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 "runtime/vframe.hpp" 44 #include "utilities/dtrace.hpp" 45 #include "utilities/events.hpp" 46 #include "utilities/preserveException.hpp" 47 48 #if defined(__GNUC__) && !defined(PPC64) 49 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 50 #define NOINLINE __attribute__((noinline)) 51 #else 52 #define NOINLINE 53 #endif 54 55 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 56 57 // The "core" versions of monitor enter and exit reside in this file. 58 // The interpreter and compilers contain specialized transliterated 59 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 60 // for instance. If you make changes here, make sure to modify the 61 // interpreter, and both C1 and C2 fast-path inline locking code emission. 62 // 63 // ----------------------------------------------------------------------------- 64 65 #ifdef DTRACE_ENABLED 66 67 // Only bother with this argument setup if dtrace is available 68 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 69 70 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 71 char* bytes = NULL; \ 72 int len = 0; \ 73 jlong jtid = SharedRuntime::get_java_tid(thread); \ 74 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 75 if (klassname != NULL) { \ 76 bytes = (char*)klassname->bytes(); \ 77 len = klassname->utf8_length(); \ 78 } 79 80 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 81 { \ 82 if (DTraceMonitorProbes) { \ 83 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 84 HOTSPOT_MONITOR_WAIT(jtid, \ 85 (uintptr_t)(monitor), bytes, len, (millis)); \ 86 } \ 87 } 88 89 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 90 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 91 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 92 93 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 94 { \ 95 if (DTraceMonitorProbes) { \ 96 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 97 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 98 (uintptr_t)(monitor), bytes, len); \ 99 } \ 100 } 101 102 #else // ndef DTRACE_ENABLED 103 104 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 105 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 106 107 #endif // ndef DTRACE_ENABLED 108 109 // This exists only as a workaround of dtrace bug 6254741 110 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 111 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 112 return 0; 113 } 114 115 #define NINFLATIONLOCKS 256 116 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS]; 117 118 // global list of blocks of monitors 119 // gBlockList is really PaddedEnd<ObjectMonitor> *, but we don't 120 // want to expose the PaddedEnd template more than necessary. 121 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL; 122 // global monitor free list 123 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 124 // global monitor in-use list, for moribund threads, 125 // monitors they inflated need to be scanned for deflation 126 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 127 // count of entries in gOmInUseList 128 int ObjectSynchronizer::gOmInUseCount = 0; 129 130 static volatile intptr_t gListLock = 0; // protects global monitor lists 131 static volatile int gMonitorFreeCount = 0; // # on gFreeList 132 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation 133 134 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 135 136 137 // =====================> Quick functions 138 139 // The quick_* forms are special fast-path variants used to improve 140 // performance. In the simplest case, a "quick_*" implementation could 141 // simply return false, in which case the caller will perform the necessary 142 // state transitions and call the slow-path form. 143 // The fast-path is designed to handle frequently arising cases in an efficient 144 // manner and is just a degenerate "optimistic" variant of the slow-path. 145 // returns true -- to indicate the call was satisfied. 146 // returns false -- to indicate the call needs the services of the slow-path. 147 // A no-loitering ordinance is in effect for code in the quick_* family 148 // operators: safepoints or indefinite blocking (blocking that might span a 149 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 150 // entry. 151 // 152 // Consider: An interesting optimization is to have the JIT recognize the 153 // following common idiom: 154 // synchronized (someobj) { .... ; notify(); } 155 // That is, we find a notify() or notifyAll() call that immediately precedes 156 // the monitorexit operation. In that case the JIT could fuse the operations 157 // into a single notifyAndExit() runtime primitive. 158 159 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) { 160 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 161 assert(self->is_Java_thread(), "invariant"); 162 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 163 No_Safepoint_Verifier nsv; 164 if (obj == NULL) return false; // slow-path for invalid obj 165 const markOop mark = obj->mark(); 166 167 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) { 168 // Degenerate notify 169 // stack-locked by caller so by definition the implied waitset is empty. 170 return true; 171 } 172 173 if (mark->has_monitor()) { 174 ObjectMonitor * const mon = mark->monitor(); 175 assert(mon->object() == obj, "invariant"); 176 if (mon->owner() != self) return false; // slow-path for IMS exception 177 178 if (mon->first_waiter() != NULL) { 179 // We have one or more waiters. Since this is an inflated monitor 180 // that we own, we can transfer one or more threads from the waitset 181 // to the entrylist here and now, avoiding the slow-path. 182 if (all) { 183 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 184 } else { 185 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 186 } 187 int tally = 0; 188 do { 189 mon->INotify(self); 190 ++tally; 191 } while (mon->first_waiter() != NULL && all); 192 OM_PERFDATA_OP(Notifications, inc(tally)); 193 } 194 return true; 195 } 196 197 // biased locking and any other IMS exception states take the slow-path 198 return false; 199 } 200 201 202 // The LockNode emitted directly at the synchronization site would have 203 // been too big if it were to have included support for the cases of inflated 204 // recursive enter and exit, so they go here instead. 205 // Note that we can't safely call AsyncPrintJavaStack() from within 206 // quick_enter() as our thread state remains _in_Java. 207 208 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self, 209 BasicLock * Lock) { 210 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 211 assert(Self->is_Java_thread(), "invariant"); 212 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant"); 213 No_Safepoint_Verifier nsv; 214 if (obj == NULL) return false; // Need to throw NPE 215 const markOop mark = obj->mark(); 216 217 if (mark->has_monitor()) { 218 ObjectMonitor * const m = mark->monitor(); 219 assert(m->object() == obj, "invariant"); 220 Thread * const owner = (Thread *) m->_owner; 221 222 // Lock contention and Transactional Lock Elision (TLE) diagnostics 223 // and observability 224 // Case: light contention possibly amenable to TLE 225 // Case: TLE inimical operations such as nested/recursive synchronization 226 227 if (owner == Self) { 228 m->_recursions++; 229 return true; 230 } 231 232 if (owner == NULL && 233 Atomic::cmpxchg_ptr(Self, &(m->_owner), NULL) == NULL) { 234 assert(m->_recursions == 0, "invariant"); 235 assert(m->_owner == Self, "invariant"); 236 return true; 237 } 238 } 239 240 // Note that we could inflate in quick_enter. 241 // This is likely a useful optimization 242 // Critically, in quick_enter() we must not: 243 // -- perform bias revocation, or 244 // -- block indefinitely, or 245 // -- reach a safepoint 246 247 return false; // revert to slow-path 248 } 249 250 // ----------------------------------------------------------------------------- 251 // Fast Monitor Enter/Exit 252 // This the fast monitor enter. The interpreter and compiler use 253 // some assembly copies of this code. Make sure update those code 254 // if the following function is changed. The implementation is 255 // extremely sensitive to race condition. Be careful. 