1 /* 2 * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "memory/metaspaceShared.hpp" 28 #include "memory/padded.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "oops/markOop.hpp" 31 #include "oops/oop.inline.hpp" 32 #include "runtime/atomic.inline.hpp" 33 #include "runtime/biasedLocking.hpp" 34 #include "runtime/handles.inline.hpp" 35 #include "runtime/interfaceSupport.hpp" 36 #include "runtime/mutexLocker.hpp" 37 #include "runtime/objectMonitor.hpp" 38 #include "runtime/objectMonitor.inline.hpp" 39 #include "runtime/osThread.hpp" 40 #include "runtime/stubRoutines.hpp" 41 #include "runtime/synchronizer.hpp" 42 #include "runtime/thread.inline.hpp" 43 #include "utilities/dtrace.hpp" 44 #include "utilities/events.hpp" 45 #include "utilities/preserveException.hpp" 46 47 #if defined(__GNUC__) && !defined(PPC64) 48 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 49 #define NOINLINE __attribute__((noinline)) 50 #else 51 #define NOINLINE 52 #endif 53 54 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 55 56 // The "core" versions of monitor enter and exit reside in this file. 57 // The interpreter and compilers contain specialized transliterated 58 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 59 // for instance. If you make changes here, make sure to modify the 60 // interpreter, and both C1 and C2 fast-path inline locking code emission. 61 // 62 // ----------------------------------------------------------------------------- 63 64 #ifdef DTRACE_ENABLED 65 66 // Only bother with this argument setup if dtrace is available 67 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 68 69 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 70 char* bytes = NULL; \ 71 int len = 0; \ 72 jlong jtid = SharedRuntime::get_java_tid(thread); \ 73 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 74 if (klassname != NULL) { \ 75 bytes = (char*)klassname->bytes(); \ 76 len = klassname->utf8_length(); \ 77 } 78 79 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 80 { \ 81 if (DTraceMonitorProbes) { \ 82 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 83 HOTSPOT_MONITOR_WAIT(jtid, \ 84 (uintptr_t)(monitor), bytes, len, (millis)); \ 85 } \ 86 } 87 88 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 89 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 90 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 91 92 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 93 { \ 94 if (DTraceMonitorProbes) { \ 95 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 96 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 97 (uintptr_t)(monitor), bytes, len); \ 98 } \ 99 } 100 101 #else // ndef DTRACE_ENABLED 102 103 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 104 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 105 106 #endif // ndef DTRACE_ENABLED 107 108 // This exists only as a workaround of dtrace bug 6254741 109 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 110 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 111 return 0; 112 } 113 114 #define NINFLATIONLOCKS 256 115 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS]; 116 117 // global list of blocks of monitors 118 // gBlockList is really PaddedEnd<ObjectMonitor> *, but we don't 119 // want to expose the PaddedEnd template more than necessary. 120 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL; 121 // global monitor free list 122 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 123 // global monitor in-use list, for moribund threads, 124 // monitors they inflated need to be scanned for deflation 125 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 126 // count of entries in gOmInUseList 127 int ObjectSynchronizer::gOmInUseCount = 0; 128 129 static volatile intptr_t gListLock = 0; // protects global monitor lists 130 static volatile int gMonitorFreeCount = 0; // # on gFreeList 131 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation 132 133 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 134 135 136 // =====================> Quick functions 137 138 // The quick_* forms are special fast-path variants used to improve 139 // performance. In the simplest case, a "quick_*" implementation could 140 // simply return false, in which case the caller will perform the necessary 141 // state transitions and call the slow-path form. 142 // The fast-path is designed to handle frequently arising cases in an efficient 143 // manner and is just a degenerate "optimistic" variant of the slow-path. 144 // returns true -- to indicate the call was satisfied. 145 // returns false -- to indicate the call needs the services of the slow-path. 146 // A no-loitering ordinance is in effect for code in the quick_* family 147 // operators: safepoints or indefinite blocking (blocking that might span a 148 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 149 // entry. 150 // 151 // Consider: An interesting optimization is to have the JIT recognize the 152 // following common idiom: 153 // synchronized (someobj) { .... ; notify(); } 154 // That is, we find a notify() or notifyAll() call that immediately precedes 155 // the monitorexit operation. In that case the JIT could fuse the operations 156 // into a single notifyAndExit() runtime primitive. 157 158 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) { 159 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 160 assert(self->is_Java_thread(), "invariant"); 161 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 162 No_Safepoint_Verifier nsv; 163 if (obj == NULL) return false; // slow-path for invalid obj 164 const markOop mark = obj->mark(); 165 166 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) { 167 // Degenerate notify 168 // stack-locked by caller so by definition the implied waitset is empty. 169 return true; 170 } 171 172 if (mark->has_monitor()) { 173 ObjectMonitor * const mon = mark->monitor(); 174 assert(mon->object() == obj, "invariant"); 175 if (mon->owner() != self) return false; // slow-path for IMS exception 176 177 if (mon->first_waiter() != NULL) { 178 // We have one or more waiters. Since this is an inflated monitor 179 // that we own, we can transfer one or more threads from the waitset 180 // to the entrylist here and now, avoiding the slow-path. 181 if (all) { 182 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 183 } else { 184 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 185 } 186 int tally = 0; 187 do { 188 mon->INotify(self); 189 ++tally; 190 } while (mon->first_waiter() != NULL && all); 191 if (ObjectMonitor::_sync_Notifications != NULL) { 192 ObjectMonitor::_sync_Notifications->inc(tally); 193 } 194 } 195 return true; 196 } 197 198 // biased locking and any other IMS exception states take the slow-path 199 return false; 200 } 201 202 203 // The LockNode emitted directly at the synchronization site would have 204 // been too big if it were to have included support for the cases of inflated 205 // recursive enter and exit, so they go here instead. 206 // Note that we can't safely call AsyncPrintJavaStack() from within 207 // quick_enter() as our thread state remains _in_Java. 208 209 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self, 210 BasicLock * Lock) { 211 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 212 assert(Self->is_Java_thread(), "invariant"); 213 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant"); 214 No_Safepoint_Verifier nsv; 215 if (obj == NULL) return false; // Need to throw NPE 216 const markOop mark = obj->mark(); 217 218 if (mark->has_monitor()) { 219 ObjectMonitor * const m = mark->monitor(); 220 assert(m->object() == obj, "invariant"); 221 Thread * const owner = (Thread *) m->_owner; 222 223 // Lock contention and Transactional Lock Elision (TLE) diagnostics 224 // and observability 225 // Case: light contention possibly amenable to TLE 226 // Case: TLE inimical operations such as nested/recursive synchronization 227 228 if (owner == Self) { 229 m->_recursions++; 230 return true; 231 } 232 233 if (owner == NULL && 234 Atomic::cmpxchg_ptr(Self, &(m->_owner), NULL) == NULL) { 235 assert(m->_recursions == 0, "invariant"); 236 assert(m->_owner == Self, "invariant"); 237 return true; 238 } 239 } 240 241 // Note that we could inflate in quick_enter. 242 // This is likely a useful optimization 243 // Critically, in quick_enter() we must not: 244 // -- perform bias revocation, or 245 // -- block indefinitely, or 246 // -- reach a safepoint 247 248 return false; // revert to slow-path 249 } 250 251 // ----------------------------------------------------------------------------- 252 // Fast Monitor Enter/Exit 253 // This the fast monitor enter. The interpreter and compiler use 254 // some assembly copies of this code. Make sure update those code 255 // if the following function is changed. The implementation is 256 // extremely sensitive to race condition. Be careful. 