; 220static int BackOffMask = 0 ; 221 222static int Knob_FastHSSEC = 0 ; 223static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee 224static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline 225static volatile int InitDone = 0 ; 226 227 228// hashCode() generation : 229// 230// Possibilities: 231// * MD5Digest of {obj,stwRandom} 232// * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 233// * A DES- or AES-style SBox[] mechanism 234// * One of the Phi-based schemes, such as: 235// 2654435761 = 2^32 * Phi (golden ratio) 236// HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 237// * A variation of Marsaglia's shift-xor RNG scheme. 238// * (obj ^ stwRandom) is appealing, but can result 239// in undesirable regularity in the hashCode values of adjacent objects 240// (objects allocated back-to-back, in particular). This could potentially 241// result in hashtable collisions and reduced hashtable efficiency. 242// There are simple ways to "diffuse" the middle address bits over the 243// generated hashCode values: 244// 245 246static inline intptr_t get_next_hash(Thread * Self, oop obj) { 247 intptr_t value = 0 ; 248 if (hashCode == 0) { 249 // This form uses an unguarded global Park-Miller RNG, 250 // so it's possible for two threads to race and generate the same RNG. 251 // On MP system we'll have lots of RW access to a global, so the 252 // mechanism induces lots of coherency traffic. 253 value = os::random() ; 254 } else 255 if (hashCode == 1) { 256 // This variation has the property of being stable (idempotent) 257 // between STW operations. This can be useful in some of the 1-0 258 // synchronization schemes. 259 intptr_t addrBits = intptr_t(obj) >> 3 ; 260 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ; 261 } else 262 if (hashCode == 2) { 263 value = 1 ; // for sensitivity testing 264 } else 265 if (hashCode == 3) { 266 value = ++GVars.hcSequence ; 267 } else 268 if (hashCode == 4) { 269 value = intptr_t(obj) ; 270 } else { 271 // Marsaglia's xor-shift scheme with thread-specific state 272 // This is probably the best overall implementation -- we'll 273 // likely make this the default in future releases. 274 unsigned t = Self->_hashStateX ; 275 t ^= (t << 11) ; 276 Self->_hashStateX = Self->_hashStateY ; 277 Self->_hashStateY = Self->_hashStateZ ; 278 Self->_hashStateZ = Self->_hashStateW ; 279 unsigned v = Self->_hashStateW ; 280 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ; 281 Self->_hashStateW = v ; 282 value = v ; 283 } 284 285 value &= markOopDesc::hash_mask; 286 if (value == 0) value = 0xBAD ; 287 assert (value != markOopDesc::no_hash, "invariant") ; 288 TEVENT (hashCode: GENERATE) ; 289 return value; 290} 291 292void BasicLock::print_on(outputStream* st) const { 293 st->print("monitor"); 294} 295 296void BasicLock::move_to(oop obj, BasicLock* dest) { 297 // Check to see if we need to inflate the lock. This is only needed 298 // if an object is locked using "this" lightweight monitor. In that 299 // case, the displaced_header() is unlocked, because the 300 // displaced_header() contains the header for the originally unlocked 301 // object. However the object could have already been inflated. But it 302 // does not matter, the inflation will just a no-op. For other cases, 303 // the displaced header will be either 0x0 or 0x3, which are location 304 // independent, therefore the BasicLock is free to move. 305 // 306 // During OSR we may need to relocate a BasicLock (which contains a 307 // displaced word) from a location in an interpreter frame to a 308 // new location in a compiled frame. "this" refers to the source 309 // basiclock in the interpreter frame. "dest" refers to the destination 310 // basiclock in the new compiled frame. We *always* inflate in move_to(). 311 // The always-Inflate policy works properly, but in 1.5.0 it can sometimes 312 // cause performance problems in code that makes heavy use of a small # of 313 // uncontended locks. (We'd inflate during OSR, and then sync performance 314 // would subsequently plummet because the thread would be forced thru the slow-path). 315 // This problem has been made largely moot on IA32 by inlining the inflated fast-path 316 // operations in Fast_Lock and Fast_Unlock in i486.ad. 317 // 318 // Note that there is a way to safely swing the object's markword from 319 // one stack location to another. This avoids inflation. Obviously, 320 // we need to ensure that both locations refer to the current thread's stack. 321 // There are some subtle concurrency issues, however, and since the benefit is 322 // is small (given the support for inflated fast-path locking in the fast_lock, etc) 323 // we'll leave that optimization for another time. 324 325 if (displaced_header()->is_neutral()) { 326 ObjectSynchronizer::inflate_helper(obj); 327 // WARNING: We can not put check here, because the inflation 328 // will not update the displaced header. Once BasicLock is inflated, 329 // no one should ever look at its content. 330 } else { 331 // Typically the displaced header will be 0 (recursive stack lock) or 332 // unused_mark. Naively we'd like to assert that the displaced mark 333 // value is either 0, neutral, or 3. But with the advent of the 334 // store-before-CAS avoidance in fast_lock/compiler_lock_object 335 // we can find any flavor mark in the displaced mark. 336 } 337// [RGV] The next line appears to do nothing! 338 intptr_t dh = (intptr_t) displaced_header(); 339 dest->set_displaced_header(displaced_header()); 340} 341 342// ----------------------------------------------------------------------------- 343 344// standard constructor, allows locking failures 345ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 346 _dolock = doLock; 347 _thread = thread; 348 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 349 _obj = obj; 350 351 if (_dolock) { 352 TEVENT (ObjectLocker) ; 353 354 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 355 } 356} 357 358ObjectLocker::~ObjectLocker() { 359 if (_dolock) { 360 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 361 } 362} 363 364// ----------------------------------------------------------------------------- 365 366 367PerfCounter * ObjectSynchronizer::_sync_Inflations = NULL ; 368PerfCounter * ObjectSynchronizer::_sync_Deflations = NULL ; 369PerfCounter * ObjectSynchronizer::_sync_ContendedLockAttempts = NULL ; 370PerfCounter * ObjectSynchronizer::_sync_FutileWakeups = NULL ; 371PerfCounter * ObjectSynchronizer::_sync_Parks = NULL ; 372PerfCounter * ObjectSynchronizer::_sync_EmptyNotifications = NULL ; 373PerfCounter * ObjectSynchronizer::_sync_Notifications = NULL ; 374PerfCounter * ObjectSynchronizer::_sync_PrivateA = NULL ; 375PerfCounter * ObjectSynchronizer::_sync_PrivateB = NULL ; 376PerfCounter * ObjectSynchronizer::_sync_SlowExit = NULL ; 377PerfCounter * ObjectSynchronizer::_sync_SlowEnter = NULL ; 378PerfCounter * ObjectSynchronizer::_sync_SlowNotify = NULL ; 379PerfCounter * ObjectSynchronizer::_sync_SlowNotifyAll = NULL ; 380PerfCounter * ObjectSynchronizer::_sync_FailedSpins = NULL ; 381PerfCounter * ObjectSynchronizer::_sync_SuccessfulSpins = NULL ; 382PerfCounter * ObjectSynchronizer::_sync_MonInCirculation = NULL ; 383PerfCounter * ObjectSynchronizer::_sync_MonScavenged = NULL ; 384PerfLongVariable * ObjectSynchronizer::_sync_MonExtant = NULL ; 385 386// One-shot global initialization for the sync subsystem. 387// We could also defer initialization and initialize on-demand 388// the first time we call inflate(). Initialization would 389// be protected - like so many things - by the MonitorCache_lock. 390 391void ObjectSynchronizer::Initialize () { 392 static int InitializationCompleted = 0 ; 393 assert (InitializationCompleted == 0, "invariant") ; 394 InitializationCompleted = 1 ; 395 if (UsePerfData) { 396 EXCEPTION_MARK ; 397 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 398 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 399 NEWPERFCOUNTER(_sync_Inflations) ; 400 NEWPERFCOUNTER(_sync_Deflations) ; 401 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; 402 NEWPERFCOUNTER(_sync_FutileWakeups) ; 403 NEWPERFCOUNTER(_sync_Parks) ; 404 NEWPERFCOUNTER(_sync_EmptyNotifications) ; 405 NEWPERFCOUNTER(_sync_Notifications) ; 406 NEWPERFCOUNTER(_sync_SlowEnter) ; 407 NEWPERFCOUNTER(_sync_SlowExit) ; 408 NEWPERFCOUNTER(_sync_SlowNotify) ; 409 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; 410 NEWPERFCOUNTER(_sync_FailedSpins) ; 411 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; 412 NEWPERFCOUNTER(_sync_PrivateA) ; 413 NEWPERFCOUNTER(_sync_PrivateB) ; 414 NEWPERFCOUNTER(_sync_MonInCirculation) ; 415 NEWPERFCOUNTER(_sync_MonScavenged) ; 416 NEWPERFVARIABLE(_sync_MonExtant) ; 417 #undef NEWPERFCOUNTER 418 } 419} 420 421// Compile-time asserts 422// When possible, it's better to catch errors deterministically at 423// compile-time than at runtime. The down-side to using compile-time 424// asserts is that error message -- often something about negative array 425// indices -- is opaque. 426 427#define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 428 429void ObjectMonitor::ctAsserts() { 430 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 431} 432 433static int Adjust (volatile int * adr, int dx) { 434 int v ; 435 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; 436 return v ; 437} 438 439// Ad-hoc mutual exclusion primitives: SpinLock and Mux 440// 441// We employ SpinLocks _only for low-contention, fixed-length 442// short-duration critical sections where we're concerned 443// about native mutex_t or HotSpot Mutex:: latency. 444// The mux construct provides a spin-then-block mutual exclusion 445// mechanism. 446// 447// Testing has shown that contention on the ListLock guarding gFreeList 448// is common. If we implement ListLock as a simple SpinLock it's common 449// for the JVM to devolve to yielding with little progress. This is true 450// despite the fact that the critical sections protected by ListLock are 451// extremely short. 452// 453// TODO-FIXME: ListLock should be of type SpinLock. 454// We should make this a 1st-class type, integrated into the lock 455// hierarchy as leaf-locks. Critically, the SpinLock structure 456// should have sufficient padding to avoid false-sharing and excessive 457// cache-coherency traffic. 458 459 460typedef volatile int SpinLockT ; 461 462void Thread::SpinAcquire (volatile int * adr, const char * LockName) { 463 if (Atomic::cmpxchg (1, adr, 0) == 0) { 464 return ; // normal fast-path return 465 } 466 467 // Slow-path : We've encountered contention -- Spin/Yield/Block strategy. 468 TEVENT (SpinAcquire - ctx) ; 469 int ctr = 0 ; 470 int Yields = 0 ; 471 for (;;) { 472 while (*adr != 0) { 473 ++ctr ; 474 if ((ctr & 0xFFF) == 0 || !os::is_MP()) { 475 if (Yields > 5) { 476 // Consider using a simple NakedSleep() instead. 477 // Then SpinAcquire could be called by non-JVM threads 478 Thread::current()->_ParkEvent->park(1) ; 479 } else { 480 os::NakedYield() ; 481 ++Yields ; 482 } 483 } else { 484 SpinPause() ; 485 } 486 } 487 if (Atomic::cmpxchg (1, adr, 0) == 0) return ; 488 } 489} 490 491void Thread::SpinRelease (volatile int * adr) { 492 assert (*adr != 0, "invariant") ; 493 OrderAccess::fence() ; // guarantee at least release consistency. 494 // Roach-motel semantics. 495 // It's safe if subsequent LDs and STs float "up" into the critical section, 496 // but prior LDs and STs within the critical section can't be allowed 497 // to reorder or float past the ST that releases the lock. 498 *adr = 0 ; 499} 500 501// muxAcquire and muxRelease: 502// 503// * muxAcquire and muxRelease support a single-word lock-word construct. 504// The LSB of the word is set IFF the lock is held. 505// The remainder of the word points to the head of a singly-linked list 506// of threads blocked on the lock. 507// 508// * The current implementation of muxAcquire-muxRelease uses its own 509// dedicated Thread._MuxEvent instance. If we're interested in 510// minimizing the peak number of extant ParkEvent instances then 511// we could eliminate _MuxEvent and "borrow" _ParkEvent as long 512// as certain invariants were satisfied. Specifically, care would need 513// to be taken with regards to consuming unpark() "permits". 514// A safe rule of thumb is that a thread would never call muxAcquire() 515// if it's enqueued (cxq, EntryList, WaitList, etc) and will subsequently 516// park(). Otherwise the _ParkEvent park() operation in muxAcquire() could 517// consume an unpark() permit intended for monitorenter, for instance. 518// One way around this would be to widen the restricted-range semaphore 519// implemented in park(). Another alternative would be to provide 520// multiple instances of the PlatformEvent() for each thread. One 521// instance would be dedicated to muxAcquire-muxRelease, for instance. 522// 523// * Usage: 524// -- Only as leaf locks 525// -- for short-term locking only as muxAcquire does not perform 526// thread state transitions. 527// 528// Alternatives: 529// * We could implement muxAcquire and muxRelease with MCS or CLH locks 530// but with parking or spin-then-park instead of pure spinning. 531// * Use Taura-Oyama-Yonenzawa locks. 532// * It's possible to construct a 1-0 lock if we encode the lockword as 533// (List,LockByte). Acquire will CAS the full lockword while Release 534// will STB 0 into the LockByte. The 1-0 scheme admits stranding, so 535// acquiring threads use timers (ParkTimed) to detect and recover from 536// the stranding window. Thread/Node structures must be aligned on 256-byte 537// boundaries by using placement-new. 538// * Augment MCS with advisory back-link fields maintained with CAS(). 539// Pictorially: LockWord -> T1 <-> T2 <-> T3 <-> ... <-> Tn <-> Owner. 540// The validity of the backlinks must be ratified before we trust the value. 541// If the backlinks are invalid the exiting thread must back-track through the 542// the forward links, which are always trustworthy. 543// * Add a successor indication. The LockWord is currently encoded as 544// (List, LOCKBIT:1). We could also add a SUCCBIT or an explicit _succ variable 545// to provide the usual futile-wakeup optimization. 546// See RTStt for details. 547// * Consider schedctl.sc_nopreempt to cover the critical section. 548// 549 550 551typedef volatile intptr_t MutexT ; // Mux Lock-word 552enum MuxBits { LOCKBIT = 1 } ; 553 554void Thread::muxAcquire (volatile intptr_t * Lock, const char * LockName) { 555 intptr_t w = Atomic::cmpxchg_ptr (LOCKBIT, Lock, 0) ; 556 if (w == 0) return ; 557 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { 558 return ; 559 } 560 561 TEVENT (muxAcquire - Contention) ; 562 ParkEvent * const Self = Thread::current()->_MuxEvent ; 563 assert ((intptr_t(Self) & LOCKBIT) == 0, "invariant") ; 564 for (;;) { 565 int its = (os::is_MP() ? 100 : 0) + 1 ; 566 567 // Optional spin phase: spin-then-park strategy 568 while (--its >= 0) { 569 w = *Lock ; 570 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { 571 return ; 572 } 573 } 574 575 Self->reset() ; 576 Self->OnList = intptr_t(Lock) ; 577 // The following fence() isn't _strictly necessary as the subsequent 578 // CAS() both serializes execution and ratifies the fetched *Lock value. 579 OrderAccess::fence(); 580 for (;;) { 581 w = *Lock ; 582 if ((w & LOCKBIT) == 0) { 583 if (Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { 584 Self->OnList = 0 ; // hygiene - allows stronger asserts 585 return ; 586 } 587 continue ; // Interference -- *Lock changed -- Just retry 588 } 589 assert (w & LOCKBIT, "invariant") ; 590 Self->ListNext = (ParkEvent *) (w & ~LOCKBIT ); 591 if (Atomic::cmpxchg_ptr (intptr_t(Self)|LOCKBIT, Lock, w) == w) break ; 592 } 593 594 while (Self->OnList != 0) { 595 Self->park() ; 596 } 597 } 598} 599 600void Thread::muxAcquireW (volatile intptr_t * Lock, ParkEvent * ev) { 601 intptr_t w = Atomic::cmpxchg_ptr (LOCKBIT, Lock, 0) ; 602 if (w == 0) return ; 603 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { 604 return ; 605 } 606 607 TEVENT (muxAcquire - Contention) ; 608 ParkEvent * ReleaseAfter = NULL ; 609 if (ev == NULL) { 610 ev = ReleaseAfter = ParkEvent::Allocate (NULL) ; 611 } 612 assert ((intptr_t(ev) & LOCKBIT) == 0, "invariant") ; 613 for (;;) { 614 guarantee (ev->OnList == 0, "invariant") ; 615 int its = (os::is_MP() ? 100 : 0) + 1 ; 616 617 // Optional spin phase: spin-then-park strategy 618 while (--its >= 0) { 619 w = *Lock ; 620 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { 621 if (ReleaseAfter != NULL) { 622 ParkEvent::Release (ReleaseAfter) ; 623 } 624 return ; 625 } 626 } 627 628 ev->reset() ; 629 ev->OnList = intptr_t(Lock) ; 630 // The following fence() isn't _strictly necessary as the subsequent 631 // CAS() both serializes execution and ratifies the fetched *Lock value. 632 OrderAccess::fence(); 633 for (;;) { 634 w = *Lock ; 635 if ((w & LOCKBIT) == 0) { 636 if (Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { 637 ev->OnList = 0 ; 638 // We call ::Release while holding the outer lock, thus 639 // artificially lengthening the critical section. 640 // Consider deferring the ::Release() until the subsequent unlock(), 641 // after we've dropped the outer lock. 642 if (ReleaseAfter != NULL) { 643 ParkEvent::Release (ReleaseAfter) ; 644 } 645 return ; 646 } 647 continue ; // Interference -- *Lock changed -- Just retry 648 } 649 assert (w & LOCKBIT, "invariant") ; 650 ev->ListNext = (ParkEvent *) (w & ~LOCKBIT ); 651 if (Atomic::cmpxchg_ptr (intptr_t(ev)|LOCKBIT, Lock, w) == w) break ; 652 } 653 654 while (ev->OnList != 0) { 655 ev->park() ; 656 } 657 } 658} 659 660// Release() must extract a successor from the list and then wake that thread. 661// It can "pop" the front of the list or use a detach-modify-reattach (DMR) scheme 662// similar to that used by ParkEvent::Allocate() and ::Release(). DMR-based 663// Release() would : 664// (A) CAS() or swap() null to *Lock, releasing the lock and detaching the list. 665// (B) Extract a successor from the private list "in-hand" 666// (C) attempt to CAS() the residual back into *Lock over null. 667// If there were any newly arrived threads and the CAS() would fail. 668// In that case Release() would detach the RATs, re-merge the list in-hand 669// with the RATs and repeat as needed. Alternately, Release() might 670// detach and extract a successor, but then pass the residual list to the wakee. 671// The wakee would be responsible for reattaching and remerging before it 672// competed for the lock. 673// 674// Both "pop" and DMR are immune from ABA corruption -- there can be 675// multiple concurrent pushers, but only one popper or detacher. 676// This implementation pops from the head of the list. This is unfair, 677// but tends to provide excellent throughput as hot threads remain hot. 678// (We wake recently run threads first). 679 680void Thread::muxRelease (volatile intptr_t * Lock) { 681 for (;;) { 682 const intptr_t w = Atomic::cmpxchg_ptr (0, Lock, LOCKBIT) ; 683 assert (w & LOCKBIT, "invariant") ; 684 if (w == LOCKBIT) return ; 685 ParkEvent * List = (ParkEvent *) (w & ~LOCKBIT) ; 686 assert (List != NULL, "invariant") ; 687 assert (List->OnList == intptr_t(Lock), "invariant") ; 688 ParkEvent * nxt = List->ListNext ; 689 690 // The following CAS() releases the lock and pops the head element. 691 if (Atomic::cmpxchg_ptr (intptr_t(nxt), Lock, w) != w) { 692 continue ; 693 } 694 List->OnList = 0 ; 695 OrderAccess::fence() ; 696 List->unpark () ; 697 return ; 698 } 699} 700 701// ObjectMonitor Lifecycle 702// ----------------------- 703// Inflation unlinks monitors from the global gFreeList and 704// associates them with objects. Deflation -- which occurs at 705// STW-time -- disassociates idle monitors from objects. Such 706// scavenged monitors are returned to the gFreeList. 707// 708// The global list is protected by ListLock. All the critical sections 709// are short and operate in constant-time. 710// 711// ObjectMonitors reside in type-stable memory (TSM) and are immortal. 712// 713// Lifecycle: 714// -- unassigned and on the global free list 715// -- unassigned and on a thread's private omFreeList 716// -- assigned to an object. The object is inflated and the mark refers 717// to the objectmonitor. 718// 719// TODO-FIXME: 720// 721// * We currently protect the gFreeList with a simple lock. 722// An alternate lock-free scheme would be to pop elements from the gFreeList 723// with CAS. This would be safe from ABA corruption as long we only 724// recycled previously appearing elements onto the list in deflate_idle_monitors() 725// at STW-time. Completely new elements could always be pushed onto the gFreeList 726// with CAS. Elements that appeared previously on the list could only 727// be installed at STW-time. 728// 729// * For efficiency and to help reduce the store-before-CAS penalty 730// the objectmonitors on gFreeList or local free lists should be ready to install 731// with the exception of _header and _object. _object can be set after inflation. 732// In particular, keep all objectMonitors on a thread's private list in ready-to-install 733// state with m.Owner set properly. 734// 735// * We could all diffuse contention by using multiple global (FreeList, Lock) 736// pairs -- threads could use trylock() and a cyclic-scan strategy to search for 737// an unlocked free list. 738// 739// * Add lifecycle tags and assert()s. 740// 741// * Be more consistent about when we clear an objectmonitor's fields: 742// A. After extracting the objectmonitor from a free list. 743// B. After adding an objectmonitor to a free list. 744// 745 746ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ; 747ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ; 748static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache 749#define CHAINMARKER ((oop)-1) 750 751ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) { 752 // A large MAXPRIVATE value reduces both list lock contention 753 // and list coherency traffic, but also tends to increase the 754 // number of objectMonitors in circulation as well as the STW 755 // scavenge costs. As usual, we lean toward time in space-time 756 // tradeoffs. 757 const int MAXPRIVATE = 1024 ; 758 for (;;) { 759 ObjectMonitor * m ; 760 761 // 1: try to allocate from the thread's local omFreeList. 