1 /* 2 * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "oops/markOop.hpp" 30 #include "oops/oop.inline.hpp" 31 #include "runtime/biasedLocking.hpp" 32 #include "runtime/handles.inline.hpp" 33 #include "runtime/interfaceSupport.hpp" 34 #include "runtime/mutexLocker.hpp" 35 #include "runtime/objectMonitor.hpp" 36 #include "runtime/objectMonitor.inline.hpp" 37 #include "runtime/osThread.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #include "runtime/synchronizer.hpp" 40 #include "runtime/thread.inline.hpp" 41 #include "utilities/dtrace.hpp" 42 #include "utilities/events.hpp" 43 #include "utilities/preserveException.hpp" 44 #ifdef TARGET_OS_FAMILY_linux 45 # include "os_linux.inline.hpp" 46 #endif 47 #ifdef TARGET_OS_FAMILY_solaris 48 # include "os_solaris.inline.hpp" 49 #endif 50 #ifdef TARGET_OS_FAMILY_windows 51 # include "os_windows.inline.hpp" 52 #endif 53 #ifdef TARGET_OS_FAMILY_bsd 54 # include "os_bsd.inline.hpp" 55 #endif 56 57 #if defined(__GNUC__) && !defined(PPC64) 58 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 59 #define ATTR __attribute__((noinline)) 60 #else 61 #define ATTR 62 #endif 63 64 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 65 66 // The "core" versions of monitor enter and exit reside in this file. 67 // The interpreter and compilers contain specialized transliterated 68 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 69 // for instance. If you make changes here, make sure to modify the 70 // interpreter, and both C1 and C2 fast-path inline locking code emission. 71 // 72 // 73 // ----------------------------------------------------------------------------- 74 75 #ifdef DTRACE_ENABLED 76 77 // Only bother with this argument setup if dtrace is available 78 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 79 80 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 81 char* bytes = NULL; \ 82 int len = 0; \ 83 jlong jtid = SharedRuntime::get_java_tid(thread); \ 84 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 85 if (klassname != NULL) { \ 86 bytes = (char*)klassname->bytes(); \ 87 len = klassname->utf8_length(); \ 88 } 89 90 #ifndef USDT2 91 HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait, 92 jlong, uintptr_t, char*, int, long); 93 HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited, 94 jlong, uintptr_t, char*, int); 95 96 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 97 { \ 98 if (DTraceMonitorProbes) { \ 99 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 100 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ 101 (monitor), bytes, len, (millis)); \ 102 } \ 103 } 104 105 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 106 { \ 107 if (DTraceMonitorProbes) { \ 108 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 109 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ 110 (uintptr_t)(monitor), bytes, len); \ 111 } \ 112 } 113 114 #else /* USDT2 */ 115 116 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 117 { \ 118 if (DTraceMonitorProbes) { \ 119 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 120 HOTSPOT_MONITOR_WAIT(jtid, \ 121 (uintptr_t)(monitor), bytes, len, (millis)); \ 122 } \ 123 } 124 125 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 126 127 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 128 { \ 129 if (DTraceMonitorProbes) { \ 130 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 131 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 132 (uintptr_t)(monitor), bytes, len); \ 133 } \ 134 } 135 136 #endif /* USDT2 */ 137 #else // ndef DTRACE_ENABLED 138 139 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 140 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 141 142 #endif // ndef DTRACE_ENABLED 143 144 // This exists only as a workaround of dtrace bug 6254741 145 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 146 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 147 return 0; 148 } 149 150 #define NINFLATIONLOCKS 256 151 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ; 152 153 ObjectMonitor * volatile ObjectSynchronizer::gBlockList = NULL; 154 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ; 155 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ; 156 int ObjectSynchronizer::gOmInUseCount = 0; 157 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache 158 static volatile int MonitorFreeCount = 0 ; // # on gFreeList 159 static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation 160 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 161 162 // ----------------------------------------------------------------------------- 163 // Fast Monitor Enter/Exit 164 // This the fast monitor enter. The interpreter and compiler use 165 // some assembly copies of this code. Make sure update those code 166 // if the following function is changed. The implementation is 167 // extremely sensitive to race condition. Be careful. 168 169 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) { 170 if (UseBiasedLocking) { 171 if (!SafepointSynchronize::is_at_safepoint()) { 172 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 173 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 174 return; 175 } 176 } else { 177 assert(!attempt_rebias, "can not rebias toward VM thread"); 178 BiasedLocking::revoke_at_safepoint(obj); 179 } 180 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 181 } 182 183 slow_enter (obj, lock, THREAD) ; 184 } 185 186 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 187 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); 188 // if displaced header is null, the previous enter is recursive enter, no-op 189 markOop dhw = lock->displaced_header(); 190 markOop mark ; 191 if (dhw == NULL) { 192 // Recursive stack-lock. 193 // Diagnostics -- Could be: stack-locked, inflating, inflated. 194 mark = object->mark() ; 195 assert (!mark->is_neutral(), "invariant") ; 196 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { 197 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ; 198 } 199 if (mark->has_monitor()) { 200 ObjectMonitor * m = mark->monitor() ; 201 assert(((oop)(m->object()))->mark() == mark, "invariant") ; 202 assert(m->is_entered(THREAD), "invariant") ; 203 } 204 return ; 205 } 206 207 mark = object->mark() ; 208 209 // If the object is stack-locked by the current thread, try to 210 // swing the displaced header from the box back to the mark. 211 if (mark == (markOop) lock) { 212 assert (dhw->is_neutral(), "invariant") ; 213 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { 214 TEVENT (fast_exit: release stacklock) ; 215 return; 216 } 217 } 218 219 ObjectSynchronizer::inflate(THREAD, object)->exit (true, THREAD) ; 220 } 221 222 // ----------------------------------------------------------------------------- 223 // Interpreter/Compiler Slow Case 224 // This routine is used to handle interpreter/compiler slow case 225 // We don't need to use fast path here, because it must have been 226 // failed in the interpreter/compiler code. 227 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 228 markOop mark = obj->mark(); 229 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 230 231 if (mark->is_neutral()) { 232 // Anticipate successful CAS -- the ST of the displaced mark must 233 // be visible <= the ST performed by the CAS. 234 lock->set_displaced_header(mark); 235 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { 236 TEVENT (slow_enter: release stacklock) ; 237 return ; 238 } 239 // Fall through to inflate() ... 240 } else 241 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 242 assert(lock != mark->locker(), "must not re-lock the same lock"); 243 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 244 lock->set_displaced_header(NULL); 245 return; 246 } 247 248 #if 0 249 // The following optimization isn't particularly useful. 250 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) { 251 lock->set_displaced_header (NULL) ; 252 return ; 253 } 254 #endif 255 256 // The object header will never be displaced to this lock, 257 // so it does not matter what the value is, except that it 258 // must be non-zero to avoid looking like a re-entrant lock, 259 // and must not look locked either. 