257 258 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 259 bool attempt_rebias, TRAPS) { 260 if (UseBiasedLocking) { 261 if (!SafepointSynchronize::is_at_safepoint()) { 262 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 263 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 264 return; 265 } 266 } else { 267 assert(!attempt_rebias, "can not rebias toward VM thread"); 268 BiasedLocking::revoke_at_safepoint(obj); 269 } 270 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 271 } 272 273 slow_enter(obj, lock, THREAD); 274 } 275 276 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 277 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); 278 // if displaced header is null, the previous enter is recursive enter, no-op 279 markOop dhw = lock->displaced_header(); 280 markOop mark; 281 if (dhw == NULL) { 282 // Recursive stack-lock. 283 // Diagnostics -- Could be: stack-locked, inflating, inflated. 284 mark = object->mark(); 285 assert(!mark->is_neutral(), "invariant"); 286 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { 287 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant"); 288 } 289 if (mark->has_monitor()) { 290 ObjectMonitor * m = mark->monitor(); 291 assert(((oop)(m->object()))->mark() == mark, "invariant"); 292 assert(m->is_entered(THREAD), "invariant"); 293 } 294 return; 295 } 296 297 mark = object->mark(); 298 299 // If the object is stack-locked by the current thread, try to 300 // swing the displaced header from the box back to the mark. 301 if (mark == (markOop) lock) { 302 assert(dhw->is_neutral(), "invariant"); 303 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { 304 TEVENT(fast_exit: release stacklock); 305 return; 306 } 307 } 308 309 ObjectSynchronizer::inflate(THREAD, object)->exit(true, THREAD); 310 } 311 312 // ----------------------------------------------------------------------------- 313 // Interpreter/Compiler Slow Case 314 // This routine is used to handle interpreter/compiler slow case 315 // We don't need to use fast path here, because it must have been 316 // failed in the interpreter/compiler code. 317 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 318 markOop mark = obj->mark(); 319 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 320 321 if (mark->is_neutral()) { 322 // Anticipate successful CAS -- the ST of the displaced mark must 323 // be visible <= the ST performed by the CAS. 324 lock->set_displaced_header(mark); 325 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { 326 TEVENT(slow_enter: release stacklock); 327 return; 328 } 329 // Fall through to inflate() ... 330 } else if (mark->has_locker() && 331 THREAD->is_lock_owned((address)mark->locker())) { 332 assert(lock != mark->locker(), "must not re-lock the same lock"); 333 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 334 lock->set_displaced_header(NULL); 335 return; 336 } 337 338 // The object header will never be displaced to this lock, 339 // so it does not matter what the value is, except that it 340 // must be non-zero to avoid looking like a re-entrant lock, 341 // and must not look locked either. 342 lock->set_displaced_header(markOopDesc::unused_mark()); 343 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 344 } 345 346 // This routine is used to handle interpreter/compiler slow case 347 // We don't need to use fast path here, because it must have 348 // failed in the interpreter/compiler code. Simply use the heavy 349 // weight monitor should be ok, unless someone find otherwise. 350 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 351 fast_exit(object, lock, THREAD); 352 } 353 354 // ----------------------------------------------------------------------------- 355 // Class Loader support to workaround deadlocks on the class loader lock objects 356 // Also used by GC 357 // complete_exit()/reenter() are used to wait on a nested lock 358 // i.e. to give up an outer lock completely and then re-enter 359 // Used when holding nested locks - lock acquisition order: lock1 then lock2 360 // 1) complete_exit lock1 - saving recursion count 361 // 2) wait on lock2 362 // 3) when notified on lock2, unlock lock2 363 // 4) reenter lock1 with original recursion count 364 // 5) lock lock2 365 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 366 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 367 TEVENT(complete_exit); 368 if (UseBiasedLocking) { 369 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 370 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 371 } 372 373 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 374 375 return monitor->complete_exit(THREAD); 376 } 377 378 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 379 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 380 TEVENT(reenter); 381 if (UseBiasedLocking) { 382 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 383 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 384 } 385 386 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 387 388 monitor->reenter(recursion, THREAD); 389 } 390 // ----------------------------------------------------------------------------- 391 // JNI locks on java objects 392 // NOTE: must use heavy weight monitor to handle jni monitor enter 393 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 394 // the current locking is from JNI instead of Java code 395 TEVENT(jni_enter); 396 if (UseBiasedLocking) { 397 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 398 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 399 } 400 THREAD->set_current_pending_monitor_is_from_java(false); 401 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 402 THREAD->set_current_pending_monitor_is_from_java(true); 403 } 404 405 // NOTE: must use heavy weight monitor to handle jni monitor exit 406 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 407 TEVENT(jni_exit); 408 if (UseBiasedLocking) { 409 Handle h_obj(THREAD, obj); 410 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 411 obj = h_obj(); 412 } 413 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 414 415 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); 416 // If this thread has locked the object, exit the monitor. Note: can't use 417 // monitor->check(CHECK); must exit even if an exception is pending. 418 if (monitor->check(THREAD)) { 419 monitor->exit(true, THREAD); 420 } 421 } 422 423 // ----------------------------------------------------------------------------- 424 // Internal VM locks on java objects 425 // standard constructor, allows locking failures 426 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 427 _dolock = doLock; 428 _thread = thread; 429 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 430 _obj = obj; 431 432 if (_dolock) { 433 TEVENT(ObjectLocker); 434 435 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 436 } 437 } 438 439 ObjectLocker::~ObjectLocker() { 440 if (_dolock) { 441 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 442 } 443 } 444 445 446 // ----------------------------------------------------------------------------- 447 // Wait/Notify/NotifyAll 448 // NOTE: must use heavy weight monitor to handle wait() 449 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 450 if (UseBiasedLocking) { 451 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 452 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 453 } 454 if (millis < 0) { 455 TEVENT(wait - throw IAX); 456 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 457 } 458 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 459 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 460 monitor->wait(millis, true, THREAD); 461 462 // This dummy call is in place to get around dtrace bug 6254741. Once 463 // that's fixed we can uncomment the following line, remove the call 464 // and change this function back into a "void" func. 465 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 466 return dtrace_waited_probe(monitor, obj, THREAD); 467 } 468 469 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 470 if (UseBiasedLocking) { 471 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 472 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 473 } 474 if (millis < 0) { 475 TEVENT(wait - throw IAX); 476 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 477 } 478 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD); 479 } 480 481 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 482 if (UseBiasedLocking) { 483 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 484 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 485 } 486 487 markOop mark = obj->mark(); 488 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 489 return; 490 } 491 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); 492 } 493 494 // NOTE: see comment of notify() 495 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 496 if (UseBiasedLocking) { 497 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 498 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 499 } 500 501 markOop mark = obj->mark(); 502 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 503 return; 504 } 505 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); 506 } 507 508 // ----------------------------------------------------------------------------- 509 // Hash Code handling 510 // 511 // Performance concern: 512 // OrderAccess::storestore() calls release() which at one time stored 0 513 // into the global volatile OrderAccess::dummy variable. This store was 514 // unnecessary for correctness. Many threads storing into a common location 515 // causes considerable cache migration or "sloshing" on large SMP systems. 