762 // Threads will attempt to allocate first from their local list, then 763 // from the global list, and only after those attempts fail will the thread 764 // attempt to instantiate new monitors. Thread-local free lists take 765 // heat off the ListLock and improve allocation latency, as well as reducing 766 // coherency traffic on the shared global list. 767 m = Self->omFreeList ; 768 if (m != NULL) { 769 Self->omFreeList = m->FreeNext ; 770 Self->omFreeCount -- ; 771 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 772 guarantee (m->object() == NULL, "invariant") ; 773 return m ; 774 } 775 776 // 2: try to allocate from the global gFreeList 777 // CONSIDER: use muxTry() instead of muxAcquire(). 778 // If the muxTry() fails then drop immediately into case 3. 779 // If we're using thread-local free lists then try 780 // to reprovision the caller's free list. 781 if (gFreeList != NULL) { 782 // Reprovision the thread's omFreeList. 783 // Use bulk transfers to reduce the allocation rate and heat 784 // on various locks. 785 Thread::muxAcquire (&ListLock, "omAlloc") ; 786 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) { 787 ObjectMonitor * take = gFreeList ; 788 gFreeList = take->FreeNext ; 789 guarantee (take->object() == NULL, "invariant") ; 790 guarantee (!take->is_busy(), "invariant") ; 791 take->Recycle() ; 792 omRelease (Self, take) ; 793 } 794 Thread::muxRelease (&ListLock) ; 795 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ; 796 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ; 797 TEVENT (omFirst - reprovision) ; 798 continue ; 799 } 800 801 // 3: allocate a block of new ObjectMonitors 802 // Both the local and global free lists are empty -- resort to malloc(). 803 // In the current implementation objectMonitors are TSM - immortal. 804 assert (_BLOCKSIZE > 1, "invariant") ; 805 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; 806 807 // NOTE: (almost) no way to recover if allocation failed. 808 // We might be able to induce a STW safepoint and scavenge enough 809 // objectMonitors to permit progress. 810 if (temp == NULL) { 811 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ; 812 } 813 814 // Format the block. 815 // initialize the linked list, each monitor points to its next 816 // forming the single linked free list, the very first monitor 817 // will points to next block, which forms the block list. 818 // The trick of using the 1st element in the block as gBlockList 819 // linkage should be reconsidered. A better implementation would 820 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 821 822 for (int i = 1; i < _BLOCKSIZE ; i++) { 823 temp[i].FreeNext = &temp[i+1]; 824 } 825 826 // terminate the last monitor as the end of list 827 temp[_BLOCKSIZE - 1].FreeNext = NULL ; 828 829 // Element [0] is reserved for global list linkage 830 temp[0].set_object(CHAINMARKER); 831 832 // Consider carving out this thread's current request from the 833 // block in hand. This avoids some lock traffic and redundant 834 // list activity. 835 836 // Acquire the ListLock to manipulate BlockList and FreeList. 837 // An Oyama-Taura-Yonezawa scheme might be more efficient. 838 Thread::muxAcquire (&ListLock, "omAlloc [2]") ; 839 840 // Add the new block to the list of extant blocks (gBlockList). 841 // The very first objectMonitor in a block is reserved and dedicated. 842 // It serves as blocklist "next" linkage. 843 temp[0].FreeNext = gBlockList; 844 gBlockList = temp; 845 846 // Add the new string of objectMonitors to the global free list 847 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ; 848 gFreeList = temp + 1; 849 Thread::muxRelease (&ListLock) ; 850 TEVENT (Allocate block of monitors) ; 851 } 852} 853 854// Place "m" on the caller's private per-thread omFreeList. 855// In practice there's no need to clamp or limit the number of 856// monitors on a thread's omFreeList as the only time we'll call 857// omRelease is to return a monitor to the free list after a CAS 858// attempt failed. This doesn't allow unbounded #s of monitors to 859// accumulate on a thread's free list. 860// 861// In the future the usage of omRelease() might change and monitors 862// could migrate between free lists. In that case to avoid excessive 863// accumulation we could limit omCount to (omProvision*2), otherwise return 864// the objectMonitor to the global list. We should drain (return) in reasonable chunks. 865// That is, *not* one-at-a-time. 866 867 868void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m) { 869 guarantee (m->object() == NULL, "invariant") ; 870 m->FreeNext = Self->omFreeList ; 871 Self->omFreeList = m ; 872 Self->omFreeCount ++ ; 873} 874 875// Return the monitors of a moribund thread's local free list to 876// the global free list. Typically a thread calls omFlush() when 877// it's dying. We could also consider having the VM thread steal 878// monitors from threads that have not run java code over a few 879// consecutive STW safepoints. Relatedly, we might decay 880// omFreeProvision at STW safepoints. 881// 882// We currently call omFlush() from the Thread:: dtor _after the thread 883// has been excised from the thread list and is no longer a mutator. 884// That means that omFlush() can run concurrently with a safepoint and 885// the scavenge operator. Calling omFlush() from JavaThread::exit() might 886// be a better choice as we could safely reason that that the JVM is 887// not at a safepoint at the time of the call, and thus there could 888// be not inopportune interleavings between omFlush() and the scavenge 889// operator. 890 891void ObjectSynchronizer::omFlush (Thread * Self) { 892 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL 893 Self->omFreeList = NULL ; 894 if (List == NULL) return ; 895 ObjectMonitor * Tail = NULL ; 896 ObjectMonitor * s ; 897 for (s = List ; s != NULL ; s = s->FreeNext) { 898 Tail = s ; 899 guarantee (s->object() == NULL, "invariant") ; 900 guarantee (!s->is_busy(), "invariant") ; 901 s->set_owner (NULL) ; // redundant but good hygiene 902 TEVENT (omFlush - Move one) ; 903 } 904 905 guarantee (Tail != NULL && List != NULL, "invariant") ; 906 Thread::muxAcquire (&ListLock, "omFlush") ; 907 Tail->FreeNext = gFreeList ; 908 gFreeList = List ; 909 Thread::muxRelease (&ListLock) ; 910 TEVENT (omFlush) ; 911} 912 913 914// Get the next block in the block list. 915static inline ObjectMonitor* next(ObjectMonitor* block) { 916 assert(block->object() == CHAINMARKER, "must be a block header"); 917 block = block->FreeNext ; 918 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 919 return block; 920} 921 922// Fast path code shared by multiple functions 923ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 924 markOop mark = obj->mark(); 925 if (mark->has_monitor()) { 926 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 927 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 928 return mark->monitor(); 929 } 930 return ObjectSynchronizer::inflate(Thread::current(), obj); 931} 932 933// Note that we could encounter some performance loss through false-sharing as 934// multiple locks occupy the same $ line. Padding might be appropriate. 935 936#define NINFLATIONLOCKS 256 937static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ; 938 939static markOop ReadStableMark (oop obj) { 940 markOop mark = obj->mark() ; 941 if (!mark->is_being_inflated()) { 942 return mark ; // normal fast-path return 943 } 944 945 int its = 0 ; 946 for (;;) { 947 markOop mark = obj->mark() ; 948 if (!mark->is_being_inflated()) { 949 return mark ; // normal fast-path return 950 } 951 952 // The object is being inflated by some other thread. 953 // The caller of ReadStableMark() must wait for inflation to complete. 954 // Avoid live-lock 955 // TODO: consider calling SafepointSynchronize::do_call_back() while 956 // spinning to see if there's a safepoint pending. If so, immediately 957 // yielding or blocking would be appropriate. Avoid spinning while 958 // there is a safepoint pending. 959 // TODO: add inflation contention performance counters. 960 // TODO: restrict the aggregate number of spinners. 961 962 ++its ; 963 if (its > 10000 || !os::is_MP()) { 964 if (its & 1) { 965 os::NakedYield() ; 966 TEVENT (Inflate: INFLATING - yield) ; 967 } else { 968 // Note that the following code attenuates the livelock problem but is not 969 // a complete remedy. A more complete solution would require that the inflating 970 // thread hold the associated inflation lock. The following code simply restricts 971 // the number of spinners to at most one. We'll have N-2 threads blocked 972 // on the inflationlock, 1 thread holding the inflation lock and using 973 // a yield/park strategy, and 1 thread in the midst of inflation. 974 // A more refined approach would be to change the encoding of INFLATING 975 // to allow encapsulation of a native thread pointer. Threads waiting for 976 // inflation to complete would use CAS to push themselves onto a singly linked 977 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 978 // and calling park(). When inflation was complete the thread that accomplished inflation 979 // would detach the list and set the markword to inflated with a single CAS and 980 // then for each thread on the list, set the flag and unpark() the thread. 981 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 982 // wakes at most one thread whereas we need to wake the entire list. 983 int ix = (intptr_t(obj) >> 5) & (NINFLATIONLOCKS-1) ; 984 int YieldThenBlock = 0 ; 985 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ; 986 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ; 987 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ; 988 while (obj->mark() == markOopDesc::INFLATING()) { 989 // Beware: NakedYield() is advisory and has almost no effect on some platforms 990 // so we periodically call Self->_ParkEvent->park(1). 991 // We use a mixed spin/yield/block mechanism. 992 if ((YieldThenBlock++) >= 16) { 993 Thread::current()->_ParkEvent->park(1) ; 994 } else { 995 os::NakedYield() ; 996 } 997 } 998 Thread::muxRelease (InflationLocks + ix ) ; 999 TEVENT (Inflate: INFLATING - yield/park) ; 1000 } 1001 } else { 1002 SpinPause() ; // SMP-polite spinning 1003 } 1004 } 1005} 1006 1007ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) { 1008 // Inflate mutates the heap ... 1009 // Relaxing assertion for bug 6320749. 1010 assert (Universe::verify_in_progress() || 1011 !SafepointSynchronize::is_at_safepoint(), "invariant") ; 1012 1013 for (;;) { 1014 const markOop mark = object->mark() ; 1015 assert (!mark->has_bias_pattern(), "invariant") ; 1016 1017 // The mark can be in one of the following states: 1018 // * Inflated - just return 1019 // * Stack-locked - coerce it to inflated 1020 // * INFLATING - busy wait for conversion to complete 1021 // * Neutral - aggressively inflate the object. 1022 // * BIASED - Illegal. We should never see this 1023 1024 // CASE: inflated 1025 if (mark->has_monitor()) { 1026 ObjectMonitor * inf = mark->monitor() ; 1027 assert (inf->header()->is_neutral(), "invariant"); 1028 assert (inf->object() == object, "invariant") ; 1029 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1030 return inf ; 1031 } 1032 1033 // CASE: inflation in progress - inflating over a stack-lock. 1034 // Some other thread is converting from stack-locked to inflated. 1035 // Only that thread can complete inflation -- other threads must wait. 1036 // The INFLATING value is transient. 1037 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1038 // We could always eliminate polling by parking the thread on some auxiliary list. 1039 if (mark == markOopDesc::INFLATING()) { 1040 TEVENT (Inflate: spin while INFLATING) ; 1041 ReadStableMark(object) ; 1042 continue ; 1043 } 1044 1045 // CASE: stack-locked 1046 // Could be stack-locked either by this thread or by some other thread. 1047 // 1048 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1049 // to install INFLATING into the mark word. We originally installed INFLATING, 1050 // allocated the objectmonitor, and then finally STed the address of the 1051 // objectmonitor into the mark. This was correct, but artificially lengthened 1052 // the interval in which INFLATED appeared in the mark, thus increasing 1053 // the odds of inflation contention. 1054 // 1055 // We now use per-thread private objectmonitor free lists. 1056 // These list are reprovisioned from the global free list outside the 1057 // critical INFLATING...ST interval. A thread can transfer 1058 // multiple objectmonitors en-mass from the global free list to its local free list. 1059 // This reduces coherency traffic and lock contention on the global free list. 1060 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1061 // before or after the CAS(INFLATING) operation. 1062 // See the comments in omAlloc(). 1063 1064 if (mark->has_locker()) { 1065 ObjectMonitor * m = omAlloc (Self) ; 1066 // Optimistically prepare the objectmonitor - anticipate successful CAS 1067 // We do this before the CAS in order to minimize the length of time 1068 // in which INFLATING appears in the mark. 1069 m->Recycle(); 1070 m->FreeNext = NULL ; 1071 m->_Responsible = NULL ; 1072 m->OwnerIsThread = 0 ; 1073 m->_recursions = 0 ; 1074 m->_SpinDuration = Knob_SpinLimit ; // Consider: maintain by type/class 1075 1076 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ; 1077 if (cmp != mark) { 1078 omRelease (Self, m) ; 1079 continue ; // Interference -- just retry 1080 } 1081 1082 // We've successfully installed INFLATING (0) into the mark-word. 1083 // This is the only case where 0 will appear in a mark-work. 1084 // Only the singular thread that successfully swings the mark-word 1085 // to 0 can perform (or more precisely, complete) inflation. 1086 // 1087 // Why do we CAS a 0 into the mark-word instead of just CASing the 1088 // mark-word from the stack-locked value directly to the new inflated state? 1089 // Consider what happens when a thread unlocks a stack-locked object. 1090 // It attempts to use CAS to swing the displaced header value from the 1091 // on-stack basiclock back into the object header. Recall also that the 1092 // header value (hashcode, etc) can reside in (a) the object header, or 1093 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1094 // header in an objectMonitor. The inflate() routine must copy the header 1095 // value from the basiclock on the owner's stack to the objectMonitor, all 1096 // the while preserving the hashCode stability invariants. If the owner 1097 // decides to release the lock while the value is 0, the unlock will fail 1098 // and control will eventually pass from slow_exit() to inflate. The owner 1099 // will then spin, waiting for the 0 value to disappear. Put another way, 1100 // the 0 causes the owner to stall if the owner happens to try to 1101 // drop the lock (restoring the header from the basiclock to the object) 1102 // while inflation is in-progress. This protocol avoids races that might 1103 // would otherwise permit hashCode values to change or "flicker" for an object. 1104 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1105 // 0 serves as a "BUSY" inflate-in-progress indicator. 1106 1107 1108 // fetch the displaced mark from the owner's stack. 1109 // The owner can't die or unwind past the lock while our INFLATING 1110 // object is in the mark. Furthermore the owner can't complete 1111 // an unlock on the object, either. 1112 markOop dmw = mark->displaced_mark_helper() ; 1113 assert (dmw->is_neutral(), "invariant") ; 1114 1115 // Setup monitor fields to proper values -- prepare the monitor 1116 m->set_header(dmw) ; 1117 1118 // Optimization: if the mark->locker stack address is associated 1119 // with this thread we could simply set m->_owner = Self and 1120 // m->OwnerIsThread = 1. Note that a thread can inflate an object 1121 // that it has stack-locked -- as might happen in wait() -- directly 1122 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1123 m->set_owner (mark->locker()); 1124 m->set_object(object); 1125 // TODO-FIXME: assert BasicLock->dhw != 0. 1126 1127 // Must preserve store ordering. The monitor state must 1128 // be stable at the time of publishing the monitor address. 1129 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ; 1130 object->release_set_mark(markOopDesc::encode(m)); 1131 1132 // Hopefully the performance counters are allocated on distinct cache lines 1133 // to avoid false sharing on MP systems ... 1134 if (_sync_Inflations != NULL) _sync_Inflations->inc() ; 1135 TEVENT(Inflate: overwrite stacklock) ; 1136 if (TraceMonitorInflation) { 1137 if (object->is_instance()) { 1138 ResourceMark rm; 1139 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1140 (intptr_t) object, (intptr_t) object->mark(), 1141 Klass::cast(object->klass())->external_name()); 1142 } 1143 } 1144 return m ; 1145 } 1146 1147 // CASE: neutral 1148 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1149 // If we know we're inflating for entry it's better to inflate by swinging a 1150 // pre-locked objectMonitor pointer into the object header. A successful 1151 // CAS inflates the object *and* confers ownership to the inflating thread. 1152 // In the current implementation we use a 2-step mechanism where we CAS() 1153 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1154 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1155 // would be useful. 1156 1157 assert (mark->is_neutral(), "invariant"); 1158 ObjectMonitor * m = omAlloc (Self) ; 1159 // prepare m for installation - set monitor to initial state 1160 m->Recycle(); 1161 m->set_header(mark); 1162 m->set_owner(NULL); 1163 m->set_object(object); 1164 m->OwnerIsThread = 1 ; 1165 m->_recursions = 0 ; 1166 m->FreeNext = NULL ; 1167 m->_Responsible = NULL ; 1168 m->_SpinDuration = Knob_SpinLimit ; // consider: keep metastats by type/class 1169 1170 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1171 m->set_object (NULL) ; 1172 m->set_owner (NULL) ; 1173 m->OwnerIsThread = 0 ; 1174 m->Recycle() ; 1175 omRelease (Self, m) ; 1176 m = NULL ; 1177 continue ; 1178 // interference - the markword changed - just retry. 1179 // The state-transitions are one-way, so there's no chance of 1180 // live-lock -- "Inflated" is an absorbing state. 1181 } 1182 1183 // Hopefully the performance counters are allocated on distinct 1184 // cache lines to avoid false sharing on MP systems ... 1185 if (_sync_Inflations != NULL) _sync_Inflations->inc() ; 1186 TEVENT(Inflate: overwrite neutral) ; 1187 if (TraceMonitorInflation) { 1188 if (object->is_instance()) { 1189 ResourceMark rm; 1190 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1191 (intptr_t) object, (intptr_t) object->mark(), 1192 Klass::cast(object->klass())->external_name()); 1193 } 1194 } 1195 return m ; 1196 } 1197} 1198 1199 1200// This the fast monitor enter. The interpreter and compiler use 1201// some assembly copies of this code. Make sure update those code 1202// if the following function is changed. The implementation is 1203// extremely sensitive to race condition. Be careful. 1204 1205void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) { 1206 if (UseBiasedLocking) { 1207 if (!SafepointSynchronize::is_at_safepoint()) { 1208 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 1209 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 1210 return; 1211 } 1212 } else { 1213 assert(!attempt_rebias, "can not rebias toward VM thread"); 1214 BiasedLocking::revoke_at_safepoint(obj); 1215 } 1216 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1217 } 1218 1219 THREAD->update_highest_lock((address)lock); 1220 slow_enter (obj, lock, THREAD) ; 1221} 1222 1223void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 1224 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); 1225 // if displaced header is null, the previous enter is recursive enter, no-op 1226 markOop dhw = lock->displaced_header(); 1227 markOop mark ; 1228 if (dhw == NULL) { 1229 // Recursive stack-lock. 1230 // Diagnostics -- Could be: stack-locked, inflating, inflated. 1231 mark = object->mark() ; 1232 assert (!mark->is_neutral(), "invariant") ; 1233 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { 1234 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ; 1235 } 1236 if (mark->has_monitor()) { 1237 ObjectMonitor * m = mark->monitor() ; 1238 assert(((oop)(m->object()))->mark() == mark, "invariant") ; 1239 assert(m->is_entered(THREAD), "invariant") ; 1240 } 1241 return ; 1242 } 1243 1244 mark = object->mark() ; 1245 1246 // If the object is stack-locked by the current thread, try to 1247 // swing the displaced header from the box back to the mark. 1248 if (mark == (markOop) lock) { 1249 assert (dhw->is_neutral(), "invariant") ; 1250 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { 1251 TEVENT (fast_exit: release stacklock) ; 1252 return; 1253 } 1254 } 1255 1256 ObjectSynchronizer::inflate(THREAD, object)->exit (THREAD) ; 1257} 1258 1259// This routine is used to handle interpreter/compiler slow case 1260// We don't need to use fast path here, because it must have been 1261// failed in the interpreter/compiler code. 1262void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 1263 markOop mark = obj->mark(); 1264 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 1265 1266 if (mark->is_neutral()) { 1267 // Anticipate successful CAS -- the ST of the displaced mark must 1268 // be visible <= the ST performed by the CAS. 1269 lock->set_displaced_header(mark); 1270 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { 1271 TEVENT (slow_enter: release stacklock) ; 1272 return ; 1273 } 1274 // Fall through to inflate() ... 1275 } else 1276 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 1277 assert(lock != mark->locker(), "must not re-lock the same lock"); 1278 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 1279 lock->set_displaced_header(NULL); 1280 return; 1281 } 1282 1283#if 0 1284 // The following optimization isn't particularly useful. 1285 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) { 1286 lock->set_displaced_header (NULL) ; 1287 return ; 1288 } 1289#endif 1290 1291 // The object header will never be displaced to this lock, 1292 // so it does not matter what the value is, except that it 1293 // must be non-zero to avoid looking like a re-entrant lock, 1294 // and must not look locked either. 1295 lock->set_displaced_header(markOopDesc::unused_mark()); 1296 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 1297} 1298 1299// This routine is used to handle interpreter/compiler slow case 1300// We don't need to use fast path here, because it must have 1301// failed in the interpreter/compiler code. Simply use the heavy 1302// weight monitor should be ok, unless someone find otherwise. 