1 /* 2 * Copyright (c) 1998, 2016, 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 "logging/log.hpp" 28 #include "memory/metaspaceShared.hpp" 29 #include "memory/padded.hpp" 30 #include "memory/resourceArea.hpp" 31 #include "oops/markOop.hpp" 32 #include "oops/oop.inline.hpp" 33 #include "runtime/atomic.inline.hpp" 34 #include "runtime/biasedLocking.hpp" 35 #include "runtime/handles.inline.hpp" 36 #include "runtime/interfaceSupport.hpp" 37 #include "runtime/mutexLocker.hpp" 38 #include "runtime/objectMonitor.hpp" 39 #include "runtime/objectMonitor.inline.hpp" 40 #include "runtime/osThread.hpp" 41 #include "runtime/stubRoutines.hpp" 42 #include "runtime/synchronizer.hpp" 43 #include "runtime/thread.inline.hpp" 44 #include "runtime/vframe.hpp" 45 #include "trace/traceMacros.hpp" 46 #include "trace/tracing.hpp" 47 #include "utilities/dtrace.hpp" 48 #include "utilities/events.hpp" 49 #include "utilities/preserveException.hpp" 50 51 // The "core" versions of monitor enter and exit reside in this file. 52 // The interpreter and compilers contain specialized transliterated 53 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 54 // for instance. If you make changes here, make sure to modify the 55 // interpreter, and both C1 and C2 fast-path inline locking code emission. 56 // 57 // ----------------------------------------------------------------------------- 58 59 #ifdef DTRACE_ENABLED 60 61 // Only bother with this argument setup if dtrace is available 62 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 63 64 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 65 char* bytes = NULL; \ 66 int len = 0; \ 67 jlong jtid = SharedRuntime::get_java_tid(thread); \ 68 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 69 if (klassname != NULL) { \ 70 bytes = (char*)klassname->bytes(); \ 71 len = klassname->utf8_length(); \ 72 } 73 74 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 75 { \ 76 if (DTraceMonitorProbes) { \ 77 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 78 HOTSPOT_MONITOR_WAIT(jtid, \ 79 (uintptr_t)(monitor), bytes, len, (millis)); \ 80 } \ 81 } 82 83 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 84 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 86 87 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 88 { \ 89 if (DTraceMonitorProbes) { \ 90 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 91 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 92 (uintptr_t)(monitor), bytes, len); \ 93 } \ 94 } 95 96 #else // ndef DTRACE_ENABLED 97 98 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 99 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 100 101 #endif // ndef DTRACE_ENABLED 102 103 // This exists only as a workaround of dtrace bug 6254741 104 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 105 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 106 return 0; 107 } 108 109 #define NINFLATIONLOCKS 256 110 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS]; 111 112 // global list of blocks of monitors 113 // gBlockList is really PaddedEnd<ObjectMonitor> *, but we don't 114 // want to expose the PaddedEnd template more than necessary. 115 ObjectMonitor * volatile ObjectSynchronizer::gBlockList = NULL; 116 // global monitor free list 117 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 118 // global monitor in-use list, for moribund threads, 119 // monitors they inflated need to be scanned for deflation 120 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 121 // count of entries in gOmInUseList 122 int ObjectSynchronizer::gOmInUseCount = 0; 123 124 static volatile intptr_t gListLock = 0; // protects global monitor lists 125 static volatile int gMonitorFreeCount = 0; // # on gFreeList 126 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation 127 128 static void post_monitor_inflate_event(EventJavaMonitorInflate&, 129 const oop, 130 const ObjectSynchronizer::InflateCause); 131 132 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 133 134 135 // =====================> Quick functions 136 137 // The quick_* forms are special fast-path variants used to improve 138 // performance. In the simplest case, a "quick_*" implementation could 139 // simply return false, in which case the caller will perform the necessary 140 // state transitions and call the slow-path form. 141 // The fast-path is designed to handle frequently arising cases in an efficient 142 // manner and is just a degenerate "optimistic" variant of the slow-path. 143 // returns true -- to indicate the call was satisfied. 