260 lock->set_displaced_header(markOopDesc::unused_mark()); 261 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 262 } 263 264 // This routine is used to handle interpreter/compiler slow case 265 // We don't need to use fast path here, because it must have 266 // failed in the interpreter/compiler code. Simply use the heavy 267 // weight monitor should be ok, unless someone find otherwise. 268 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 269 fast_exit (object, lock, THREAD) ; 270 } 271 272 // ----------------------------------------------------------------------------- 273 // Class Loader support to workaround deadlocks on the class loader lock objects 274 // Also used by GC 275 // complete_exit()/reenter() are used to wait on a nested lock 276 // i.e. to give up an outer lock completely and then re-enter 277 // Used when holding nested locks - lock acquisition order: lock1 then lock2 278 // 1) complete_exit lock1 - saving recursion count 279 // 2) wait on lock2 280 // 3) when notified on lock2, unlock lock2 281 // 4) reenter lock1 with original recursion count 282 // 5) lock lock2 283 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 284 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 285 TEVENT (complete_exit) ; 286 if (UseBiasedLocking) { 287 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 288 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 289 } 290 291 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 292 293 return monitor->complete_exit(THREAD); 294 } 295 296 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 297 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 298 TEVENT (reenter) ; 299 if (UseBiasedLocking) { 300 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 301 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 302 } 303 304 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 305 306 monitor->reenter(recursion, THREAD); 307 } 308 // ----------------------------------------------------------------------------- 309 // JNI locks on java objects 310 // NOTE: must use heavy weight monitor to handle jni monitor enter 311 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter 312 // the current locking is from JNI instead of Java code 313 TEVENT (jni_enter) ; 314 if (UseBiasedLocking) { 315 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 316 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 317 } 318 THREAD->set_current_pending_monitor_is_from_java(false); 319 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 320 THREAD->set_current_pending_monitor_is_from_java(true); 321 } 322 323 // NOTE: must use heavy weight monitor to handle jni monitor enter 324 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) { 325 if (UseBiasedLocking) { 326 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 327 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 328 } 329 330 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj()); 331 return monitor->try_enter(THREAD); 332 } 333 334 335 // NOTE: must use heavy weight monitor to handle jni monitor exit 336 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 337 TEVENT (jni_exit) ; 338 if (UseBiasedLocking) { 339 Handle h_obj(THREAD, obj); 340 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 341 obj = h_obj(); 342 } 343 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 344 345 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); 346 // If this thread has locked the object, exit the monitor. Note: can't use 347 // monitor->check(CHECK); must exit even if an exception is pending. 348 if (monitor->check(THREAD)) { 349 monitor->exit(true, THREAD); 350 } 351 } 352 353 // ----------------------------------------------------------------------------- 354 // Internal VM locks on java objects 355 // standard constructor, allows locking failures 356 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 357 _dolock = doLock; 358 _thread = thread; 359 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 360 _obj = obj; 361 362 if (_dolock) { 363 TEVENT (ObjectLocker) ; 364 365 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 366 } 367 } 368 369 ObjectLocker::~ObjectLocker() { 370 if (_dolock) { 371 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 372 } 373 } 374 375 376 // ----------------------------------------------------------------------------- 377 // Wait/Notify/NotifyAll 378 // NOTE: must use heavy weight monitor to handle wait() 379 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 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 if (millis < 0) { 385 TEVENT (wait - throw IAX) ; 386 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 387 } 388 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 389 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 390 monitor->wait(millis, true, THREAD); 391 392 /* This dummy call is in place to get around dtrace bug 6254741. Once 393 that's fixed we can uncomment the following line and remove the call */ 394 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 395 dtrace_waited_probe(monitor, obj, THREAD); 396 } 397 398 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) { 399 if (UseBiasedLocking) { 400 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 401 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 402 } 403 if (millis < 0) { 404 TEVENT (wait - throw IAX) ; 405 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 406 } 407 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ; 408 } 409 410 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 411 if (UseBiasedLocking) { 412 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 413 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 414 } 415 416 markOop mark = obj->mark(); 417 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 418 return; 419 } 420 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); 421 } 422 423 // NOTE: see comment of notify() 424 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 425 if (UseBiasedLocking) { 426 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 427 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 428 } 429 430 markOop mark = obj->mark(); 431 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 432 return; 433 } 434 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); 435 } 436 437 // ----------------------------------------------------------------------------- 438 // Hash Code handling 439 // 440 // Performance concern: 441 // OrderAccess::storestore() calls release() which at one time stored 0 442 // into the global volatile OrderAccess::dummy variable. This store was 443 // unnecessary for correctness. Many threads storing into a common location 444 // causes considerable cache migration or "sloshing" on large SMP systems. 445 // As such, I avoided using OrderAccess::storestore(). In some cases 446 // OrderAccess::fence() -- which incurs local latency on the executing 447 // processor -- is a better choice as it scales on SMP systems. 448 // 449 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 450 // a discussion of coherency costs. Note that all our current reference 451 // platforms provide strong ST-ST order, so the issue is moot on IA32, 452 // x64, and SPARC. 453 // 454 // As a general policy we use "volatile" to control compiler-based reordering 455 // and explicit fences (barriers) to control for architectural reordering 456 // performed by the CPU(s) or platform. 457 458 struct SharedGlobals { 459 // These are highly shared mostly-read variables. 460 // To avoid false-sharing they need to be the sole occupants of a $ line. 461 double padPrefix [8]; 462 volatile int stwRandom ; 463 volatile int stwCycle ; 464 465 // Hot RW variables -- Sequester to avoid false-sharing 466 double padSuffix [16]; 467 volatile int hcSequence ; 468 double padFinal [8] ; 469 } ; 470 471 static SharedGlobals GVars ; 472 static int MonitorScavengeThreshold = 1000000 ; 473 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending 474 475 static markOop ReadStableMark (oop obj) { 476 markOop mark = obj->mark() ; 477 if (!mark->is_being_inflated()) { 478 return mark ; // normal fast-path return 479 } 480 481 int its = 0 ; 482 for (;;) { 483 markOop mark = obj->mark() ; 484 if (!mark->is_being_inflated()) { 485 return mark ; // normal fast-path return 486 } 487 488 // The object is being inflated by some other thread. 