256 257 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 258 bool attempt_rebias, TRAPS) { 259 if (UseBiasedLocking) { 260 if (!SafepointSynchronize::is_at_safepoint()) { 261 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 262 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 263 return; 264 } 265 } else { 266 assert(!attempt_rebias, "can not rebias toward VM thread"); 267 BiasedLocking::revoke_at_safepoint(obj); 268 } 269 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 270 } 271 272 slow_enter(obj, lock, THREAD); 273 } 274 275 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 276 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); 277 // if displaced header is null, the previous enter is recursive enter, no-op 278 markOop dhw = lock->displaced_header(); 279 markOop mark; 280 if (dhw == NULL) { 281 // Recursive stack-lock. 282 // Diagnostics -- Could be: stack-locked, inflating, inflated. 283 mark = object->mark(); 284 assert(!mark->is_neutral(), "invariant"); 285 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { 286 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant"); 287 } 288 if (mark->has_monitor()) { 289 ObjectMonitor * m = mark->monitor(); 290 assert(((oop)(m->object()))->mark() == mark, "invariant"); 291 assert(m->is_entered(THREAD), "invariant"); 292 } 293 return; 294 } 295 296 mark = object->mark(); 297 298 // If the object is stack-locked by the current thread, try to 299 // swing the displaced header from the box back to the mark. 300 if (mark == (markOop) lock) { 301 assert(dhw->is_neutral(), "invariant"); 302 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { 303 TEVENT(fast_exit: release stacklock); 304 return; 305 } 306 } 307 308 ObjectSynchronizer::inflate(THREAD, object)->exit(true, THREAD); 309 } 310 311 // ----------------------------------------------------------------------------- 312 // Interpreter/Compiler Slow Case 313 // This routine is used to handle interpreter/compiler slow case 314 // We don't need to use fast path here, because it must have been 315 // failed in the interpreter/compiler code. 316 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 317 markOop mark = obj->mark(); 318 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 319 320 if (mark->is_neutral()) { 321 // Anticipate successful CAS -- the ST of the displaced mark must 322 // be visible <= the ST performed by the CAS. 323 lock->set_displaced_header(mark); 324 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { 325 TEVENT(slow_enter: release stacklock); 326 return; 327 } 328 // Fall through to inflate() ... 329 } else if (mark->has_locker() && 330 THREAD->is_lock_owned((address)mark->locker())) { 331 assert(lock != mark->locker(), "must not re-lock the same lock"); 332 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 333 lock->set_displaced_header(NULL); 334 return; 335 } 336 337 // The object header will never be displaced to this lock, 338 // so it does not matter what the value is, except that it 339 // must be non-zero to avoid looking like a re-entrant lock, 340 // and must not look locked either. 341 lock->set_displaced_header(markOopDesc::unused_mark()); 342 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 343 } 344 345 // This routine is used to handle interpreter/compiler slow case 346 // We don't need to use fast path here, because it must have 347 // failed in the interpreter/compiler code. Simply use the heavy 348 // weight monitor should be ok, unless someone find otherwise. 349 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 350 fast_exit(object, lock, THREAD); 351 } 352 353 // ----------------------------------------------------------------------------- 354 // Class Loader support to workaround deadlocks on the class loader lock objects 355 // Also used by GC 356 // complete_exit()/reenter() are used to wait on a nested lock 357 // i.e. to give up an outer lock completely and then re-enter 358 // Used when holding nested locks - lock acquisition order: lock1 then lock2 359 // 1) complete_exit lock1 - saving recursion count 360 // 2) wait on lock2 361 // 3) when notified on lock2, unlock lock2 362 // 4) reenter lock1 with original recursion count 363 // 5) lock lock2 364 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 365 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 366 TEVENT(complete_exit); 367 if (UseBiasedLocking) { 368 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 369 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 370 } 371 372 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 373 374 return monitor->complete_exit(THREAD); 375 } 376 377 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 378 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 379 TEVENT(reenter); 380 if (UseBiasedLocking) { 381 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 382 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 383 } 384 385 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 386 387 monitor->reenter(recursion, THREAD); 388 } 389 // ----------------------------------------------------------------------------- 390 // JNI locks on java objects 391 // NOTE: must use heavy weight monitor to handle jni monitor enter 392 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 393 // the current locking is from JNI instead of Java code 394 TEVENT(jni_enter); 395 if (UseBiasedLocking) { 396 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 397 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 398 } 399 THREAD->set_current_pending_monitor_is_from_java(false); 400 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 401 THREAD->set_current_pending_monitor_is_from_java(true); 402 } 403 404 // NOTE: must use heavy weight monitor to handle jni monitor exit 405 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 406 TEVENT(jni_exit); 407 if (UseBiasedLocking) { 408 Handle h_obj(THREAD, obj); 409 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 410 obj = h_obj(); 411 } 412 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 413 414 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); 415 // If this thread has locked the object, exit the monitor. Note: can't use 416 // monitor->check(CHECK); must exit even if an exception is pending. 417 if (monitor->check(THREAD)) { 418 monitor->exit(true, THREAD); 419 } 420 } 421 422 // ----------------------------------------------------------------------------- 423 // Internal VM locks on java objects 424 // standard constructor, allows locking failures 425 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 426 _dolock = doLock; 427 _thread = thread; 428 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 429 _obj = obj; 430 431 if (_dolock) { 432 TEVENT(ObjectLocker); 433 434 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 435 } 436 } 437 438 ObjectLocker::~ObjectLocker() { 439 if (_dolock) { 440 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 441 } 442 } 443 444 445 // ----------------------------------------------------------------------------- 446 // Wait/Notify/NotifyAll 447 // NOTE: must use heavy weight monitor to handle wait() 448 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 449 if (UseBiasedLocking) { 450 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 451 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 452 } 453 if (millis < 0) { 454 TEVENT(wait - throw IAX); 455 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 456 } 457 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 458 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 459 monitor->wait(millis, true, THREAD); 460 461 // This dummy call is in place to get around dtrace bug 6254741. Once 462 // that's fixed we can uncomment the following line, remove the call 463 // and change this function back into a "void" func. 464 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 465 return dtrace_waited_probe(monitor, obj, THREAD); 466 } 467 468 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 469 if (UseBiasedLocking) { 470 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 471 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 472 } 473 if (millis < 0) { 474 TEVENT(wait - throw IAX); 475 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 476 } 477 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD); 478 } 479 480 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 481 if (UseBiasedLocking) { 482 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 483 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 484 } 485 486 markOop mark = obj->mark(); 487 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 488 return; 489 } 490 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); 491 } 492 493 // NOTE: see comment of notify() 494 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 495 if (UseBiasedLocking) { 496 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 497 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 498 } 499 500 markOop mark = obj->mark(); 501 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 502 return; 503 } 504 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); 505 } 506 507 // ----------------------------------------------------------------------------- 508 // Hash Code handling 509 // 510 // Performance concern: 511 // OrderAccess::storestore() calls release() which at one time stored 0 512 // into the global volatile OrderAccess::dummy variable. This store was 513 // unnecessary for correctness. Many threads storing into a common location 514 // causes considerable cache migration or "sloshing" on large SMP systems. 