144 // returns false -- to indicate the call needs the services of the slow-path. 145 // A no-loitering ordinance is in effect for code in the quick_* family 146 // operators: safepoints or indefinite blocking (blocking that might span a 147 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 148 // entry. 149 // 150 // Consider: An interesting optimization is to have the JIT recognize the 151 // following common idiom: 152 // synchronized (someobj) { .... ; notify(); } 153 // That is, we find a notify() or notifyAll() call that immediately precedes 154 // the monitorexit operation. In that case the JIT could fuse the operations 155 // into a single notifyAndExit() runtime primitive. 156 157 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) { 158 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 159 assert(self->is_Java_thread(), "invariant"); 160 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 161 NoSafepointVerifier nsv; 162 if (obj == NULL) return false; // slow-path for invalid obj 163 const markOop mark = obj->mark(); 164 165 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) { 166 // Degenerate notify 167 // stack-locked by caller so by definition the implied waitset is empty. 168 return true; 169 } 170 171 if (mark->has_monitor()) { 172 ObjectMonitor * const mon = mark->monitor(); 173 assert(mon->object() == obj, "invariant"); 174 if (mon->owner() != self) return false; // slow-path for IMS exception 175 176 if (mon->first_waiter() != NULL) { 177 // We have one or more waiters. Since this is an inflated monitor 178 // that we own, we can transfer one or more threads from the waitset 179 // to the entrylist here and now, avoiding the slow-path. 180 if (all) { 181 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 182 } else { 183 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 184 } 185 int tally = 0; 186 do { 187 mon->INotify(self); 188 ++tally; 189 } while (mon->first_waiter() != NULL && all); 190 OM_PERFDATA_OP(Notifications, inc(tally)); 191 } 192 return true; 193 } 194 195 // biased locking and any other IMS exception states take the slow-path 196 return false; 197 } 198 199 200 // The LockNode emitted directly at the synchronization site would have 201 // been too big if it were to have included support for the cases of inflated 202 // recursive enter and exit, so they go here instead. 203 // Note that we can't safely call AsyncPrintJavaStack() from within 204 // quick_enter() as our thread state remains _in_Java. 205 206 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self, 207 BasicLock * Lock) { 208 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 209 assert(Self->is_Java_thread(), "invariant"); 210 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant"); 211 NoSafepointVerifier nsv; 212 if (obj == NULL) return false; // Need to throw NPE 213 const markOop mark = obj->mark(); 214 215 if (mark->has_monitor()) { 216 ObjectMonitor * const m = mark->monitor(); 217 assert(m->object() == obj, "invariant"); 218 Thread * const owner = (Thread *) m->_owner; 219 220 // Lock contention and Transactional Lock Elision (TLE) diagnostics 221 // and observability 222 // Case: light contention possibly amenable to TLE 223 // Case: TLE inimical operations such as nested/recursive synchronization 224 225 if (owner == Self) { 226 m->_recursions++; 227 return true; 228 } 229 230 if (owner == NULL && 231 Atomic::cmpxchg_ptr(Self, &(m->_owner), NULL) == NULL) { 232 assert(m->_recursions == 0, "invariant"); 233 assert(m->_owner == Self, "invariant"); 234 return true; 235 } 236 } 237 238 // Note that we could inflate in quick_enter. 239 // This is likely a useful optimization 240 // Critically, in quick_enter() we must not: 241 // -- perform bias revocation, or 242 // -- block indefinitely, or 243 // -- reach a safepoint 244 245 return false; // revert to slow-path 246 } 247 248 // ----------------------------------------------------------------------------- 249 // Fast Monitor Enter/Exit 250 // This the fast monitor enter. The interpreter and compiler use 251 // some assembly copies of this code. Make sure update those code 252 // if the following function is changed. The implementation is 253 // extremely sensitive to race condition. Be careful. 254 255 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 256 bool attempt_rebias, TRAPS) { 257 if (UseBiasedLocking) { 258 if (!SafepointSynchronize::is_at_safepoint()) { 259 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 260 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 261 return; 262 } 263 } else { 264 assert(!attempt_rebias, "can not rebias toward VM thread"); 265 BiasedLocking::revoke_at_safepoint(obj); 266 } 267 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 268 } 269 270 slow_enter(obj, lock, THREAD); 271 } 272 273 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 274 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); 275 // if displaced header is null, the previous enter is recursive enter, no-op 276 markOop dhw = lock->displaced_header(); 277 markOop mark; 278 if (dhw == NULL) { 279 // Recursive stack-lock. 280 // Diagnostics -- Could be: stack-locked, inflating, inflated. 