516 // As such, I avoided using OrderAccess::storestore(). In some cases 517 // OrderAccess::fence() -- which incurs local latency on the executing 518 // processor -- is a better choice as it scales on SMP systems. 519 // 520 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 521 // a discussion of coherency costs. Note that all our current reference 522 // platforms provide strong ST-ST order, so the issue is moot on IA32, 523 // x64, and SPARC. 524 // 525 // As a general policy we use "volatile" to control compiler-based reordering 526 // and explicit fences (barriers) to control for architectural reordering 527 // performed by the CPU(s) or platform. 528 529 struct SharedGlobals { 530 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE]; 531 // These are highly shared mostly-read variables. 532 // To avoid false-sharing they need to be the sole occupants of a cache line. 533 volatile int stwRandom; 534 volatile int stwCycle; 535 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 536 // Hot RW variable -- Sequester to avoid false-sharing 537 volatile int hcSequence; 538 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int)); 539 }; 540 541 static SharedGlobals GVars; 542 static int MonitorScavengeThreshold = 1000000; 543 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 544 545 static markOop ReadStableMark(oop obj) { 546 markOop mark = obj->mark(); 547 if (!mark->is_being_inflated()) { 548 return mark; // normal fast-path return 549 } 550 551 int its = 0; 552 for (;;) { 553 markOop mark = obj->mark(); 554 if (!mark->is_being_inflated()) { 555 return mark; // normal fast-path return 556 } 557 558 // The object is being inflated by some other thread. 559 // The caller of ReadStableMark() must wait for inflation to complete. 560 // Avoid live-lock 561 // TODO: consider calling SafepointSynchronize::do_call_back() while 562 // spinning to see if there's a safepoint pending. If so, immediately 563 // yielding or blocking would be appropriate. Avoid spinning while 564 // there is a safepoint pending. 565 // TODO: add inflation contention performance counters. 566 // TODO: restrict the aggregate number of spinners. 567 568 ++its; 569 if (its > 10000 || !os::is_MP()) { 570 if (its & 1) { 571 os::naked_yield(); 572 TEVENT(Inflate: INFLATING - yield); 573 } else { 574 // Note that the following code attenuates the livelock problem but is not 575 // a complete remedy. A more complete solution would require that the inflating 576 // thread hold the associated inflation lock. The following code simply restricts 577 // the number of spinners to at most one. We'll have N-2 threads blocked 578 // on the inflationlock, 1 thread holding the inflation lock and using 579 // a yield/park strategy, and 1 thread in the midst of inflation. 580 // A more refined approach would be to change the encoding of INFLATING 581 // to allow encapsulation of a native thread pointer. Threads waiting for 582 // inflation to complete would use CAS to push themselves onto a singly linked 583 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 584 // and calling park(). When inflation was complete the thread that accomplished inflation 585 // would detach the list and set the markword to inflated with a single CAS and 586 // then for each thread on the list, set the flag and unpark() the thread. 587 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 588 // wakes at most one thread whereas we need to wake the entire list. 589 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 590 int YieldThenBlock = 0; 591 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 592 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 593 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 594 while (obj->mark() == markOopDesc::INFLATING()) { 595 // Beware: NakedYield() is advisory and has almost no effect on some platforms 596 // so we periodically call Self->_ParkEvent->park(1). 597 // We use a mixed spin/yield/block mechanism. 598 if ((YieldThenBlock++) >= 16) { 599 Thread::current()->_ParkEvent->park(1); 600 } else { 601 os::naked_yield(); 602 } 603 } 604 Thread::muxRelease(gInflationLocks + ix); 605 TEVENT(Inflate: INFLATING - yield/park); 606 } 607 } else { 608 SpinPause(); // SMP-polite spinning 609 } 610 } 611 } 612 613 // hashCode() generation : 614 // 615 // Possibilities: 616 // * MD5Digest of {obj,stwRandom} 617 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 618 // * A DES- or AES-style SBox[] mechanism 619 // * One of the Phi-based schemes, such as: 620 // 2654435761 = 2^32 * Phi (golden ratio) 621 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 622 // * A variation of Marsaglia's shift-xor RNG scheme. 623 // * (obj ^ stwRandom) is appealing, but can result 624 // in undesirable regularity in the hashCode values of adjacent objects 625 // (objects allocated back-to-back, in particular). This could potentially 626 // result in hashtable collisions and reduced hashtable efficiency. 627 // There are simple ways to "diffuse" the middle address bits over the 628 // generated hashCode values: 629 630 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 631 intptr_t value = 0; 632 if (hashCode == 0) { 633 // This form uses an unguarded global Park-Miller RNG, 634 // so it's possible for two threads to race and generate the same RNG. 635 // On MP system we'll have lots of RW access to a global, so the 636 // mechanism induces lots of coherency traffic. 637 value = os::random(); 638 } else if (hashCode == 1) { 639 // This variation has the property of being stable (idempotent) 640 // between STW operations. This can be useful in some of the 1-0 641 // synchronization schemes. 642 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 643 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 644 } else if (hashCode == 2) { 645 value = 1; // for sensitivity testing 646 } else if (hashCode == 3) { 647 value = ++GVars.hcSequence; 648 } else if (hashCode == 4) { 649 value = cast_from_oop<intptr_t>(obj); 650 } else { 651 // Marsaglia's xor-shift scheme with thread-specific state 652 // This is probably the best overall implementation -- we'll 653 // likely make this the default in future releases. 654 unsigned t = Self->_hashStateX; 655 t ^= (t << 11); 656 Self->_hashStateX = Self->_hashStateY; 657 Self->_hashStateY = Self->_hashStateZ; 658 Self->_hashStateZ = Self->_hashStateW; 659 unsigned v = Self->_hashStateW; 660 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 661 Self->_hashStateW = v; 662 value = v; 663 } 664 665 value &= markOopDesc::hash_mask; 666 if (value == 0) value = 0xBAD; 667 assert(value != markOopDesc::no_hash, "invariant"); 668 TEVENT(hashCode: GENERATE); 669 return value; 670 } 671 672 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 673 if (UseBiasedLocking) { 674 // NOTE: many places throughout the JVM do not expect a safepoint 675 // to be taken here, in particular most operations on perm gen 676 // objects. However, we only ever bias Java instances and all of 677 // the call sites of identity_hash that might revoke biases have 678 // been checked to make sure they can handle a safepoint. The 679 // added check of the bias pattern is to avoid useless calls to 680 // thread-local storage. 681 if (obj->mark()->has_bias_pattern()) { 682 // Handle for oop obj in case of STW safepoint 683 Handle hobj(Self, obj); 684 // Relaxing assertion for bug 6320749. 685 assert(Universe::verify_in_progress() || 686 !SafepointSynchronize::is_at_safepoint(), 687 "biases should not be seen by VM thread here"); 688 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 689 obj = hobj(); 690 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 691 } 692 } 693 694 // hashCode() is a heap mutator ... 