515 // As such, I avoided using OrderAccess::storestore(). In some cases 516 // OrderAccess::fence() -- which incurs local latency on the executing 517 // processor -- is a better choice as it scales on SMP systems. 518 // 519 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 520 // a discussion of coherency costs. Note that all our current reference 521 // platforms provide strong ST-ST order, so the issue is moot on IA32, 522 // x64, and SPARC. 523 // 524 // As a general policy we use "volatile" to control compiler-based reordering 525 // and explicit fences (barriers) to control for architectural reordering 526 // performed by the CPU(s) or platform. 527 528 struct SharedGlobals { 529 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE]; 530 // These are highly shared mostly-read variables. 531 // To avoid false-sharing they need to be the sole occupants of a cache line. 532 volatile int stwRandom; 533 volatile int stwCycle; 534 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 535 // Hot RW variable -- Sequester to avoid false-sharing 536 volatile int hcSequence; 537 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int)); 538 }; 539 540 static SharedGlobals GVars; 541 static int MonitorScavengeThreshold = 1000000; 542 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 543 544 static markOop ReadStableMark(oop obj) { 545 markOop mark = obj->mark(); 546 if (!mark->is_being_inflated()) { 547 return mark; // normal fast-path return 548 } 549 550 int its = 0; 551 for (;;) { 552 markOop mark = obj->mark(); 553 if (!mark->is_being_inflated()) { 554 return mark; // normal fast-path return 555 } 556 557 // The object is being inflated by some other thread. 558 // The caller of ReadStableMark() must wait for inflation to complete. 559 // Avoid live-lock 560 // TODO: consider calling SafepointSynchronize::do_call_back() while 561 // spinning to see if there's a safepoint pending. If so, immediately 562 // yielding or blocking would be appropriate. Avoid spinning while 563 // there is a safepoint pending. 564 // TODO: add inflation contention performance counters. 565 // TODO: restrict the aggregate number of spinners. 566 567 ++its; 568 if (its > 10000 || !os::is_MP()) { 569 if (its & 1) { 570 os::naked_yield(); 571 TEVENT(Inflate: INFLATING - yield); 572 } else { 573 // Note that the following code attenuates the livelock problem but is not 574 // a complete remedy. A more complete solution would require that the inflating 575 // thread hold the associated inflation lock. The following code simply restricts 576 // the number of spinners to at most one. We'll have N-2 threads blocked 577 // on the inflationlock, 1 thread holding the inflation lock and using 578 // a yield/park strategy, and 1 thread in the midst of inflation. 579 // A more refined approach would be to change the encoding of INFLATING 580 // to allow encapsulation of a native thread pointer. Threads waiting for 581 // inflation to complete would use CAS to push themselves onto a singly linked 582 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 583 // and calling park(). When inflation was complete the thread that accomplished inflation 584 // would detach the list and set the markword to inflated with a single CAS and 585 // then for each thread on the list, set the flag and unpark() the thread. 586 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 587 // wakes at most one thread whereas we need to wake the entire list. 588 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 589 int YieldThenBlock = 0; 590 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 591 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 592 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 593 while (obj->mark() == markOopDesc::INFLATING()) { 594 // Beware: NakedYield() is advisory and has almost no effect on some platforms 595 // so we periodically call Self->_ParkEvent->park(1). 596 // We use a mixed spin/yield/block mechanism. 597 if ((YieldThenBlock++) >= 16) { 598 Thread::current()->_ParkEvent->park(1); 599 } else { 600 os::naked_yield(); 601 } 602 } 603 Thread::muxRelease(gInflationLocks + ix); 604 TEVENT(Inflate: INFLATING - yield/park); 605 } 606 } else { 607 SpinPause(); // SMP-polite spinning 608 } 609 } 610 } 611 612 // hashCode() generation : 613 // 614 // Possibilities: 615 // * MD5Digest of {obj,stwRandom} 616 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 617 // * A DES- or AES-style SBox[] mechanism 618 // * One of the Phi-based schemes, such as: 619 // 2654435761 = 2^32 * Phi (golden ratio) 620 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 621 // * A variation of Marsaglia's shift-xor RNG scheme. 622 // * (obj ^ stwRandom) is appealing, but can result 623 // in undesirable regularity in the hashCode values of adjacent objects 624 // (objects allocated back-to-back, in particular). This could potentially 625 // result in hashtable collisions and reduced hashtable efficiency. 626 // There are simple ways to "diffuse" the middle address bits over the 627 // generated hashCode values: 628 629 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 630 intptr_t value = 0; 631 if (hashCode == 0) { 632 // This form uses an unguarded global Park-Miller RNG, 633 // so it's possible for two threads to race and generate the same RNG. 634 // On MP system we'll have lots of RW access to a global, so the 635 // mechanism induces lots of coherency traffic. 636 value = os::random(); 637 } else if (hashCode == 1) { 638 // This variation has the property of being stable (idempotent) 639 // between STW operations. This can be useful in some of the 1-0 640 // synchronization schemes. 641 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 642 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 643 } else if (hashCode == 2) { 644 value = 1; // for sensitivity testing 645 } else if (hashCode == 3) { 646 value = ++GVars.hcSequence; 647 } else if (hashCode == 4) { 648 value = cast_from_oop<intptr_t>(obj); 649 } else { 650 // Marsaglia's xor-shift scheme with thread-specific state 651 // This is probably the best overall implementation -- we'll 652 // likely make this the default in future releases. 653 unsigned t = Self->_hashStateX; 654 t ^= (t << 11); 655 Self->_hashStateX = Self->_hashStateY; 656 Self->_hashStateY = Self->_hashStateZ; 657 Self->_hashStateZ = Self->_hashStateW; 658 unsigned v = Self->_hashStateW; 659 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 660 Self->_hashStateW = v; 661 value = v; 662 } 663 664 value &= markOopDesc::hash_mask; 665 if (value == 0) value = 0xBAD; 666 assert(value != markOopDesc::no_hash, "invariant"); 667 TEVENT(hashCode: GENERATE); 668 return value; 669 } 670 671 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 672 if (UseBiasedLocking) { 673 // NOTE: many places throughout the JVM do not expect a safepoint 674 // to be taken here, in particular most operations on perm gen 675 // objects. However, we only ever bias Java instances and all of 676 // the call sites of identity_hash that might revoke biases have 677 // been checked to make sure they can handle a safepoint. The 678 // added check of the bias pattern is to avoid useless calls to 679 // thread-local storage. 680 if (obj->mark()->has_bias_pattern()) { 681 // Handle for oop obj in case of STW safepoint 682 Handle hobj(Self, obj); 683 // Relaxing assertion for bug 6320749. 684 assert(Universe::verify_in_progress() || 685 !SafepointSynchronize::is_at_safepoint(), 686 "biases should not be seen by VM thread here"); 687 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 688 obj = hobj(); 689 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 690 } 691 } 692 693 // hashCode() is a heap mutator ... 694 // Relaxing assertion for bug 6320749. 