1303void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 1304 fast_exit (object, lock, THREAD) ; 1305} 1306 1307// NOTE: must use heavy weight monitor to handle jni monitor enter 1308void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter 1309 // the current locking is from JNI instead of Java code 1310 TEVENT (jni_enter) ; 1311 if (UseBiasedLocking) { 1312 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1313 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1314 } 1315 THREAD->set_current_pending_monitor_is_from_java(false); 1316 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 1317 THREAD->set_current_pending_monitor_is_from_java(true); 1318} 1319 1320// NOTE: must use heavy weight monitor to handle jni monitor enter 1321bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) { 1322 if (UseBiasedLocking) { 1323 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1324 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1325 } 1326 1327 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj()); 1328 return monitor->try_enter(THREAD); 1329} 1330 1331 1332// NOTE: must use heavy weight monitor to handle jni monitor exit 1333void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 1334 TEVENT (jni_exit) ; 1335 if (UseBiasedLocking) { 1336 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1337 } 1338 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1339 1340 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); 1341 // If this thread has locked the object, exit the monitor. Note: can't use 1342 // monitor->check(CHECK); must exit even if an exception is pending. 1343 if (monitor->check(THREAD)) { 1344 monitor->exit(THREAD); 1345 } 1346} 1347 1348// complete_exit()/reenter() are used to wait on a nested lock 1349// i.e. to give up an outer lock completely and then re-enter 1350// Used when holding nested locks - lock acquisition order: lock1 then lock2 1351// 1) complete_exit lock1 - saving recursion count 1352// 2) wait on lock2 1353// 3) when notified on lock2, unlock lock2 1354// 4) reenter lock1 with original recursion count 1355// 5) lock lock2 1356// NOTE: must use heavy weight monitor to handle complete_exit/reenter() 1357intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 1358 TEVENT (complete_exit) ; 1359 if (UseBiasedLocking) { 1360 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1361 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1362 } 1363 1364 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 1365 1366 return monitor->complete_exit(THREAD); 1367} 1368 1369// NOTE: must use heavy weight monitor to handle complete_exit/reenter() 1370void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 1371 TEVENT (reenter) ; 1372 if (UseBiasedLocking) { 1373 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1374 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1375 } 1376 1377 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 1378 1379 monitor->reenter(recursion, THREAD); 1380} 1381 1382// This exists only as a workaround of dtrace bug 6254741 1383int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 1384 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 1385 return 0; 1386} 1387 1388// NOTE: must use heavy weight monitor to handle wait() 1389void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 1390 if (UseBiasedLocking) { 1391 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1392 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1393 } 1394 if (millis < 0) { 1395 TEVENT (wait - throw IAX) ; 1396 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 1397 } 1398 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 1399 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 1400 monitor->wait(millis, true, THREAD); 1401 1402 /* This dummy call is in place to get around dtrace bug 6254741. Once 1403 that's fixed we can uncomment the following line and remove the call */ 1404 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 1405 dtrace_waited_probe(monitor, obj, THREAD); 1406} 1407 1408void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) { 1409 if (UseBiasedLocking) { 1410 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1411 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1412 } 1413 if (millis < 0) { 1414 TEVENT (wait - throw IAX) ; 1415 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 1416 } 1417 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ; 1418} 1419 1420void ObjectSynchronizer::notify(Handle obj, TRAPS) { 1421 if (UseBiasedLocking) { 1422 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1423 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1424 } 1425 1426 markOop mark = obj->mark(); 1427 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 1428 return; 1429 } 1430 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); 1431} 1432 1433// NOTE: see comment of notify() 1434void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 1435 if (UseBiasedLocking) { 1436 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 1437 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1438 } 1439 1440 markOop mark = obj->mark(); 1441 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 1442 return; 1443 } 1444 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); 1445} 1446 1447intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { 1448 if (UseBiasedLocking) { 1449 // NOTE: many places throughout the JVM do not expect a safepoint 1450 // to be taken here, in particular most operations on perm gen 1451 // objects. However, we only ever bias Java instances and all of 1452 // the call sites of identity_hash that might revoke biases have 1453 // been checked to make sure they can handle a safepoint. The 1454 // added check of the bias pattern is to avoid useless calls to 1455 // thread-local storage. 1456 if (obj->mark()->has_bias_pattern()) { 1457 // Box and unbox the raw reference just in case we cause a STW safepoint. 1458 Handle hobj (Self, obj) ; 1459 // Relaxing assertion for bug 6320749. 1460 assert (Universe::verify_in_progress() || 1461 !SafepointSynchronize::is_at_safepoint(), 1462 "biases should not be seen by VM thread here"); 1463 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 1464 obj = hobj() ; 1465 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1466 } 1467 } 1468 1469 // hashCode() is a heap mutator ... 1470 // Relaxing assertion for bug 6320749. 1471 assert (Universe::verify_in_progress() || 1472 !SafepointSynchronize::is_at_safepoint(), "invariant") ; 1473 assert (Universe::verify_in_progress() || 1474 Self->is_Java_thread() , "invariant") ; 1475 assert (Universe::verify_in_progress() || 1476 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 1477 1478 ObjectMonitor* monitor = NULL; 1479 markOop temp, test; 1480 intptr_t hash; 1481 markOop mark = ReadStableMark (obj); 1482 1483 // object should remain ineligible for biased locking 1484 assert (!mark->has_bias_pattern(), "invariant") ; 1485 1486 if (mark->is_neutral()) { 1487 hash = mark->hash(); // this is a normal header 1488 if (hash) { // if it has hash, just return it 1489 return hash; 1490 } 1491 hash = get_next_hash(Self, obj); // allocate a new hash code 1492 temp = mark->copy_set_hash(hash); // merge the hash code into header 1493 // use (machine word version) atomic operation to install the hash 1494 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 1495 if (test == mark) { 1496 return hash; 1497 } 1498 // If atomic operation failed, we must inflate the header 1499 // into heavy weight monitor. We could add more code here 1500 // for fast path, but it does not worth the complexity. 1501 } else if (mark->has_monitor()) { 1502 monitor = mark->monitor(); 1503 temp = monitor->header(); 1504 assert (temp->is_neutral(), "invariant") ; 1505 hash = temp->hash(); 1506 if (hash) { 1507 return hash; 1508 } 1509 // Skip to the following code to reduce code size 1510 } else if (Self->is_lock_owned((address)mark->locker())) { 1511 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 1512 assert (temp->is_neutral(), "invariant") ; 1513 hash = temp->hash(); // by current thread, check if the displaced 1514 if (hash) { // header contains hash code 1515 return hash; 1516 } 1517 // WARNING: 1518 // The displaced header is strictly immutable. 1519 // It can NOT be changed in ANY cases. So we have 1520 // to inflate the header into heavyweight monitor 1521 // even the current thread owns the lock. The reason 1522 // is the BasicLock (stack slot) will be asynchronously 1523 // read by other threads during the inflate() function. 1524 // Any change to stack may not propagate to other threads 1525 // correctly. 1526 } 1527 1528 // Inflate the monitor to set hash code 1529 monitor = ObjectSynchronizer::inflate(Self, obj); 1530 // Load displaced header and check it has hash code 1531 mark = monitor->header(); 1532 assert (mark->is_neutral(), "invariant") ; 1533 hash = mark->hash(); 1534 if (hash == 0) { 1535 hash = get_next_hash(Self, obj); 1536 temp = mark->copy_set_hash(hash); // merge hash code into header 1537 assert (temp->is_neutral(), "invariant") ; 1538 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 1539 if (test != mark) { 1540 // The only update to the header in the monitor (outside GC) 1541 // is install the hash code. If someone add new usage of 1542 // displaced header, please update this code 1543 hash = test->hash(); 1544 assert (test->is_neutral(), "invariant") ; 1545 assert (hash != 0, "Trivial unexpected object/monitor header usage."); 1546 } 1547 } 1548 // We finally get the hash 1549 return hash; 1550} 1551 1552// Deprecated -- use FastHashCode() instead. 1553 1554intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 1555 return FastHashCode (Thread::current(), obj()) ; 1556} 1557 1558bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 1559 Handle h_obj) { 1560 if (UseBiasedLocking) { 1561 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 1562 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1563 } 1564 1565 assert(thread == JavaThread::current(), "Can only be called on current thread"); 1566 oop obj = h_obj(); 1567 1568 markOop mark = ReadStableMark (obj) ; 1569 1570 // Uncontended case, header points to stack 1571 if (mark->has_locker()) { 1572 return thread->is_lock_owned((address)mark->locker()); 1573 } 1574 // Contended case, header points to ObjectMonitor (tagged pointer) 1575 if (mark->has_monitor()) { 1576 ObjectMonitor* monitor = mark->monitor(); 1577 return monitor->is_entered(thread) != 0 ; 1578 } 1579 // Unlocked case, header in place 1580 assert(mark->is_neutral(), "sanity check"); 1581 return false; 1582} 1583 1584// Be aware of this method could revoke bias of the lock object. 1585// This method querys the ownership of the lock handle specified by 'h_obj'. 1586// If the current thread owns the lock, it returns owner_self. If no 1587// thread owns the lock, it returns owner_none. Otherwise, it will return 1588// ower_other. 1589ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 1590(JavaThread *self, Handle h_obj) { 1591 // The caller must beware this method can revoke bias, and 1592 // revocation can result in a safepoint. 1593 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 1594 assert (self->thread_state() != _thread_blocked , "invariant") ; 1595 1596 // Possible mark states: neutral, biased, stack-locked, inflated 1597 1598 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 1599 // CASE: biased 1600 BiasedLocking::revoke_and_rebias(h_obj, false, self); 1601 assert(!h_obj->mark()->has_bias_pattern(), 1602 "biases should be revoked by now"); 1603 } 1604 1605 assert(self == JavaThread::current(), "Can only be called on current thread"); 1606 oop obj = h_obj(); 1607 markOop mark = ReadStableMark (obj) ; 1608 1609 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 1610 if (mark->has_locker()) { 1611 return self->is_lock_owned((address)mark->locker()) ? 1612 owner_self : owner_other; 1613 } 1614 1615 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 1616 // The Object:ObjectMonitor relationship is stable as long as we're 1617 // not at a safepoint. 1618 if (mark->has_monitor()) { 1619 void * owner = mark->monitor()->_owner ; 1620 if (owner == NULL) return owner_none ; 1621 return (owner == self || 1622 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 1623 } 1624 1625 // CASE: neutral 1626 assert(mark->is_neutral(), "sanity check"); 1627 return owner_none ; // it's unlocked 1628} 1629 1630// FIXME: jvmti should call this 1631JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 1632 if (UseBiasedLocking) { 1633 if (SafepointSynchronize::is_at_safepoint()) { 1634 BiasedLocking::revoke_at_safepoint(h_obj); 1635 } else { 1636 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 1637 } 1638 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 1639 } 1640 1641 oop obj = h_obj(); 1642 address owner = NULL; 1643 1644 markOop mark = ReadStableMark (obj) ; 1645 1646 // Uncontended case, header points to stack 1647 if (mark->has_locker()) { 1648 owner = (address) mark->locker(); 1649 } 1650 1651 // Contended case, header points to ObjectMonitor (tagged pointer) 1652 if (mark->has_monitor()) { 1653 ObjectMonitor* monitor = mark->monitor(); 1654 assert(monitor != NULL, "monitor should be non-null"); 1655 owner = (address) monitor->owner(); 1656 } 1657 1658 if (owner != NULL) { 1659 return Threads::owning_thread_from_monitor_owner(owner, doLock); 1660 } 1661 1662 // Unlocked case, header in place 1663 // Cannot have assertion since this object may have been 1664 // locked by another thread when reaching here. 1665 // assert(mark->is_neutral(), "sanity check"); 1666 1667 return NULL; 1668} 1669 1670// Iterate through monitor cache and attempt to release thread's monitors 1671// Gives up on a particular monitor if an exception occurs, but continues 1672// the overall iteration, swallowing the exception. 1673class ReleaseJavaMonitorsClosure: public MonitorClosure { 1674private: 1675 TRAPS; 1676 1677public: 1678 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1679 void do_monitor(ObjectMonitor* mid) { 1680 if (mid->owner() == THREAD) { 1681 (void)mid->complete_exit(CHECK); 1682 } 1683 } 1684}; 1685 1686// Release all inflated monitors owned by THREAD. Lightweight monitors are 1687// ignored. This is meant to be called during JNI thread detach which assumes 1688// all remaining monitors are heavyweight. All exceptions are swallowed. 1689// Scanning the extant monitor list can be time consuming. 1690// A simple optimization is to add a per-thread flag that indicates a thread 1691// called jni_monitorenter() during its lifetime. 1692// 1693// Instead of No_Savepoint_Verifier it might be cheaper to 1694// use an idiom of the form: 1695// auto int tmp = SafepointSynchronize::_safepoint_counter ; 1696// <code that must not run at safepoint> 1697// guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1698// Since the tests are extremely cheap we could leave them enabled 1699// for normal product builds. 1700 1701void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1702 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1703 No_Safepoint_Verifier nsv ; 1704 ReleaseJavaMonitorsClosure rjmc(THREAD); 1705 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); 1706 ObjectSynchronizer::monitors_iterate(&rjmc); 1707 Thread::muxRelease(&ListLock); 1708 THREAD->clear_pending_exception(); 1709} 1710 1711// Visitors ... 1712 1713void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 1714 ObjectMonitor* block = gBlockList; 1715 ObjectMonitor* mid; 1716 while (block) { 1717 assert(block->object() == CHAINMARKER, "must be a block header"); 1718 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 1719 mid = block + i; 1720 oop object = (oop) mid->object(); 1721 if (object != NULL) { 1722 closure->do_monitor(mid); 1723 } 1724 } 1725 block = (ObjectMonitor*) block->FreeNext; 1726 } 1727} 1728 1729void ObjectSynchronizer::oops_do(OopClosure* f) { 1730 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1731 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { 1732 assert(block->object() == CHAINMARKER, "must be a block header"); 1733 for (int i = 1; i < _BLOCKSIZE; i++) { 1734 ObjectMonitor* mid = &block[i]; 1735 if (mid->object() != NULL) { 1736 f->do_oop((oop*)mid->object_addr()); 1737 } 1738 } 1739 } 1740} 1741 1742// Deflate_idle_monitors() is called at all safepoints, immediately 1743// after all mutators are stopped, but before any objects have moved. 1744// It traverses the list of known monitors, deflating where possible. 1745// The scavenged monitor are returned to the monitor free list. 1746// 1747// Beware that we scavenge at *every* stop-the-world point. 1748// Having a large number of monitors in-circulation negatively 1749// impacts the performance of some applications (e.g., PointBase). 1750// Broadly, we want to minimize the # of monitors in circulation. 1751// Alternately, we could partition the active monitors into sub-lists 1752// of those that need scanning and those that do not. 1753// Specifically, we would add a new sub-list of objectmonitors 1754// that are in-circulation and potentially active. deflate_idle_monitors() 1755// would scan only that list. Other monitors could reside on a quiescent 1756// list. Such sequestered monitors wouldn't need to be scanned by 1757// deflate_idle_monitors(). omAlloc() would first check the global free list, 1758// then the quiescent list, and, failing those, would allocate a new block. 1759// Deflate_idle_monitors() would scavenge and move monitors to the 1760// quiescent list. 1761// 1762// Perversely, the heap size -- and thus the STW safepoint rate -- 1763// typically drives the scavenge rate. Large heaps can mean infrequent GC, 1764// which in turn can mean large(r) numbers of objectmonitors in circulation. 1765// This is an unfortunate aspect of this design. 1766// 1767// Another refinement would be to refrain from calling deflate_idle_monitors() 1768// except at stop-the-world points associated with garbage collections. 1769// 1770// An even better solution would be to deflate on-the-fly, aggressively, 1771// at monitorexit-time as is done in EVM's metalock or Relaxed Locks. 1772 1773void ObjectSynchronizer::deflate_idle_monitors() { 1774 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1775 int nInuse = 0 ; // currently associated with objects 1776 int nInCirculation = 0 ; // extant 1777 int nScavenged = 0 ; // reclaimed 1778 1779 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors 1780 ObjectMonitor * FreeTail = NULL ; 1781 1782 // Iterate over all extant monitors - Scavenge all idle monitors. 1783 TEVENT (deflate_idle_monitors) ; 1784 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { 1785 assert(block->object() == CHAINMARKER, "must be a block header"); 1786 nInCirculation += _BLOCKSIZE ; 1787 for (int i = 1 ; i < _BLOCKSIZE; i++) { 1788 ObjectMonitor* mid = &block[i]; 1789 oop obj = (oop) mid->object(); 1790 1791 if (obj == NULL) { 1792 // The monitor is not associated with an object. 1793 // The monitor should either be a thread-specific private 1794 // free list or the global free list. 1795 // obj == NULL IMPLIES mid->is_busy() == 0 1796 guarantee (!mid->is_busy(), "invariant") ; 1797 continue ; 1798 } 1799 1800 // Normal case ... The monitor is associated with obj. 1801 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ; 1802 guarantee (mid == obj->mark()->monitor(), "invariant"); 1803 guarantee (mid->header()->is_neutral(), "invariant"); 1804 1805 if (mid->is_busy()) { 1806 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ; 1807 nInuse ++ ; 1808 } else { 1809 // Deflate the monitor if it is no longer being used 1810 // It's idle - scavenge and return to the global free list 1811 // plain old deflation ... 1812 TEVENT (deflate_idle_monitors - scavenge1) ; 1813 if (TraceMonitorInflation) { 1814 if (obj->is_instance()) { 1815 ResourceMark rm; 1816 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1817 (intptr_t) obj, (intptr_t) obj->mark(), Klass::cast(obj->klass())->external_name()); 1818 } 1819 } 1820 1821 // Restore the header back to obj 1822 obj->release_set_mark(mid->header()); 1823 mid->clear(); 1824 1825 assert (mid->object() == NULL, "invariant") ; 1826 1827 // Move the object to the working free list defined by FreeHead,FreeTail. 1828 mid->FreeNext = NULL ; 1829 if (FreeHead == NULL) FreeHead = mid ; 1830 if (FreeTail != NULL) FreeTail->FreeNext = mid ; 1831 FreeTail = mid ; 1832 nScavenged ++ ; 1833 } 1834 } 1835 } 1836 1837 // Move the scavenged monitors back to the global free list. 1838 // In theory we don't need the freelist lock as we're at a STW safepoint. 1839 // omAlloc() and omFree() can only be called while a thread is _not in safepoint state. 1840 // But it's remotely possible that omFlush() or release_monitors_owned_by_thread() 1841 // might be called while not at a global STW safepoint. In the interest of 1842 // safety we protect the following access with ListLock. 1843 // An even more conservative and prudent approach would be to guard 1844 // the main loop in scavenge_idle_monitors() with ListLock. 1845 if (FreeHead != NULL) { 1846 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ; 1847 assert (FreeTail->FreeNext == NULL, "invariant") ; 1848 // constant-time list splice - prepend scavenged segment to gFreeList 1849 Thread::muxAcquire (&ListLock, "scavenge - return") ; 1850 FreeTail->FreeNext = gFreeList ; 1851 gFreeList = FreeHead ; 1852 Thread::muxRelease (&ListLock) ; 1853 } 1854 1855 if (_sync_Deflations != NULL) _sync_Deflations->inc(nScavenged) ; 1856 if (_sync_MonExtant != NULL) _sync_MonExtant ->set_value(nInCirculation); 1857 1858 // TODO: Add objectMonitor leak detection. 1859 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1860 GVars.stwRandom = os::random() ; 1861 GVars.stwCycle ++ ; 1862} 1863 1864// A macro is used below because there may already be a pending 1865// exception which should not abort the execution of the routines 1866// which use this (which is why we don't put this into check_slow and 1867// call it with a CHECK argument). 1868 1869#define CHECK_OWNER() \ 1870 do { \ 1871 if (THREAD != _owner) { \ 1872 if (THREAD->is_lock_owned((address) _owner)) { \ 1873 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ 1874 _recursions = 0; \ 1875 OwnerIsThread = 1 ; \ 1876 } else { \ 1877 TEVENT (Throw IMSX) ; \ 1878 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1879 } \ 1880 } \ 1881 } while (false) 1882 1883// TODO-FIXME: eliminate ObjectWaiters. Replace this visitor/enumerator 1884// interface with a simple FirstWaitingThread(), NextWaitingThread() interface. 