489 // The caller of ReadStableMark() must wait for inflation to complete. 490 // Avoid live-lock 491 // TODO: consider calling SafepointSynchronize::do_call_back() while 492 // spinning to see if there's a safepoint pending. If so, immediately 493 // yielding or blocking would be appropriate. Avoid spinning while 494 // there is a safepoint pending. 495 // TODO: add inflation contention performance counters. 496 // TODO: restrict the aggregate number of spinners. 497 498 ++its ; 499 if (its > 10000 || !os::is_MP()) { 500 if (its & 1) { 501 os::NakedYield() ; 502 TEVENT (Inflate: INFLATING - yield) ; 503 } else { 504 // Note that the following code attenuates the livelock problem but is not 505 // a complete remedy. A more complete solution would require that the inflating 506 // thread hold the associated inflation lock. The following code simply restricts 507 // the number of spinners to at most one. We'll have N-2 threads blocked 508 // on the inflationlock, 1 thread holding the inflation lock and using 509 // a yield/park strategy, and 1 thread in the midst of inflation. 510 // A more refined approach would be to change the encoding of INFLATING 511 // to allow encapsulation of a native thread pointer. Threads waiting for 512 // inflation to complete would use CAS to push themselves onto a singly linked 513 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 514 // and calling park(). When inflation was complete the thread that accomplished inflation 515 // would detach the list and set the markword to inflated with a single CAS and 516 // then for each thread on the list, set the flag and unpark() the thread. 517 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 518 // wakes at most one thread whereas we need to wake the entire list. 519 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1) ; 520 int YieldThenBlock = 0 ; 521 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ; 522 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ; 523 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ; 524 while (obj->mark() == markOopDesc::INFLATING()) { 525 // Beware: NakedYield() is advisory and has almost no effect on some platforms 526 // so we periodically call Self->_ParkEvent->park(1). 527 // We use a mixed spin/yield/block mechanism. 528 if ((YieldThenBlock++) >= 16) { 529 Thread::current()->_ParkEvent->park(1) ; 530 } else { 531 os::NakedYield() ; 532 } 533 } 534 Thread::muxRelease (InflationLocks + ix ) ; 535 TEVENT (Inflate: INFLATING - yield/park) ; 536 } 537 } else { 538 SpinPause() ; // SMP-polite spinning 539 } 540 } 541 } 542 543 // hashCode() generation : 544 // 545 // Possibilities: 546 // * MD5Digest of {obj,stwRandom} 547 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 548 // * A DES- or AES-style SBox[] mechanism 549 // * One of the Phi-based schemes, such as: 550 // 2654435761 = 2^32 * Phi (golden ratio) 551 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 552 // * A variation of Marsaglia's shift-xor RNG scheme. 553 // * (obj ^ stwRandom) is appealing, but can result 554 // in undesirable regularity in the hashCode values of adjacent objects 555 // (objects allocated back-to-back, in particular). This could potentially 556 // result in hashtable collisions and reduced hashtable efficiency. 557 // There are simple ways to "diffuse" the middle address bits over the 558 // generated hashCode values: 559 // 560 561 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 562 intptr_t value = 0 ; 563 if (hashCode == 0) { 564 // This form uses an unguarded global Park-Miller RNG, 565 // so it's possible for two threads to race and generate the same RNG. 566 // On MP system we'll have lots of RW access to a global, so the 567 // mechanism induces lots of coherency traffic. 568 value = os::random() ; 569 } else 570 if (hashCode == 1) { 571 // This variation has the property of being stable (idempotent) 572 // between STW operations. This can be useful in some of the 1-0 573 // synchronization schemes. 574 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3 ; 575 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ; 576 } else 577 if (hashCode == 2) { 578 value = 1 ; // for sensitivity testing 579 } else 580 if (hashCode == 3) { 581 value = ++GVars.hcSequence ; 582 } else 583 if (hashCode == 4) { 584 value = cast_from_oop<intptr_t>(obj) ; 585 } else { 586 // Marsaglia's xor-shift scheme with thread-specific state 587 // This is probably the best overall implementation -- we'll 588 // likely make this the default in future releases. 589 unsigned t = Self->_hashStateX ; 590 t ^= (t << 11) ; 591 Self->_hashStateX = Self->_hashStateY ; 592 Self->_hashStateY = Self->_hashStateZ ; 593 Self->_hashStateZ = Self->_hashStateW ; 594 unsigned v = Self->_hashStateW ; 595 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ; 596 Self->_hashStateW = v ; 597 value = v ; 598 } 599 600 value &= markOopDesc::hash_mask; 601 if (value == 0) value = 0xBAD ; 602 assert (value != markOopDesc::no_hash, "invariant") ; 603 TEVENT (hashCode: GENERATE) ; 604 return value; 605 } 606 // 607 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { 608 if (UseBiasedLocking) { 609 // NOTE: many places throughout the JVM do not expect a safepoint 610 // to be taken here, in particular most operations on perm gen 611 // objects. However, we only ever bias Java instances and all of 612 // the call sites of identity_hash that might revoke biases have 613 // been checked to make sure they can handle a safepoint. The 614 // added check of the bias pattern is to avoid useless calls to 615 // thread-local storage. 616 if (obj->mark()->has_bias_pattern()) { 617 // Box and unbox the raw reference just in case we cause a STW safepoint. 618 Handle hobj (Self, obj) ; 619 // Relaxing assertion for bug 6320749. 620 assert (Universe::verify_in_progress() || 621 !SafepointSynchronize::is_at_safepoint(), 622 "biases should not be seen by VM thread here"); 623 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 624 obj = hobj() ; 625 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 626 } 627 } 628 629 // hashCode() is a heap mutator ... 630 // Relaxing assertion for bug 6320749. 631 assert (Universe::verify_in_progress() || 632 !SafepointSynchronize::is_at_safepoint(), "invariant") ; 633 assert (Universe::verify_in_progress() || 634 Self->is_Java_thread() , "invariant") ; 635 assert (Universe::verify_in_progress() || 636 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; 637 638 ObjectMonitor* monitor = NULL; 639 markOop temp, test; 640 intptr_t hash; 641 markOop mark = ReadStableMark (obj); 642 643 // object should remain ineligible for biased locking 644 assert (!mark->has_bias_pattern(), "invariant") ; 645 646 if (mark->is_neutral()) { 647 hash = mark->hash(); // this is a normal header 648 if (hash) { // if it has hash, just return it 649 return hash; 650 } 651 hash = get_next_hash(Self, obj); // allocate a new hash code 652 temp = mark->copy_set_hash(hash); // merge the hash code into header 653 // use (machine word version) atomic operation to install the hash 654 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 655 if (test == mark) { 656 return hash; 657 } 658 // If atomic operation failed, we must inflate the header 659 // into heavy weight monitor. We could add more code here 660 // for fast path, but it does not worth the complexity. 661 } else if (mark->has_monitor()) { 662 monitor = mark->monitor(); 663 temp = monitor->header(); 664 assert (temp->is_neutral(), "invariant") ; 665 hash = temp->hash(); 666 if (hash) { 667 return hash; 668 } 669 // Skip to the following code to reduce code size 670 } else if (Self->is_lock_owned((address)mark->locker())) { 671 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 672 assert (temp->is_neutral(), "invariant") ; 673 hash = temp->hash(); // by current thread, check if the displaced 674 if (hash) { // header contains hash code 675 return hash; 676 } 677 // WARNING: 678 // The displaced header is strictly immutable. 679 // It can NOT be changed in ANY cases. So we have 680 // to inflate the header into heavyweight monitor 681 // even the current thread owns the lock. The reason 682 // is the BasicLock (stack slot) will be asynchronously 683 // read by other threads during the inflate() function. 