695 assert(Universe::verify_in_progress() || DumpSharedSpaces || 696 !SafepointSynchronize::is_at_safepoint(), "invariant"); 697 assert(Universe::verify_in_progress() || DumpSharedSpaces || 698 Self->is_Java_thread() , "invariant"); 699 assert(Universe::verify_in_progress() || DumpSharedSpaces || 700 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 701 702 ObjectMonitor* monitor = NULL; 703 markOop temp, test; 704 intptr_t hash; 705 markOop mark = ReadStableMark(obj); 706 707 // object should remain ineligible for biased locking 708 assert(!mark->has_bias_pattern(), "invariant"); 709 710 if (mark->is_neutral()) { 711 hash = mark->hash(); // this is a normal header 712 if (hash) { // if it has hash, just return it 713 return hash; 714 } 715 hash = get_next_hash(Self, obj); // allocate a new hash code 716 temp = mark->copy_set_hash(hash); // merge the hash code into header 717 // use (machine word version) atomic operation to install the hash 718 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 719 if (test == mark) { 720 return hash; 721 } 722 // If atomic operation failed, we must inflate the header 723 // into heavy weight monitor. We could add more code here 724 // for fast path, but it does not worth the complexity. 725 } else if (mark->has_monitor()) { 726 monitor = mark->monitor(); 727 temp = monitor->header(); 728 assert(temp->is_neutral(), "invariant"); 729 hash = temp->hash(); 730 if (hash) { 731 return hash; 732 } 733 // Skip to the following code to reduce code size 734 } else if (Self->is_lock_owned((address)mark->locker())) { 735 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 736 assert(temp->is_neutral(), "invariant"); 737 hash = temp->hash(); // by current thread, check if the displaced 738 if (hash) { // header contains hash code 739 return hash; 740 } 741 // WARNING: 742 // The displaced header is strictly immutable. 743 // It can NOT be changed in ANY cases. So we have 744 // to inflate the header into heavyweight monitor 745 // even the current thread owns the lock. The reason 746 // is the BasicLock (stack slot) will be asynchronously 747 // read by other threads during the inflate() function. 748 // Any change to stack may not propagate to other threads 749 // correctly. 750 } 751 752 // Inflate the monitor to set hash code 753 monitor = ObjectSynchronizer::inflate(Self, obj); 754 // Load displaced header and check it has hash code 755 mark = monitor->header(); 756 assert(mark->is_neutral(), "invariant"); 757 hash = mark->hash(); 758 if (hash == 0) { 759 hash = get_next_hash(Self, obj); 760 temp = mark->copy_set_hash(hash); // merge hash code into header 761 assert(temp->is_neutral(), "invariant"); 762 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 763 if (test != mark) { 764 // The only update to the header in the monitor (outside GC) 765 // is install the hash code. If someone add new usage of 766 // displaced header, please update this code 767 hash = test->hash(); 768 assert(test->is_neutral(), "invariant"); 769 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 770 } 771 } 772 // We finally get the hash 773 return hash; 774 } 775 776 // Deprecated -- use FastHashCode() instead. 777 778 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 779 return FastHashCode(Thread::current(), obj()); 780 } 781 782 783 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 784 Handle h_obj) { 785 if (UseBiasedLocking) { 786 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 787 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 788 } 789 790 assert(thread == JavaThread::current(), "Can only be called on current thread"); 791 oop obj = h_obj(); 792 793 markOop mark = ReadStableMark(obj); 794 795 // Uncontended case, header points to stack 796 if (mark->has_locker()) { 797 return thread->is_lock_owned((address)mark->locker()); 798 } 799 // Contended case, header points to ObjectMonitor (tagged pointer) 800 if (mark->has_monitor()) { 801 ObjectMonitor* monitor = mark->monitor(); 802 return monitor->is_entered(thread) != 0; 803 } 804 // Unlocked case, header in place 805 assert(mark->is_neutral(), "sanity check"); 806 return false; 807 } 808 809 // Be aware of this method could revoke bias of the lock object. 810 // This method queries the ownership of the lock handle specified by 'h_obj'. 811 // If the current thread owns the lock, it returns owner_self. If no 812 // thread owns the lock, it returns owner_none. Otherwise, it will return 813 // owner_other. 814 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 815 (JavaThread *self, Handle h_obj) { 816 // The caller must beware this method can revoke bias, and 817 // revocation can result in a safepoint. 818 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 819 assert(self->thread_state() != _thread_blocked, "invariant"); 820 821 // Possible mark states: neutral, biased, stack-locked, inflated 822 823 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 824 // CASE: biased 825 BiasedLocking::revoke_and_rebias(h_obj, false, self); 826 assert(!h_obj->mark()->has_bias_pattern(), 827 "biases should be revoked by now"); 828 } 829 830 assert(self == JavaThread::current(), "Can only be called on current thread"); 831 oop obj = h_obj(); 832 markOop mark = ReadStableMark(obj); 833 834 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 835 if (mark->has_locker()) { 836 return self->is_lock_owned((address)mark->locker()) ? 837 owner_self : owner_other; 838 } 839 840 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 841 // The Object:ObjectMonitor relationship is stable as long as we're 842 // not at a safepoint. 843 if (mark->has_monitor()) { 844 void * owner = mark->monitor()->_owner; 845 if (owner == NULL) return owner_none; 846 return (owner == self || 847 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 848 } 849 850 // CASE: neutral 851 assert(mark->is_neutral(), "sanity check"); 852 return owner_none; // it's unlocked 853 } 854 855 // FIXME: jvmti should call this 856 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 857 if (UseBiasedLocking) { 858 if (SafepointSynchronize::is_at_safepoint()) { 859 BiasedLocking::revoke_at_safepoint(h_obj); 860 } else { 861 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 862 } 863 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 864 } 865 866 oop obj = h_obj(); 867 address owner = NULL; 868 869 markOop mark = ReadStableMark(obj); 870 871 // Uncontended case, header points to stack 872 if (mark->has_locker()) { 873 owner = (address) mark->locker(); 874 } 875 876 // Contended case, header points to ObjectMonitor (tagged pointer) 877 if (mark->has_monitor()) { 878 ObjectMonitor* monitor = mark->monitor(); 879 assert(monitor != NULL, "monitor should be non-null"); 880 owner = (address) monitor->owner(); 881 } 882 883 if (owner != NULL) { 884 // owning_thread_from_monitor_owner() may also return NULL here 885 return Threads::owning_thread_from_monitor_owner(owner, doLock); 886 } 887 888 // Unlocked case, header in place 889 // Cannot have assertion since this object may have been 890 // locked by another thread when reaching here. 891 // assert(mark->is_neutral(), "sanity check"); 892 893 return NULL; 894 } 895 // Visitors ... 896 897 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 898 PaddedEnd<ObjectMonitor> * block = (PaddedEnd<ObjectMonitor> *)gBlockList; 899 ObjectMonitor* mid; 900 while (block) { 901 assert(block->object() == CHAINMARKER, "must be a block header"); 902 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 903 mid = (ObjectMonitor *)(block + i); 904 oop object = (oop) mid->object(); 905 if (object != NULL) { 906 closure->do_monitor(mid); 907 } 908 } 909 block = (PaddedEnd<ObjectMonitor> *) block->FreeNext; 910 } 911 } 912 913 // Get the next block in the block list. 914 static inline ObjectMonitor* next(ObjectMonitor* block) { 915 assert(block->object() == CHAINMARKER, "must be a block header"); 916 block = block->FreeNext; 917 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 918 return block; 919 } 920 921 922 void ObjectSynchronizer::oops_do(OopClosure* f) { 923 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 924 for (PaddedEnd<ObjectMonitor> * block = 925 (PaddedEnd<ObjectMonitor> *)gBlockList; block != NULL; 926 block = (PaddedEnd<ObjectMonitor> *)next(block)) { 927 assert(block->object() == CHAINMARKER, "must be a block header"); 928 for (int i = 1; i < _BLOCKSIZE; i++) { 929 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 930 if (mid->object() != NULL) { 931 f->do_oop((oop*)mid->object_addr()); 932 } 933 } 934 } 935 } 936 937 938 // ----------------------------------------------------------------------------- 939 // ObjectMonitor Lifecycle 940 // ----------------------- 941 // Inflation unlinks monitors from the global gFreeList and 942 // associates them with objects. Deflation -- which occurs at 943 // STW-time -- disassociates idle monitors from objects. Such 944 // scavenged monitors are returned to the gFreeList. 