1885 1886ObjectWaiter* ObjectMonitor::first_waiter() { 1887 return _WaitSet; 1888} 1889 1890ObjectWaiter* ObjectMonitor::next_waiter(ObjectWaiter* o) { 1891 return o->_next; 1892} 1893 1894Thread* ObjectMonitor::thread_of_waiter(ObjectWaiter* o) { 1895 return o->_thread; 1896} 1897 1898// initialize the monitor, exception the semaphore, all other fields 1899// are simple integers or pointers 1900ObjectMonitor::ObjectMonitor() { 1901 _header = NULL; 1902 _count = 0; 1903 _waiters = 0, 1904 _recursions = 0; 1905 _object = NULL; 1906 _owner = NULL; 1907 _WaitSet = NULL; 1908 _WaitSetLock = 0 ; 1909 _Responsible = NULL ; 1910 _succ = NULL ; 1911 _cxq = NULL ; 1912 FreeNext = NULL ; 1913 _EntryList = NULL ; 1914 _SpinFreq = 0 ; 1915 _SpinClock = 0 ; 1916 OwnerIsThread = 0 ; 1917} 1918 1919ObjectMonitor::~ObjectMonitor() { 1920 // TODO: Add asserts ... 1921 // _cxq == 0 _succ == NULL _owner == NULL _waiters == 0 1922 // _count == 0 _EntryList == NULL etc 1923} 1924 1925intptr_t ObjectMonitor::is_busy() const { 1926 // TODO-FIXME: merge _count and _waiters. 1927 // TODO-FIXME: assert _owner == null implies _recursions = 0 1928 // TODO-FIXME: assert _WaitSet != null implies _count > 0 1929 return _count|_waiters|intptr_t(_owner)|intptr_t(_cxq)|intptr_t(_EntryList ) ; 1930} 1931 1932void ObjectMonitor::Recycle () { 1933 // TODO: add stronger asserts ... 1934 // _cxq == 0 _succ == NULL _owner == NULL _waiters == 0 1935 // _count == 0 EntryList == NULL 1936 // _recursions == 0 _WaitSet == NULL 1937 // TODO: assert (is_busy()|_recursions) == 0 1938 _succ = NULL ; 1939 _EntryList = NULL ; 1940 _cxq = NULL ; 1941 _WaitSet = NULL ; 1942 _recursions = 0 ; 1943 _SpinFreq = 0 ; 1944 _SpinClock = 0 ; 1945 OwnerIsThread = 0 ; 1946} 1947 1948// WaitSet management ... 1949 1950inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 1951 assert(node != NULL, "should not dequeue NULL node"); 1952 assert(node->_prev == NULL, "node already in list"); 1953 assert(node->_next == NULL, "node already in list"); 1954 // put node at end of queue (circular doubly linked list) 1955 if (_WaitSet == NULL) { 1956 _WaitSet = node; 1957 node->_prev = node; 1958 node->_next = node; 1959 } else { 1960 ObjectWaiter* head = _WaitSet ; 1961 ObjectWaiter* tail = head->_prev; 1962 assert(tail->_next == head, "invariant check"); 1963 tail->_next = node; 1964 head->_prev = node; 1965 node->_next = head; 1966 node->_prev = tail; 1967 } 1968} 1969 1970inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 1971 // dequeue the very first waiter 1972 ObjectWaiter* waiter = _WaitSet; 1973 if (waiter) { 1974 DequeueSpecificWaiter(waiter); 1975 } 1976 return waiter; 1977} 1978 1979inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 1980 assert(node != NULL, "should not dequeue NULL node"); 1981 assert(node->_prev != NULL, "node already removed from list"); 1982 assert(node->_next != NULL, "node already removed from list"); 1983 // when the waiter has woken up because of interrupt, 1984 // timeout or other spurious wake-up, dequeue the 1985 // waiter from waiting list 1986 ObjectWaiter* next = node->_next; 1987 if (next == node) { 1988 assert(node->_prev == node, "invariant check"); 1989 _WaitSet = NULL; 1990 } else { 1991 ObjectWaiter* prev = node->_prev; 1992 assert(prev->_next == node, "invariant check"); 1993 assert(next->_prev == node, "invariant check"); 1994 next->_prev = prev; 1995 prev->_next = next; 1996 if (_WaitSet == node) { 1997 _WaitSet = next; 1998 } 1999 } 2000 node->_next = NULL; 2001 node->_prev = NULL; 2002} 2003 2004static char * kvGet (char * kvList, const char * Key) { 2005 if (kvList == NULL) return NULL ; 2006 size_t n = strlen (Key) ; 2007 char * Search ; 2008 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { 2009 if (strncmp (Search, Key, n) == 0) { 2010 if (Search[n] == '=') return Search + n + 1 ; 2011 if (Search[n] == 0) return (char *) "1" ; 2012 } 2013 } 2014 return NULL ; 2015} 2016 2017static int kvGetInt (char * kvList, const char * Key, int Default) { 2018 char * v = kvGet (kvList, Key) ; 2019 int rslt = v ? ::strtol (v, NULL, 0) : Default ; 2020 if (Knob_ReportSettings && v != NULL) { 2021 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2022 ::fflush (stdout) ; 2023 } 2024 return rslt ; 2025} 2026 2027// By convention we unlink a contending thread from EntryList|cxq immediately 2028// after the thread acquires the lock in ::enter(). Equally, we could defer 2029// unlinking the thread until ::exit()-time. 2030 2031void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 2032{ 2033 assert (_owner == Self, "invariant") ; 2034 assert (SelfNode->_thread == Self, "invariant") ; 2035 2036 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 2037 // Normal case: remove Self from the DLL EntryList . 2038 // This is a constant-time operation. 2039 ObjectWaiter * nxt = SelfNode->_next ; 2040 ObjectWaiter * prv = SelfNode->_prev ; 2041 if (nxt != NULL) nxt->_prev = prv ; 2042 if (prv != NULL) prv->_next = nxt ; 2043 if (SelfNode == _EntryList ) _EntryList = nxt ; 2044 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; 2045 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; 2046 TEVENT (Unlink from EntryList) ; 2047 } else { 2048 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; 2049 // Inopportune interleaving -- Self is still on the cxq. 2050 // This usually means the enqueue of self raced an exiting thread. 2051 // Normally we'll find Self near the front of the cxq, so 2052 // dequeueing is typically fast. If needbe we can accelerate 2053 // this with some MCS/CHL-like bidirectional list hints and advisory 2054 // back-links so dequeueing from the interior will normally operate 2055 // in constant-time. 2056 // Dequeue Self from either the head (with CAS) or from the interior 2057 // with a linear-time scan and normal non-atomic memory operations. 2058 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 2059 // and then unlink Self from EntryList. We have to drain eventually, 2060 // so it might as well be now. 2061 2062 ObjectWaiter * v = _cxq ; 2063 assert (v != NULL, "invariant") ; 2064 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 2065 // The CAS above can fail from interference IFF a "RAT" arrived. 2066 // In that case Self must be in the interior and can no longer be 2067 // at the head of cxq. 2068 if (v == SelfNode) { 2069 assert (_cxq != v, "invariant") ; 2070 v = _cxq ; // CAS above failed - start scan at head of list 2071 } 2072 ObjectWaiter * p ; 2073 ObjectWaiter * q = NULL ; 2074 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { 2075 q = p ; 2076 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; 2077 } 2078 assert (v != SelfNode, "invariant") ; 2079 assert (p == SelfNode, "Node not found on cxq") ; 2080 assert (p != _cxq, "invariant") ; 2081 assert (q != NULL, "invariant") ; 2082 assert (q->_next == p, "invariant") ; 2083 q->_next = p->_next ; 2084 } 2085 TEVENT (Unlink from cxq) ; 2086 } 2087 2088 // Diagnostic hygiene ... 2089 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; 2090 SelfNode->_next = (ObjectWaiter *) 0xBAD ; 2091 SelfNode->TState = ObjectWaiter::TS_RUN ; 2092} 2093 2094// Caveat: TryLock() is not necessarily serializing if it returns failure. 2095// Callers must compensate as needed. 2096 2097int ObjectMonitor::TryLock (Thread * Self) { 2098 for (;;) { 2099 void * own = _owner ; 2100 if (own != NULL) return 0 ; 2101 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 2102 // Either guarantee _recursions == 0 or set _recursions = 0. 2103 assert (_recursions == 0, "invariant") ; 2104 assert (_owner == Self, "invariant") ; 2105 // CONSIDER: set or assert that OwnerIsThread == 1 2106 return 1 ; 2107 } 2108 // The lock had been free momentarily, but we lost the race to the lock. 2109 // Interference -- the CAS failed. 2110 // We can either return -1 or retry. 2111 // Retry doesn't make as much sense because the lock was just acquired. 2112 if (true) return -1 ; 2113 } 2114} 2115 2116// NotRunnable() -- informed spinning 2117// 2118// Don't bother spinning if the owner is not eligible to drop the lock. 2119// Peek at the owner's schedctl.sc_state and Thread._thread_values and 2120// spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2121// The thread must be runnable in order to drop the lock in timely fashion. 2122// If the _owner is not runnable then spinning will not likely be 2123// successful (profitable). 2124// 2125// Beware -- the thread referenced by _owner could have died 2126// so a simply fetch from _owner->_thread_state might trap. 2127// Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2128// Because of the lifecycle issues the schedctl and _thread_state values 2129// observed by NotRunnable() might be garbage. NotRunnable must 2130// tolerate this and consider the observed _thread_state value 2131// as advisory. 2132// 2133// Beware too, that _owner is sometimes a BasicLock address and sometimes 2134// a thread pointer. We differentiate the two cases with OwnerIsThread. 2135// Alternately, we might tag the type (thread pointer vs basiclock pointer) 2136// with the LSB of _owner. Another option would be to probablistically probe 2137// the putative _owner->TypeTag value. 2138// 2139// Checking _thread_state isn't perfect. Even if the thread is 2140// in_java it might be blocked on a page-fault or have been preempted 2141// and sitting on a ready/dispatch queue. _thread state in conjunction 2142// with schedctl.sc_state gives us a good picture of what the 2143// thread is doing, however. 2144// 2145// TODO: check schedctl.sc_state. 2146// We'll need to use SafeFetch32() to read from the schedctl block. 2147// See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2148// 2149// The return value from NotRunnable() is *advisory* -- the 2150// result is based on sampling and is not necessarily coherent. 2151// The caller must tolerate false-negative and false-positive errors. 2152// Spinning, in general, is probabilistic anyway. 2153 2154 2155int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2156 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2157 if (!OwnerIsThread) return 0 ; 2158 2159 if (ox == NULL) return 0 ; 2160 2161 // Avoid transitive spinning ... 2162 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2163 // Immediately after T1 acquires L it's possible that T2, also 2164 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2165 // This occurs transiently after T1 acquired L but before 2166 // T1 managed to clear T1.Stalled. T2 does not need to abort 2167 // its spin in this circumstance. 2168 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; 2169 2170 if (BlockedOn == 1) return 1 ; 2171 if (BlockedOn != 0) { 2172 return BlockedOn != intptr_t(this) && _owner == ox ; 2173 } 2174 2175 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; 2176 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; 2177 // consider also: jst != _thread_in_Java -- but that's overspecific. 2178 return jst == _thread_blocked || jst == _thread_in_native ; 2179} 2180 2181 2182// Adaptive spin-then-block - rational spinning 2183// 2184// Note that we spin "globally" on _owner with a classic SMP-polite TATAS 2185// algorithm. On high order SMP systems it would be better to start with 2186// a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 2187// a contending thread could enqueue itself on the cxq and then spin locally 2188// on a thread-specific variable such as its ParkEvent._Event flag. 2189// That's left as an exercise for the reader. Note that global spinning is 2190// not problematic on Niagara, as the L2$ serves the interconnect and has both 2191// low latency and massive bandwidth. 2192// 2193// Broadly, we can fix the spin frequency -- that is, the % of contended lock 2194// acquisition attempts where we opt to spin -- at 100% and vary the spin count 2195// (duration) or we can fix the count at approximately the duration of 2196// a context switch and vary the frequency. Of course we could also 2197// vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 2198// See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 2199// 2200// This implementation varies the duration "D", where D varies with 2201// the success rate of recent spin attempts. (D is capped at approximately 2202// length of a round-trip context switch). The success rate for recent 2203// spin attempts is a good predictor of the success rate of future spin 2204// attempts. The mechanism adapts automatically to varying critical 2205// section length (lock modality), system load and degree of parallelism. 2206// D is maintained per-monitor in _SpinDuration and is initialized 2207// optimistically. Spin frequency is fixed at 100%. 2208// 2209// Note that _SpinDuration is volatile, but we update it without locks 2210// or atomics. The code is designed so that _SpinDuration stays within 2211// a reasonable range even in the presence of races. The arithmetic 2212// operations on _SpinDuration are closed over the domain of legal values, 2213// so at worst a race will install and older but still legal value. 2214// At the very worst this introduces some apparent non-determinism. 2215// We might spin when we shouldn't or vice-versa, but since the spin 2216// count are relatively short, even in the worst case, the effect is harmless. 2217// 2218// Care must be taken that a low "D" value does not become an 2219// an absorbing state. Transient spinning failures -- when spinning 2220// is overall profitable -- should not cause the system to converge 2221// on low "D" values. We want spinning to be stable and predictable 2222// and fairly responsive to change and at the same time we don't want 2223// it to oscillate, become metastable, be "too" non-deterministic, 2224// or converge on or enter undesirable stable absorbing states. 2225// 2226// We implement a feedback-based control system -- using past behavior 2227// to predict future behavior. We face two issues: (a) if the 2228// input signal is random then the spin predictor won't provide optimal 2229// results, and (b) if the signal frequency is too high then the control 2230// system, which has some natural response lag, will "chase" the signal. 2231// (b) can arise from multimodal lock hold times. Transient preemption 2232// can also result in apparent bimodal lock hold times. 2233// Although sub-optimal, neither condition is particularly harmful, as 2234// in the worst-case we'll spin when we shouldn't or vice-versa. 2235// The maximum spin duration is rather short so the failure modes aren't bad. 2236// To be conservative, I've tuned the gain in system to bias toward 2237// _not spinning. Relatedly, the system can sometimes enter a mode where it 2238// "rings" or oscillates between spinning and not spinning. This happens 2239// when spinning is just on the cusp of profitability, however, so the 2240// situation is not dire. The state is benign -- there's no need to add 2241// hysteresis control to damp the transition rate between spinning and 2242// not spinning. 2243// 2244// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 2245// 2246// Spin-then-block strategies ... 2247// 2248// Thoughts on ways to improve spinning : 2249// 2250// * Periodically call {psr_}getloadavg() while spinning, and 2251// permit unbounded spinning if the load average is < 2252// the number of processors. Beware, however, that getloadavg() 2253// is exceptionally fast on solaris (about 1/10 the cost of a full 2254// spin cycle, but quite expensive on linux. Beware also, that 2255// multiple JVMs could "ring" or oscillate in a feedback loop. 2256// Sufficient damping would solve that problem. 2257// 2258// * We currently use spin loops with iteration counters to approximate 2259// spinning for some interval. Given the availability of high-precision 2260// time sources such as gethrtime(), %TICK, %STICK, RDTSC, etc., we should 2261// someday reimplement the spin loops to duration-based instead of iteration-based. 2262// 2263// * Don't spin if there are more than N = (CPUs/2) threads 2264// currently spinning on the monitor (or globally). 2265// That is, limit the number of concurrent spinners. 2266// We might also limit the # of spinners in the JVM, globally. 2267// 2268// * If a spinning thread observes _owner change hands it should 2269// abort the spin (and park immediately) or at least debit 2270// the spin counter by a large "penalty". 2271// 2272// * Classically, the spin count is either K*(CPUs-1) or is a 2273// simple constant that approximates the length of a context switch. 2274// We currently use a value -- computed by a special utility -- that 2275// approximates round-trip context switch times. 2276// 2277// * Normally schedctl_start()/_stop() is used to advise the kernel 2278// to avoid preempting threads that are running in short, bounded 2279// critical sections. We could use the schedctl hooks in an inverted 2280// sense -- spinners would set the nopreempt flag, but poll the preempt 2281// pending flag. If a spinner observed a pending preemption it'd immediately 2282// abort the spin and park. As such, the schedctl service acts as 2283// a preemption warning mechanism. 2284// 2285// * In lieu of spinning, if the system is running below saturation 2286// (that is, loadavg() << #cpus), we can instead suppress futile 2287// wakeup throttling, or even wake more than one successor at exit-time. 2288// The net effect is largely equivalent to spinning. In both cases, 2289// contending threads go ONPROC and opportunistically attempt to acquire 2290// the lock, decreasing lock handover latency at the expense of wasted 2291// cycles and context switching. 2292// 2293// * We might to spin less after we've parked as the thread will 2294// have less $ and TLB affinity with the processor. 2295// Likewise, we might spin less if we come ONPROC on a different 2296// processor or after a long period (>> rechose_interval). 2297// 2298// * A table-driven state machine similar to Solaris' dispadmin scheduling 2299// tables might be a better design. Instead of encoding information in 2300// _SpinDuration, _SpinFreq and _SpinClock we'd just use explicit, 2301// discrete states. Success or failure during a spin would drive 2302// state transitions, and each state node would contain a spin count. 2303// 2304// * If the processor is operating in a mode intended to conserve power 2305// (such as Intel's SpeedStep) or to reduce thermal output (thermal 2306// step-down mode) then the Java synchronization subsystem should 2307// forgo spinning. 2308// 2309// * The minimum spin duration should be approximately the worst-case 2310// store propagation latency on the platform. That is, the time 2311// it takes a store on CPU A to become visible on CPU B, where A and 2312// B are "distant". 2313// 2314// * We might want to factor a thread's priority in the spin policy. 2315// Threads with a higher priority might spin for slightly longer. 2316// Similarly, if we use back-off in the TATAS loop, lower priority 2317// threads might back-off longer. We don't currently use a 2318// thread's priority when placing it on the entry queue. We may 2319// want to consider doing so in future releases. 2320// 2321// * We might transiently drop a thread's scheduling priority while it spins. 2322// SCHED_BATCH on linux and FX scheduling class at priority=0 on Solaris 2323// would suffice. We could even consider letting the thread spin indefinitely at 2324// a depressed or "idle" priority. This brings up fairness issues, however -- 2325// in a saturated system a thread would with a reduced priority could languish 2326// for extended periods on the ready queue. 2327// 2328// * While spinning try to use the otherwise wasted time to help the VM make 2329// progress: 2330// 2331// -- YieldTo() the owner, if the owner is OFFPROC but ready 2332// Done our remaining quantum directly to the ready thread. 2333// This helps "push" the lock owner through the critical section. 2334// It also tends to improve affinity/locality as the lock 2335// "migrates" less frequently between CPUs. 2336// -- Walk our own stack in anticipation of blocking. Memoize the roots. 2337// -- Perform strand checking for other thread. Unpark potential strandees. 2338// -- Help GC: trace or mark -- this would need to be a bounded unit of work. 2339// Unfortunately this will pollute our $ and TLBs. Recall that we 2340// spin to avoid context switching -- context switching has an 2341// immediate cost in latency, a disruptive cost to other strands on a CMT 2342// processor, and an amortized cost because of the D$ and TLB cache 2343// reload transient when the thread comes back ONPROC and repopulates 2344// $s and TLBs. 2345// -- call getloadavg() to see if the system is saturated. It'd probably 2346// make sense to call getloadavg() half way through the spin. 2347// If the system isn't at full capacity the we'd simply reset 2348// the spin counter to and extend the spin attempt. 2349// -- Doug points out that we should use the same "helping" policy 2350// in thread.yield(). 2351// 2352// * Try MONITOR-MWAIT on systems that support those instructions. 2353// 2354// * The spin statistics that drive spin decisions & frequency are 2355// maintained in the objectmonitor structure so if we deflate and reinflate 2356// we lose spin state. In practice this is not usually a concern 2357// as the default spin state after inflation is aggressive (optimistic) 2358// and tends toward spinning. So in the worst case for a lock where 2359// spinning is not profitable we may spin unnecessarily for a brief 2360// period. But then again, if a lock is contended it'll tend not to deflate 2361// in the first place. 2362 2363 2364intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; 2365int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; 2366 2367// Spinning: Fixed frequency (100%), vary duration 2368 2369int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 2370 2371 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2372 int ctr = Knob_FixedSpin ; 2373 if (ctr != 0) { 2374 while (--ctr >= 0) { 2375 if (TryLock (Self) > 0) return 1 ; 2376 SpinPause () ; 2377 } 2378 return 0 ; 2379 } 2380 2381 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { 2382 if (TryLock(Self) > 0) { 2383 // Increase _SpinDuration ... 