684 // Any change to stack may not propagate to other threads 685 // correctly. 686 } 687 688 // Inflate the monitor to set hash code 689 monitor = ObjectSynchronizer::inflate(Self, obj); 690 // Load displaced header and check it has hash code 691 mark = monitor->header(); 692 assert (mark->is_neutral(), "invariant") ; 693 hash = mark->hash(); 694 if (hash == 0) { 695 hash = get_next_hash(Self, obj); 696 temp = mark->copy_set_hash(hash); // merge hash code into header 697 assert (temp->is_neutral(), "invariant") ; 698 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 699 if (test != mark) { 700 // The only update to the header in the monitor (outside GC) 701 // is install the hash code. If someone add new usage of 702 // displaced header, please update this code 703 hash = test->hash(); 704 assert (test->is_neutral(), "invariant") ; 705 assert (hash != 0, "Trivial unexpected object/monitor header usage."); 706 } 707 } 708 // We finally get the hash 709 return hash; 710 } 711 712 // Deprecated -- use FastHashCode() instead. 713 714 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 715 return FastHashCode (Thread::current(), obj()) ; 716 } 717 718 719 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 720 Handle h_obj) { 721 if (UseBiasedLocking) { 722 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 723 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 724 } 725 726 assert(thread == JavaThread::current(), "Can only be called on current thread"); 727 oop obj = h_obj(); 728 729 markOop mark = ReadStableMark (obj) ; 730 731 // Uncontended case, header points to stack 732 if (mark->has_locker()) { 733 return thread->is_lock_owned((address)mark->locker()); 734 } 735 // Contended case, header points to ObjectMonitor (tagged pointer) 736 if (mark->has_monitor()) { 737 ObjectMonitor* monitor = mark->monitor(); 738 return monitor->is_entered(thread) != 0 ; 739 } 740 // Unlocked case, header in place 741 assert(mark->is_neutral(), "sanity check"); 742 return false; 743 } 744 745 // Be aware of this method could revoke bias of the lock object. 746 // This method querys the ownership of the lock handle specified by 'h_obj'. 747 // If the current thread owns the lock, it returns owner_self. If no 748 // thread owns the lock, it returns owner_none. Otherwise, it will return 749 // ower_other. 750 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 751 (JavaThread *self, Handle h_obj) { 752 // The caller must beware this method can revoke bias, and 753 // revocation can result in a safepoint. 754 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; 755 assert (self->thread_state() != _thread_blocked , "invariant") ; 756 757 // Possible mark states: neutral, biased, stack-locked, inflated 758 759 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 760 // CASE: biased 761 BiasedLocking::revoke_and_rebias(h_obj, false, self); 762 assert(!h_obj->mark()->has_bias_pattern(), 763 "biases should be revoked by now"); 764 } 765 766 assert(self == JavaThread::current(), "Can only be called on current thread"); 767 oop obj = h_obj(); 768 markOop mark = ReadStableMark (obj) ; 769 770 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 771 if (mark->has_locker()) { 772 return self->is_lock_owned((address)mark->locker()) ? 773 owner_self : owner_other; 774 } 775 776 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 777 // The Object:ObjectMonitor relationship is stable as long as we're 778 // not at a safepoint. 779 if (mark->has_monitor()) { 780 void * owner = mark->monitor()->_owner ; 781 if (owner == NULL) return owner_none ; 782 return (owner == self || 783 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 784 } 785 786 // CASE: neutral 787 assert(mark->is_neutral(), "sanity check"); 788 return owner_none ; // it's unlocked 789 } 790 791 // FIXME: jvmti should call this 792 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 793 if (UseBiasedLocking) { 794 if (SafepointSynchronize::is_at_safepoint()) { 795 BiasedLocking::revoke_at_safepoint(h_obj); 796 } else { 797 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 798 } 799 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 800 } 801 802 oop obj = h_obj(); 803 address owner = NULL; 804 805 markOop mark = ReadStableMark (obj) ; 806 807 // Uncontended case, header points to stack 808 if (mark->has_locker()) { 809 owner = (address) mark->locker(); 810 } 811 812 // Contended case, header points to ObjectMonitor (tagged pointer) 813 if (mark->has_monitor()) { 814 ObjectMonitor* monitor = mark->monitor(); 815 assert(monitor != NULL, "monitor should be non-null"); 816 owner = (address) monitor->owner(); 817 } 818 819 if (owner != NULL) { 820 // owning_thread_from_monitor_owner() may also return NULL here 821 return Threads::owning_thread_from_monitor_owner(owner, doLock); 822 } 823 824 // Unlocked case, header in place 825 // Cannot have assertion since this object may have been 826 // locked by another thread when reaching here. 827 // assert(mark->is_neutral(), "sanity check"); 828 829 return NULL; 830 } 831 // Visitors ... 832 833 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 834 ObjectMonitor* block = 835 (ObjectMonitor*)OrderAccess::load_ptr_acquire(&gBlockList); 836 while (block != NULL) { 837 assert(block->object() == CHAINMARKER, "must be a block header"); 838 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 839 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 840 oop object = (oop)mid->object(); 841 if (object != NULL) { 842 closure->do_monitor(mid); 843 } 844 } 845 block = (ObjectMonitor*)block->FreeNext; 846 } 847 } 848 849 // Get the next block in the block list. 850 static inline ObjectMonitor* next(ObjectMonitor* block) { 851 assert(block->object() == CHAINMARKER, "must be a block header"); 852 block = block->FreeNext ; 853 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 854 return block; 855 } 856 857 858 void ObjectSynchronizer::oops_do(OopClosure* f) { 859 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 860 ObjectMonitor* block = 861 (ObjectMonitor*)OrderAccess::load_ptr_acquire(&gBlockList); 862 for (; block != NULL; block = (ObjectMonitor *)next(block)) { 863 assert(block->object() == CHAINMARKER, "must be a block header"); 864 for (int i = 1; i < _BLOCKSIZE; i++) { 865 ObjectMonitor* mid = &block[i]; 866 if (mid->object() != NULL) { 867 f->do_oop((oop*)mid->object_addr()); 868 } 869 } 870 } 871 } 872 873 874 // ----------------------------------------------------------------------------- 875 // ObjectMonitor Lifecycle 876 // ----------------------- 877 // Inflation unlinks monitors from the global gFreeList and 878 // associates them with objects. Deflation -- which occurs at 879 // STW-time -- disassociates idle monitors from objects. Such 880 // scavenged monitors are returned to the gFreeList. 881 // 882 // The global list is protected by ListLock. All the critical sections 883 // are short and operate in constant-time. 884 // 885 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 886 // 887 // Lifecycle: 888 // -- unassigned and on the global free list 889 // -- unassigned and on a thread's private omFreeList 890 // -- assigned to an object. The object is inflated and the mark refers 891 // to the objectmonitor. 892 // 893 894 895 // Constraining monitor pool growth via MonitorBound ... 896 // 897 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 898 // the rate of scavenging is driven primarily by GC. As such, we can find 899 // an inordinate number of monitors in circulation. 900 // To avoid that scenario we can artificially induce a STW safepoint 901 // if the pool appears to be growing past some reasonable bound. 902 // Generally we favor time in space-time tradeoffs, but as there's no 903 // natural back-pressure on the # of extant monitors we need to impose some 904 // type of limit. Beware that if MonitorBound is set to too low a value 905 // we could just loop. In addition, if MonitorBound is set to a low value 906 // we'll incur more safepoints, which are harmful to performance. 907 // See also: GuaranteedSafepointInterval 908 // 909 // The current implementation uses asynchronous VM operations. 