945 // 946 // The global list is protected by gListLock. All the critical sections 947 // are short and operate in constant-time. 948 // 949 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 950 // 951 // Lifecycle: 952 // -- unassigned and on the global free list 953 // -- unassigned and on a thread's private omFreeList 954 // -- assigned to an object. The object is inflated and the mark refers 955 // to the objectmonitor. 956 957 958 // Constraining monitor pool growth via MonitorBound ... 959 // 960 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 961 // the rate of scavenging is driven primarily by GC. As such, we can find 962 // an inordinate number of monitors in circulation. 963 // To avoid that scenario we can artificially induce a STW safepoint 964 // if the pool appears to be growing past some reasonable bound. 965 // Generally we favor time in space-time tradeoffs, but as there's no 966 // natural back-pressure on the # of extant monitors we need to impose some 967 // type of limit. Beware that if MonitorBound is set to too low a value 968 // we could just loop. In addition, if MonitorBound is set to a low value 969 // we'll incur more safepoints, which are harmful to performance. 970 // See also: GuaranteedSafepointInterval 971 // 972 // The current implementation uses asynchronous VM operations. 973 974 static void InduceScavenge(Thread * Self, const char * Whence) { 975 // Induce STW safepoint to trim monitors 976 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 977 // More precisely, trigger an asynchronous STW safepoint as the number 978 // of active monitors passes the specified threshold. 979 // TODO: assert thread state is reasonable 980 981 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 982 if (ObjectMonitor::Knob_Verbose) { 983 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)", 984 Whence, ForceMonitorScavenge) ; 985 tty->flush(); 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 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)", 995 Whence, ForceMonitorScavenge) ; 996 tty->flush(); 997 } 998 } 999 } 1000 1001 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1002 ObjectMonitor* mid; 1003 int in_use_tally = 0; 1004 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1005 in_use_tally++; 1006 } 1007 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1008 1009 int free_tally = 0; 1010 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1011 free_tally++; 1012 } 1013 assert(free_tally == Self->omFreeCount, "free count off"); 1014 } 1015 1016 ObjectMonitor * NOINLINE ObjectSynchronizer::omAlloc(Thread * Self) { 1017 // A large MAXPRIVATE value reduces both list lock contention 1018 // and list coherency traffic, but also tends to increase the 1019 // number of objectMonitors in circulation as well as the STW 1020 // scavenge costs. As usual, we lean toward time in space-time 1021 // tradeoffs. 1022 const int MAXPRIVATE = 1024; 1023 for (;;) { 1024 ObjectMonitor * m; 1025 1026 // 1: try to allocate from the thread's local omFreeList. 1027 // Threads will attempt to allocate first from their local list, then 1028 // from the global list, and only after those attempts fail will the thread 1029 // attempt to instantiate new monitors. Thread-local free lists take 1030 // heat off the gListLock and improve allocation latency, as well as reducing 1031 // coherency traffic on the shared global list. 1032 m = Self->omFreeList; 1033 if (m != NULL) { 1034 Self->omFreeList = m->FreeNext; 1035 Self->omFreeCount--; 1036 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1037 guarantee(m->object() == NULL, "invariant"); 1038 if (MonitorInUseLists) { 1039 m->FreeNext = Self->omInUseList; 1040 Self->omInUseList = m; 1041 Self->omInUseCount++; 1042 if (ObjectMonitor::Knob_VerifyInUse) { 1043 verifyInUse(Self); 1044 } 1045 } else { 1046 m->FreeNext = NULL; 1047 } 1048 return m; 1049 } 1050 1051 // 2: try to allocate from the global gFreeList 1052 // CONSIDER: use muxTry() instead of muxAcquire(). 1053 // If the muxTry() fails then drop immediately into case 3. 1054 // If we're using thread-local free lists then try 1055 // to reprovision the caller's free list. 1056 if (gFreeList != NULL) { 1057 // Reprovision the thread's omFreeList. 1058 // Use bulk transfers to reduce the allocation rate and heat 1059 // on various locks. 1060 Thread::muxAcquire(&gListLock, "omAlloc"); 1061 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1062 gMonitorFreeCount--; 1063 ObjectMonitor * take = gFreeList; 1064 gFreeList = take->FreeNext; 1065 guarantee(take->object() == NULL, "invariant"); 1066 guarantee(!take->is_busy(), "invariant"); 1067 take->Recycle(); 1068 omRelease(Self, take, false); 1069 } 1070 Thread::muxRelease(&gListLock); 1071 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1072 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1073 TEVENT(omFirst - reprovision); 1074 1075 const int mx = MonitorBound; 1076 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1077 // We can't safely induce a STW safepoint from omAlloc() as our thread 1078 // state may not be appropriate for such activities and callers may hold 1079 // naked oops, so instead we defer the action. 1080 InduceScavenge(Self, "omAlloc"); 1081 } 1082 continue; 1083 } 1084 1085 // 3: allocate a block of new ObjectMonitors 1086 // Both the local and global free lists are empty -- resort to malloc(). 1087 // In the current implementation objectMonitors are TSM - immortal. 1088 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1089 // each ObjectMonitor to start at the beginning of a cache line, 1090 // so we use align_size_up(). 1091 // A better solution would be to use C++ placement-new. 1092 // BEWARE: As it stands currently, we don't run the ctors! 1093 assert(_BLOCKSIZE > 1, "invariant"); 1094 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1095 PaddedEnd<ObjectMonitor> * temp; 1096 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1097 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1098 mtInternal); 1099 temp = (PaddedEnd<ObjectMonitor> *) 1100 align_size_up((intptr_t)real_malloc_addr, 1101 DEFAULT_CACHE_LINE_SIZE); 1102 1103 // NOTE: (almost) no way to recover if allocation failed. 1104 // We might be able to induce a STW safepoint and scavenge enough 1105 // objectMonitors to permit progress. 1106 if (temp == NULL) { 1107 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1108 "Allocate ObjectMonitors"); 1109 } 1110 (void)memset((void *) temp, 0, neededsize); 1111 1112 // Format the block. 1113 // initialize the linked list, each monitor points to its next 1114 // forming the single linked free list, the very first monitor 1115 // will points to next block, which forms the block list. 1116 // The trick of using the 1st element in the block as gBlockList 1117 // linkage should be reconsidered. A better implementation would 1118 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1119 1120 for (int i = 1; i < _BLOCKSIZE; i++) { 1121 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1122 } 1123 1124 // terminate the last monitor as the end of list 1125 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1126 1127 // Element [0] is reserved for global list linkage 1128 temp[0].set_object(CHAINMARKER); 1129 1130 // Consider carving out this thread's current request from the 1131 // block in hand. This avoids some lock traffic and redundant 1132 // list activity. 1133 1134 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1135 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1136 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1137 gMonitorPopulation += _BLOCKSIZE-1; 1138 gMonitorFreeCount += _BLOCKSIZE-1; 1139 1140 // Add the new block to the list of extant blocks (gBlockList). 1141 // The very first objectMonitor in a block is reserved and dedicated. 1142 // It serves as blocklist "next" linkage. 1143 temp[0].FreeNext = gBlockList; 1144 gBlockList = temp; 1145 1146 // Add the new string of objectMonitors to the global free list 1147 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1148 gFreeList = temp + 1; 1149 Thread::muxRelease(&gListLock); 1150 TEVENT(Allocate block of monitors); 1151 } 1152 } 1153 1154 // Place "m" on the caller's private per-thread omFreeList. 