2384 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2385 // Raising _SpurDuration to the poverty line is key. 2386 int x = _SpinDuration ; 2387 if (x < Knob_SpinLimit) { 2388 if (x < Knob_Poverty) x = Knob_Poverty ; 2389 _SpinDuration = x + Knob_BonusB ; 2390 } 2391 return 1 ; 2392 } 2393 SpinPause () ; 2394 } 2395 2396 // Admission control - verify preconditions for spinning 2397 // 2398 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2399 // becoming an absorbing state. Put another way, we spin briefly to 2400 // sample, just in case the system load, parallelism, contention, or lock 2401 // modality changed. 2402 // 2403 // Consider the following alternative: 2404 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2405 // spin attempt. "Periodically" might mean after a tally of 2406 // the # of failed spin attempts (or iterations) reaches some threshold. 2407 // This takes us into the realm of 1-out-of-N spinning, where we 2408 // hold the duration constant but vary the frequency. 2409 2410 ctr = _SpinDuration ; 2411 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; 2412 if (ctr <= 0) return 0 ; 2413 2414 if (Knob_SuccRestrict && _succ != NULL) return 0 ; 2415 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2416 TEVENT (Spin abort - notrunnable [TOP]); 2417 return 0 ; 2418 } 2419 2420 int MaxSpin = Knob_MaxSpinners ; 2421 if (MaxSpin >= 0) { 2422 if (_Spinner > MaxSpin) { 2423 TEVENT (Spin abort -- too many spinners) ; 2424 return 0 ; 2425 } 2426 // Slighty racy, but benign ... 2427 Adjust (&_Spinner, 1) ; 2428 } 2429 2430 // We're good to spin ... spin ingress. 2431 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2432 // when preparing to LD...CAS _owner, etc and the CAS is likely 2433 // to succeed. 2434 int hits = 0 ; 2435 int msk = 0 ; 2436 int caspty = Knob_CASPenalty ; 2437 int oxpty = Knob_OXPenalty ; 2438 int sss = Knob_SpinSetSucc ; 2439 if (sss && _succ == NULL ) _succ = Self ; 2440 Thread * prv = NULL ; 2441 2442 // There are three ways to exit the following loop: 2443 // 1. A successful spin where this thread has acquired the lock. 2444 // 2. Spin failure with prejudice 2445 // 3. Spin failure without prejudice 2446 2447 while (--ctr >= 0) { 2448 2449 // Periodic polling -- Check for pending GC 2450 // Threads may spin while they're unsafe. 2451 // We don't want spinning threads to delay the JVM from reaching 2452 // a stop-the-world safepoint or to steal cycles from GC. 2453 // If we detect a pending safepoint we abort in order that 2454 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2455 // this thread, if safe, doesn't steal cycles from GC. 2456 // This is in keeping with the "no loitering in runtime" rule. 2457 // We periodically check to see if there's a safepoint pending. 2458 if ((ctr & 0xFF) == 0) { 2459 if (SafepointSynchronize::do_call_back()) { 2460 TEVENT (Spin: safepoint) ; 2461 goto Abort ; // abrupt spin egress 2462 } 2463 if (Knob_UsePause & 1) SpinPause () ; 2464 2465 int (*scb)(intptr_t,int) = SpinCallbackFunction ; 2466 if (hits > 50 && scb != NULL) { 2467 int abend = (*scb)(SpinCallbackArgument, 0) ; 2468 } 2469 } 2470 2471 if (Knob_UsePause & 2) SpinPause() ; 2472 2473 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2474 // This is useful on classic SMP systems, but is of less utility on 2475 // N1-style CMT platforms. 2476 // 2477 // Trade-off: lock acquisition latency vs coherency bandwidth. 2478 // Lock hold times are typically short. A histogram 2479 // of successful spin attempts shows that we usually acquire 2480 // the lock early in the spin. That suggests we want to 2481 // sample _owner frequently in the early phase of the spin, 2482 // but then back-off and sample less frequently as the spin 2483 // progresses. The back-off makes a good citizen on SMP big 2484 // SMP systems. Oversampling _owner can consume excessive 2485 // coherency bandwidth. Relatedly, if we _oversample _owner we 2486 // can inadvertently interfere with the the ST m->owner=null. 2487 // executed by the lock owner. 2488 if (ctr & msk) continue ; 2489 ++hits ; 2490 if ((hits & 0xF) == 0) { 2491 // The 0xF, above, corresponds to the exponent. 2492 // Consider: (msk+1)|msk 2493 msk = ((msk << 2)|3) & BackOffMask ; 2494 } 2495 2496 // Probe _owner with TATAS 2497 // If this thread observes the monitor transition or flicker 2498 // from locked to unlocked to locked, then the odds that this 2499 // thread will acquire the lock in this spin attempt go down 2500 // considerably. The same argument applies if the CAS fails 2501 // or if we observe _owner change from one non-null value to 2502 // another non-null value. In such cases we might abort 2503 // the spin without prejudice or apply a "penalty" to the 2504 // spin count-down variable "ctr", reducing it by 100, say. 2505 2506 Thread * ox = (Thread *) _owner ; 2507 if (ox == NULL) { 2508 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 2509 if (ox == NULL) { 2510 // The CAS succeeded -- this thread acquired ownership 2511 // Take care of some bookkeeping to exit spin state. 2512 if (sss && _succ == Self) { 2513 _succ = NULL ; 2514 } 2515 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; 2516 2517 // Increase _SpinDuration : 2518 // The spin was successful (profitable) so we tend toward 2519 // longer spin attempts in the future. 2520 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2521 // If we acquired the lock early in the spin cycle it 2522 // makes sense to increase _SpinDuration proportionally. 2523 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2524 int x = _SpinDuration ; 2525 if (x < Knob_SpinLimit) { 2526 if (x < Knob_Poverty) x = Knob_Poverty ; 2527 _SpinDuration = x + Knob_Bonus ; 2528 } 2529 return 1 ; 2530 } 2531 2532 // The CAS failed ... we can take any of the following actions: 2533 // * penalize: ctr -= Knob_CASPenalty 2534 // * exit spin with prejudice -- goto Abort; 2535 // * exit spin without prejudice. 2536 // * Since CAS is high-latency, retry again immediately. 2537 prv = ox ; 2538 TEVENT (Spin: cas failed) ; 2539 if (caspty == -2) break ; 2540 if (caspty == -1) goto Abort ; 2541 ctr -= caspty ; 2542 continue ; 2543 } 2544 2545 // Did lock ownership change hands ? 2546 if (ox != prv && prv != NULL ) { 2547 TEVENT (spin: Owner changed) 2548 if (oxpty == -2) break ; 2549 if (oxpty == -1) goto Abort ; 2550 ctr -= oxpty ; 2551 } 2552 prv = ox ; 2553 2554 // Abort the spin if the owner is not executing. 2555 // The owner must be executing in order to drop the lock. 2556 // Spinning while the owner is OFFPROC is idiocy. 2557 // Consider: ctr -= RunnablePenalty ; 2558 if (Knob_OState && NotRunnable (Self, ox)) { 2559 TEVENT (Spin abort - notrunnable); 2560 goto Abort ; 2561 } 2562 if (sss && _succ == NULL ) _succ = Self ; 2563 } 2564 2565 // Spin failed with prejudice -- reduce _SpinDuration. 2566 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2567 // AIMD is globally stable. 2568 TEVENT (Spin failure) ; 2569 { 2570 int x = _SpinDuration ; 2571 if (x > 0) { 2572 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2573 // This is globally sample and tends to damp the response. 2574 x -= Knob_Penalty ; 2575 if (x < 0) x = 0 ; 2576 _SpinDuration = x ; 2577 } 2578 } 2579 2580 Abort: 2581 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; 2582 if (sss && _succ == Self) { 2583 _succ = NULL ; 2584 // Invariant: after setting succ=null a contending thread 2585 // must recheck-retry _owner before parking. This usually happens 2586 // in the normal usage of TrySpin(), but it's safest 2587 // to make TrySpin() as foolproof as possible. 2588 OrderAccess::fence() ; 2589 if (TryLock(Self) > 0) return 1 ; 2590 } 2591 return 0 ; 2592} 2593 2594#define TrySpin TrySpin_VaryDuration 2595 2596static void DeferredInitialize () { 2597 if (InitDone > 0) return ; 2598 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2599 while (InitDone != 1) ; 2600 return ; 2601 } 2602 2603 // One-shot global initialization ... 2604 // The initialization is idempotent, so we don't need locks. 2605 // In the future consider doing this via os::init_2(). 2606 // SyncKnobs consist of <Key>=<Value> pairs in the style 2607 // of environment variables. Start by converting ':' to NUL. 2608 2609 if (SyncKnobs == NULL) SyncKnobs = "" ; 2610 2611 size_t sz = strlen (SyncKnobs) ; 2612 char * knobs = (char *) malloc (sz + 2) ; 2613 if (knobs == NULL) { 2614 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; 2615 guarantee (0, "invariant") ; 2616 } 2617 strcpy (knobs, SyncKnobs) ; 2618 knobs[sz+1] = 0 ; 2619 for (char * p = knobs ; *p ; p++) { 2620 if (*p == ':') *p = 0 ; 2621 } 2622 2623 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2624 SETKNOB(ReportSettings) ; 2625 SETKNOB(Verbose) ; 2626 SETKNOB(FixedSpin) ; 2627 SETKNOB(SpinLimit) ; 2628 SETKNOB(SpinBase) ; 2629 SETKNOB(SpinBackOff); 2630 SETKNOB(CASPenalty) ; 2631 SETKNOB(OXPenalty) ; 2632 SETKNOB(LogSpins) ; 2633 SETKNOB(SpinSetSucc) ; 2634 SETKNOB(SuccEnabled) ; 2635 SETKNOB(SuccRestrict) ; 2636 SETKNOB(Penalty) ; 2637 SETKNOB(Bonus) ; 2638 SETKNOB(BonusB) ; 2639 SETKNOB(Poverty) ; 2640 SETKNOB(SpinAfterFutile) ; 2641 SETKNOB(UsePause) ; 2642 SETKNOB(SpinEarly) ; 2643 SETKNOB(OState) ; 2644 SETKNOB(MaxSpinners) ; 2645 SETKNOB(PreSpin) ; 2646 SETKNOB(ExitPolicy) ; 2647 SETKNOB(QMode); 2648 SETKNOB(ResetEvent) ; 2649 SETKNOB(MoveNotifyee) ; 2650 SETKNOB(FastHSSEC) ; 2651 #undef SETKNOB 2652 2653 if (os::is_MP()) { 2654 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; 2655 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; 2656 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2657 } else { 2658 Knob_SpinLimit = 0 ; 2659 Knob_SpinBase = 0 ; 2660 Knob_PreSpin = 0 ; 2661 Knob_FixedSpin = -1 ; 2662 } 2663 2664 if (Knob_LogSpins == 0) { 2665 ObjectSynchronizer::_sync_FailedSpins = NULL ; 2666 } 2667 2668 free (knobs) ; 2669 OrderAccess::fence() ; 2670 InitDone = 1 ; 2671} 2672 2673// Theory of operations -- Monitors lists, thread residency, etc: 2674// 2675// * A thread acquires ownership of a monitor by successfully 2676// CAS()ing the _owner field from null to non-null. 2677// 2678// * Invariant: A thread appears on at most one monitor list -- 2679// cxq, EntryList or WaitSet -- at any one time. 2680// 2681// * Contending threads "push" themselves onto the cxq with CAS 2682// and then spin/park. 2683// 2684// * After a contending thread eventually acquires the lock it must 2685// dequeue itself from either the EntryList or the cxq. 2686// 2687// * The exiting thread identifies and unparks an "heir presumptive" 2688// tentative successor thread on the EntryList. Critically, the 2689// exiting thread doesn't unlink the successor thread from the EntryList. 2690// After having been unparked, the wakee will recontend for ownership of 2691// the monitor. The successor (wakee) will either acquire the lock or 2692// re-park itself. 2693// 2694// Succession is provided for by a policy of competitive handoff. 2695// The exiting thread does _not_ grant or pass ownership to the 2696// successor thread. (This is also referred to as "handoff" succession"). 2697// Instead the exiting thread releases ownership and possibly wakes 2698// a successor, so the successor can (re)compete for ownership of the lock. 2699// If the EntryList is empty but the cxq is populated the exiting 2700// thread will drain the cxq into the EntryList. It does so by 2701// by detaching the cxq (installing null with CAS) and folding 2702// the threads from the cxq into the EntryList. The EntryList is 2703// doubly linked, while the cxq is singly linked because of the 2704// CAS-based "push" used to enqueue recently arrived threads (RATs). 2705// 2706// * Concurrency invariants: 2707// 2708// -- only the monitor owner may access or mutate the EntryList. 2709// The mutex property of the monitor itself protects the EntryList 2710// from concurrent interference. 2711// -- Only the monitor owner may detach the cxq. 2712// 2713// * The monitor entry list operations avoid locks, but strictly speaking 2714// they're not lock-free. Enter is lock-free, exit is not. 2715// See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 2716// 2717// * The cxq can have multiple concurrent "pushers" but only one concurrent 2718// detaching thread. This mechanism is immune from the ABA corruption. 2719// More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 2720// 2721// * Taken together, the cxq and the EntryList constitute or form a 2722// single logical queue of threads stalled trying to acquire the lock. 2723// We use two distinct lists to improve the odds of a constant-time 2724// dequeue operation after acquisition (in the ::enter() epilog) and 2725// to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 2726// A key desideratum is to minimize queue & monitor metadata manipulation 2727// that occurs while holding the monitor lock -- that is, we want to 2728// minimize monitor lock holds times. Note that even a small amount of 2729// fixed spinning will greatly reduce the # of enqueue-dequeue operations 2730// on EntryList|cxq. That is, spinning relieves contention on the "inner" 2731// locks and monitor metadata. 2732// 2733// Cxq points to the the set of Recently Arrived Threads attempting entry. 2734// Because we push threads onto _cxq with CAS, the RATs must take the form of 2735// a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 2736// the unlocking thread notices that EntryList is null but _cxq is != null. 2737// 2738// The EntryList is ordered by the prevailing queue discipline and 2739// can be organized in any convenient fashion, such as a doubly-linked list or 2740// a circular doubly-linked list. Critically, we want insert and delete operations 2741// to operate in constant-time. If we need a priority queue then something akin 2742// to Solaris' sleepq would work nicely. Viz., 2743// http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 2744// Queue discipline is enforced at ::exit() time, when the unlocking thread 2745// drains the cxq into the EntryList, and orders or reorders the threads on the 2746// EntryList accordingly. 2747// 2748// Barring "lock barging", this mechanism provides fair cyclic ordering, 2749// somewhat similar to an elevator-scan. 2750// 2751// * The monitor synchronization subsystem avoids the use of native 2752// synchronization primitives except for the narrow platform-specific 2753// park-unpark abstraction. See the comments in os_solaris.cpp regarding 2754// the semantics of park-unpark. Put another way, this monitor implementation 2755// depends only on atomic operations and park-unpark. The monitor subsystem 2756// manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 2757// underlying OS manages the READY<->RUN transitions. 2758// 2759// * Waiting threads reside on the WaitSet list -- wait() puts 2760// the caller onto the WaitSet. 2761// 2762// * notify() or notifyAll() simply transfers threads from the WaitSet to 2763// either the EntryList or cxq. Subsequent exit() operations will 2764// unpark the notifyee. Unparking a notifee in notify() is inefficient - 2765// it's likely the notifyee would simply impale itself on the lock held 2766// by the notifier. 2767// 2768// * An interesting alternative is to encode cxq as (List,LockByte) where 2769// the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 2770// variable, like _recursions, in the scheme. The threads or Events that form 2771// the list would have to be aligned in 256-byte addresses. A thread would 2772// try to acquire the lock or enqueue itself with CAS, but exiting threads 2773// could use a 1-0 protocol and simply STB to set the LockByte to 0. 2774// Note that is is *not* word-tearing, but it does presume that full-word 2775// CAS operations are coherent with intermix with STB operations. That's true 2776// on most common processors. 2777// 2778// * See also http://blogs.sun.com/dave 2779 2780 2781void ATTR ObjectMonitor::EnterI (TRAPS) { 2782 Thread * Self = THREAD ; 2783 assert (Self->is_Java_thread(), "invariant") ; 2784 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; 2785 2786 // Try the lock - TATAS 2787 if (TryLock (Self) > 0) { 2788 assert (_succ != Self , "invariant") ; 2789 assert (_owner == Self , "invariant") ; 2790 assert (_Responsible != Self , "invariant") ; 2791 return ; 2792 } 2793 2794 DeferredInitialize () ; 2795 2796 // We try one round of spinning *before* enqueueing Self. 2797 // 2798 // If the _owner is ready but OFFPROC we could use a YieldTo() 2799 // operation to donate the remainder of this thread's quantum 2800 // to the owner. This has subtle but beneficial affinity 2801 // effects. 2802 2803 if (TrySpin (Self) > 0) { 2804 assert (_owner == Self , "invariant") ; 2805 assert (_succ != Self , "invariant") ; 2806 assert (_Responsible != Self , "invariant") ; 2807 return ; 2808 } 2809 2810 // The Spin failed -- Enqueue and park the thread ... 2811 assert (_succ != Self , "invariant") ; 2812 assert (_owner != Self , "invariant") ; 2813 assert (_Responsible != Self , "invariant") ; 2814 2815 // Enqueue "Self" on ObjectMonitor's _cxq. 2816 // 2817 // Node acts as a proxy for Self. 2818 // As an aside, if were to ever rewrite the synchronization code mostly 2819 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 2820 // Java objects. This would avoid awkward lifecycle and liveness issues, 2821 // as well as eliminate a subset of ABA issues. 2822 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 2823 // 2824 2825 ObjectWaiter node(Self) ; 2826 Self->_ParkEvent->reset() ; 2827 node._prev = (ObjectWaiter *) 0xBAD ; 2828 node.TState = ObjectWaiter::TS_CXQ ; 2829 2830 // Push "Self" onto the front of the _cxq. 2831 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 2832 // Note that spinning tends to reduce the rate at which threads 2833 // enqueue and dequeue on EntryList|cxq. 2834 ObjectWaiter * nxt ; 2835 for (;;) { 2836 node._next = nxt = _cxq ; 2837 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; 2838 2839 // Interference - the CAS failed because _cxq changed. Just retry. 2840 // As an optional optimization we retry the lock. 2841 if (TryLock (Self) > 0) { 2842 assert (_succ != Self , "invariant") ; 2843 assert (_owner == Self , "invariant") ; 2844 assert (_Responsible != Self , "invariant") ; 2845 return ; 2846 } 2847 } 2848 2849 // Check for cxq|EntryList edge transition to non-null. This indicates 2850 // the onset of contention. While contention persists exiting threads 2851 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 2852 // operations revert to the faster 1-0 mode. This enter operation may interleave 2853 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 2854 // arrange for one of the contending thread to use a timed park() operations 2855 // to detect and recover from the race. (Stranding is form of progress failure 2856 // where the monitor is unlocked but all the contending threads remain parked). 2857 // That is, at least one of the contended threads will periodically poll _owner. 2858 // One of the contending threads will become the designated "Responsible" thread. 2859 // The Responsible thread uses a timed park instead of a normal indefinite park 2860 // operation -- it periodically wakes and checks for and recovers from potential 2861 // strandings admitted by 1-0 exit operations. We need at most one Responsible 2862 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 2863 // be responsible for a monitor. 2864 // 2865 // Currently, one of the contended threads takes on the added role of "Responsible". 2866 // A viable alternative would be to use a dedicated "stranding checker" thread 2867 // that periodically iterated over all the threads (or active monitors) and unparked 2868 // successors where there was risk of stranding. This would help eliminate the 2869 // timer scalability issues we see on some platforms as we'd only have one thread 2870 // -- the checker -- parked on a timer. 2871 2872 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 2873 // Try to assume the role of responsible thread for the monitor. 2874 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 2875 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 2876 } 2877 2878 // The lock have been released while this thread was occupied queueing 2879 // itself onto _cxq. To close the race and avoid "stranding" and 2880 // progress-liveness failure we must resample-retry _owner before parking. 2881 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 2882 // In this case the ST-MEMBAR is accomplished with CAS(). 2883 // 2884 // TODO: Defer all thread state transitions until park-time. 2885 // Since state transitions are heavy and inefficient we'd like 2886 // to defer the state transitions until absolutely necessary, 2887 // and in doing so avoid some transitions ... 2888 2889 TEVENT (Inflated enter - Contention) ; 2890 int nWakeups = 0 ; 2891 int RecheckInterval = 1 ; 2892 2893 for (;;) { 2894 2895 if (TryLock (Self) > 0) break ; 2896 assert (_owner != Self, "invariant") ; 2897 2898 if ((SyncFlags & 2) && _Responsible == NULL) { 2899 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; 2900 } 2901 2902 // park self 2903 if (_Responsible == Self || (SyncFlags & 1)) { 2904 TEVENT (Inflated enter - park TIMED) ; 2905 Self->_ParkEvent->park ((jlong) RecheckInterval) ; 2906 // Increase the RecheckInterval, but clamp the value. 2907 RecheckInterval *= 8 ; 2908 if (RecheckInterval > 1000) RecheckInterval = 1000 ; 2909 } else { 2910 TEVENT (Inflated enter - park UNTIMED) ; 2911 Self->_ParkEvent->park() ; 2912 } 2913 2914 if (TryLock(Self) > 0) break ; 2915 2916 // The lock is still contested. 2917 // Keep a tally of the # of futile wakeups. 2918 // Note that the counter is not protected by a lock or updated by atomics. 2919 // That is by design - we trade "lossy" counters which are exposed to 2920 // races during updates for a lower probe effect. 2921 TEVENT (Inflated enter - Futile wakeup) ; 2922 if (ObjectSynchronizer::_sync_FutileWakeups != NULL) { 2923 ObjectSynchronizer::_sync_FutileWakeups->inc() ; 2924 } 2925 ++ nWakeups ; 2926 2927 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 2928 // We can defer clearing _succ until after the spin completes 2929 // TrySpin() must tolerate being called with _succ == Self. 2930 // Try yet another round of adaptive spinning. 