281 mark = object->mark(); 282 assert(!mark->is_neutral(), "invariant"); 283 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { 284 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant"); 285 } 286 if (mark->has_monitor()) { 287 ObjectMonitor * m = mark->monitor(); 288 assert(((oop)(m->object()))->mark() == mark, "invariant"); 289 assert(m->is_entered(THREAD), "invariant"); 290 } 291 return; 292 } 293 294 mark = object->mark(); 295 296 // If the object is stack-locked by the current thread, try to 297 // swing the displaced header from the box back to the mark. 298 if (mark == (markOop) lock) { 299 assert(dhw->is_neutral(), "invariant"); 300 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { 301 TEVENT(fast_exit: release stacklock); 302 return; 303 } 304 } 305 306 ObjectSynchronizer::inflate(THREAD, 307 object, 308 inflate_cause_vm_internal)->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, 343 obj(), 344 inflate_cause_monitor_enter)->enter(THREAD); 345 } 346 347 // This routine is used to handle interpreter/compiler slow case 348 // We don't need to use fast path here, because it must have 349 // failed in the interpreter/compiler code. Simply use the heavy 350 // weight monitor should be ok, unless someone find otherwise. 351 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 352 fast_exit(object, lock, THREAD); 353 } 354 355 // ----------------------------------------------------------------------------- 356 // Class Loader support to workaround deadlocks on the class loader lock objects 357 // Also used by GC 358 // complete_exit()/reenter() are used to wait on a nested lock 359 // i.e. to give up an outer lock completely and then re-enter 360 // Used when holding nested locks - lock acquisition order: lock1 then lock2 361 // 1) complete_exit lock1 - saving recursion count 362 // 2) wait on lock2 363 // 3) when notified on lock2, unlock lock2 364 // 4) reenter lock1 with original recursion count 365 // 5) lock lock2 366 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 367 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 368 TEVENT(complete_exit); 369 if (UseBiasedLocking) { 370 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 371 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 372 } 373 374 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 375 obj(), 376 inflate_cause_vm_internal); 377 378 return monitor->complete_exit(THREAD); 379 } 380 381 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 382 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 383 TEVENT(reenter); 384 if (UseBiasedLocking) { 385 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 386 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 387 } 388 389 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 390 obj(), 391 inflate_cause_vm_internal); 392 393 monitor->reenter(recursion, THREAD); 394 } 395 // ----------------------------------------------------------------------------- 396 // JNI locks on java objects 397 // NOTE: must use heavy weight monitor to handle jni monitor enter 398 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 399 // the current locking is from JNI instead of Java code 400 TEVENT(jni_enter); 401 if (UseBiasedLocking) { 402 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 403 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 404 } 405 THREAD->set_current_pending_monitor_is_from_java(false); 406 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD); 407 THREAD->set_current_pending_monitor_is_from_java(true); 408 } 409 410 // NOTE: must use heavy weight monitor to handle jni monitor exit 411 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 412 TEVENT(jni_exit); 413 if (UseBiasedLocking) { 414 Handle h_obj(THREAD, obj); 415 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 416 obj = h_obj(); 417 } 418 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 419 420 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 421 obj, 422 inflate_cause_jni_exit); 423 // If this thread has locked the object, exit the monitor. Note: can't use 424 // monitor->check(CHECK); must exit even if an exception is pending. 