1155 // In practice there's no need to clamp or limit the number of 1156 // monitors on a thread's omFreeList as the only time we'll call 1157 // omRelease is to return a monitor to the free list after a CAS 1158 // attempt failed. This doesn't allow unbounded #s of monitors to 1159 // accumulate on a thread's free list. 1160 // 1161 // Key constraint: all ObjectMonitors on a thread's free list and the global 1162 // free list must have their object field set to null. This prevents the 1163 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1164 1165 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1166 bool fromPerThreadAlloc) { 1167 guarantee(m->object() == NULL, "invariant"); 1168 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1169 // Remove from omInUseList 1170 if (MonitorInUseLists && fromPerThreadAlloc) { 1171 ObjectMonitor* cur_mid_in_use = NULL; 1172 bool extracted = false; 1173 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1174 if (m == mid) { 1175 // extract from per-thread in-use list 1176 if (mid == Self->omInUseList) { 1177 Self->omInUseList = mid->FreeNext; 1178 } else if (cur_mid_in_use != NULL) { 1179 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1180 } 1181 extracted = true; 1182 Self->omInUseCount--; 1183 if (ObjectMonitor::Knob_VerifyInUse) { 1184 verifyInUse(Self); 1185 } 1186 break; 1187 } 1188 } 1189 assert(extracted, "Should have extracted from in-use list"); 1190 } 1191 1192 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1193 m->FreeNext = Self->omFreeList; 1194 Self->omFreeList = m; 1195 Self->omFreeCount++; 1196 } 1197 1198 // Return the monitors of a moribund thread's local free list to 1199 // the global free list. Typically a thread calls omFlush() when 1200 // it's dying. We could also consider having the VM thread steal 1201 // monitors from threads that have not run java code over a few 1202 // consecutive STW safepoints. Relatedly, we might decay 1203 // omFreeProvision at STW safepoints. 1204 // 1205 // Also return the monitors of a moribund thread's omInUseList to 1206 // a global gOmInUseList under the global list lock so these 1207 // will continue to be scanned. 1208 // 1209 // We currently call omFlush() from the Thread:: dtor _after the thread 1210 // has been excised from the thread list and is no longer a mutator. 1211 // That means that omFlush() can run concurrently with a safepoint and 1212 // the scavenge operator. Calling omFlush() from JavaThread::exit() might 1213 // be a better choice as we could safely reason that that the JVM is 1214 // not at a safepoint at the time of the call, and thus there could 1215 // be not inopportune interleavings between omFlush() and the scavenge 1216 // operator. 1217 1218 void ObjectSynchronizer::omFlush(Thread * Self) { 1219 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1220 Self->omFreeList = NULL; 1221 ObjectMonitor * tail = NULL; 1222 int tally = 0; 1223 if (list != NULL) { 1224 ObjectMonitor * s; 1225 // The thread is going away, the per-thread free monitors 1226 // are freed via set_owner(NULL) 1227 // Link them to tail, which will be linked into the global free list 1228 // gFreeList below, under the gListLock 1229 for (s = list; s != NULL; s = s->FreeNext) { 1230 tally++; 1231 tail = s; 1232 guarantee(s->object() == NULL, "invariant"); 1233 guarantee(!s->is_busy(), "invariant"); 1234 s->set_owner(NULL); // redundant but good hygiene 1235 TEVENT(omFlush - Move one); 1236 } 1237 guarantee(tail != NULL && list != NULL, "invariant"); 1238 } 1239 1240 ObjectMonitor * inUseList = Self->omInUseList; 1241 ObjectMonitor * inUseTail = NULL; 1242 int inUseTally = 0; 1243 if (inUseList != NULL) { 1244 Self->omInUseList = NULL; 1245 ObjectMonitor *cur_om; 1246 // The thread is going away, however the omInUseList inflated 1247 // monitors may still be in-use by other threads. 1248 // Link them to inUseTail, which will be linked into the global in-use list 1249 // gOmInUseList below, under the gListLock 1250 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1251 inUseTail = cur_om; 1252 inUseTally++; 1253 } 1254 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1255 Self->omInUseCount = 0; 1256 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1257 } 1258 1259 Thread::muxAcquire(&gListLock, "omFlush"); 1260 if (tail != NULL) { 1261 tail->FreeNext = gFreeList; 1262 gFreeList = list; 1263 gMonitorFreeCount += tally; 1264 } 1265 1266 if (inUseTail != NULL) { 1267 inUseTail->FreeNext = gOmInUseList; 1268 gOmInUseList = inUseList; 1269 gOmInUseCount += inUseTally; 1270 } 1271 1272 Thread::muxRelease(&gListLock); 1273 TEVENT(omFlush); 1274 } 1275 1276 // Fast path code shared by multiple functions 1277 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1278 markOop mark = obj->mark(); 1279 if (mark->has_monitor()) { 1280 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1281 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1282 return mark->monitor(); 1283 } 1284 return ObjectSynchronizer::inflate(Thread::current(), obj); 1285 } 1286 1287 1288 ObjectMonitor * NOINLINE ObjectSynchronizer::inflate(Thread * Self, 1289 oop object) { 1290 // Inflate mutates the heap ... 1291 // Relaxing assertion for bug 6320749. 1292 assert(Universe::verify_in_progress() || 1293 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1294 1295 for (;;) { 1296 const markOop mark = object->mark(); 1297 assert(!mark->has_bias_pattern(), "invariant"); 1298 1299 // The mark can be in one of the following states: 1300 // * Inflated - just return 1301 // * Stack-locked - coerce it to inflated 1302 // * INFLATING - busy wait for conversion to complete 1303 // * Neutral - aggressively inflate the object. 1304 // * BIASED - Illegal. We should never see this 1305 1306 // CASE: inflated 1307 if (mark->has_monitor()) { 1308 ObjectMonitor * inf = mark->monitor(); 1309 assert(inf->header()->is_neutral(), "invariant"); 1310 assert(inf->object() == object, "invariant"); 1311 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1312 return inf; 1313 } 1314 1315 // CASE: inflation in progress - inflating over a stack-lock. 1316 // Some other thread is converting from stack-locked to inflated. 1317 // Only that thread can complete inflation -- other threads must wait. 1318 // The INFLATING value is transient. 1319 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1320 // We could always eliminate polling by parking the thread on some auxiliary list. 1321 if (mark == markOopDesc::INFLATING()) { 1322 TEVENT(Inflate: spin while INFLATING); 1323 ReadStableMark(object); 1324 continue; 1325 } 1326 1327 // CASE: stack-locked 1328 // Could be stack-locked either by this thread or by some other thread. 1329 // 1330 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1331 // to install INFLATING into the mark word. We originally installed INFLATING, 1332 // allocated the objectmonitor, and then finally STed the address of the 1333 // objectmonitor into the mark. This was correct, but artificially lengthened 1334 // the interval in which INFLATED appeared in the mark, thus increasing 1335 // the odds of inflation contention. 1336 // 1337 // We now use per-thread private objectmonitor free lists. 1338 // These list are reprovisioned from the global free list outside the 1339 // critical INFLATING...ST interval. A thread can transfer 1340 // multiple objectmonitors en-mass from the global free list to its local free list. 1341 // This reduces coherency traffic and lock contention on the global free list. 1342 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1343 // before or after the CAS(INFLATING) operation. 1344 // See the comments in omAlloc(). 1345 1346 if (mark->has_locker()) { 1347 ObjectMonitor * m = omAlloc(Self); 1348 // Optimistically prepare the objectmonitor - anticipate successful CAS 1349 // We do this before the CAS in order to minimize the length of time 1350 // in which INFLATING appears in the mark. 1351 m->Recycle(); 1352 m->_Responsible = NULL; 1353 m->_recursions = 0; 1354 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1355 1356 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark); 1357 if (cmp != mark) { 1358 omRelease(Self, m, true); 1359 continue; // Interference -- just retry 1360 } 1361 1362 // We've successfully installed INFLATING (0) into the mark-word. 1363 // This is the only case where 0 will appear in a mark-word. 1364 // Only the singular thread that successfully swings the mark-word 1365 // to 0 can perform (or more precisely, complete) inflation. 1366 // 1367 // Why do we CAS a 0 into the mark-word instead of just CASing the 1368 // mark-word from the stack-locked value directly to the new inflated state? 1369 // Consider what happens when a thread unlocks a stack-locked object. 