2931 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; 2932 2933 // We can find that we were unpark()ed and redesignated _succ while 2934 // we were spinning. That's harmless. If we iterate and call park(), 2935 // park() will consume the event and return immediately and we'll 2936 // just spin again. This pattern can repeat, leaving _succ to simply 2937 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 2938 // Alternately, we can sample fired() here, and if set, forgo spinning 2939 // in the next iteration. 2940 2941 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 2942 Self->_ParkEvent->reset() ; 2943 OrderAccess::fence() ; 2944 } 2945 if (_succ == Self) _succ = NULL ; 2946 2947 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 2948 OrderAccess::fence() ; 2949 } 2950 2951 // Egress : 2952 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 2953 // Normally we'll find Self on the EntryList . 2954 // From the perspective of the lock owner (this thread), the 2955 // EntryList is stable and cxq is prepend-only. 2956 // The head of cxq is volatile but the interior is stable. 2957 // In addition, Self.TState is stable. 2958 2959 assert (_owner == Self , "invariant") ; 2960 assert (object() != NULL , "invariant") ; 2961 // I'd like to write: 2962 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 2963 // but as we're at a safepoint that's not safe. 2964 2965 UnlinkAfterAcquire (Self, &node) ; 2966 if (_succ == Self) _succ = NULL ; 2967 2968 assert (_succ != Self, "invariant") ; 2969 if (_Responsible == Self) { 2970 _Responsible = NULL ; 2971 // Dekker pivot-point. 2972 // Consider OrderAccess::storeload() here 2973 2974 // We may leave threads on cxq|EntryList without a designated 2975 // "Responsible" thread. This is benign. When this thread subsequently 2976 // exits the monitor it can "see" such preexisting "old" threads -- 2977 // threads that arrived on the cxq|EntryList before the fence, above -- 2978 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 2979 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 2980 // non-null and elect a new "Responsible" timer thread. 2981 // 2982 // This thread executes: 2983 // ST Responsible=null; MEMBAR (in enter epilog - here) 2984 // LD cxq|EntryList (in subsequent exit) 2985 // 2986 // Entering threads in the slow/contended path execute: 2987 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 2988 // The (ST cxq; MEMBAR) is accomplished with CAS(). 2989 // 2990 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 2991 // exit operation from floating above the ST Responsible=null. 2992 // 2993 // In *practice* however, EnterI() is always followed by some atomic 2994 // operation such as the decrement of _count in ::enter(). Those atomics 2995 // obviate the need for the explicit MEMBAR, above. 2996 } 2997 2998 // We've acquired ownership with CAS(). 2999 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 3000 // But since the CAS() this thread may have also stored into _succ, 3001 // EntryList, cxq or Responsible. These meta-data updates must be 3002 // visible __before this thread subsequently drops the lock. 3003 // Consider what could occur if we didn't enforce this constraint -- 3004 // STs to monitor meta-data and user-data could reorder with (become 3005 // visible after) the ST in exit that drops ownership of the lock. 3006 // Some other thread could then acquire the lock, but observe inconsistent 3007 // or old monitor meta-data and heap data. That violates the JMM. 3008 // To that end, the 1-0 exit() operation must have at least STST|LDST 3009 // "release" barrier semantics. Specifically, there must be at least a 3010 // STST|LDST barrier in exit() before the ST of null into _owner that drops 3011 // the lock. The barrier ensures that changes to monitor meta-data and data 3012 // protected by the lock will be visible before we release the lock, and 3013 // therefore before some other thread (CPU) has a chance to acquire the lock. 3014 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 3015 // 3016 // Critically, any prior STs to _succ or EntryList must be visible before 3017 // the ST of null into _owner in the *subsequent* (following) corresponding 3018 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 3019 // execute a serializing instruction. 3020 3021 if (SyncFlags & 8) { 3022 OrderAccess::fence() ; 3023 } 3024 return ; 3025} 3026 3027// ExitSuspendEquivalent: 3028// A faster alternate to handle_special_suspend_equivalent_condition() 3029// 3030// handle_special_suspend_equivalent_condition() unconditionally 3031// acquires the SR_lock. On some platforms uncontended MutexLocker() 3032// operations have high latency. Note that in ::enter() we call HSSEC 3033// while holding the monitor, so we effectively lengthen the critical sections. 3034// 3035// There are a number of possible solutions: 3036// 3037// A. To ameliorate the problem we might also defer state transitions 3038// to as late as possible -- just prior to parking. 3039// Given that, we'd call HSSEC after having returned from park(), 3040// but before attempting to acquire the monitor. This is only a 3041// partial solution. It avoids calling HSSEC while holding the 3042// monitor (good), but it still increases successor reacquisition latency -- 3043// the interval between unparking a successor and the time the successor 3044// resumes and retries the lock. See ReenterI(), which defers state transitions. 3045// If we use this technique we can also avoid EnterI()-exit() loop 3046// in ::enter() where we iteratively drop the lock and then attempt 3047// to reacquire it after suspending. 3048// 3049// B. In the future we might fold all the suspend bits into a 3050// composite per-thread suspend flag and then update it with CAS(). 3051// Alternately, a Dekker-like mechanism with multiple variables 3052// would suffice: 3053// ST Self->_suspend_equivalent = false 3054// MEMBAR 3055// LD Self_>_suspend_flags 3056// 3057 3058 3059bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 3060 int Mode = Knob_FastHSSEC ; 3061 if (Mode && !jSelf->is_external_suspend()) { 3062 assert (jSelf->is_suspend_equivalent(), "invariant") ; 3063 jSelf->clear_suspend_equivalent() ; 3064 if (2 == Mode) OrderAccess::storeload() ; 3065 if (!jSelf->is_external_suspend()) return false ; 3066 // We raced a suspension -- fall thru into the slow path 3067 TEVENT (ExitSuspendEquivalent - raced) ; 3068 jSelf->set_suspend_equivalent() ; 3069 } 3070 return jSelf->handle_special_suspend_equivalent_condition() ; 3071} 3072 3073 3074// ReenterI() is a specialized inline form of the latter half of the 3075// contended slow-path from EnterI(). We use ReenterI() only for 3076// monitor reentry in wait(). 3077// 3078// In the future we should reconcile EnterI() and ReenterI(), adding 3079// Knob_Reset and Knob_SpinAfterFutile support and restructuring the 3080// loop accordingly. 3081 3082void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 3083 assert (Self != NULL , "invariant") ; 3084 assert (SelfNode != NULL , "invariant") ; 3085 assert (SelfNode->_thread == Self , "invariant") ; 3086 assert (_waiters > 0 , "invariant") ; 3087 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; 3088 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 3089 JavaThread * jt = (JavaThread *) Self ; 3090 3091 int nWakeups = 0 ; 3092 for (;;) { 3093 ObjectWaiter::TStates v = SelfNode->TState ; 3094 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 3095 assert (_owner != Self, "invariant") ; 3096 3097 if (TryLock (Self) > 0) break ; 3098 if (TrySpin (Self) > 0) break ; 3099 3100 TEVENT (Wait Reentry - parking) ; 3101 3102 // State transition wrappers around park() ... 3103 // ReenterI() wisely defers state transitions until 3104 // it's clear we must park the thread. 3105 { 3106 OSThreadContendState osts(Self->osthread()); 3107 ThreadBlockInVM tbivm(jt); 3108 3109 // cleared by handle_special_suspend_equivalent_condition() 3110 // or java_suspend_self() 3111 jt->set_suspend_equivalent(); 3112 if (SyncFlags & 1) { 3113 Self->_ParkEvent->park ((jlong)1000) ; 3114 } else { 3115 Self->_ParkEvent->park () ; 3116 } 3117 3118 // were we externally suspended while we were waiting? 3119 for (;;) { 3120 if (!ExitSuspendEquivalent (jt)) break ; 3121 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 3122 jt->java_suspend_self(); 3123 jt->set_suspend_equivalent(); 3124 } 3125 } 3126 3127 // Try again, but just so we distinguish between futile wakeups and 3128 // successful wakeups. The following test isn't algorithmically 3129 // necessary, but it helps us maintain sensible statistics. 3130 if (TryLock(Self) > 0) break ; 3131 3132 // The lock is still contested. 3133 // Keep a tally of the # of futile wakeups. 3134 // Note that the counter is not protected by a lock or updated by atomics. 3135 // That is by design - we trade "lossy" counters which are exposed to 3136 // races during updates for a lower probe effect. 3137 TEVENT (Wait Reentry - futile wakeup) ; 3138 ++ nWakeups ; 3139 3140 // Assuming this is not a spurious wakeup we'll normally 3141 // find that _succ == Self. 3142 if (_succ == Self) _succ = NULL ; 3143 3144 // Invariant: after clearing _succ a contending thread 3145 // *must* retry _owner before parking. 3146 OrderAccess::fence() ; 3147 3148 if (ObjectSynchronizer::_sync_FutileWakeups != NULL) { 3149 ObjectSynchronizer::_sync_FutileWakeups->inc() ; 3150 } 3151 } 3152 3153 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 3154 // Normally we'll find Self on the EntryList. 3155 // Unlinking from the EntryList is constant-time and atomic-free. 3156 // From the perspective of the lock owner (this thread), the 3157 // EntryList is stable and cxq is prepend-only. 3158 // The head of cxq is volatile but the interior is stable. 3159 // In addition, Self.TState is stable. 3160 3161 assert (_owner == Self, "invariant") ; 3162 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 3163 UnlinkAfterAcquire (Self, SelfNode) ; 3164 if (_succ == Self) _succ = NULL ; 3165 assert (_succ != Self, "invariant") ; 3166 SelfNode->TState = ObjectWaiter::TS_RUN ; 3167 OrderAccess::fence() ; // see comments at the end of EnterI() 3168} 3169 3170bool ObjectMonitor::try_enter(Thread* THREAD) { 3171 if (THREAD != _owner) { 3172 if (THREAD->is_lock_owned ((address)_owner)) { 3173 assert(_recursions == 0, "internal state error"); 3174 _owner = THREAD ; 3175 _recursions = 1 ; 3176 OwnerIsThread = 1 ; 3177 return true; 3178 } 3179 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 3180 return false; 3181 } 3182 return true; 3183 } else { 3184 _recursions++; 3185 return true; 3186 } 3187} 3188 3189void ATTR ObjectMonitor::enter(TRAPS) { 3190 // The following code is ordered to check the most common cases first 3191 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 3192 Thread * const Self = THREAD ; 3193 void * cur ; 3194 3195 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; 3196 if (cur == NULL) { 3197 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 3198 assert (_recursions == 0 , "invariant") ; 3199 assert (_owner == Self, "invariant") ; 3200 // CONSIDER: set or assert OwnerIsThread == 1 3201 return ; 3202 } 3203 3204 if (cur == Self) { 3205 // TODO-FIXME: check for integer overflow! BUGID 6557169. 3206 _recursions ++ ; 3207 return ; 3208 } 3209 3210 if (Self->is_lock_owned ((address)cur)) { 3211 assert (_recursions == 0, "internal state error"); 3212 _recursions = 1 ; 3213 // Commute owner from a thread-specific on-stack BasicLockObject address to 3214 // a full-fledged "Thread *". 3215 _owner = Self ; 3216 OwnerIsThread = 1 ; 3217 return ; 3218 } 3219 3220 // We've encountered genuine contention. 3221 assert (Self->_Stalled == 0, "invariant") ; 3222 Self->_Stalled = intptr_t(this) ; 3223 3224 // Try one round of spinning *before* enqueueing Self 3225 // and before going through the awkward and expensive state 3226 // transitions. The following spin is strictly optional ... 3227 // Note that if we acquire the monitor from an initial spin 3228 // we forgo posting JVMTI events and firing DTRACE probes. 3229 if (Knob_SpinEarly && TrySpin (Self) > 0) { 3230 assert (_owner == Self , "invariant") ; 3231 assert (_recursions == 0 , "invariant") ; 3232 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 3233 Self->_Stalled = 0 ; 3234 return ; 3235 } 3236 3237 assert (_owner != Self , "invariant") ; 3238 assert (_succ != Self , "invariant") ; 3239 assert (Self->is_Java_thread() , "invariant") ; 3240 JavaThread * jt = (JavaThread *) Self ; 3241 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 3242 assert (jt->thread_state() != _thread_blocked , "invariant") ; 3243 assert (this->object() != NULL , "invariant") ; 3244 assert (_count >= 0, "invariant") ; 3245 3246 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 3247 // Ensure the object-monitor relationship remains stable while there's contention. 3248 Atomic::inc_ptr(&_count); 3249 3250 { // Change java thread status to indicate blocked on monitor enter. 3251 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 3252 3253 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 3254 if (JvmtiExport::should_post_monitor_contended_enter()) { 3255 JvmtiExport::post_monitor_contended_enter(jt, this); 3256 } 3257 3258 OSThreadContendState osts(Self->osthread()); 3259 ThreadBlockInVM tbivm(jt); 3260 3261 Self->set_current_pending_monitor(this); 3262 3263 // TODO-FIXME: change the following for(;;) loop to straight-line code. 3264 for (;;) { 3265 jt->set_suspend_equivalent(); 3266 // cleared by handle_special_suspend_equivalent_condition() 3267 // or java_suspend_self() 3268 3269 EnterI (THREAD) ; 3270 3271 if (!ExitSuspendEquivalent(jt)) break ; 3272 3273 // 3274 // We have acquired the contended monitor, but while we were 3275 // waiting another thread suspended us. We don't want to enter 3276 // the monitor while suspended because that would surprise the 3277 // thread that suspended us. 3278 // 3279 _recursions = 0 ; 3280 _succ = NULL ; 3281 exit (Self) ; 3282 3283 jt->java_suspend_self(); 3284 } 3285 Self->set_current_pending_monitor(NULL); 3286 } 3287 3288 Atomic::dec_ptr(&_count); 3289 assert (_count >= 0, "invariant") ; 3290 Self->_Stalled = 0 ; 3291 3292 // Must either set _recursions = 0 or ASSERT _recursions == 0. 3293 assert (_recursions == 0 , "invariant") ; 3294 assert (_owner == Self , "invariant") ; 3295 assert (_succ != Self , "invariant") ; 3296 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 3297 3298 // The thread -- now the owner -- is back in vm mode. 3299 // Report the glorious news via TI,DTrace and jvmstat. 3300 // The probe effect is non-trivial. All the reportage occurs 3301 // while we hold the monitor, increasing the length of the critical 3302 // section. Amdahl's parallel speedup law comes vividly into play. 3303 // 3304 // Another option might be to aggregate the events (thread local or 3305 // per-monitor aggregation) and defer reporting until a more opportune 3306 // time -- such as next time some thread encounters contention but has 3307 // yet to acquire the lock. While spinning that thread could 3308 // spinning we could increment JVMStat counters, etc. 3309 3310 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 3311 if (JvmtiExport::should_post_monitor_contended_entered()) { 3312 JvmtiExport::post_monitor_contended_entered(jt, this); 3313 } 3314 if (ObjectSynchronizer::_sync_ContendedLockAttempts != NULL) { 3315 ObjectSynchronizer::_sync_ContendedLockAttempts->inc() ; 3316 } 3317} 3318 3319void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 3320 assert (_owner == Self, "invariant") ; 3321 3322 // Exit protocol: 3323 // 1. ST _succ = wakee 3324 // 2. membar #loadstore|#storestore; 3325 // 2. ST _owner = NULL 3326 // 3. unpark(wakee) 3327 3328 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; 3329 ParkEvent * Trigger = Wakee->_event ; 3330 3331 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 3332 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 3333 // out-of-scope (non-extant). 3334 Wakee = NULL ; 3335 3336 // Drop the lock 3337 OrderAccess::release_store_ptr (&_owner, NULL) ; 3338 OrderAccess::fence() ; // ST _owner vs LD in unpark() 3339 3340 // TODO-FIXME: 3341 // If there's a safepoint pending the best policy would be to 3342 // get _this thread to a safepoint and only wake the successor 3343 // after the safepoint completed. monitorexit uses a "leaf" 3344 // state transition, however, so this thread can't become 3345 // safe at this point in time. (Its stack isn't walkable). 3346 // The next best thing is to defer waking the successor by 3347 // adding to a list of thread to be unparked after at the 3348 // end of the forthcoming STW). 3349 if (SafepointSynchronize::do_call_back()) { 3350 TEVENT (unpark before SAFEPOINT) ; 3351 } 3352 3353 // Possible optimizations ... 3354 // 3355 // * Consider: set Wakee->UnparkTime = timeNow() 3356 // When the thread wakes up it'll compute (timeNow() - Self->UnparkTime()). 3357 // By measuring recent ONPROC latency we can approximate the 3358 // system load. In turn, we can feed that information back 3359 // into the spinning & succession policies. 3360 // (ONPROC latency correlates strongly with load). 3361 // 3362 // * Pull affinity: 3363 // If the wakee is cold then transiently setting it's affinity 3364 // to the current CPU is a good idea. 3365 // See http://j2se.east/~dice/PERSIST/050624-PullAffinity.txt 3366 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 3367 Trigger->unpark() ; 3368 3369 // Maintain stats and report events to JVMTI 3370 if (ObjectSynchronizer::_sync_Parks != NULL) { 3371 ObjectSynchronizer::_sync_Parks->inc() ; 3372 } 3373} 3374 3375 3376// exit() 3377// ~~~~~~ 3378// Note that the collector can't reclaim the objectMonitor or deflate 3379// the object out from underneath the thread calling ::exit() as the 3380// thread calling ::exit() never transitions to a stable state. 3381// This inhibits GC, which in turn inhibits asynchronous (and 3382// inopportune) reclamation of "this". 3383// 3384// We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 3385// There's one exception to the claim above, however. EnterI() can call 3386// exit() to drop a lock if the acquirer has been externally suspended. 3387// In that case exit() is called with _thread_state as _thread_blocked, 3388// but the monitor's _count field is > 0, which inhibits reclamation. 3389// 3390// 1-0 exit 3391// ~~~~~~~~ 3392// ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 3393// the fast-path operators have been optimized so the common ::exit() 3394// operation is 1-0. See i486.ad fast_unlock(), for instance. 3395// The code emitted by fast_unlock() elides the usual MEMBAR. This 3396// greatly improves latency -- MEMBAR and CAS having considerable local 3397// latency on modern processors -- but at the cost of "stranding". Absent the 3398// MEMBAR, a thread in fast_unlock() can race a thread in the slow 3399// ::enter() path, resulting in the entering thread being stranding 3400// and a progress-liveness failure. Stranding is extremely rare. 3401// We use timers (timed park operations) & periodic polling to detect 3402// and recover from stranding. Potentially stranded threads periodically 3403// wake up and poll the lock. See the usage of the _Responsible variable. 3404// 3405// The CAS() in enter provides for safety and exclusion, while the CAS or 3406// MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 3407// eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 3408// We detect and recover from stranding with timers. 3409// 3410// If a thread transiently strands it'll park until (a) another 3411// thread acquires the lock and then drops the lock, at which time the 3412// exiting thread will notice and unpark the stranded thread, or, (b) 3413// the timer expires. If the lock is high traffic then the stranding latency 3414// will be low due to (a). If the lock is low traffic then the odds of 3415// stranding are lower, although the worst-case stranding latency 3416// is longer. Critically, we don't want to put excessive load in the 3417// platform's timer subsystem. We want to minimize both the timer injection 3418// rate (timers created/sec) as well as the number of timers active at 3419// any one time. (more precisely, we want to minimize timer-seconds, which is 3420// the integral of the # of active timers at any instant over time). 3421// Both impinge on OS scalability. Given that, at most one thread parked on 3422// a monitor will use a timer. 3423 3424void ATTR ObjectMonitor::exit(TRAPS) { 3425 Thread * Self = THREAD ; 3426 if (THREAD != _owner) { 3427 if (THREAD->is_lock_owned((address) _owner)) { 3428 // Transmute _owner from a BasicLock pointer to a Thread address. 3429 // We don't need to hold _mutex for this transition. 3430 // Non-null to Non-null is safe as long as all readers can 3431 // tolerate either flavor. 3432 assert (_recursions == 0, "invariant") ; 3433 _owner = THREAD ; 3434 _recursions = 0 ; 3435 OwnerIsThread = 1 ; 3436 } else { 3437 // NOTE: we need to handle unbalanced monitor enter/exit 3438 // in native code by throwing an exception. 3439 // TODO: Throw an IllegalMonitorStateException ? 3440 TEVENT (Exit - Throw IMSX) ; 3441 assert(false, "Non-balanced monitor enter/exit!"); 3442 if (false) { 3443 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 3444 } 3445 return; 3446 } 3447 } 3448 3449 if (_recursions != 0) { 3450 _recursions--; // this is simple recursive enter 3451 TEVENT (Inflated exit - recursive) ; 3452 return ; 3453 } 3454 3455 // Invariant: after setting Responsible=null an thread must execute 3456 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 3457 if ((SyncFlags & 4) == 0) { 3458 _Responsible = NULL ; 3459 } 3460 3461 for (;;) { 3462 assert (THREAD == _owner, "invariant") ; 3463 3464 // Fast-path monitor exit: 3465 // 3466 // Observe the Dekker/Lamport duality: 3467 // A thread in ::exit() executes: 3468 // ST Owner=null; MEMBAR; LD EntryList|cxq. 3469 // A thread in the contended ::enter() path executes the complementary: 3470 // ST EntryList|cxq = nonnull; MEMBAR; LD Owner. 3471 // 3472 // Note that there's a benign race in the exit path. We can drop the 3473 // lock, another thread can reacquire the lock immediately, and we can 3474 // then wake a thread unnecessarily (yet another flavor of futile wakeup). 3475 // This is benign, and we've structured the code so the windows are short 3476 // and the frequency of such futile wakeups is low. 3477 // 3478 // We could eliminate the race by encoding both the "LOCKED" state and 3479 // the queue head in a single word. Exit would then use either CAS to 3480 // clear the LOCKED bit/byte. This precludes the desirable 1-0 optimization, 3481 // however. 