695 // Relaxing assertion for bug 6320749. 696 assert(Universe::verify_in_progress() || DumpSharedSpaces || 697 !SafepointSynchronize::is_at_safepoint(), "invariant"); 698 assert(Universe::verify_in_progress() || DumpSharedSpaces || 699 Self->is_Java_thread() , "invariant"); 700 assert(Universe::verify_in_progress() || DumpSharedSpaces || 701 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 702 703 ObjectMonitor* monitor = NULL; 704 markOop temp, test; 705 intptr_t hash; 706 markOop mark = ReadStableMark(obj); 707 708 // object should remain ineligible for biased locking 709 assert(!mark->has_bias_pattern(), "invariant"); 710 711 if (mark->is_neutral()) { 712 hash = mark->hash(); // this is a normal header 713 if (hash) { // if it has hash, just return it 714 return hash; 715 } 716 hash = get_next_hash(Self, obj); // allocate a new hash code 717 temp = mark->copy_set_hash(hash); // merge the hash code into header 718 // use (machine word version) atomic operation to install the hash 719 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 720 if (test == mark) { 721 return hash; 722 } 723 // If atomic operation failed, we must inflate the header 724 // into heavy weight monitor. We could add more code here 725 // for fast path, but it does not worth the complexity. 726 } else if (mark->has_monitor()) { 727 monitor = mark->monitor(); 728 temp = monitor->header(); 729 assert(temp->is_neutral(), "invariant"); 730 hash = temp->hash(); 731 if (hash) { 732 return hash; 733 } 734 // Skip to the following code to reduce code size 735 } else if (Self->is_lock_owned((address)mark->locker())) { 736 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 737 assert(temp->is_neutral(), "invariant"); 738 hash = temp->hash(); // by current thread, check if the displaced 739 if (hash) { // header contains hash code 740 return hash; 741 } 742 // WARNING: 743 // The displaced header is strictly immutable. 744 // It can NOT be changed in ANY cases. So we have 745 // to inflate the header into heavyweight monitor 746 // even the current thread owns the lock. The reason 747 // is the BasicLock (stack slot) will be asynchronously 748 // read by other threads during the inflate() function. 749 // Any change to stack may not propagate to other threads 750 // correctly. 751 } 752 753 // Inflate the monitor to set hash code 754 monitor = ObjectSynchronizer::inflate(Self, obj); 755 // Load displaced header and check it has hash code 756 mark = monitor->header(); 757 assert(mark->is_neutral(), "invariant"); 758 hash = mark->hash(); 759 if (hash == 0) { 760 hash = get_next_hash(Self, obj); 761 temp = mark->copy_set_hash(hash); // merge hash code into header 762 assert(temp->is_neutral(), "invariant"); 763 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 764 if (test != mark) { 765 // The only update to the header in the monitor (outside GC) 766 // is install the hash code. If someone add new usage of 767 // displaced header, please update this code 768 hash = test->hash(); 769 assert(test->is_neutral(), "invariant"); 770 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 771 } 772 } 773 // We finally get the hash 774 return hash; 775 } 776 777 // Deprecated -- use FastHashCode() instead. 778 779 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 780 return FastHashCode(Thread::current(), obj()); 781 } 782 783 784 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 785 Handle h_obj) { 786 if (UseBiasedLocking) { 787 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 788 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 789 } 790 791 assert(thread == JavaThread::current(), "Can only be called on current thread"); 792 oop obj = h_obj(); 793 794 markOop mark = ReadStableMark(obj); 795 796 // Uncontended case, header points to stack 797 if (mark->has_locker()) { 798 return thread->is_lock_owned((address)mark->locker()); 799 } 800 // Contended case, header points to ObjectMonitor (tagged pointer) 801 if (mark->has_monitor()) { 802 ObjectMonitor* monitor = mark->monitor(); 803 return monitor->is_entered(thread) != 0; 804 } 805 // Unlocked case, header in place 806 assert(mark->is_neutral(), "sanity check"); 807 return false; 808 } 809 810 // Be aware of this method could revoke bias of the lock object. 811 // This method queries the ownership of the lock handle specified by 'h_obj'. 812 // If the current thread owns the lock, it returns owner_self. If no 813 // thread owns the lock, it returns owner_none. Otherwise, it will return 814 // owner_other. 815 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 816 (JavaThread *self, Handle h_obj) { 817 // The caller must beware this method can revoke bias, and 818 // revocation can result in a safepoint. 819 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 820 assert(self->thread_state() != _thread_blocked, "invariant"); 821 822 // Possible mark states: neutral, biased, stack-locked, inflated 823 824 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 825 // CASE: biased 826 BiasedLocking::revoke_and_rebias(h_obj, false, self); 827 assert(!h_obj->mark()->has_bias_pattern(), 828 "biases should be revoked by now"); 829 } 830 831 assert(self == JavaThread::current(), "Can only be called on current thread"); 832 oop obj = h_obj(); 833 markOop mark = ReadStableMark(obj); 834 835 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 836 if (mark->has_locker()) { 837 return self->is_lock_owned((address)mark->locker()) ? 838 owner_self : owner_other; 839 } 840 841 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 842 // The Object:ObjectMonitor relationship is stable as long as we're 843 // not at a safepoint. 844 if (mark->has_monitor()) { 845 void * owner = mark->monitor()->_owner; 846 if (owner == NULL) return owner_none; 847 return (owner == self || 848 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 849 } 850 851 // CASE: neutral 852 assert(mark->is_neutral(), "sanity check"); 853 return owner_none; // it's unlocked 854 } 855 856 // FIXME: jvmti should call this 857 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 858 if (UseBiasedLocking) { 859 if (SafepointSynchronize::is_at_safepoint()) { 860 BiasedLocking::revoke_at_safepoint(h_obj); 861 } else { 862 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 863 } 864 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 865 } 866 867 oop obj = h_obj(); 868 address owner = NULL; 869 870 markOop mark = ReadStableMark(obj); 871 872 // Uncontended case, header points to stack 873 if (mark->has_locker()) { 874 owner = (address) mark->locker(); 875 } 876 877 // Contended case, header points to ObjectMonitor (tagged pointer) 878 if (mark->has_monitor()) { 879 ObjectMonitor* monitor = mark->monitor(); 880 assert(monitor != NULL, "monitor should be non-null"); 881 owner = (address) monitor->owner(); 882 } 883 884 if (owner != NULL) { 885 // owning_thread_from_monitor_owner() may also return NULL here 886 return Threads::owning_thread_from_monitor_owner(owner, doLock); 887 } 888 889 // Unlocked case, header in place 890 // Cannot have assertion since this object may have been 891 // locked by another thread when reaching here. 892 // assert(mark->is_neutral(), "sanity check"); 893 894 return NULL; 895 } 896 // Visitors ... 897 898 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 899 PaddedEnd<ObjectMonitor> * block = (PaddedEnd<ObjectMonitor> *)gBlockList; 900 ObjectMonitor* mid; 901 while (block) { 902 assert(block->object() == CHAINMARKER, "must be a block header"); 903 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 904 mid = (ObjectMonitor *)(block + i); 905 oop object = (oop) mid->object(); 906 if (object != NULL) { 907 closure->do_monitor(mid); 908 } 909 } 910 block = (PaddedEnd<ObjectMonitor> *) block->FreeNext; 911 } 912 } 913 914 // Get the next block in the block list. 915 static 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 } 921 922 923 void ObjectSynchronizer::oops_do(OopClosure* f) { 924 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 925 for (PaddedEnd<ObjectMonitor> * block = 926 (PaddedEnd<ObjectMonitor> *)gBlockList; block != NULL; 927 block = (PaddedEnd<ObjectMonitor> *)next(block)) { 928 assert(block->object() == CHAINMARKER, "must be a block header"); 929 for (int i = 1; i < _BLOCKSIZE; i++) { 930 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 931 if (mid->object() != NULL) { 932 f->do_oop((oop*)mid->object_addr()); 933 } 934 } 935 } 936 } 937 938 939 // ----------------------------------------------------------------------------- 940 // ObjectMonitor Lifecycle 941 // ----------------------- 942 // Inflation unlinks monitors from the global gFreeList and 943 // associates them with objects. Deflation -- which occurs at 944 // STW-time -- disassociates idle monitors from objects. Such 945 // scavenged monitors are returned to the gFreeList. 