1370 // It attempts to use CAS to swing the displaced header value from the 1371 // on-stack basiclock back into the object header. Recall also that the 1372 // header value (hashcode, etc) can reside in (a) the object header, or 1373 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1374 // header in an objectMonitor. The inflate() routine must copy the header 1375 // value from the basiclock on the owner's stack to the objectMonitor, all 1376 // the while preserving the hashCode stability invariants. If the owner 1377 // decides to release the lock while the value is 0, the unlock will fail 1378 // and control will eventually pass from slow_exit() to inflate. The owner 1379 // will then spin, waiting for the 0 value to disappear. Put another way, 1380 // the 0 causes the owner to stall if the owner happens to try to 1381 // drop the lock (restoring the header from the basiclock to the object) 1382 // while inflation is in-progress. This protocol avoids races that might 1383 // would otherwise permit hashCode values to change or "flicker" for an object. 1384 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1385 // 0 serves as a "BUSY" inflate-in-progress indicator. 1386 1387 1388 // fetch the displaced mark from the owner's stack. 1389 // The owner can't die or unwind past the lock while our INFLATING 1390 // object is in the mark. Furthermore the owner can't complete 1391 // an unlock on the object, either. 1392 markOop dmw = mark->displaced_mark_helper(); 1393 assert(dmw->is_neutral(), "invariant"); 1394 1395 // Setup monitor fields to proper values -- prepare the monitor 1396 m->set_header(dmw); 1397 1398 // Optimization: if the mark->locker stack address is associated 1399 // with this thread we could simply set m->_owner = Self. 1400 // Note that a thread can inflate an object 1401 // that it has stack-locked -- as might happen in wait() -- directly 1402 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1403 m->set_owner(mark->locker()); 1404 m->set_object(object); 1405 // TODO-FIXME: assert BasicLock->dhw != 0. 1406 1407 // Must preserve store ordering. The monitor state must 1408 // be stable at the time of publishing the monitor address. 1409 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1410 object->release_set_mark(markOopDesc::encode(m)); 1411 1412 // Hopefully the performance counters are allocated on distinct cache lines 1413 // to avoid false sharing on MP systems ... 1414 OM_PERFDATA_OP(Inflations, inc()); 1415 TEVENT(Inflate: overwrite stacklock); 1416 if (TraceMonitorInflation) { 1417 if (object->is_instance()) { 1418 ResourceMark rm; 1419 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1420 (void *) object, (intptr_t) object->mark(), 1421 object->klass()->external_name()); 1422 } 1423 } 1424 return m; 1425 } 1426 1427 // CASE: neutral 1428 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1429 // If we know we're inflating for entry it's better to inflate by swinging a 1430 // pre-locked objectMonitor pointer into the object header. A successful 1431 // CAS inflates the object *and* confers ownership to the inflating thread. 1432 // In the current implementation we use a 2-step mechanism where we CAS() 1433 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1434 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1435 // would be useful. 1436 1437 assert(mark->is_neutral(), "invariant"); 1438 ObjectMonitor * m = omAlloc(Self); 1439 // prepare m for installation - set monitor to initial state 1440 m->Recycle(); 1441 m->set_header(mark); 1442 m->set_owner(NULL); 1443 m->set_object(object); 1444 m->_recursions = 0; 1445 m->_Responsible = NULL; 1446 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1447 1448 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1449 m->set_object(NULL); 1450 m->set_owner(NULL); 1451 m->Recycle(); 1452 omRelease(Self, m, true); 1453 m = NULL; 1454 continue; 1455 // interference - the markword changed - just retry. 1456 // The state-transitions are one-way, so there's no chance of 1457 // live-lock -- "Inflated" is an absorbing state. 1458 } 1459 1460 // Hopefully the performance counters are allocated on distinct 1461 // cache lines to avoid false sharing on MP systems ... 1462 OM_PERFDATA_OP(Inflations, inc()); 1463 TEVENT(Inflate: overwrite neutral); 1464 if (TraceMonitorInflation) { 1465 if (object->is_instance()) { 1466 ResourceMark rm; 1467 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1468 (void *) object, (intptr_t) object->mark(), 1469 object->klass()->external_name()); 1470 } 1471 } 1472 return m; 1473 } 1474 } 1475 1476 1477 // Deflate_idle_monitors() is called at all safepoints, immediately 1478 // after all mutators are stopped, but before any objects have moved. 1479 // It traverses the list of known monitors, deflating where possible. 1480 // The scavenged monitor are returned to the monitor free list. 1481 // 1482 // Beware that we scavenge at *every* stop-the-world point. 1483 // Having a large number of monitors in-circulation negatively 1484 // impacts the performance of some applications (e.g., PointBase). 1485 // Broadly, we want to minimize the # of monitors in circulation. 1486 // 1487 // We have added a flag, MonitorInUseLists, which creates a list 1488 // of active monitors for each thread. deflate_idle_monitors() 1489 // only scans the per-thread in-use lists. omAlloc() puts all 1490 // assigned monitors on the per-thread list. deflate_idle_monitors() 1491 // returns the non-busy monitors to the global free list. 1492 // When a thread dies, omFlush() adds the list of active monitors for 1493 // that thread to a global gOmInUseList acquiring the 1494 // global list lock. deflate_idle_monitors() acquires the global 1495 // list lock to scan for non-busy monitors to the global free list. 1496 // An alternative could have used a single global in-use list. The 1497 // downside would have been the additional cost of acquiring the global list lock 1498 // for every omAlloc(). 1499 // 1500 // Perversely, the heap size -- and thus the STW safepoint rate -- 1501 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1502 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1503 // This is an unfortunate aspect of this design. 1504 1505 enum ManifestConstants { 1506 ClearResponsibleAtSTW = 0 1507 }; 1508 1509 // Deflate a single monitor if not in-use 1510 // Return true if deflated, false if in-use 1511 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1512 ObjectMonitor** freeHeadp, 1513 ObjectMonitor** freeTailp) { 1514 bool deflated; 1515 // Normal case ... The monitor is associated with obj. 1516 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1517 guarantee(mid == obj->mark()->monitor(), "invariant"); 1518 guarantee(mid->header()->is_neutral(), "invariant"); 1519 1520 if (mid->is_busy()) { 1521 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1522 deflated = false; 1523 } else { 1524 // Deflate the monitor if it is no longer being used 1525 // It's idle - scavenge and return to the global free list 1526 // plain old deflation ... 1527 TEVENT(deflate_idle_monitors - scavenge1); 1528 if (TraceMonitorInflation) { 1529 if (obj->is_instance()) { 1530 ResourceMark rm; 1531 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1532 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name()); 1533 } 1534 } 1535 1536 // Restore the header back to obj 1537 obj->release_set_mark(mid->header()); 1538 mid->clear(); 1539 1540 assert(mid->object() == NULL, "invariant"); 1541 1542 // Move the object to the working free list defined by freeHeadp, freeTailp 1543 if (*freeHeadp == NULL) *freeHeadp = mid; 1544 if (*freeTailp != NULL) { 1545 ObjectMonitor * prevtail = *freeTailp; 1546 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1547 prevtail->FreeNext = mid; 1548 } 1549 *freeTailp = mid; 1550 deflated = true; 1551 } 1552 return deflated; 1553 } 1554 1555 // Walk a given monitor list, and deflate idle monitors 1556 // The given list could be a per-thread list or a global list 1557 // Caller acquires gListLock 1558 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1559 ObjectMonitor** freeHeadp, 1560 ObjectMonitor** freeTailp) { 1561 ObjectMonitor* mid; 1562 ObjectMonitor* next; 1563 ObjectMonitor* cur_mid_in_use = NULL; 1564 int deflated_count = 0; 1565 1566 for (mid = *listHeadp; mid != NULL;) { 1567 oop obj = (oop) mid->object(); 1568 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1569 // if deflate_monitor succeeded, 1570 // extract from per-thread in-use list 1571 if (mid == *listHeadp) { 1572 *listHeadp = mid->FreeNext; 1573 } else if (cur_mid_in_use != NULL) { 1574 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1575 } 1576 next = mid->FreeNext; 1577 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1578 mid = next; 1579 deflated_count++; 1580 } else { 1581 cur_mid_in_use = mid; 1582 mid = mid->FreeNext; 1583 } 1584 } 1585 return deflated_count; 1586 } 1587 1588 void ObjectSynchronizer::deflate_idle_monitors() { 1589 