3482 // 3483 // Possible fast-path ::exit() optimization: 3484 // The current fast-path exit implementation fetches both cxq and EntryList. 3485 // See also i486.ad fast_unlock(). Testing has shown that two LDs 3486 // isn't measurably slower than a single LD on any platforms. 3487 // Still, we could reduce the 2 LDs to one or zero by one of the following: 3488 // 3489 // - Use _count instead of cxq|EntryList 3490 // We intend to eliminate _count, however, when we switch 3491 // to on-the-fly deflation in ::exit() as is used in 3492 // Metalocks and RelaxedLocks. 3493 // 3494 // - Establish the invariant that cxq == null implies EntryList == null. 3495 // set cxq == EMPTY (1) to encode the state where cxq is empty 3496 // by EntryList != null. EMPTY is a distinguished value. 3497 // The fast-path exit() would fetch cxq but not EntryList. 3498 // 3499 // - Encode succ as follows: 3500 // succ = t : Thread t is the successor -- t is ready or is spinning. 3501 // Exiting thread does not need to wake a successor. 3502 // succ = 0 : No successor required -> (EntryList|cxq) == null 3503 // Exiting thread does not need to wake a successor 3504 // succ = 1 : Successor required -> (EntryList|cxq) != null and 3505 // logically succ == null. 3506 // Exiting thread must wake a successor. 3507 // 3508 // The 1-1 fast-exit path would appear as : 3509 // _owner = null ; membar ; 3510 // if (_succ == 1 && CAS (&_owner, null, Self) == null) goto SlowPath 3511 // goto FastPathDone ; 3512 // 3513 // and the 1-0 fast-exit path would appear as: 3514 // if (_succ == 1) goto SlowPath 3515 // Owner = null ; 3516 // goto FastPathDone 3517 // 3518 // - Encode the LSB of _owner as 1 to indicate that exit() 3519 // must use the slow-path and make a successor ready. 3520 // (_owner & 1) == 0 IFF succ != null || (EntryList|cxq) == null 3521 // (_owner & 1) == 0 IFF succ == null && (EntryList|cxq) != null (obviously) 3522 // The 1-0 fast exit path would read: 3523 // if (_owner != Self) goto SlowPath 3524 // _owner = null 3525 // goto FastPathDone 3526 3527 if (Knob_ExitPolicy == 0) { 3528 // release semantics: prior loads and stores from within the critical section 3529 // must not float (reorder) past the following store that drops the lock. 3530 // On SPARC that requires MEMBAR #loadstore|#storestore. 3531 // But of course in TSO #loadstore|#storestore is not required. 3532 // I'd like to write one of the following: 3533 // A. OrderAccess::release() ; _owner = NULL 3534 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 3535 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 3536 // store into a _dummy variable. That store is not needed, but can result 3537 // in massive wasteful coherency traffic on classic SMP systems. 3538 // Instead, I use release_store(), which is implemented as just a simple 3539 // ST on x64, x86 and SPARC. 3540 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 3541 OrderAccess::storeload() ; // See if we need to wake a successor 3542 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 3543 TEVENT (Inflated exit - simple egress) ; 3544 return ; 3545 } 3546 TEVENT (Inflated exit - complex egress) ; 3547 3548 // Normally the exiting thread is responsible for ensuring succession, 3549 // but if other successors are ready or other entering threads are spinning 3550 // then this thread can simply store NULL into _owner and exit without 3551 // waking a successor. The existence of spinners or ready successors 3552 // guarantees proper succession (liveness). Responsibility passes to the 3553 // ready or running successors. The exiting thread delegates the duty. 3554 // More precisely, if a successor already exists this thread is absolved 3555 // of the responsibility of waking (unparking) one. 3556 // 3557 // The _succ variable is critical to reducing futile wakeup frequency. 3558 // _succ identifies the "heir presumptive" thread that has been made 3559 // ready (unparked) but that has not yet run. We need only one such 3560 // successor thread to guarantee progress. 3561 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 3562 // section 3.3 "Futile Wakeup Throttling" for details. 3563 // 3564 // Note that spinners in Enter() also set _succ non-null. 3565 // In the current implementation spinners opportunistically set 3566 // _succ so that exiting threads might avoid waking a successor. 3567 // Another less appealing alternative would be for the exiting thread 3568 // to drop the lock and then spin briefly to see if a spinner managed 3569 // to acquire the lock. If so, the exiting thread could exit 3570 // immediately without waking a successor, otherwise the exiting 3571 // thread would need to dequeue and wake a successor. 3572 // (Note that we'd need to make the post-drop spin short, but no 3573 // shorter than the worst-case round-trip cache-line migration time. 3574 // The dropped lock needs to become visible to the spinner, and then 3575 // the acquisition of the lock by the spinner must become visible to 3576 // the exiting thread). 3577 // 3578 3579 // It appears that an heir-presumptive (successor) must be made ready. 3580 // Only the current lock owner can manipulate the EntryList or 3581 // drain _cxq, so we need to reacquire the lock. If we fail 3582 // to reacquire the lock the responsibility for ensuring succession 3583 // falls to the new owner. 3584 // 3585 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 3586 return ; 3587 } 3588 TEVENT (Exit - Reacquired) ; 3589 } else { 3590 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 3591 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock 3592 OrderAccess::storeload() ; 3593 // Ratify the previously observed values. 3594 if (_cxq == NULL || _succ != NULL) { 3595 TEVENT (Inflated exit - simple egress) ; 3596 return ; 3597 } 3598 3599 // inopportune interleaving -- the exiting thread (this thread) 3600 // in the fast-exit path raced an entering thread in the slow-enter 3601 // path. 3602 // We have two choices: 3603 // A. Try to reacquire the lock. 3604 // If the CAS() fails return immediately, otherwise 3605 // we either restart/rerun the exit operation, or simply 3606 // fall-through into the code below which wakes a successor. 3607 // B. If the elements forming the EntryList|cxq are TSM 3608 // we could simply unpark() the lead thread and return 3609 // without having set _succ. 3610 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 3611 TEVENT (Inflated exit - reacquired succeeded) ; 3612 return ; 3613 } 3614 TEVENT (Inflated exit - reacquired failed) ; 3615 } else { 3616 TEVENT (Inflated exit - complex egress) ; 3617 } 3618 } 3619 3620 guarantee (_owner == THREAD, "invariant") ; 3621 3622 // Select an appropriate successor ("heir presumptive") from the EntryList 3623 // and make it ready. Generally we just wake the head of EntryList . 3624 // There's no algorithmic constraint that we use the head - it's just 3625 // a policy decision. Note that the thread at head of the EntryList 3626 // remains at the head until it acquires the lock. This means we'll 3627 // repeatedly wake the same thread until it manages to grab the lock. 3628 // This is generally a good policy - if we're seeing lots of futile wakeups 3629 // at least we're waking/rewaking a thread that's like to be hot or warm 3630 // (have residual D$ and TLB affinity). 3631 // 3632 // "Wakeup locality" optimization: 3633 // http://j2se.east/~dice/PERSIST/040825-WakeLocality.txt 3634 // In the future we'll try to bias the selection mechanism 3635 // to preferentially pick a thread that recently ran on 3636 // a processor element that shares cache with the CPU on which 3637 // the exiting thread is running. We need access to Solaris' 3638 // schedctl.sc_cpu to make that work. 3639 // 3640 ObjectWaiter * w = NULL ; 3641 int QMode = Knob_QMode ; 3642 3643 if (QMode == 2 && _cxq != NULL) { 3644 // QMode == 2 : cxq has precedence over EntryList. 3645 // Try to directly wake a successor from the cxq. 3646 // If successful, the successor will need to unlink itself from cxq. 3647 w = _cxq ; 3648 assert (w != NULL, "invariant") ; 3649 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 3650 ExitEpilog (Self, w) ; 3651 return ; 3652 } 3653 3654 if (QMode == 3 && _cxq != NULL) { 3655 // Aggressively drain cxq into EntryList at the first opportunity. 3656 // This policy ensure that recently-run threads live at the head of EntryList. 3657 // Drain _cxq into EntryList - bulk transfer. 3658 // First, detach _cxq. 3659 // The following loop is tantamount to: w = swap (&cxq, NULL) 3660 w = _cxq ; 3661 for (;;) { 3662 assert (w != NULL, "Invariant") ; 3663 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 3664 if (u == w) break ; 3665 w = u ; 3666 } 3667 assert (w != NULL , "invariant") ; 3668 3669 ObjectWaiter * q = NULL ; 3670 ObjectWaiter * p ; 3671 for (p = w ; p != NULL ; p = p->_next) { 3672 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 3673 p->TState = ObjectWaiter::TS_ENTER ; 3674 p->_prev = q ; 3675 q = p ; 3676 } 3677 3678 // Append the RATs to the EntryList 3679 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 3680 ObjectWaiter * Tail ; 3681 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; 3682 if (Tail == NULL) { 3683 _EntryList = w ; 3684 } else { 3685 Tail->_next = w ; 3686 w->_prev = Tail ; 3687 } 3688 3689 // Fall thru into code that tries to wake a successor from EntryList 3690 } 3691 3692 if (QMode == 4 && _cxq != NULL) { 3693 // Aggressively drain cxq into EntryList at the first opportunity. 3694 // This policy ensure that recently-run threads live at the head of EntryList. 3695 3696 // Drain _cxq into EntryList - bulk transfer. 3697 // First, detach _cxq. 3698 // The following loop is tantamount to: w = swap (&cxq, NULL) 3699 w = _cxq ; 3700 for (;;) { 3701 assert (w != NULL, "Invariant") ; 3702 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 3703 if (u == w) break ; 3704 w = u ; 3705 } 3706 assert (w != NULL , "invariant") ; 3707 3708 ObjectWaiter * q = NULL ; 3709 ObjectWaiter * p ; 3710 for (p = w ; p != NULL ; p = p->_next) { 3711 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 3712 p->TState = ObjectWaiter::TS_ENTER ; 3713 p->_prev = q ; 3714 q = p ; 3715 } 3716 3717 // Prepend the RATs to the EntryList 3718 if (_EntryList != NULL) { 3719 q->_next = _EntryList ; 3720 _EntryList->_prev = q ; 3721 } 3722 _EntryList = w ; 3723 3724 // Fall thru into code that tries to wake a successor from EntryList 3725 } 3726 3727 w = _EntryList ; 3728 if (w != NULL) { 3729 // I'd like to write: guarantee (w->_thread != Self). 3730 // But in practice an exiting thread may find itself on the EntryList. 3731 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 3732 // then calls exit(). Exit release the lock by setting O._owner to NULL. 3733 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 3734 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 3735 // release the lock "O". T2 resumes immediately after the ST of null into 3736 // _owner, above. T2 notices that the EntryList is populated, so it 3737 // reacquires the lock and then finds itself on the EntryList. 3738 // Given all that, we have to tolerate the circumstance where "w" is 3739 // associated with Self. 3740 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 3741 ExitEpilog (Self, w) ; 3742 return ; 3743 } 3744 3745 // If we find that both _cxq and EntryList are null then just 3746 // re-run the exit protocol from the top. 3747 w = _cxq ; 3748 if (w == NULL) continue ; 3749 3750 // Drain _cxq into EntryList - bulk transfer. 3751 // First, detach _cxq. 3752 // The following loop is tantamount to: w = swap (&cxq, NULL) 3753 for (;;) { 3754 assert (w != NULL, "Invariant") ; 3755 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; 3756 if (u == w) break ; 3757 w = u ; 3758 } 3759 TEVENT (Inflated exit - drain cxq into EntryList) ; 3760 3761 assert (w != NULL , "invariant") ; 3762 assert (_EntryList == NULL , "invariant") ; 3763 3764 // Convert the LIFO SLL anchored by _cxq into a DLL. 3765 // The list reorganization step operates in O(LENGTH(w)) time. 3766 // It's critical that this step operate quickly as 3767 // "Self" still holds the outer-lock, restricting parallelism 3768 // and effectively lengthening the critical section. 3769 // Invariant: s chases t chases u. 3770 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 3771 // we have faster access to the tail. 3772 3773 if (QMode == 1) { 3774 // QMode == 1 : drain cxq to EntryList, reversing order 3775 // We also reverse the order of the list. 3776 ObjectWaiter * s = NULL ; 3777 ObjectWaiter * t = w ; 3778 ObjectWaiter * u = NULL ; 3779 while (t != NULL) { 3780 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; 3781 t->TState = ObjectWaiter::TS_ENTER ; 3782 u = t->_next ; 3783 t->_prev = u ; 3784 t->_next = s ; 3785 s = t; 3786 t = u ; 3787 } 3788 _EntryList = s ; 3789 assert (s != NULL, "invariant") ; 3790 } else { 3791 // QMode == 0 or QMode == 2 3792 _EntryList = w ; 3793 ObjectWaiter * q = NULL ; 3794 ObjectWaiter * p ; 3795 for (p = w ; p != NULL ; p = p->_next) { 3796 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; 3797 p->TState = ObjectWaiter::TS_ENTER ; 3798 p->_prev = q ; 3799 q = p ; 3800 } 3801 } 3802 3803 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 3804 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 3805 3806 // See if we can abdicate to a spinner instead of waking a thread. 3807 // A primary goal of the implementation is to reduce the 3808 // context-switch rate. 3809 if (_succ != NULL) continue; 3810 3811 w = _EntryList ; 3812 if (w != NULL) { 3813 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; 3814 ExitEpilog (Self, w) ; 3815 return ; 3816 } 3817 } 3818} 3819// complete_exit exits a lock returning recursion count 3820// complete_exit/reenter operate as a wait without waiting 3821// complete_exit requires an inflated monitor 3822// The _owner field is not always the Thread addr even with an 3823// inflated monitor, e.g. the monitor can be inflated by a non-owning 3824// thread due to contention. 3825intptr_t ObjectMonitor::complete_exit(TRAPS) { 3826 Thread * const Self = THREAD; 3827 assert(Self->is_Java_thread(), "Must be Java thread!"); 3828 JavaThread *jt = (JavaThread *)THREAD; 3829 3830 DeferredInitialize(); 3831 3832 if (THREAD != _owner) { 3833 if (THREAD->is_lock_owned ((address)_owner)) { 3834 assert(_recursions == 0, "internal state error"); 3835 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ 3836 _recursions = 0 ; 3837 OwnerIsThread = 1 ; 3838 } 3839 } 3840 3841 guarantee(Self == _owner, "complete_exit not owner"); 3842 intptr_t save = _recursions; // record the old recursion count 3843 _recursions = 0; // set the recursion level to be 0 3844 exit (Self) ; // exit the monitor 3845 guarantee (_owner != Self, "invariant"); 3846 return save; 3847} 3848 3849// reenter() enters a lock and sets recursion count 3850// complete_exit/reenter operate as a wait without waiting 3851void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 3852 Thread * const Self = THREAD; 3853 assert(Self->is_Java_thread(), "Must be Java thread!"); 3854 JavaThread *jt = (JavaThread *)THREAD; 3855 3856 guarantee(_owner != Self, "reenter already owner"); 3857 enter (THREAD); // enter the monitor 3858 guarantee (_recursions == 0, "reenter recursion"); 3859 _recursions = recursions; 3860 return; 3861} 3862 3863// Note: a subset of changes to ObjectMonitor::wait() 3864// will need to be replicated in complete_exit above 3865void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 3866 Thread * const Self = THREAD ; 3867 assert(Self->is_Java_thread(), "Must be Java thread!"); 3868 JavaThread *jt = (JavaThread *)THREAD; 3869 3870 DeferredInitialize () ; 3871 3872 // Throw IMSX or IEX. 3873 CHECK_OWNER(); 3874 3875 // check for a pending interrupt 3876 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 3877 // post monitor waited event. Note that this is past-tense, we are done waiting. 3878 if (JvmtiExport::should_post_monitor_waited()) { 3879 // Note: 'false' parameter is passed here because the 3880 // wait was not timed out due to thread interrupt. 3881 JvmtiExport::post_monitor_waited(jt, this, false); 3882 } 3883 TEVENT (Wait - Throw IEX) ; 3884 THROW(vmSymbols::java_lang_InterruptedException()); 3885 return ; 3886 } 3887 TEVENT (Wait) ; 3888 3889 assert (Self->_Stalled == 0, "invariant") ; 3890 Self->_Stalled = intptr_t(this) ; 3891 jt->set_current_waiting_monitor(this); 3892 3893 // create a node to be put into the queue 3894 // Critically, after we reset() the event but prior to park(), we must check 3895 // for a pending interrupt. 3896 ObjectWaiter node(Self); 3897 node.TState = ObjectWaiter::TS_WAIT ; 3898 Self->_ParkEvent->reset() ; 3899 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 3900 3901 // Enter the waiting queue, which is a circular doubly linked list in this case 3902 // but it could be a priority queue or any data structure. 3903 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 3904 // by the the owner of the monitor *except* in the case where park() 3905 // returns because of a timeout of interrupt. Contention is exceptionally rare 3906 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 3907 3908 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; 3909 AddWaiter (&node) ; 3910 Thread::SpinRelease (&_WaitSetLock) ; 3911 3912 if ((SyncFlags & 4) == 0) { 3913 _Responsible = NULL ; 3914 } 3915 intptr_t save = _recursions; // record the old recursion count 3916 _waiters++; // increment the number of waiters 3917 _recursions = 0; // set the recursion level to be 1 3918 exit (Self) ; // exit the monitor 3919 guarantee (_owner != Self, "invariant") ; 3920 3921 // As soon as the ObjectMonitor's ownership is dropped in the exit() 3922 // call above, another thread can enter() the ObjectMonitor, do the 3923 // notify(), and exit() the ObjectMonitor. If the other thread's 3924 // exit() call chooses this thread as the successor and the unpark() 3925 // call happens to occur while this thread is posting a 3926 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event 3927 // handler using RawMonitors and consuming the unpark(). 3928 // 3929 // To avoid the problem, we re-post the event. This does no harm 3930 // even if the original unpark() was not consumed because we are the 3931 // chosen successor for this monitor. 3932 if (node._notified != 0 && _succ == Self) { 3933 node._event->unpark(); 3934 } 3935 3936 // The thread is on the WaitSet list - now park() it. 3937 // On MP systems it's conceivable that a brief spin before we park 3938 // could be profitable. 3939 // 3940 // TODO-FIXME: change the following logic to a loop of the form 3941 // while (!timeout && !interrupted && _notified == 0) park() 3942 3943 int ret = OS_OK ; 3944 int WasNotified = 0 ; 3945 { // State transition wrappers 3946 OSThread* osthread = Self->osthread(); 3947 OSThreadWaitState osts(osthread, true); 3948 { 3949 ThreadBlockInVM tbivm(jt); 3950 // Thread is in thread_blocked state and oop access is unsafe. 3951 jt->set_suspend_equivalent(); 3952 3953 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 3954 // Intentionally empty 3955 } else 3956 if (node._notified == 0) { 3957 if (millis <= 0) { 3958 Self->_ParkEvent->park () ; 3959 } else { 3960 ret = Self->_ParkEvent->park (millis) ; 3961 } 3962 } 3963 3964 // were we externally suspended while we were waiting? 3965 if (ExitSuspendEquivalent (jt)) { 3966 // TODO-FIXME: add -- if succ == Self then succ = null. 3967 jt->java_suspend_self(); 3968 } 3969 3970 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 3971 3972 3973 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 3974 // from the WaitSet to the EntryList. 3975 // See if we need to remove Node from the WaitSet. 3976 // We use double-checked locking to avoid grabbing _WaitSetLock 3977 // if the thread is not on the wait queue. 3978 // 3979 // Note that we don't need a fence before the fetch of TState. 3980 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 3981 // written by the is thread. (perhaps the fetch might even be satisfied 3982 // by a look-aside into the processor's own store buffer, although given 3983 // the length of the code path between the prior ST and this load that's 3984 // highly unlikely). If the following LD fetches a stale TS_WAIT value 3985 // then we'll acquire the lock and then re-fetch a fresh TState value. 3986 // That is, we fail toward safety. 3987 3988 if (node.TState == ObjectWaiter::TS_WAIT) { 3989 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; 3990 if (node.TState == ObjectWaiter::TS_WAIT) { 3991 DequeueSpecificWaiter (&node) ; // unlink from WaitSet 3992 assert(node._notified == 0, "invariant"); 3993 node.TState = ObjectWaiter::TS_RUN ; 3994 } 3995 Thread::SpinRelease (&_WaitSetLock) ; 3996 } 3997 3998 // The thread is now either on off-list (TS_RUN), 3999 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 4000 // The Node's TState variable is stable from the perspective of this thread. 4001 // No other threads will asynchronously modify TState. 4002 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; 4003 OrderAccess::loadload() ; 4004 if (_succ == Self) _succ = NULL ; 4005 WasNotified = node._notified ; 4006 4007 // Reentry phase -- reacquire the monitor. 4008 // re-enter contended monitor after object.wait(). 4009 // retain OBJECT_WAIT state until re-enter successfully completes 4010 // Thread state is thread_in_vm and oop access is again safe, 4011 // although the raw address of the object may have changed. 4012 // (Don't cache naked oops over safepoints, of course). 4013 4014 // post monitor waited event. Note that this is past-tense, we are done waiting. 4015 if (JvmtiExport::should_post_monitor_waited()) { 4016 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 4017 } 4018 OrderAccess::fence() ; 4019 4020 assert (Self->_Stalled != 0, "invariant") ; 4021 Self->_Stalled = 0 ; 4022 4023 assert (_owner != Self, "invariant") ; 4024 ObjectWaiter::TStates v = node.TState ; 4025 if (v == ObjectWaiter::TS_RUN) { 4026 enter (Self) ; 4027 } else { 4028 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; 4029 ReenterI (Self, &node) ; 4030 node.wait_reenter_end(this); 4031 } 4032 4033 // Self has reacquired the lock. 4034 // Lifecycle - the node representing Self must not appear on any queues. 