946 // 947 // The global list is protected by gListLock. All the critical sections 948 // are short and operate in constant-time. 949 // 950 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 951 // 952 // Lifecycle: 953 // -- unassigned and on the global free list 954 // -- unassigned and on a thread's private omFreeList 955 // -- assigned to an object. The object is inflated and the mark refers 956 // to the objectmonitor. 957 958 959 // Constraining monitor pool growth via MonitorBound ... 960 // 961 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 962 // the rate of scavenging is driven primarily by GC. As such, we can find 963 // an inordinate number of monitors in circulation. 964 // To avoid that scenario we can artificially induce a STW safepoint 965 // if the pool appears to be growing past some reasonable bound. 966 // Generally we favor time in space-time tradeoffs, but as there's no 967 // natural back-pressure on the # of extant monitors we need to impose some 968 // type of limit. Beware that if MonitorBound is set to too low a value 969 // we could just loop. In addition, if MonitorBound is set to a low value 970 // we'll incur more safepoints, which are harmful to performance. 971 // See also: GuaranteedSafepointInterval 972 // 973 // The current implementation uses asynchronous VM operations. 974 975 static void InduceScavenge(Thread * Self, const char * Whence) { 976 // Induce STW safepoint to trim monitors 977 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 978 // More precisely, trigger an asynchronous STW safepoint as the number 979 // of active monitors passes the specified threshold. 980 // TODO: assert thread state is reasonable 981 982 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 983 if (ObjectMonitor::Knob_Verbose) { 984 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; 985 ::fflush(stdout); 986 } 987 // Induce a 'null' safepoint to scavenge monitors 988 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 989 // to the VMthread and have a lifespan longer than that of this activation record. 990 // The VMThread will delete the op when completed. 991 VMThread::execute(new VM_ForceAsyncSafepoint()); 992 993 if (ObjectMonitor::Knob_Verbose) { 994 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; 995 ::fflush(stdout); 996 } 997 } 998 } 999 1000 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1001 ObjectMonitor* mid; 1002 int in_use_tally = 0; 1003 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1004 in_use_tally++; 1005 } 1006 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1007 1008 int free_tally = 0; 1009 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1010 free_tally++; 1011 } 1012 assert(free_tally == Self->omFreeCount, "free count off"); 1013 } 1014 1015 ObjectMonitor * NOINLINE ObjectSynchronizer::omAlloc(Thread * Self) { 1016 // A large MAXPRIVATE value reduces both list lock contention 1017 // and list coherency traffic, but also tends to increase the 1018 // number of objectMonitors in circulation as well as the STW 1019 // scavenge costs. As usual, we lean toward time in space-time 1020 // tradeoffs. 1021 const int MAXPRIVATE = 1024; 1022 for (;;) { 1023 ObjectMonitor * m; 1024 1025 // 1: try to allocate from the thread's local omFreeList. 1026 // Threads will attempt to allocate first from their local list, then 1027 // from the global list, and only after those attempts fail will the thread 1028 // attempt to instantiate new monitors. Thread-local free lists take 1029 // heat off the gListLock and improve allocation latency, as well as reducing 1030 // coherency traffic on the shared global list. 1031 m = Self->omFreeList; 1032 if (m != NULL) { 1033 Self->omFreeList = m->FreeNext; 1034 Self->omFreeCount--; 1035 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1036 guarantee(m->object() == NULL, "invariant"); 1037 if (MonitorInUseLists) { 1038 m->FreeNext = Self->omInUseList; 1039 Self->omInUseList = m; 1040 Self->omInUseCount++; 1041 if (ObjectMonitor::Knob_VerifyInUse) { 1042 verifyInUse(Self); 1043 } 1044 } else { 1045 m->FreeNext = NULL; 1046 } 1047 return m; 1048 } 1049 1050 // 2: try to allocate from the global gFreeList 1051 // CONSIDER: use muxTry() instead of muxAcquire(). 1052 // If the muxTry() fails then drop immediately into case 3. 1053 // If we're using thread-local free lists then try 1054 // to reprovision the caller's free list. 1055 if (gFreeList != NULL) { 1056 // Reprovision the thread's omFreeList. 1057 // Use bulk transfers to reduce the allocation rate and heat 1058 // on various locks. 1059 Thread::muxAcquire(&gListLock, "omAlloc"); 1060 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1061 gMonitorFreeCount--; 1062 ObjectMonitor * take = gFreeList; 1063 gFreeList = take->FreeNext; 1064 guarantee(take->object() == NULL, "invariant"); 1065 guarantee(!take->is_busy(), "invariant"); 1066 take->Recycle(); 1067 omRelease(Self, take, false); 1068 } 1069 Thread::muxRelease(&gListLock); 1070 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1071 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1072 TEVENT(omFirst - reprovision); 1073 1074 const int mx = MonitorBound; 1075 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1076 // We can't safely induce a STW safepoint from omAlloc() as our thread 1077 // state may not be appropriate for such activities and callers may hold 1078 // naked oops, so instead we defer the action. 1079 InduceScavenge(Self, "omAlloc"); 1080 } 1081 continue; 1082 } 1083 1084 // 3: allocate a block of new ObjectMonitors 1085 // Both the local and global free lists are empty -- resort to malloc(). 1086 // In the current implementation objectMonitors are TSM - immortal. 1087 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1088 // each ObjectMonitor to start at the beginning of a cache line, 1089 // so we use align_size_up(). 1090 // A better solution would be to use C++ placement-new. 1091 // BEWARE: As it stands currently, we don't run the ctors! 1092 assert(_BLOCKSIZE > 1, "invariant"); 1093 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1094 PaddedEnd<ObjectMonitor> * temp; 1095 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1096 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1097 mtInternal); 1098 temp = (PaddedEnd<ObjectMonitor> *) 1099 align_size_up((intptr_t)real_malloc_addr, 1100 DEFAULT_CACHE_LINE_SIZE); 1101 1102 // NOTE: (almost) no way to recover if allocation failed. 1103 // We might be able to induce a STW safepoint and scavenge enough 1104 // objectMonitors to permit progress. 1105 if (temp == NULL) { 1106 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1107 "Allocate ObjectMonitors"); 1108 } 1109 (void)memset((void *) temp, 0, neededsize); 1110 1111 // Format the block. 1112 // initialize the linked list, each monitor points to its next 1113 // forming the single linked free list, the very first monitor 1114 // will points to next block, which forms the block list. 1115 // The trick of using the 1st element in the block as gBlockList 1116 // linkage should be reconsidered. A better implementation would 1117 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1118 1119 for (int i = 1; i < _BLOCKSIZE; i++) { 1120 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1121 } 1122 1123 // terminate the last monitor as the end of list 1124 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1125 1126 // Element [0] is reserved for global list linkage 1127 temp[0].set_object(CHAINMARKER); 1128 1129 // Consider carving out this thread's current request from the 1130 // block in hand. This avoids some lock traffic and redundant 1131 // list activity. 1132 1133 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1134 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1135 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1136 gMonitorPopulation += _BLOCKSIZE-1; 1137 gMonitorFreeCount += _BLOCKSIZE-1; 1138 1139 // Add the new block to the list of extant blocks (gBlockList). 1140 // The very first objectMonitor in a block is reserved and dedicated. 1141 // It serves as blocklist "next" linkage. 1142 temp[0].FreeNext = gBlockList; 1143 gBlockList = temp; 1144 1145 // Add the new string of objectMonitors to the global free list 1146 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1147 gFreeList = temp + 1; 1148 Thread::muxRelease(&gListLock); 1149 TEVENT(Allocate block of monitors); 1150 } 1151 } 1152 1153 // Place "m" on the caller's private per-thread omFreeList. 