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1590 int nInuse = 0; // currently associated with objects 1591 int nInCirculation = 0; // extant 1592 int nScavenged = 0; // reclaimed 1593 bool deflated = false; 1594 1595 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1596 ObjectMonitor * freeTailp = NULL; 1597 1598 TEVENT(deflate_idle_monitors); 1599 // Prevent omFlush from changing mids in Thread dtor's during deflation 1600 // And in case the vm thread is acquiring a lock during a safepoint 1601 // See e.g. 6320749 1602 Thread::muxAcquire(&gListLock, "scavenge - return"); 1603 1604 if (MonitorInUseLists) { 1605 int inUse = 0; 1606 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { 1607 nInCirculation+= cur->omInUseCount; 1608 int deflated_count = deflate_monitor_list(cur->omInUseList_addr(), &freeHeadp, &freeTailp); 1609 cur->omInUseCount-= deflated_count; 1610 if (ObjectMonitor::Knob_VerifyInUse) { 1611 verifyInUse(cur); 1612 } 1613 nScavenged += deflated_count; 1614 nInuse += cur->omInUseCount; 1615 } 1616 1617 // For moribund threads, scan gOmInUseList 1618 if (gOmInUseList) { 1619 nInCirculation += gOmInUseCount; 1620 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1621 gOmInUseCount-= deflated_count; 1622 nScavenged += deflated_count; 1623 nInuse += gOmInUseCount; 1624 } 1625 1626 } else for (PaddedEnd<ObjectMonitor> * block = 1627 (PaddedEnd<ObjectMonitor> *)gBlockList; block != NULL; 1628 block = (PaddedEnd<ObjectMonitor> *)next(block)) { 1629 // Iterate over all extant monitors - Scavenge all idle monitors. 1630 assert(block->object() == CHAINMARKER, "must be a block header"); 1631 nInCirculation += _BLOCKSIZE; 1632 for (int i = 1; i < _BLOCKSIZE; i++) { 1633 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1634 oop obj = (oop) mid->object(); 1635 1636 if (obj == NULL) { 1637 // The monitor is not associated with an object. 1638 // The monitor should either be a thread-specific private 1639 // free list or the global free list. 1640 // obj == NULL IMPLIES mid->is_busy() == 0 1641 guarantee(!mid->is_busy(), "invariant"); 1642 continue; 1643 } 1644 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1645 1646 if (deflated) { 1647 mid->FreeNext = NULL; 1648 nScavenged++; 1649 } else { 1650 nInuse++; 1651 } 1652 } 1653 } 1654 1655 gMonitorFreeCount += nScavenged; 1656 1657 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1658 1659 if (ObjectMonitor::Knob_Verbose) { 1660 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d " 1661 "ForceMonitorScavenge=%d : pop=%d free=%d", 1662 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, 1663 gMonitorPopulation, gMonitorFreeCount); 1664 tty->flush(); 1665 } 1666 1667 ForceMonitorScavenge = 0; // Reset 1668 1669 // Move the scavenged monitors back to the global free list. 1670 if (freeHeadp != NULL) { 1671 guarantee(freeTailp != NULL && nScavenged > 0, "invariant"); 1672 assert(freeTailp->FreeNext == NULL, "invariant"); 1673 // constant-time list splice - prepend scavenged segment to gFreeList 1674 freeTailp->FreeNext = gFreeList; 1675 gFreeList = freeHeadp; 1676 } 1677 Thread::muxRelease(&gListLock); 1678 1679 OM_PERFDATA_OP(Deflations, inc(nScavenged)); 1680 OM_PERFDATA_OP(MonExtant, set_value(nInCirculation)); 1681 1682 // TODO: Add objectMonitor leak detection. 1683 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1684 GVars.stwRandom = os::random(); 1685 GVars.stwCycle++; 1686 } 1687 1688 // Monitor cleanup on JavaThread::exit 1689 1690 // Iterate through monitor cache and attempt to release thread's monitors 1691 // Gives up on a particular monitor if an exception occurs, but continues 1692 // the overall iteration, swallowing the exception. 1693 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1694 private: 1695 TRAPS; 1696 1697 public: 1698 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1699 void do_monitor(ObjectMonitor* mid) { 1700 if (mid->owner() == THREAD) { 1701 if (ObjectMonitor::Knob_VerifyMatch != 0) { 1702 Handle obj((oop) mid->object()); 1703 tty->print("INFO: unexpected locked object:"); 1704 javaVFrame::print_locked_object_class_name(tty, obj, "locked"); 1705 fatal("exiting JavaThread=" INTPTR_FORMAT 1706 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT, 1707 THREAD, mid); 1708 } 1709 (void)mid->complete_exit(CHECK); 1710 } 1711 } 1712 }; 1713 1714 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1715 // ignored. This is meant to be called during JNI thread detach which assumes 1716 // all remaining monitors are heavyweight. All exceptions are swallowed. 1717 // Scanning the extant monitor list can be time consuming. 1718 // A simple optimization is to add a per-thread flag that indicates a thread 1719 // called jni_monitorenter() during its lifetime. 1720 // 1721 // Instead of No_Savepoint_Verifier it might be cheaper to 1722 // use an idiom of the form: 1723 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1724 // <code that must not run at safepoint> 1725 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1726 // Since the tests are extremely cheap we could leave them enabled 1727 // for normal product builds. 1728 1729 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1730 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1731 No_Safepoint_Verifier nsv; 1732 ReleaseJavaMonitorsClosure rjmc(THREAD); 1733 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1734 ObjectSynchronizer::monitors_iterate(&rjmc); 1735 Thread::muxRelease(&gListLock); 1736 THREAD->clear_pending_exception(); 1737 } 1738 1739 //------------------------------------------------------------------------------ 1740 // Debugging code 1741 1742 void ObjectSynchronizer::sanity_checks(const bool verbose, 1743 const uint cache_line_size, 1744 int *error_cnt_ptr, 1745 int *warning_cnt_ptr) { 1746 u_char *addr_begin = (u_char*)&GVars; 1747 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1748 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1749 1750 if (verbose) { 1751 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1752 sizeof(SharedGlobals)); 1753 } 1754 1755 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1756 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1757 1758 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1759 if (verbose) { 1760 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1761 } 1762 1763 if (cache_line_size != 0) { 1764 // We were able to determine the L1 data cache line size so 1765 // do some cache line specific sanity checks 1766 1767 if (offset_stwRandom < cache_line_size) { 1768 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1769 "to the struct beginning than a cache line which permits " 1770 "false sharing."); 1771 (*warning_cnt_ptr)++; 1772 } 1773 1774 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1775 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1776 "SharedGlobals.hcSequence fields are closer than a cache " 1777 "line which permits false sharing."); 1778 (*warning_cnt_ptr)++; 1779 } 1780 1781 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1782 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1783 "to the struct end than a cache line which permits false " 1784 "sharing."); 1785 (*warning_cnt_ptr)++; 1786 } 1787 } 1788 } 1789 1790 #ifndef PRODUCT 1791 1792 // Verify all monitors in the monitor cache, the verification is weak. 1793 void ObjectSynchronizer::verify() { 1794 PaddedEnd<ObjectMonitor> * block = (PaddedEnd<ObjectMonitor> *)gBlockList; 1795 ObjectMonitor* mid; 1796 while (block) { 1797 assert(block->object() == CHAINMARKER, "must be a block header"); 1798 for (int i = 1; i < _BLOCKSIZE; i++) { 1799 mid = (ObjectMonitor *)(block + i); 1800 oop object = (oop) mid->object(); 1801 if (object != NULL) { 1802 mid->verify(); 1803 } 1804 } 1805 block = (PaddedEnd<ObjectMonitor> *) block->FreeNext; 1806 } 1807 } 1808 1809 // Check if monitor belongs to the monitor cache 1810 // The list is grow-only so it's *relatively* safe to traverse 1811 // the list of extant blocks without taking a lock. 1812 1813 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1814 PaddedEnd<ObjectMonitor> * block = (PaddedEnd<ObjectMonitor> *)gBlockList; 1815 1816 while (block) { 1817 assert(block->object() == CHAINMARKER, "must be a block header"); 1818 if (monitor > (ObjectMonitor *)&block[0] && 1819 monitor < (ObjectMonitor *)&block[_BLOCKSIZE]) { 1820 address mon = (address) monitor; 1821 address blk = (address) block; 1822 size_t diff = mon - blk; 1823 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "check"); 1824 return 1; 1825 } 1826 block = (PaddedEnd<ObjectMonitor> *) block->FreeNext; 1827 } 1828 return 0; 1829 } 1830 1831 #endif