910 // 911 912 static void InduceScavenge (Thread * Self, const char * Whence) { 913 // Induce STW safepoint to trim monitors 914 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 915 // More precisely, trigger an asynchronous STW safepoint as the number 916 // of active monitors passes the specified threshold. 917 // TODO: assert thread state is reasonable 918 919 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 920 if (ObjectMonitor::Knob_Verbose) { 921 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; 922 ::fflush(stdout) ; 923 } 924 // Induce a 'null' safepoint to scavenge monitors 925 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 926 // to the VMthread and have a lifespan longer than that of this activation record. 927 // The VMThread will delete the op when completed. 928 VMThread::execute (new VM_ForceAsyncSafepoint()) ; 929 930 if (ObjectMonitor::Knob_Verbose) { 931 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; 932 ::fflush(stdout) ; 933 } 934 } 935 } 936 /* Too slow for general assert or debug 937 void ObjectSynchronizer::verifyInUse (Thread *Self) { 938 ObjectMonitor* mid; 939 int inusetally = 0; 940 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 941 inusetally ++; 942 } 943 assert(inusetally == Self->omInUseCount, "inuse count off"); 944 945 int freetally = 0; 946 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 947 freetally ++; 948 } 949 assert(freetally == Self->omFreeCount, "free count off"); 950 } 951 */ 952 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) { 953 // A large MAXPRIVATE value reduces both list lock contention 954 // and list coherency traffic, but also tends to increase the 955 // number of objectMonitors in circulation as well as the STW 956 // scavenge costs. As usual, we lean toward time in space-time 957 // tradeoffs. 958 const int MAXPRIVATE = 1024 ; 959 for (;;) { 960 ObjectMonitor * m ; 961 962 // 1: try to allocate from the thread's local omFreeList. 963 // Threads will attempt to allocate first from their local list, then 964 // from the global list, and only after those attempts fail will the thread 965 // attempt to instantiate new monitors. Thread-local free lists take 966 // heat off the ListLock and improve allocation latency, as well as reducing 967 // coherency traffic on the shared global list. 968 m = Self->omFreeList ; 969 if (m != NULL) { 970 Self->omFreeList = m->FreeNext ; 971 Self->omFreeCount -- ; 972 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 973 guarantee (m->object() == NULL, "invariant") ; 974 if (MonitorInUseLists) { 975 m->FreeNext = Self->omInUseList; 976 Self->omInUseList = m; 977 Self->omInUseCount ++; 978 // verifyInUse(Self); 979 } else { 980 m->FreeNext = NULL; 981 } 982 return m ; 983 } 984 985 // 2: try to allocate from the global gFreeList 986 // CONSIDER: use muxTry() instead of muxAcquire(). 987 // If the muxTry() fails then drop immediately into case 3. 988 // If we're using thread-local free lists then try 989 // to reprovision the caller's free list. 990 if (gFreeList != NULL) { 991 // Reprovision the thread's omFreeList. 992 // Use bulk transfers to reduce the allocation rate and heat 993 // on various locks. 994 Thread::muxAcquire (&ListLock, "omAlloc") ; 995 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) { 996 MonitorFreeCount --; 997 ObjectMonitor * take = gFreeList ; 998 gFreeList = take->FreeNext ; 999 guarantee (take->object() == NULL, "invariant") ; 1000 guarantee (!take->is_busy(), "invariant") ; 1001 take->Recycle() ; 1002 omRelease (Self, take, false) ; 1003 } 1004 Thread::muxRelease (&ListLock) ; 1005 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ; 1006 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ; 1007 TEVENT (omFirst - reprovision) ; 1008 1009 const int mx = MonitorBound ; 1010 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) { 1011 // We can't safely induce a STW safepoint from omAlloc() as our thread 1012 // state may not be appropriate for such activities and callers may hold 1013 // naked oops, so instead we defer the action. 1014 InduceScavenge (Self, "omAlloc") ; 1015 } 1016 continue; 1017 } 1018 1019 // 3: allocate a block of new ObjectMonitors 1020 // Both the local and global free lists are empty -- resort to malloc(). 1021 // In the current implementation objectMonitors are TSM - immortal. 1022 assert (_BLOCKSIZE > 1, "invariant") ; 1023 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; 1024 1025 // NOTE: (almost) no way to recover if allocation failed. 1026 // We might be able to induce a STW safepoint and scavenge enough 1027 // objectMonitors to permit progress. 1028 if (temp == NULL) { 1029 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), OOM_MALLOC_ERROR, 1030 "Allocate ObjectMonitors"); 1031 } 1032 1033 // Format the block. 1034 // initialize the linked list, each monitor points to its next 1035 // forming the single linked free list, the very first monitor 1036 // will points to next block, which forms the block list. 1037 // The trick of using the 1st element in the block as gBlockList 1038 // linkage should be reconsidered. A better implementation would 1039 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1040 1041 for (int i = 1; i < _BLOCKSIZE ; i++) { 1042 temp[i].FreeNext = &temp[i+1]; 1043 } 1044 1045 // terminate the last monitor as the end of list 1046 temp[_BLOCKSIZE - 1].FreeNext = NULL ; 1047 1048 // Element [0] is reserved for global list linkage 1049 temp[0].set_object(CHAINMARKER); 1050 1051 // Consider carving out this thread's current request from the 1052 // block in hand. This avoids some lock traffic and redundant 1053 // list activity. 1054 1055 // Acquire the ListLock to manipulate BlockList and FreeList. 1056 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1057 Thread::muxAcquire (&ListLock, "omAlloc [2]") ; 1058 MonitorPopulation += _BLOCKSIZE-1; 1059 MonitorFreeCount += _BLOCKSIZE-1; 1060 1061 // Add the new block to the list of extant blocks (gBlockList). 1062 // The very first objectMonitor in a block is reserved and dedicated. 1063 // It serves as blocklist "next" linkage. 1064 temp[0].FreeNext = gBlockList; 1065 // There are lock-free uses of gBlockList so make sure that 1066 // the previous stores happen before we update gBlockList. 1067 OrderAccess::release_store_ptr(&gBlockList, temp); 1068 1069 // Add the new string of objectMonitors to the global free list 1070 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ; 1071 gFreeList = temp + 1; 1072 Thread::muxRelease (&ListLock) ; 1073 TEVENT (Allocate block of monitors) ; 1074 } 1075 } 1076 1077 // Place "m" on the caller's private per-thread omFreeList. 1078 // In practice there's no need to clamp or limit the number of 1079 // monitors on a thread's omFreeList as the only time we'll call 1080 // omRelease is to return a monitor to the free list after a CAS 1081 // attempt failed. This doesn't allow unbounded #s of monitors to 1082 // accumulate on a thread's free list. 1083 // 1084 1085 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) { 1086 guarantee (m->object() == NULL, "invariant") ; 1087 1088 // Remove from omInUseList 1089 if (MonitorInUseLists && fromPerThreadAlloc) { 1090 ObjectMonitor* curmidinuse = NULL; 1091 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) { 1092 if (m == mid) { 1093 // extract from per-thread in-use-list 1094 if (mid == Self->omInUseList) { 1095 Self->omInUseList = mid->FreeNext; 1096 } else if (curmidinuse != NULL) { 1097 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist 1098 } 1099 Self->omInUseCount --; 1100 // verifyInUse(Self); 1101 break; 1102 } else { 1103 curmidinuse = mid; 1104 mid = mid->FreeNext; 1105 } 1106 } 1107 } 1108 1109 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new 1110 m->FreeNext = Self->omFreeList ; 1111 Self->omFreeList = m ; 1112 Self->omFreeCount ++ ; 1113 } 1114 1115 // Return the monitors of a moribund thread's local free list to 1116 // the global free list. Typically a thread calls omFlush() when 1117 // it's dying. We could also consider having the VM thread steal 1118 // monitors from threads that have not run java code over a few 1119 // consecutive STW safepoints. Relatedly, we might decay 1120 // omFreeProvision at STW safepoints. 