1154 // In practice there's no need to clamp or limit the number of 1155 // monitors on a thread's omFreeList as the only time we'll call 1156 // omRelease is to return a monitor to the free list after a CAS 1157 // attempt failed. This doesn't allow unbounded #s of monitors to 1158 // accumulate on a thread's free list. 1159 // 1160 // Key constraint: all ObjectMonitors on a thread's free list and the global 1161 // free list must have their object field set to null. This prevents the 1162 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1163 1164 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1165 bool fromPerThreadAlloc) { 1166 guarantee(m->object() == NULL, "invariant"); 1167 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1168 // Remove from omInUseList 1169 if (MonitorInUseLists && fromPerThreadAlloc) { 1170 ObjectMonitor* cur_mid_in_use = NULL; 1171 bool extracted = false; 1172 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1173 if (m == mid) { 1174 // extract from per-thread in-use list 1175 if (mid == Self->omInUseList) { 1176 Self->omInUseList = mid->FreeNext; 1177 } else if (cur_mid_in_use != NULL) { 1178 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1179 } 1180 extracted = true; 1181 Self->omInUseCount--; 1182 if (ObjectMonitor::Knob_VerifyInUse) { 1183 verifyInUse(Self); 1184 } 1185 break; 1186 } 1187 } 1188 assert(extracted, "Should have extracted from in-use list"); 1189 } 1190 1191 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1192 m->FreeNext = Self->omFreeList; 1193 Self->omFreeList = m; 1194 Self->omFreeCount++; 1195 } 1196 1197 // Return the monitors of a moribund thread's local free list to 1198 // the global free list. Typically a thread calls omFlush() when 1199 // it's dying. We could also consider having the VM thread steal 1200 // monitors from threads that have not run java code over a few 1201 // consecutive STW safepoints. Relatedly, we might decay 1202 // omFreeProvision at STW safepoints. 1203 // 1204 // Also return the monitors of a moribund thread's omInUseList to 1205 // a global gOmInUseList under the global list lock so these 1206 // will continue to be scanned. 1207 // 1208 // We currently call omFlush() from the Thread:: dtor _after the thread 1209 // has been excised from the thread list and is no longer a mutator. 1210 // That means that omFlush() can run concurrently with a safepoint and 1211 // the scavenge operator. Calling omFlush() from JavaThread::exit() might 1212 // be a better choice as we could safely reason that that the JVM is 1213 // not at a safepoint at the time of the call, and thus there could 1214 // be not inopportune interleavings between omFlush() and the scavenge 1215 // operator. 1216 1217 void ObjectSynchronizer::omFlush(Thread * Self) { 1218 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1219 Self->omFreeList = NULL; 1220 ObjectMonitor * tail = NULL; 1221 int tally = 0; 1222 if (list != NULL) { 1223 ObjectMonitor * s; 1224 // The thread is going away, the per-thread free monitors 1225 // are freed via set_owner(NULL) 1226 // Link them to tail, which will be linked into the global free list 1227 // gFreeList below, under the gListLock 1228 for (s = list; s != NULL; s = s->FreeNext) { 1229 tally++; 1230 tail = s; 1231 guarantee(s->object() == NULL, "invariant"); 1232 guarantee(!s->is_busy(), "invariant"); 1233 s->set_owner(NULL); // redundant but good hygiene 1234 TEVENT(omFlush - Move one); 1235 } 1236 guarantee(tail != NULL && list != NULL, "invariant"); 1237 } 1238 1239 ObjectMonitor * inUseList = Self->omInUseList; 1240 ObjectMonitor * inUseTail = NULL; 1241 int inUseTally = 0; 1242 if (inUseList != NULL) { 1243 Self->omInUseList = NULL; 1244 ObjectMonitor *cur_om; 1245 // The thread is going away, however the omInUseList inflated 1246 // monitors may still be in-use by other threads. 1247 // Link them to inUseTail, which will be linked into the global in-use list 1248 // gOmInUseList below, under the gListLock 1249 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1250 inUseTail = cur_om; 1251 inUseTally++; 1252 } 1253 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1254 Self->omInUseCount = 0; 1255 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1256 } 1257 1258 Thread::muxAcquire(&gListLock, "omFlush"); 1259 if (tail != NULL) { 1260 tail->FreeNext = gFreeList; 1261 gFreeList = list; 1262 gMonitorFreeCount += tally; 1263 } 1264 1265 if (inUseTail != NULL) { 1266 inUseTail->FreeNext = gOmInUseList; 1267 gOmInUseList = inUseList; 1268 gOmInUseCount += inUseTally; 1269 } 1270 1271 Thread::muxRelease(&gListLock); 1272 TEVENT(omFlush); 1273 } 1274 1275 // Fast path code shared by multiple functions 1276 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1277 markOop mark = obj->mark(); 1278 if (mark->has_monitor()) { 1279 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1280 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1281 return mark->monitor(); 1282 } 1283 return ObjectSynchronizer::inflate(Thread::current(), obj); 1284 } 1285 1286 1287 ObjectMonitor * NOINLINE ObjectSynchronizer::inflate(Thread * Self, 1288 oop object) { 1289 // Inflate mutates the heap ... 1290 // Relaxing assertion for bug 6320749. 1291 assert(Universe::verify_in_progress() || 1292 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1293 1294 for (;;) { 1295 const markOop mark = object->mark(); 1296 assert(!mark->has_bias_pattern(), "invariant"); 1297 1298 // The mark can be in one of the following states: 1299 // * Inflated - just return 1300 // * Stack-locked - coerce it to inflated 1301 // * INFLATING - busy wait for conversion to complete 1302 // * Neutral - aggressively inflate the object. 1303 // * BIASED - Illegal. We should never see this 1304 1305 // CASE: inflated 1306 if (mark->has_monitor()) { 1307 ObjectMonitor * inf = mark->monitor(); 1308 assert(inf->header()->is_neutral(), "invariant"); 1309 assert(inf->object() == object, "invariant"); 1310 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1311 return inf; 1312 } 1313 1314 // CASE: inflation in progress - inflating over a stack-lock. 1315 // Some other thread is converting from stack-locked to inflated. 1316 // Only that thread can complete inflation -- other threads must wait. 1317 // The INFLATING value is transient. 1318 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1319 // We could always eliminate polling by parking the thread on some auxiliary list. 1320 if (mark == markOopDesc::INFLATING()) { 1321 TEVENT(Inflate: spin while INFLATING); 1322 ReadStableMark(object); 1323 continue; 1324 } 1325 1326 // CASE: stack-locked 1327 // Could be stack-locked either by this thread or by some other thread. 1328 // 1329 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1330 // to install INFLATING into the mark word. We originally installed INFLATING, 1331 // allocated the objectmonitor, and then finally STed the address of the 1332 // objectmonitor into the mark. This was correct, but artificially lengthened 1333 // the interval in which INFLATED appeared in the mark, thus increasing 1334 // the odds of inflation contention. 1335 // 1336 // We now use per-thread private objectmonitor free lists. 1337 // These list are reprovisioned from the global free list outside the 1338 // critical INFLATING...ST interval. A thread can transfer 1339 // multiple objectmonitors en-mass from the global free list to its local free list. 1340 // This reduces coherency traffic and lock contention on the global free list. 1341 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1342 // before or after the CAS(INFLATING) operation. 1343 // See the comments in omAlloc(). 1344 1345 if (mark->has_locker()) { 1346 ObjectMonitor * m = omAlloc(Self); 1347 // Optimistically prepare the objectmonitor - anticipate successful CAS 1348 // We do this before the CAS in order to minimize the length of time 1349 // in which INFLATING appears in the mark. 1350 m->Recycle(); 1351 m->_Responsible = NULL; 1352 m->_recursions = 0; 1353 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1354 1355 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark); 1356 if (cmp != mark) { 1357 omRelease(Self, m, true); 1358 continue; // Interference -- just retry 1359 } 1360 1361 // We've successfully installed INFLATING (0) into the mark-word. 1362 // This is the only case where 0 will appear in a mark-work. 1363 // Only the singular thread that successfully swings the mark-word 1364 // to 0 can perform (or more precisely, complete) inflation. 1365 // 1366 // Why do we CAS a 0 into the mark-word instead of just CASing the 1367 // mark-word from the stack-locked value directly to the new inflated state? 