425 if (monitor->check(THREAD)) { 426 monitor->exit(true, THREAD); 427 } 428 } 429 430 // ----------------------------------------------------------------------------- 431 // Internal VM locks on java objects 432 // standard constructor, allows locking failures 433 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 434 _dolock = doLock; 435 _thread = thread; 436 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 437 _obj = obj; 438 439 if (_dolock) { 440 TEVENT(ObjectLocker); 441 442 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 443 } 444 } 445 446 ObjectLocker::~ObjectLocker() { 447 if (_dolock) { 448 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 449 } 450 } 451 452 453 // ----------------------------------------------------------------------------- 454 // Wait/Notify/NotifyAll 455 // NOTE: must use heavy weight monitor to handle wait() 456 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 457 if (UseBiasedLocking) { 458 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 459 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 460 } 461 if (millis < 0) { 462 TEVENT(wait - throw IAX); 463 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 464 } 465 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 466 obj(), 467 inflate_cause_wait); 468 469 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 470 monitor->wait(millis, true, THREAD); 471 472 // This dummy call is in place to get around dtrace bug 6254741. Once 473 // that's fixed we can uncomment the following line, remove the call 474 // and change this function back into a "void" func. 475 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 476 return dtrace_waited_probe(monitor, obj, THREAD); 477 } 478 479 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 480 if (UseBiasedLocking) { 481 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 482 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 483 } 484 if (millis < 0) { 485 TEVENT(wait - throw IAX); 486 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 487 } 488 ObjectSynchronizer::inflate(THREAD, 489 obj(), 490 inflate_cause_wait)->wait(millis, false, THREAD); 491 } 492 493 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 494 if (UseBiasedLocking) { 495 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 496 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 497 } 498 499 markOop mark = obj->mark(); 500 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 501 return; 502 } 503 ObjectSynchronizer::inflate(THREAD, 504 obj(), 505 inflate_cause_notify)->notify(THREAD); 506 } 507 508 // NOTE: see comment of notify() 509 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 510 if (UseBiasedLocking) { 511 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 512 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 513 } 514 515 markOop mark = obj->mark(); 516 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 517 return; 518 } 519 ObjectSynchronizer::inflate(THREAD, 520 obj(), 521 inflate_cause_notify)->notifyAll(THREAD); 522 } 523 524 // ----------------------------------------------------------------------------- 525 // Hash Code handling 526 // 527 // Performance concern: 528 // OrderAccess::storestore() calls release() which at one time stored 0 529 // into the global volatile OrderAccess::dummy variable. This store was 530 // unnecessary for correctness. Many threads storing into a common location 531 // causes considerable cache migration or "sloshing" on large SMP systems. 532 // As such, I avoided using OrderAccess::storestore(). In some cases 533 // OrderAccess::fence() -- which incurs local latency on the executing 534 // processor -- is a better choice as it scales on SMP systems. 535 // 536 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 537 // a discussion of coherency costs. Note that all our current reference 538 // platforms provide strong ST-ST order, so the issue is moot on IA32, 539 // x64, and SPARC. 540 // 541 // As a general policy we use "volatile" to control compiler-based reordering 542 // and explicit fences (barriers) to control for architectural reordering 543 // performed by the CPU(s) or platform. 544 545 struct SharedGlobals { 546 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE]; 547 // These are highly shared mostly-read variables. 