1121 // 1122 // Also return the monitors of a moribund thread"s omInUseList to 1123 // a global gOmInUseList under the global list lock so these 1124 // will continue to be scanned. 1125 // 1126 // We currently call omFlush() from the Thread:: dtor _after the thread 1127 // has been excised from the thread list and is no longer a mutator. 1128 // That means that omFlush() can run concurrently with a safepoint and 1129 // the scavenge operator. Calling omFlush() from JavaThread::exit() might 1130 // be a better choice as we could safely reason that that the JVM is 1131 // not at a safepoint at the time of the call, and thus there could 1132 // be not inopportune interleavings between omFlush() and the scavenge 1133 // operator. 1134 1135 void ObjectSynchronizer::omFlush (Thread * Self) { 1136 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL 1137 Self->omFreeList = NULL ; 1138 ObjectMonitor * Tail = NULL ; 1139 int Tally = 0; 1140 if (List != NULL) { 1141 ObjectMonitor * s ; 1142 for (s = List ; s != NULL ; s = s->FreeNext) { 1143 Tally ++ ; 1144 Tail = s ; 1145 guarantee (s->object() == NULL, "invariant") ; 1146 guarantee (!s->is_busy(), "invariant") ; 1147 s->set_owner (NULL) ; // redundant but good hygiene 1148 TEVENT (omFlush - Move one) ; 1149 } 1150 guarantee (Tail != NULL && List != NULL, "invariant") ; 1151 } 1152 1153 ObjectMonitor * InUseList = Self->omInUseList; 1154 ObjectMonitor * InUseTail = NULL ; 1155 int InUseTally = 0; 1156 if (InUseList != NULL) { 1157 Self->omInUseList = NULL; 1158 ObjectMonitor *curom; 1159 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) { 1160 InUseTail = curom; 1161 InUseTally++; 1162 } 1163 // TODO debug 1164 assert(Self->omInUseCount == InUseTally, "inuse count off"); 1165 Self->omInUseCount = 0; 1166 guarantee (InUseTail != NULL && InUseList != NULL, "invariant"); 1167 } 1168 1169 Thread::muxAcquire (&ListLock, "omFlush") ; 1170 if (Tail != NULL) { 1171 Tail->FreeNext = gFreeList ; 1172 gFreeList = List ; 1173 MonitorFreeCount += Tally; 1174 } 1175 1176 if (InUseTail != NULL) { 1177 InUseTail->FreeNext = gOmInUseList; 1178 gOmInUseList = InUseList; 1179 gOmInUseCount += InUseTally; 1180 } 1181 1182 Thread::muxRelease (&ListLock) ; 1183 TEVENT (omFlush) ; 1184 } 1185 1186 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1187 const oop obj) { 1188 assert(event != NULL, "invariant"); 1189 assert(event->should_commit(), "invariant"); 1190 event->set_monitorClass(obj->klass()); 1191 event->set_address((uintptr_t)(void*)obj); 1192 event->commit(); 1193 } 1194 1195 // Fast path code shared by multiple functions 1196 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1197 markOop mark = obj->mark(); 1198 if (mark->has_monitor()) { 1199 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1200 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1201 return mark->monitor(); 1202 } 1203 return ObjectSynchronizer::inflate(Thread::current(), obj); 1204 } 1205 1206 1207 // Note that we could encounter some performance loss through false-sharing as 1208 // multiple locks occupy the same $ line. Padding might be appropriate. 1209 1210 1211 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) { 1212 // Inflate mutates the heap ... 1213 // Relaxing assertion for bug 6320749. 1214 assert (Universe::verify_in_progress() || 1215 !SafepointSynchronize::is_at_safepoint(), "invariant") ; 1216 1217 EventJavaMonitorInflate event; 1218 1219 for (;;) { 1220 const markOop mark = object->mark() ; 1221 assert (!mark->has_bias_pattern(), "invariant") ; 1222 1223 // The mark can be in one of the following states: 1224 // * Inflated - just return 1225 // * Stack-locked - coerce it to inflated 1226 // * INFLATING - busy wait for conversion to complete 1227 // * Neutral - aggressively inflate the object. 1228 // * BIASED - Illegal. We should never see this 1229 1230 // CASE: inflated 1231 if (mark->has_monitor()) { 1232 ObjectMonitor * inf = mark->monitor() ; 1233 assert (inf->header()->is_neutral(), "invariant"); 1234 assert (inf->object() == object, "invariant") ; 1235 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1236 return inf ; 1237 } 1238 1239 // CASE: inflation in progress - inflating over a stack-lock. 1240 // Some other thread is converting from stack-locked to inflated. 1241 // Only that thread can complete inflation -- other threads must wait. 1242 // The INFLATING value is transient. 1243 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1244 // We could always eliminate polling by parking the thread on some auxiliary list. 1245 if (mark == markOopDesc::INFLATING()) { 1246 TEVENT (Inflate: spin while INFLATING) ; 1247 ReadStableMark(object) ; 1248 continue ; 1249 } 1250 1251 // CASE: stack-locked 1252 // Could be stack-locked either by this thread or by some other thread. 1253 // 1254 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1255 // to install INFLATING into the mark word. We originally installed INFLATING, 1256 // allocated the objectmonitor, and then finally STed the address of the 1257 // objectmonitor into the mark. This was correct, but artificially lengthened 1258 // the interval in which INFLATED appeared in the mark, thus increasing 1259 // the odds of inflation contention. 1260 // 1261 // We now use per-thread private objectmonitor free lists. 1262 // These list are reprovisioned from the global free list outside the 1263 // critical INFLATING...ST interval. A thread can transfer 1264 // multiple objectmonitors en-mass from the global free list to its local free list. 1265 // This reduces coherency traffic and lock contention on the global free list. 1266 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1267 // before or after the CAS(INFLATING) operation. 1268 // See the comments in omAlloc(). 1269 1270 if (mark->has_locker()) { 1271 ObjectMonitor * m = omAlloc (Self) ; 1272 // Optimistically prepare the objectmonitor - anticipate successful CAS 1273 // We do this before the CAS in order to minimize the length of time 1274 // in which INFLATING appears in the mark. 1275 m->Recycle(); 1276 m->_Responsible = NULL ; 1277 m->OwnerIsThread = 0 ; 1278 m->_recursions = 0 ; 1279 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // Consider: maintain by type/class 1280 1281 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ; 1282 if (cmp != mark) { 1283 omRelease (Self, m, true) ; 1284 continue ; // Interference -- just retry 1285 } 1286 1287 // We've successfully installed INFLATING (0) into the mark-word. 1288 // This is the only case where 0 will appear in a mark-work. 1289 // Only the singular thread that successfully swings the mark-word 1290 // to 0 can perform (or more precisely, complete) inflation. 1291 // 1292 // Why do we CAS a 0 into the mark-word instead of just CASing the 1293 // mark-word from the stack-locked value directly to the new inflated state? 1294 // Consider what happens when a thread unlocks a stack-locked object. 1295 // It attempts to use CAS to swing the displaced header value from the 1296 // on-stack basiclock back into the object header. Recall also that the 1297 // header value (hashcode, etc) can reside in (a) the object header, or 1298 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1299 // header in an objectMonitor. The inflate() routine must copy the header 1300 // value from the basiclock on the owner's stack to the objectMonitor, all 1301 // the while preserving the hashCode stability invariants. If the owner 1302 // decides to release the lock while the value is 0, the unlock will fail 1303 // and control will eventually pass from slow_exit() to inflate. The owner 1304 // will then spin, waiting for the 0 value to disappear. Put another way, 1305 // the 0 causes the owner to stall if the owner happens to try to 1306 // drop the lock (restoring the header from the basiclock to the object) 1307 // while inflation is in-progress. This protocol avoids races that might 1308 // would otherwise permit hashCode values to change or "flicker" for an object. 1309 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1310 // 0 serves as a "BUSY" inflate-in-progress indicator. 1311 1312 1313 // fetch the displaced mark from the owner's stack. 