1368 // Consider what happens when a thread unlocks a stack-locked object. 1369 // It attempts to use CAS to swing the displaced header value from the 1370 // on-stack basiclock back into the object header. Recall also that the 1371 // header value (hashcode, etc) can reside in (a) the object header, or 1372 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1373 // header in an objectMonitor. The inflate() routine must copy the header 1374 // value from the basiclock on the owner's stack to the objectMonitor, all 1375 // the while preserving the hashCode stability invariants. If the owner 1376 // decides to release the lock while the value is 0, the unlock will fail 1377 // and control will eventually pass from slow_exit() to inflate. The owner 1378 // will then spin, waiting for the 0 value to disappear. Put another way, 1379 // the 0 causes the owner to stall if the owner happens to try to 1380 // drop the lock (restoring the header from the basiclock to the object) 1381 // while inflation is in-progress. This protocol avoids races that might 1382 // would otherwise permit hashCode values to change or "flicker" for an object. 1383 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1384 // 0 serves as a "BUSY" inflate-in-progress indicator. 1385 1386 1387 // fetch the displaced mark from the owner's stack. 1388 // The owner can't die or unwind past the lock while our INFLATING 1389 // object is in the mark. Furthermore the owner can't complete 1390 // an unlock on the object, either. 1391 markOop dmw = mark->displaced_mark_helper(); 1392 assert(dmw->is_neutral(), "invariant"); 1393 1394 // Setup monitor fields to proper values -- prepare the monitor 1395 m->set_header(dmw); 1396 1397 // Optimization: if the mark->locker stack address is associated 1398 // with this thread we could simply set m->_owner = Self. 1399 // Note that a thread can inflate an object 1400 // that it has stack-locked -- as might happen in wait() -- directly 1401 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1402 m->set_owner(mark->locker()); 1403 m->set_object(object); 1404 // TODO-FIXME: assert BasicLock->dhw != 0. 1405 1406 // Must preserve store ordering. The monitor state must 1407 // be stable at the time of publishing the monitor address. 1408 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1409 object->release_set_mark(markOopDesc::encode(m)); 1410 1411 // Hopefully the performance counters are allocated on distinct cache lines 1412 // to avoid false sharing on MP systems ... 1413 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc(); 1414 TEVENT(Inflate: overwrite stacklock); 1415 if (TraceMonitorInflation) { 1416 if (object->is_instance()) { 1417 ResourceMark rm; 1418 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1419 (void *) object, (intptr_t) object->mark(), 1420 object->klass()->external_name()); 1421 } 1422 } 1423 return m; 1424 } 1425 1426 // CASE: neutral 1427 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1428 // If we know we're inflating for entry it's better to inflate by swinging a 1429 // pre-locked objectMonitor pointer into the object header. A successful 1430 // CAS inflates the object *and* confers ownership to the inflating thread. 1431 // In the current implementation we use a 2-step mechanism where we CAS() 1432 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1433 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1434 // would be useful. 1435 1436 assert(mark->is_neutral(), "invariant"); 1437 ObjectMonitor * m = omAlloc(Self); 1438 // prepare m for installation - set monitor to initial state 1439 m->Recycle(); 1440 m->set_header(mark); 1441 m->set_owner(NULL); 1442 m->set_object(object); 1443 m->_recursions = 0; 1444 m->_Responsible = NULL; 1445 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1446 1447 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1448 m->set_object(NULL); 1449 m->set_owner(NULL); 1450 m->Recycle(); 1451 omRelease(Self, m, true); 1452 m = NULL; 1453 continue; 1454 // interference - the markword changed - just retry. 1455 // The state-transitions are one-way, so there's no chance of 1456 // live-lock -- "Inflated" is an absorbing state. 1457 } 1458 1459 // Hopefully the performance counters are allocated on distinct 1460 // cache lines to avoid false sharing on MP systems ... 1461 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc(); 1462 TEVENT(Inflate: overwrite neutral); 1463 if (TraceMonitorInflation) { 1464 if (object->is_instance()) { 1465 ResourceMark rm; 1466 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1467 (void *) object, (intptr_t) object->mark(), 1468 object->klass()->external_name()); 1469 } 1470 } 1471 return m; 1472 } 1473 } 1474 1475 1476 // Deflate_idle_monitors() is called at all safepoints, immediately 1477 // after all mutators are stopped, but before any objects have moved. 1478 // It traverses the list of known monitors, deflating where possible. 1479 // The scavenged monitor are returned to the monitor free list. 1480 // 1481 // Beware that we scavenge at *every* stop-the-world point. 1482 // Having a large number of monitors in-circulation negatively 1483 // impacts the performance of some applications (e.g., PointBase). 1484 // Broadly, we want to minimize the # of monitors in circulation. 1485 // 1486 // We have added a flag, MonitorInUseLists, which creates a list 1487 // of active monitors for each thread. deflate_idle_monitors() 1488 // only scans the per-thread in-use lists. omAlloc() puts all 1489 // assigned monitors on the per-thread list. deflate_idle_monitors() 1490 // returns the non-busy monitors to the global free list. 1491 // When a thread dies, omFlush() adds the list of active monitors for 1492 // that thread to a global gOmInUseList acquiring the 1493 // global list lock. deflate_idle_monitors() acquires the global 1494 // list lock to scan for non-busy monitors to the global free list. 1495 // An alternative could have used a single global in-use list. The 1496 // downside would have been the additional cost of acquiring the global list lock 1497 // for every omAlloc(). 1498 // 1499 // Perversely, the heap size -- and thus the STW safepoint rate -- 1500 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1501 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1502 // This is an unfortunate aspect of this design. 1503 1504 enum ManifestConstants { 1505 ClearResponsibleAtSTW = 0 1506 }; 1507 1508 // Deflate a single monitor if not in-use 1509 // Return true if deflated, false if in-use 1510 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1511 ObjectMonitor** freeHeadp, 1512 ObjectMonitor** freeTailp) { 1513 bool deflated; 1514 // Normal case ... The monitor is associated with obj. 1515 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1516 guarantee(mid == obj->mark()->monitor(), "invariant"); 1517 guarantee(mid->header()->is_neutral(), "invariant"); 1518 1519 if (mid->is_busy()) { 1520 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1521 deflated = false; 1522 } else { 1523 // Deflate the monitor if it is no longer being used 1524 // It's idle - scavenge and return to the global free list 1525 // plain old deflation ... 1526 TEVENT(deflate_idle_monitors - scavenge1); 1527 if (TraceMonitorInflation) { 1528 if (obj->is_instance()) { 1529 ResourceMark rm; 1530 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1531 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name()); 1532 } 1533 } 1534 1535 // Restore the header back to obj 1536 obj->release_set_mark(mid->header()); 1537 mid->clear(); 1538 1539 assert(mid->object() == NULL, "invariant"); 1540 1541 // Move the object to the working free list defined by freeHeadp, freeTailp 1542 if (*freeHeadp == NULL) *freeHeadp = mid; 1543 if (*freeTailp != NULL) { 1544 ObjectMonitor * prevtail = *freeTailp; 1545 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1546 prevtail->FreeNext = mid; 1547 } 1548 *freeTailp = mid; 1549 deflated = true; 1550 } 1551 return deflated; 1552 } 1553 1554 // Walk a given monitor list, and deflate idle monitors 1555 // The given list could be a per-thread list or a global list 1556 // Caller acquires gListLock 1557 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1558 ObjectMonitor** freeHeadp, 1559 ObjectMonitor** freeTailp) { 1560 ObjectMonitor* mid; 1561 ObjectMonitor* next; 1562 ObjectMonitor* cur_mid_in_use = NULL; 1563 int deflated_count = 0; 1564 1565 for (mid = *listHeadp; mid != NULL;) { 1566 oop obj = (oop) mid->object(); 1567 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1568 // if deflate_monitor succeeded, 1569 // extract from per-thread in-use list 1570 if (mid == *listHeadp) { 1571 *listHeadp = mid->FreeNext; 1572 } else