4035 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 4036 // want residual elements associated with this thread left on any lists. 4037 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; 4038 assert (_owner == Self, "invariant") ; 4039 assert (_succ != Self , "invariant") ; 4040 } // OSThreadWaitState() 4041 4042 jt->set_current_waiting_monitor(NULL); 4043 4044 guarantee (_recursions == 0, "invariant") ; 4045 _recursions = save; // restore the old recursion count 4046 _waiters--; // decrement the number of waiters 4047 4048 // Verify a few postconditions 4049 assert (_owner == Self , "invariant") ; 4050 assert (_succ != Self , "invariant") ; 4051 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 4052 4053 if (SyncFlags & 32) { 4054 OrderAccess::fence() ; 4055 } 4056 4057 // check if the notification happened 4058 if (!WasNotified) { 4059 // no, it could be timeout or Thread.interrupt() or both 4060 // check for interrupt event, otherwise it is timeout 4061 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 4062 TEVENT (Wait - throw IEX from epilog) ; 4063 THROW(vmSymbols::java_lang_InterruptedException()); 4064 } 4065 } 4066 4067 // NOTE: Spurious wake up will be consider as timeout. 4068 // Monitor notify has precedence over thread interrupt. 4069} 4070 4071 4072// Consider: 4073// If the lock is cool (cxq == null && succ == null) and we're on an MP system 4074// then instead of transferring a thread from the WaitSet to the EntryList 4075// we might just dequeue a thread from the WaitSet and directly unpark() it. 4076 4077void ObjectMonitor::notify(TRAPS) { 4078 CHECK_OWNER(); 4079 if (_WaitSet == NULL) { 4080 TEVENT (Empty-Notify) ; 4081 return ; 4082 } 4083 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 4084 4085 int Policy = Knob_MoveNotifyee ; 4086 4087 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; 4088 ObjectWaiter * iterator = DequeueWaiter() ; 4089 if (iterator != NULL) { 4090 TEVENT (Notify1 - Transfer) ; 4091 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 4092 guarantee (iterator->_notified == 0, "invariant") ; 4093 // Disposition - what might we do with iterator ? 4094 // a. add it directly to the EntryList - either tail or head. 4095 // b. push it onto the front of the _cxq. 4096 // For now we use (a). 4097 if (Policy != 4) { 4098 iterator->TState = ObjectWaiter::TS_ENTER ; 4099 } 4100 iterator->_notified = 1 ; 4101 4102 ObjectWaiter * List = _EntryList ; 4103 if (List != NULL) { 4104 assert (List->_prev == NULL, "invariant") ; 4105 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 4106 assert (List != iterator, "invariant") ; 4107 } 4108 4109 if (Policy == 0) { // prepend to EntryList 4110 if (List == NULL) { 4111 iterator->_next = iterator->_prev = NULL ; 4112 _EntryList = iterator ; 4113 } else { 4114 List->_prev = iterator ; 4115 iterator->_next = List ; 4116 iterator->_prev = NULL ; 4117 _EntryList = iterator ; 4118 } 4119 } else 4120 if (Policy == 1) { // append to EntryList 4121 if (List == NULL) { 4122 iterator->_next = iterator->_prev = NULL ; 4123 _EntryList = iterator ; 4124 } else { 4125 // CONSIDER: finding the tail currently requires a linear-time walk of 4126 // the EntryList. We can make tail access constant-time by converting to 4127 // a CDLL instead of using our current DLL. 4128 ObjectWaiter * Tail ; 4129 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 4130 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 4131 Tail->_next = iterator ; 4132 iterator->_prev = Tail ; 4133 iterator->_next = NULL ; 4134 } 4135 } else 4136 if (Policy == 2) { // prepend to cxq 4137 // prepend to cxq 4138 if (List == NULL) { 4139 iterator->_next = iterator->_prev = NULL ; 4140 _EntryList = iterator ; 4141 } else { 4142 iterator->TState = ObjectWaiter::TS_CXQ ; 4143 for (;;) { 4144 ObjectWaiter * Front = _cxq ; 4145 iterator->_next = Front ; 4146 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 4147 break ; 4148 } 4149 } 4150 } 4151 } else 4152 if (Policy == 3) { // append to cxq 4153 iterator->TState = ObjectWaiter::TS_CXQ ; 4154 for (;;) { 4155 ObjectWaiter * Tail ; 4156 Tail = _cxq ; 4157 if (Tail == NULL) { 4158 iterator->_next = NULL ; 4159 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 4160 break ; 4161 } 4162 } else { 4163 while (Tail->_next != NULL) Tail = Tail->_next ; 4164 Tail->_next = iterator ; 4165 iterator->_prev = Tail ; 4166 iterator->_next = NULL ; 4167 break ; 4168 } 4169 } 4170 } else { 4171 ParkEvent * ev = iterator->_event ; 4172 iterator->TState = ObjectWaiter::TS_RUN ; 4173 OrderAccess::fence() ; 4174 ev->unpark() ; 4175 } 4176 4177 if (Policy < 4) { 4178 iterator->wait_reenter_begin(this); 4179 } 4180 4181 // _WaitSetLock protects the wait queue, not the EntryList. We could 4182 // move the add-to-EntryList operation, above, outside the critical section 4183 // protected by _WaitSetLock. In practice that's not useful. With the 4184 // exception of wait() timeouts and interrupts the monitor owner 4185 // is the only thread that grabs _WaitSetLock. There's almost no contention 4186 // on _WaitSetLock so it's not profitable to reduce the length of the 4187 // critical section. 4188 } 4189 4190 Thread::SpinRelease (&_WaitSetLock) ; 4191 4192 if (iterator != NULL && ObjectSynchronizer::_sync_Notifications != NULL) { 4193 ObjectSynchronizer::_sync_Notifications->inc() ; 4194 } 4195} 4196 4197 4198void ObjectMonitor::notifyAll(TRAPS) { 4199 CHECK_OWNER(); 4200 ObjectWaiter* iterator; 4201 if (_WaitSet == NULL) { 4202 TEVENT (Empty-NotifyAll) ; 4203 return ; 4204 } 4205 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 4206 4207 int Policy = Knob_MoveNotifyee ; 4208 int Tally = 0 ; 4209 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; 4210 4211 for (;;) { 4212 iterator = DequeueWaiter () ; 4213 if (iterator == NULL) break ; 4214 TEVENT (NotifyAll - Transfer1) ; 4215 ++Tally ; 4216 4217 // Disposition - what might we do with iterator ? 4218 // a. add it directly to the EntryList - either tail or head. 4219 // b. push it onto the front of the _cxq. 4220 // For now we use (a). 4221 // 4222 // TODO-FIXME: currently notifyAll() transfers the waiters one-at-a-time from the waitset 4223 // to the EntryList. This could be done more efficiently with a single bulk transfer, 4224 // but in practice it's not time-critical. Beware too, that in prepend-mode we invert the 4225 // order of the waiters. Lets say that the waitset is "ABCD" and the EntryList is "XYZ". 4226 // After a notifyAll() in prepend mode the waitset will be empty and the EntryList will 4227 // be "DCBAXYZ". 4228 4229 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; 4230 guarantee (iterator->_notified == 0, "invariant") ; 4231 iterator->_notified = 1 ; 4232 if (Policy != 4) { 4233 iterator->TState = ObjectWaiter::TS_ENTER ; 4234 } 4235 4236 ObjectWaiter * List = _EntryList ; 4237 if (List != NULL) { 4238 assert (List->_prev == NULL, "invariant") ; 4239 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; 4240 assert (List != iterator, "invariant") ; 4241 } 4242 4243 if (Policy == 0) { // prepend to EntryList 4244 if (List == NULL) { 4245 iterator->_next = iterator->_prev = NULL ; 4246 _EntryList = iterator ; 4247 } else { 4248 List->_prev = iterator ; 4249 iterator->_next = List ; 4250 iterator->_prev = NULL ; 4251 _EntryList = iterator ; 4252 } 4253 } else 4254 if (Policy == 1) { // append to EntryList 4255 if (List == NULL) { 4256 iterator->_next = iterator->_prev = NULL ; 4257 _EntryList = iterator ; 4258 } else { 4259 // CONSIDER: finding the tail currently requires a linear-time walk of 4260 // the EntryList. We can make tail access constant-time by converting to 4261 // a CDLL instead of using our current DLL. 4262 ObjectWaiter * Tail ; 4263 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; 4264 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; 4265 Tail->_next = iterator ; 4266 iterator->_prev = Tail ; 4267 iterator->_next = NULL ; 4268 } 4269 } else 4270 if (Policy == 2) { // prepend to cxq 4271 // prepend to cxq 4272 iterator->TState = ObjectWaiter::TS_CXQ ; 4273 for (;;) { 4274 ObjectWaiter * Front = _cxq ; 4275 iterator->_next = Front ; 4276 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 4277 break ; 4278 } 4279 } 4280 } else 4281 if (Policy == 3) { // append to cxq 4282 iterator->TState = ObjectWaiter::TS_CXQ ; 4283 for (;;) { 4284 ObjectWaiter * Tail ; 4285 Tail = _cxq ; 4286 if (Tail == NULL) { 4287 iterator->_next = NULL ; 4288 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 4289 break ; 4290 } 4291 } else { 4292 while (Tail->_next != NULL) Tail = Tail->_next ; 4293 Tail->_next = iterator ; 4294 iterator->_prev = Tail ; 4295 iterator->_next = NULL ; 4296 break ; 4297 } 4298 } 4299 } else { 4300 ParkEvent * ev = iterator->_event ; 4301 iterator->TState = ObjectWaiter::TS_RUN ; 4302 OrderAccess::fence() ; 4303 ev->unpark() ; 4304 } 4305 4306 if (Policy < 4) { 4307 iterator->wait_reenter_begin(this); 4308 } 4309 4310 // _WaitSetLock protects the wait queue, not the EntryList. We could 4311 // move the add-to-EntryList operation, above, outside the critical section 4312 // protected by _WaitSetLock. In practice that's not useful. With the 4313 // exception of wait() timeouts and interrupts the monitor owner 4314 // is the only thread that grabs _WaitSetLock. There's almost no contention 4315 // on _WaitSetLock so it's not profitable to reduce the length of the 4316 // critical section. 4317 } 4318 4319 Thread::SpinRelease (&_WaitSetLock) ; 4320 4321 if (Tally != 0 && ObjectSynchronizer::_sync_Notifications != NULL) { 4322 ObjectSynchronizer::_sync_Notifications->inc(Tally) ; 4323 } 4324} 4325 4326// check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 4327// TODO-FIXME: remove check_slow() -- it's likely dead. 4328 4329void ObjectMonitor::check_slow(TRAPS) { 4330 TEVENT (check_slow - throw IMSX) ; 4331 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 4332 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 4333} 4334 4335 4336// ------------------------------------------------------------------------- 4337// The raw monitor subsystem is entirely distinct from normal 4338// java-synchronization or jni-synchronization. raw monitors are not 4339// associated with objects. They can be implemented in any manner 4340// that makes sense. The original implementors decided to piggy-back 4341// the raw-monitor implementation on the existing Java objectMonitor mechanism. 4342// This flaw needs to fixed. We should reimplement raw monitors as sui-generis. 4343// Specifically, we should not implement raw monitors via java monitors. 4344// Time permitting, we should disentangle and deconvolve the two implementations 4345// and move the resulting raw monitor implementation over to the JVMTI directories. 4346// Ideally, the raw monitor implementation would be built on top of 4347// park-unpark and nothing else. 4348// 4349// raw monitors are used mainly by JVMTI 4350// The raw monitor implementation borrows the ObjectMonitor structure, 4351// but the operators are degenerate and extremely simple. 4352// 4353// Mixed use of a single objectMonitor instance -- as both a raw monitor 4354// and a normal java monitor -- is not permissible. 4355// 4356// Note that we use the single RawMonitor_lock to protect queue operations for 4357// _all_ raw monitors. This is a scalability impediment, but since raw monitor usage 4358// is deprecated and rare, this is not of concern. The RawMonitor_lock can not 4359// be held indefinitely. The critical sections must be short and bounded. 4360// 4361// ------------------------------------------------------------------------- 4362 4363int ObjectMonitor::SimpleEnter (Thread * Self) { 4364 for (;;) { 4365 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 4366 return OS_OK ; 4367 } 4368 4369 ObjectWaiter Node (Self) ; 4370 Self->_ParkEvent->reset() ; // strictly optional 4371 Node.TState = ObjectWaiter::TS_ENTER ; 4372 4373 RawMonitor_lock->lock_without_safepoint_check() ; 4374 Node._next = _EntryList ; 4375 _EntryList = &Node ; 4376 OrderAccess::fence() ; 4377 if (_owner == NULL && Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 4378 _EntryList = Node._next ; 4379 RawMonitor_lock->unlock() ; 4380 return OS_OK ; 4381 } 4382 RawMonitor_lock->unlock() ; 4383 while (Node.TState == ObjectWaiter::TS_ENTER) { 4384 Self->_ParkEvent->park() ; 4385 } 4386 } 4387} 4388 4389int ObjectMonitor::SimpleExit (Thread * Self) { 4390 guarantee (_owner == Self, "invariant") ; 4391 OrderAccess::release_store_ptr (&_owner, NULL) ; 4392 OrderAccess::fence() ; 4393 if (_EntryList == NULL) return OS_OK ; 4394 ObjectWaiter * w ; 4395 4396 RawMonitor_lock->lock_without_safepoint_check() ; 4397 w = _EntryList ; 4398 if (w != NULL) { 4399 _EntryList = w->_next ; 4400 } 4401 RawMonitor_lock->unlock() ; 4402 if (w != NULL) { 4403 guarantee (w ->TState == ObjectWaiter::TS_ENTER, "invariant") ; 4404 ParkEvent * ev = w->_event ; 4405 w->TState = ObjectWaiter::TS_RUN ; 4406 OrderAccess::fence() ; 4407 ev->unpark() ; 4408 } 4409 return OS_OK ; 4410} 4411 4412int ObjectMonitor::SimpleWait (Thread * Self, jlong millis) { 4413 guarantee (_owner == Self , "invariant") ; 4414 guarantee (_recursions == 0, "invariant") ; 4415 4416 ObjectWaiter Node (Self) ; 4417 Node._notified = 0 ; 4418 Node.TState = ObjectWaiter::TS_WAIT ; 4419 4420 RawMonitor_lock->lock_without_safepoint_check() ; 4421 Node._next = _WaitSet ; 4422 _WaitSet = &Node ; 4423 RawMonitor_lock->unlock() ; 4424 4425 SimpleExit (Self) ; 4426 guarantee (_owner != Self, "invariant") ; 4427 4428 int ret = OS_OK ; 4429 if (millis <= 0) { 4430 Self->_ParkEvent->park(); 4431 } else { 4432 ret = Self->_ParkEvent->park(millis); 4433 } 4434 4435 // If thread still resides on the waitset then unlink it. 4436 // Double-checked locking -- the usage is safe in this context 4437 // as we TState is volatile and the lock-unlock operators are 4438 // serializing (barrier-equivalent). 4439 4440 if (Node.TState == ObjectWaiter::TS_WAIT) { 4441 RawMonitor_lock->lock_without_safepoint_check() ; 4442 if (Node.TState == ObjectWaiter::TS_WAIT) { 4443 // Simple O(n) unlink, but performance isn't critical here. 4444 ObjectWaiter * p ; 4445 ObjectWaiter * q = NULL ; 4446 for (p = _WaitSet ; p != &Node; p = p->_next) { 4447 q = p ; 4448 } 4449 guarantee (p == &Node, "invariant") ; 4450 if (q == NULL) { 4451 guarantee (p == _WaitSet, "invariant") ; 4452 _WaitSet = p->_next ; 4453 } else { 4454 guarantee (p == q->_next, "invariant") ; 4455 q->_next = p->_next ; 4456 } 4457 Node.TState = ObjectWaiter::TS_RUN ; 4458 } 4459 RawMonitor_lock->unlock() ; 4460 } 4461 4462 guarantee (Node.TState == ObjectWaiter::TS_RUN, "invariant") ; 4463 SimpleEnter (Self) ; 4464 4465 guarantee (_owner == Self, "invariant") ; 4466 guarantee (_recursions == 0, "invariant") ; 4467 return ret ; 4468} 4469 4470int ObjectMonitor::SimpleNotify (Thread * Self, bool All) { 4471 guarantee (_owner == Self, "invariant") ; 4472 if (_WaitSet == NULL) return OS_OK ; 4473 4474 // We have two options: 4475 // A. Transfer the threads from the WaitSet to the EntryList 4476 // B. Remove the thread from the WaitSet and unpark() it. 4477 // 4478 // We use (B), which is crude and results in lots of futile 4479 // context switching. In particular (B) induces lots of contention. 4480 4481 ParkEvent * ev = NULL ; // consider using a small auto array ... 4482 RawMonitor_lock->lock_without_safepoint_check() ; 4483 for (;;) { 4484 ObjectWaiter * w = _WaitSet ; 4485 if (w == NULL) break ; 4486 _WaitSet = w->_next ; 4487 if (ev != NULL) { ev->unpark(); ev = NULL; } 4488 ev = w->_event ; 4489 OrderAccess::loadstore() ; 4490 w->TState = ObjectWaiter::TS_RUN ; 4491 OrderAccess::storeload(); 4492 if (!All) break ; 4493 } 4494 RawMonitor_lock->unlock() ; 4495 if (ev != NULL) ev->unpark(); 4496 return OS_OK ; 4497} 4498 4499// Any JavaThread will enter here with state _thread_blocked 4500int ObjectMonitor::raw_enter(TRAPS) { 4501 TEVENT (raw_enter) ; 4502 void * Contended ; 4503 4504 // don't enter raw monitor if thread is being externally suspended, it will 4505 // surprise the suspender if a "suspended" thread can still enter monitor 4506 JavaThread * jt = (JavaThread *)THREAD; 4507 if (THREAD->is_Java_thread()) { 4508 jt->SR_lock()->lock_without_safepoint_check(); 4509 while (jt->is_external_suspend()) { 4510 jt->SR_lock()->unlock(); 4511 jt->java_suspend_self(); 4512 jt->SR_lock()->lock_without_safepoint_check(); 4513 } 4514 // guarded by SR_lock to avoid racing with new external suspend requests. 4515 Contended = Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) ; 4516 jt->SR_lock()->unlock(); 4517 } else { 4518 Contended = Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) ; 4519 } 4520 4521 if (Contended == THREAD) { 4522 _recursions ++ ; 4523 return OM_OK ; 4524 } 4525 4526 if (Contended == NULL) { 4527 guarantee (_owner == THREAD, "invariant") ; 4528 guarantee (_recursions == 0, "invariant") ; 4529 return OM_OK ; 4530 } 4531 4532 THREAD->set_current_pending_monitor(this); 4533 4534 if (!THREAD->is_Java_thread()) { 4535 // No other non-Java threads besides VM thread would acquire 4536 // a raw monitor. 4537 assert(THREAD->is_VM_thread(), "must be VM thread"); 4538 SimpleEnter (THREAD) ; 4539 } else { 4540 guarantee (jt->thread_state() == _thread_blocked, "invariant") ; 4541 for (;;) { 4542 jt->set_suspend_equivalent(); 4543 // cleared by handle_special_suspend_equivalent_condition() or 4544 // java_suspend_self() 4545 SimpleEnter (THREAD) ; 4546 4547 // were we externally suspended while we were waiting? 4548 if (!jt->handle_special_suspend_equivalent_condition()) break ; 4549 4550 // This thread was externally suspended 4551 // 4552 // This logic isn't needed for JVMTI raw monitors, 4553 // but doesn't hurt just in case the suspend rules change. This 4554 // logic is needed for the ObjectMonitor.wait() reentry phase. 4555 // We have reentered the contended monitor, but while we were 4556 // waiting another thread suspended us. We don't want to reenter 4557 // the monitor while suspended because that would surprise the 4558 // thread that suspended us. 4559 // 4560 // Drop the lock - 4561 SimpleExit (THREAD) ; 4562 4563 jt->java_suspend_self(); 4564 } 4565 4566 assert(_owner == THREAD, "Fatal error with monitor owner!"); 4567 assert(_recursions == 0, "Fatal error with monitor recursions!"); 4568 } 4569 4570 THREAD->set_current_pending_monitor(NULL); 4571 guarantee (_recursions == 0, "invariant") ; 4572 return OM_OK; 4573} 4574 4575// Used mainly for JVMTI raw monitor implementation 4576// Also used for ObjectMonitor::wait(). 4577int ObjectMonitor::raw_exit(TRAPS) { 4578 TEVENT (raw_exit) ; 4579 if (THREAD != _owner) { 4580 return OM_ILLEGAL_MONITOR_STATE; 4581 } 4582 if (_recursions > 0) { 4583 --_recursions ; 4584 return OM_OK ; 4585 } 4586 4587 void * List = _EntryList ; 4588 SimpleExit (THREAD) ; 4589 4590 return OM_OK; 4591} 4592 4593// Used for JVMTI raw monitor implementation. 4594// All JavaThreads will enter here with state _thread_blocked 4595 4596int ObjectMonitor::raw_wait(jlong millis, bool interruptible, TRAPS) { 4597 TEVENT (raw_wait) ; 4598 if (THREAD != _owner) { 4599 return OM_ILLEGAL_MONITOR_STATE; 4600 } 4601 4602 // To avoid spurious wakeups we reset the parkevent -- This is strictly optional. 4603 // The caller must be able to tolerate spurious returns from raw_wait(). 4604 THREAD->_ParkEvent->reset() ; 4605 OrderAccess::fence() ; 4606 4607 // check interrupt event 4608 if (interruptible && Thread::is_interrupted(THREAD, true)) { 4609 return OM_INTERRUPTED; 4610 } 4611 4612 intptr_t save = _recursions ; 4613 _recursions = 0 ; 4614 _waiters ++ ; 4615 if (THREAD->is_Java_thread()) { 4616 guarantee (((JavaThread *) THREAD)->thread_state() == _thread_blocked, "invariant") ; 4617 ((JavaThread *)THREAD)->set_suspend_equivalent(); 4618 } 4619 int rv = SimpleWait (THREAD, millis) ; 4620 _recursions = save ; 4621 _waiters -- ; 4622 4623 guarantee (THREAD == _owner, "invariant") ; 4624 if (THREAD->is_Java_thread()) { 4625 JavaThread * jSelf = (JavaThread *) THREAD ; 4626 for (;;) { 4627 if (!jSelf->handle_special_suspend_equivalent_condition()) break ; 4628 SimpleExit (THREAD) ; 4629 jSelf->java_suspend_self(); 4630 SimpleEnter (THREAD) ; 4631 jSelf->set_suspend_equivalent() ; 4632 } 4633 } 4634 guarantee (THREAD == _owner, "invariant") ; 4635 4636 if (interruptible && Thread::is_interrupted(THREAD, true)) { 4637 return OM_INTERRUPTED; 4638 } 4639 return OM_OK ; 4640} 4641 4642int ObjectMonitor::raw_notify(TRAPS) { 4643 TEVENT (raw_notify) ; 4644 if (THREAD != _owner) { 4645 return OM_ILLEGAL_MONITOR_STATE; 4646 } 4647 SimpleNotify (THREAD, false) ; 4648 return OM_OK; 4649} 4650 4651int ObjectMonitor::raw_notifyAll(TRAPS) { 4652 TEVENT (raw_notifyAll) ; 4653 if (THREAD != _owner) { 4654 return OM_ILLEGAL_MONITOR_STATE; 4655 } 4656 SimpleNotify (THREAD, true) ; 4657 return OM_OK; 4658} 4659 4660#ifndef PRODUCT 4661void ObjectMonitor::verify() { 4662} 4663 4664void ObjectMonitor::print() { 4665} 4666#endif 4667 4668//------------------------------------------------------------------------------ 4669// Non-product code 4670 4671#ifndef PRODUCT 4672 4673void ObjectSynchronizer::trace_locking(Handle locking_obj, bool is_compiled, 4674 bool is_method, bool is_locking) { 4675 // Don't know what to do here 4676} 4677 4678// Verify all monitors in the monitor cache, the verification is weak. 4679void ObjectSynchronizer::verify() { 4680 ObjectMonitor* block = gBlockList; 4681 ObjectMonitor* mid; 4682 while (block) { 4683 assert(block->object() == CHAINMARKER, "must be a block header"); 4684 for (int i = 1; i < _BLOCKSIZE; i++) { 4685 mid = block + i; 4686 oop object = (oop) mid->object(); 4687 if (object != NULL) { 4688 mid->verify(); 4689 } 4690 } 4691 block = (ObjectMonitor*) block->FreeNext; 4692 } 4693} 4694 4695// Check if monitor belongs to the monitor cache 4696// The list is grow-only so it's *relatively* safe to traverse 4697// the list of extant blocks without taking a lock. 4698 4699int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 4700 ObjectMonitor* block = gBlockList; 4701 4702 while (block) { 4703 assert(block->object() == CHAINMARKER, "must be a block header"); 4704 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 4705 address mon = (address) monitor; 4706 address blk = (address) block; 4707 size_t diff = mon - blk; 4708 assert((diff % sizeof(ObjectMonitor)) == 0, "check"); 4709 return 1; 4710 } 4711 block = (ObjectMonitor*) block->FreeNext; 4712 } 4713 return 0; 4714} 4715 4716#endif 4717