1314 // The owner can't die or unwind past the lock while our INFLATING 1315 // object is in the mark. Furthermore the owner can't complete 1316 // an unlock on the object, either. 1317 markOop dmw = mark->displaced_mark_helper() ; 1318 assert (dmw->is_neutral(), "invariant") ; 1319 1320 // Setup monitor fields to proper values -- prepare the monitor 1321 m->set_header(dmw) ; 1322 1323 // Optimization: if the mark->locker stack address is associated 1324 // with this thread we could simply set m->_owner = Self and 1325 // m->OwnerIsThread = 1. Note that a thread can inflate an object 1326 // that it has stack-locked -- as might happen in wait() -- directly 1327 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1328 m->set_owner(mark->locker()); 1329 m->set_object(object); 1330 // TODO-FIXME: assert BasicLock->dhw != 0. 1331 1332 // Must preserve store ordering. The monitor state must 1333 // be stable at the time of publishing the monitor address. 1334 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ; 1335 object->release_set_mark(markOopDesc::encode(m)); 1336 1337 // Hopefully the performance counters are allocated on distinct cache lines 1338 // to avoid false sharing on MP systems ... 1339 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ; 1340 TEVENT(Inflate: overwrite stacklock) ; 1341 if (TraceMonitorInflation) { 1342 if (object->is_instance()) { 1343 ResourceMark rm; 1344 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1345 (void *) object, (intptr_t) object->mark(), 1346 object->klass()->external_name()); 1347 } 1348 } 1349 if (event.should_commit()) { 1350 post_monitor_inflate_event(&event, object); 1351 } 1352 return m ; 1353 } 1354 1355 // CASE: neutral 1356 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1357 // If we know we're inflating for entry it's better to inflate by swinging a 1358 // pre-locked objectMonitor pointer into the object header. A successful 1359 // CAS inflates the object *and* confers ownership to the inflating thread. 1360 // In the current implementation we use a 2-step mechanism where we CAS() 1361 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1362 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1363 // would be useful. 1364 1365 assert (mark->is_neutral(), "invariant"); 1366 ObjectMonitor * m = omAlloc (Self) ; 1367 // prepare m for installation - set monitor to initial state 1368 m->Recycle(); 1369 m->set_header(mark); 1370 m->set_owner(NULL); 1371 m->set_object(object); 1372 m->OwnerIsThread = 1 ; 1373 m->_recursions = 0 ; 1374 m->_Responsible = NULL ; 1375 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // consider: keep metastats by type/class 1376 1377 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1378 m->set_object (NULL) ; 1379 m->set_owner (NULL) ; 1380 m->OwnerIsThread = 0 ; 1381 m->Recycle() ; 1382 omRelease (Self, m, true) ; 1383 m = NULL ; 1384 continue ; 1385 // interference - the markword changed - just retry. 1386 // The state-transitions are one-way, so there's no chance of 1387 // live-lock -- "Inflated" is an absorbing state. 1388 } 1389 1390 // Hopefully the performance counters are allocated on distinct 1391 // cache lines to avoid false sharing on MP systems ... 1392 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ; 1393 TEVENT(Inflate: overwrite neutral) ; 1394 if (TraceMonitorInflation) { 1395 if (object->is_instance()) { 1396 ResourceMark rm; 1397 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1398 (void *) object, (intptr_t) object->mark(), 1399 object->klass()->external_name()); 1400 } 1401 } 1402 if (event.should_commit()) { 1403 post_monitor_inflate_event(&event, object); 1404 } 1405 return m ; 1406 } 1407 } 1408 1409 // Note that we could encounter some performance loss through false-sharing as 1410 // multiple locks occupy the same $ line. Padding might be appropriate. 1411 1412 1413 // Deflate_idle_monitors() is called at all safepoints, immediately 1414 // after all mutators are stopped, but before any objects have moved. 1415 // It traverses the list of known monitors, deflating where possible. 1416 // The scavenged monitor are returned to the monitor free list. 1417 // 1418 // Beware that we scavenge at *every* stop-the-world point. 1419 // Having a large number of monitors in-circulation negatively 1420 // impacts the performance of some applications (e.g., PointBase). 1421 // Broadly, we want to minimize the # of monitors in circulation. 1422 // 1423 // We have added a flag, MonitorInUseLists, which creates a list 1424 // of active monitors for each thread. deflate_idle_monitors() 1425 // only scans the per-thread inuse lists. omAlloc() puts all 1426 // assigned monitors on the per-thread list. deflate_idle_monitors() 1427 // returns the non-busy monitors to the global free list. 1428 // When a thread dies, omFlush() adds the list of active monitors for 1429 // that thread to a global gOmInUseList acquiring the 1430 // global list lock. deflate_idle_monitors() acquires the global 1431 // list lock to scan for non-busy monitors to the global free list. 1432 // An alternative could have used a single global inuse list. The 1433 // downside would have been the additional cost of acquiring the global list lock 1434 // for every omAlloc(). 1435 // 1436 // Perversely, the heap size -- and thus the STW safepoint rate -- 1437 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1438 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1439 // This is an unfortunate aspect of this design. 1440 // 1441 1442 enum ManifestConstants { 1443 ClearResponsibleAtSTW = 0, 1444 MaximumRecheckInterval = 1000 1445 } ; 1446 1447 // Deflate a single monitor if not in use 1448 // Return true if deflated, false if in use 1449 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1450 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) { 1451 bool deflated; 1452 // Normal case ... The monitor is associated with obj. 1453 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ; 1454 guarantee (mid == obj->mark()->monitor(), "invariant"); 1455 guarantee (mid->header()->is_neutral(), "invariant"); 1456 1457 if (mid->is_busy()) { 1458 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ; 1459 deflated = false; 1460 } else { 1461 // Deflate the monitor if it is no longer being used 1462 // It's idle - scavenge and return to the global free list 1463 // plain old deflation ... 1464 TEVENT (deflate_idle_monitors - scavenge1) ; 1465 if (TraceMonitorInflation) { 1466 if (obj->is_instance()) { 1467 ResourceMark rm; 1468 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1469 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name()); 1470 } 1471 } 1472 1473 // Restore the header back to obj 1474 obj->release_set_mark(mid->header()); 1475 mid->clear(); 1476 1477 assert (mid->object() == NULL, "invariant") ; 1478 1479 // Move the object to the working free list defined by FreeHead,FreeTail. 1480 if (*FreeHeadp == NULL) *FreeHeadp = mid; 1481 if (*FreeTailp != NULL) { 1482 ObjectMonitor * prevtail = *FreeTailp; 1483 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK 1484 prevtail->FreeNext = mid; 1485 } 1486 *FreeTailp = mid; 1487 deflated = true; 1488 } 1489 return deflated; 1490 } 1491 1492 // Caller acquires ListLock 1493 int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp, 1494 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) { 1495 ObjectMonitor* mid; 1496 ObjectMonitor* next; 1497 ObjectMonitor* curmidinuse = NULL; 1498 int deflatedcount = 0; 1499 1500 for (mid = *listheadp; mid != NULL; ) { 1501 oop obj = (oop) mid->object(); 1502 bool deflated = false; 1503 if (obj != NULL) { 1504 deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp); 1505 } 1506 if (deflated) { 1507 // extract from per-thread in-use-list 1508 if (mid == *listheadp) { 1509 *listheadp = mid->FreeNext; 1510 } else if (curmidinuse != NULL) { 1511 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist 1512 } 1513 next = mid->FreeNext; 1514 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list 1515 mid = next; 1516 deflatedcount++; 1517 } else { 1518 curmidinuse = mid; 1519 mid = mid->FreeNext; 1520 } 1521 } 1522 return deflatedcount; 1523 } 1524 1525 void ObjectSynchronizer::deflate_idle_monitors() { 