if (cur_mid_in_use != NULL) { 1573 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1574 } 1575 next = mid->FreeNext; 1576 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1577 mid = next; 1578 deflated_count++; 1579 } else { 1580 cur_mid_in_use = mid; 1581 mid = mid->FreeNext; 1582 } 1583 } 1584 return deflated_count; 1585 } 1586 1587 void ObjectSynchronizer::deflate_idle_monitors() { 1588 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1589 int nInuse = 0; // currently associated with objects 1590 int nInCirculation = 0; // extant 1591 int nScavenged = 0; // reclaimed 1592 bool deflated = false; 1593 1594 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1595 ObjectMonitor * freeTailp = NULL; 1596 1597 TEVENT(deflate_idle_monitors); 1598 // Prevent omFlush from changing mids in Thread dtor's during deflation 1599 // And in case the vm thread is acquiring a lock during a safepoint 1600 // See e.g. 6320749 1601 Thread::muxAcquire(&gListLock, "scavenge - return"); 1602 1603 if (MonitorInUseLists) { 1604 int inUse = 0; 1605 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { 1606 nInCirculation+= cur->omInUseCount; 1607 int deflated_count = deflate_monitor_list(cur->omInUseList_addr(), &freeHeadp, &freeTailp); 1608 cur->omInUseCount-= deflated_count; 1609 if (ObjectMonitor::Knob_VerifyInUse) { 1610 verifyInUse(cur); 1611 } 1612 nScavenged += deflated_count; 1613 nInuse += cur->omInUseCount; 1614 } 1615 1616 // For moribund threads, scan gOmInUseList 1617 if (gOmInUseList) { 1618 nInCirculation += gOmInUseCount; 1619 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1620 gOmInUseCount-= deflated_count; 1621 nScavenged += deflated_count; 1622 nInuse += gOmInUseCount; 1623 } 1624 1625 } else for (PaddedEnd<ObjectMonitor> * block = 1626 (PaddedEnd<ObjectMonitor> *)gBlockList; block != NULL; 1627 block = (PaddedEnd<ObjectMonitor> *)next(block)) { 1628 // Iterate over all extant monitors - Scavenge all idle monitors. 1629 assert(block->object() == CHAINMARKER, "must be a block header"); 1630 nInCirculation += _BLOCKSIZE; 1631 for (int i = 1; i < _BLOCKSIZE; i++) { 1632 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1633 oop obj = (oop) mid->object(); 1634 1635 if (obj == NULL) { 1636 // The monitor is not associated with an object. 1637 // The monitor should either be a thread-specific private 1638 // free list or the global free list. 1639 // obj == NULL IMPLIES mid->is_busy() == 0 1640 guarantee(!mid->is_busy(), "invariant"); 1641 continue; 1642 } 1643 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1644 1645 if (deflated) { 1646 mid->FreeNext = NULL; 1647 nScavenged++; 1648 } else { 1649 nInuse++; 1650 } 1651 } 1652 } 1653 1654 gMonitorFreeCount += nScavenged; 1655 1656 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1657 1658 if (ObjectMonitor::Knob_Verbose) { 1659 ::printf("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n", 1660 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, 1661 gMonitorPopulation, gMonitorFreeCount); 1662 ::fflush(stdout); 1663 } 1664 1665 ForceMonitorScavenge = 0; // Reset 1666 1667 // Move the scavenged monitors back to the global free list. 1668 if (freeHeadp != NULL) { 1669 guarantee(freeTailp != NULL && nScavenged > 0, "invariant"); 1670 assert(freeTailp->FreeNext == NULL, "invariant"); 1671 // constant-time list splice - prepend scavenged segment to gFreeList 1672 freeTailp->FreeNext = gFreeList; 1673 gFreeList = freeHeadp; 1674 } 1675 Thread::muxRelease(&gListLock); 1676 1677 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged); 1678 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation); 1679 1680 // TODO: Add objectMonitor leak detection. 1681 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1682 GVars.stwRandom = os::random(); 1683 GVars.stwCycle++; 1684 } 1685 1686 // Monitor cleanup on JavaThread::exit 1687 1688 // Iterate through monitor cache and attempt to release thread's monitors 1689 // Gives up on a particular monitor if an exception occurs, but continues 1690 // the overall iteration, swallowing the exception. 1691 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1692 private: 1693 TRAPS; 1694 1695 public: 1696 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1697 void do_monitor(ObjectMonitor* mid) { 1698 if (mid->owner() == THREAD) { 1699 (void)mid->complete_exit(CHECK); 1700 } 1701 } 1702 }; 1703 1704 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1705 // ignored. This is meant to be called during JNI thread detach which assumes 1706 // all remaining monitors are heavyweight. All exceptions are swallowed. 1707 // Scanning the extant monitor list can be time consuming. 1708 // A simple optimization is to add a per-thread flag that indicates a thread 1709 // called jni_monitorenter() during its lifetime. 1710 // 1711 // Instead of No_Savepoint_Verifier it might be cheaper to 1712 // use an idiom of the form: 1713 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1714 // <code that must not run at safepoint> 1715 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1716 // Since the tests are extremely cheap we could leave them enabled 1717 // for normal product builds. 1718 1719 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1720 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1721 No_Safepoint_Verifier nsv; 1722 ReleaseJavaMonitorsClosure rjmc(THREAD); 1723 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1724 ObjectSynchronizer::monitors_iterate(&rjmc); 1725 Thread::muxRelease(&gListLock); 1726 THREAD->clear_pending_exception(); 1727 } 1728 1729 //------------------------------------------------------------------------------ 1730 // Debugging code 1731 1732 void ObjectSynchronizer::sanity_checks(const bool verbose, 1733 const uint cache_line_size, 1734 int *error_cnt_ptr, 1735 int *warning_cnt_ptr) { 1736 u_char *addr_begin = (u_char*)&GVars; 1737 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1738 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1739 1740 if (verbose) { 1741 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1742 sizeof(SharedGlobals)); 1743 } 1744 1745 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1746 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1747 1748 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1749 if (verbose) { 1750 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1751 } 1752 1753 if (cache_line_size != 0) { 1754 // We were able to determine the L1 data cache line size so 1755 // do some cache line specific sanity checks 1756 1757 if (offset_stwRandom < cache_line_size) { 1758 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1759 "to the struct beginning than a cache line which permits " 1760 "false sharing."); 1761 (*warning_cnt_ptr)++; 1762 } 1763 1764 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1765 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1766 "SharedGlobals.hcSequence fields are closer than a cache " 1767 "line which permits false sharing."); 1768 (*warning_cnt_ptr)++; 1769 } 1770 1771 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1772 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1773 "to the struct end than a cache line which permits false " 1774 "sharing."); 1775 (*warning_cnt_ptr)++; 1776 } 1777 } 1778 } 1779 1780 #ifndef PRODUCT 1781 1782 // Verify all monitors in the monitor cache, the verification is weak. 1783 void ObjectSynchronizer::verify() { 1784 PaddedEnd<ObjectMonitor> * block = (PaddedEnd<ObjectMonitor> *)gBlockList; 1785 ObjectMonitor* mid; 1786 while (block) { 1787 assert(block->object() == CHAINMARKER, "must be a block header"); 1788 for (int i = 1; i < _BLOCKSIZE; i++) { 1789 mid = (ObjectMonitor *)(block + i); 1790 oop object = (oop) mid->object(); 1791 if (object != NULL) { 1792 mid->verify(); 1793 } 1794 } 1795 block = (PaddedEnd<ObjectMonitor> *) block->FreeNext; 1796 } 1797 } 1798 1799 // Check if monitor belongs to the monitor cache 1800 // The list is grow-only so it's *relatively* safe to traverse 1801 // the list of extant blocks without taking a lock. 1802 1803 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1804 PaddedEnd<ObjectMonitor> * block = (PaddedEnd<ObjectMonitor> *)gBlockList; 1805 1806 while (block) { 1807 assert(block->object() == CHAINMARKER, "must be a block header"); 1808 if (monitor > (ObjectMonitor *)&block[0] && 1809 monitor < (ObjectMonitor *)&block[_BLOCKSIZE]) { 1810 address mon = (address) monitor; 1811 address blk = (address) block; 1812 size_t diff = mon - blk; 1813 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "check"); 1814 return 1; 1815 } 1816 block = (PaddedEnd<ObjectMonitor> *) block->FreeNext; 1817 } 1818 return 0; 1819 } 1820 1821 #endif