1526 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1527 int nInuse = 0 ; // currently associated with objects 1528 int nInCirculation = 0 ; // extant 1529 int nScavenged = 0 ; // reclaimed 1530 bool deflated = false; 1531 1532 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors 1533 ObjectMonitor * FreeTail = NULL ; 1534 1535 TEVENT (deflate_idle_monitors) ; 1536 // Prevent omFlush from changing mids in Thread dtor's during deflation 1537 // And in case the vm thread is acquiring a lock during a safepoint 1538 // See e.g. 6320749 1539 Thread::muxAcquire (&ListLock, "scavenge - return") ; 1540 1541 if (MonitorInUseLists) { 1542 int inUse = 0; 1543 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { 1544 nInCirculation+= cur->omInUseCount; 1545 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail); 1546 cur->omInUseCount-= deflatedcount; 1547 // verifyInUse(cur); 1548 nScavenged += deflatedcount; 1549 nInuse += cur->omInUseCount; 1550 } 1551 1552 // For moribund threads, scan gOmInUseList 1553 if (gOmInUseList) { 1554 nInCirculation += gOmInUseCount; 1555 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail); 1556 gOmInUseCount-= deflatedcount; 1557 nScavenged += deflatedcount; 1558 nInuse += gOmInUseCount; 1559 } 1560 1561 } else { 1562 ObjectMonitor* block = 1563 (ObjectMonitor*)OrderAccess::load_ptr_acquire(&gBlockList); 1564 for (; block != NULL; block = (ObjectMonitor*)next(block)) { 1565 // Iterate over all extant monitors - Scavenge all idle monitors. 1566 assert(block->object() == CHAINMARKER, "must be a block header"); 1567 nInCirculation += _BLOCKSIZE; 1568 for (int i = 1; i < _BLOCKSIZE; i++) { 1569 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1570 oop obj = (oop)mid->object(); 1571 1572 if (obj == NULL) { 1573 // The monitor is not associated with an object. 1574 // The monitor should either be a thread-specific private 1575 // free list or the global free list. 1576 // obj == NULL IMPLIES mid->is_busy() == 0 1577 guarantee(!mid->is_busy(), "invariant"); 1578 continue; 1579 } 1580 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail); 1581 1582 if (deflated) { 1583 mid->FreeNext = NULL; 1584 nScavenged++; 1585 } else { 1586 nInuse++; 1587 } 1588 } 1589 } 1590 } 1591 1592 MonitorFreeCount += nScavenged; 1593 1594 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree. 1595 1596 if (ObjectMonitor::Knob_Verbose) { 1597 ::printf ("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n", 1598 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, 1599 MonitorPopulation, MonitorFreeCount) ; 1600 ::fflush(stdout) ; 1601 } 1602 1603 ForceMonitorScavenge = 0; // Reset 1604 1605 // Move the scavenged monitors back to the global free list. 1606 if (FreeHead != NULL) { 1607 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ; 1608 assert (FreeTail->FreeNext == NULL, "invariant") ; 1609 // constant-time list splice - prepend scavenged segment to gFreeList 1610 FreeTail->FreeNext = gFreeList ; 1611 gFreeList = FreeHead ; 1612 } 1613 Thread::muxRelease (&ListLock) ; 1614 1615 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ; 1616 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation); 1617 1618 // TODO: Add objectMonitor leak detection. 1619 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1620 GVars.stwRandom = os::random() ; 1621 GVars.stwCycle ++ ; 1622 } 1623 1624 // Monitor cleanup on JavaThread::exit 1625 1626 // Iterate through monitor cache and attempt to release thread's monitors 1627 // Gives up on a particular monitor if an exception occurs, but continues 1628 // the overall iteration, swallowing the exception. 1629 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1630 private: 1631 TRAPS; 1632 1633 public: 1634 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1635 void do_monitor(ObjectMonitor* mid) { 1636 if (mid->owner() == THREAD) { 1637 (void)mid->complete_exit(CHECK); 1638 } 1639 } 1640 }; 1641 1642 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1643 // ignored. This is meant to be called during JNI thread detach which assumes 1644 // all remaining monitors are heavyweight. All exceptions are swallowed. 1645 // Scanning the extant monitor list can be time consuming. 1646 // A simple optimization is to add a per-thread flag that indicates a thread 1647 // called jni_monitorenter() during its lifetime. 1648 // 1649 // Instead of No_Savepoint_Verifier it might be cheaper to 1650 // use an idiom of the form: 1651 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1652 // <code that must not run at safepoint> 1653 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1654 // Since the tests are extremely cheap we could leave them enabled 1655 // for normal product builds. 1656 1657 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1658 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1659 No_Safepoint_Verifier nsv ; 1660 ReleaseJavaMonitorsClosure rjmc(THREAD); 1661 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); 1662 ObjectSynchronizer::monitors_iterate(&rjmc); 1663 Thread::muxRelease(&ListLock); 1664 THREAD->clear_pending_exception(); 1665 } 1666 1667 //------------------------------------------------------------------------------ 1668 // Debugging code 1669 1670 void ObjectSynchronizer::sanity_checks(const bool verbose, 1671 const uint cache_line_size, 1672 int *error_cnt_ptr, 1673 int *warning_cnt_ptr) { 1674 u_char *addr_begin = (u_char*)&GVars; 1675 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1676 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1677 1678 if (verbose) { 1679 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1680 sizeof(SharedGlobals)); 1681 } 1682 1683 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1684 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1685 1686 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1687 if (verbose) { 1688 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1689 } 1690 1691 if (cache_line_size != 0) { 1692 // We were able to determine the L1 data cache line size so 1693 // do some cache line specific sanity checks 1694 1695 if (offset_stwRandom < cache_line_size) { 1696 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1697 "to the struct beginning than a cache line which permits " 1698 "false sharing."); 1699 (*warning_cnt_ptr)++; 1700 } 1701 1702 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1703 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1704 "SharedGlobals.hcSequence fields are closer than a cache " 1705 "line which permits false sharing."); 1706 (*warning_cnt_ptr)++; 1707 } 1708 1709 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1710 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1711 "to the struct end than a cache line which permits false " 1712 "sharing."); 1713 (*warning_cnt_ptr)++; 1714 } 1715 } 1716 } 1717 1718 #ifndef PRODUCT 1719 1720 // Verify all monitors in the monitor cache, the verification is weak. 1721 void ObjectSynchronizer::verify() { 1722 ObjectMonitor* block = 1723 (ObjectMonitor *)OrderAccess::load_ptr_acquire(&gBlockList); 1724 while (block != NULL) { 1725 assert(block->object() == CHAINMARKER, "must be a block header"); 1726 for (int i = 1; i < _BLOCKSIZE; i++) { 1727 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 1728 oop object = (oop)mid->object(); 1729 if (object != NULL) { 1730 mid->verify(); 1731 } 1732 } 1733 block = (ObjectMonitor*) block->FreeNext; 1734 } 1735 } 1736 1737 // Check if monitor belongs to the monitor cache 1738 // The list is grow-only so it's *relatively* safe to traverse 1739 // the list of extant blocks without taking a lock. 1740 1741 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1742 ObjectMonitor* block = 1743 (ObjectMonitor*)OrderAccess::load_ptr_acquire(&gBlockList); 1744 while (block != NULL) { 1745 assert(block->object() == CHAINMARKER, "must be a block header"); 1746 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 1747 address mon = (address)monitor; 1748 address blk = (address)block; 1749 size_t diff = mon - blk; 1750 assert((diff % sizeof(ObjectMonitor)) == 0, "must be aligned"); 1751 return 1; 1752 } 1753 block = (ObjectMonitor*)block->FreeNext; 1754 } 1755 return 0; 1756 } 1757 1758 #endif