548 // To avoid false-sharing they need to be the sole occupants of a cache line. 549 volatile int stwRandom; 550 volatile int stwCycle; 551 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 552 // Hot RW variable -- Sequester to avoid false-sharing 553 volatile int hcSequence; 554 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int)); 555 }; 556 557 static SharedGlobals GVars; 558 static int MonitorScavengeThreshold = 1000000; 559 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 560 561 static markOop ReadStableMark(oop obj) { 562 markOop mark = obj->mark(); 563 if (!mark->is_being_inflated()) { 564 return mark; // normal fast-path return 565 } 566 567 int its = 0; 568 for (;;) { 569 markOop mark = obj->mark(); 570 if (!mark->is_being_inflated()) { 571 return mark; // normal fast-path return 572 } 573 574 // The object is being inflated by some other thread. 575 // The caller of ReadStableMark() must wait for inflation to complete. 576 // Avoid live-lock 577 // TODO: consider calling SafepointSynchronize::do_call_back() while 578 // spinning to see if there's a safepoint pending. If so, immediately 579 // yielding or blocking would be appropriate. Avoid spinning while 580 // there is a safepoint pending. 581 // TODO: add inflation contention performance counters. 582 // TODO: restrict the aggregate number of spinners. 583 584 ++its; 585 if (its > 10000 || !os::is_MP()) { 586 if (its & 1) { 587 os::naked_yield(); 588 TEVENT(Inflate: INFLATING - yield); 589 } else { 590 // Note that the following code attenuates the livelock problem but is not 591 // a complete remedy. A more complete solution would require that the inflating 592 // thread hold the associated inflation lock. The following code simply restricts 593 // the number of spinners to at most one. We'll have N-2 threads blocked 594 // on the inflationlock, 1 thread holding the inflation lock and using 595 // a yield/park strategy, and 1 thread in the midst of inflation. 596 // A more refined approach would be to change the encoding of INFLATING 597 // to allow encapsulation of a native thread pointer. Threads waiting for 598 // inflation to complete would use CAS to push themselves onto a singly linked 599 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 600 // and calling park(). When inflation was complete the thread that accomplished inflation 601 // would detach the list and set the markword to inflated with a single CAS and 602 // then for each thread on the list, set the flag and unpark() the thread. 603 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 604 // wakes at most one thread whereas we need to wake the entire list. 605 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 606 int YieldThenBlock = 0; 607 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 608 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 609 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 610 while (obj->mark() == markOopDesc::INFLATING()) { 611 // Beware: NakedYield() is advisory and has almost no effect on some platforms 612 // so we periodically call Self->_ParkEvent->park(1). 613 // We use a mixed spin/yield/block mechanism. 614 if ((YieldThenBlock++) >= 16) { 615 Thread::current()->_ParkEvent->park(1); 616 } else { 617 os::naked_yield(); 618 } 619 } 620 Thread::muxRelease(gInflationLocks + ix); 621 TEVENT(Inflate: INFLATING - yield/park); 622 } 623 } else { 624 SpinPause(); // SMP-polite spinning 625 } 626 } 627 } 628 629 // hashCode() generation : 630 // 631 // Possibilities: 632 // * MD5Digest of {obj,stwRandom} 633 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 634 // * A DES- or AES-style SBox[] mechanism 635 // * One of the Phi-based schemes, such as: 636 // 2654435761 = 2^32 * Phi (golden ratio) 637 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 638 // * A variation of Marsaglia's shift-xor RNG scheme. 639 // * (obj ^ stwRandom) is appealing, but can result 640 // in undesirable regularity in the hashCode values of adjacent objects 641 // (objects allocated back-to-back, in particular). This could potentially 642 // result in hashtable collisions and reduced hashtable efficiency. 643 // There are simple ways to "diffuse" the middle address bits over the 644 // generated hashCode values: 645 646 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 647 intptr_t value = 0; 648 if (hashCode == 0) { 649 // This form uses an unguarded global Park-Miller RNG, 650 // so it's possible for two threads to race and generate the same RNG. 651 // On MP system we'll have lots of RW access to a global, so the 652 // mechanism induces lots of coherency traffic. 653 value = os::random(); 654 } else if (hashCode == 1) { 655 // This variation has the property of being stable (idempotent) 656 // between STW operations. This can be useful in some of the 1-0 657 // synchronization schemes. 658 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 659 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 660 } else if (hashCode == 2) { 661 value = 1; // for sensitivity testing 662 } else if (hashCode == 3) { 663 value = ++GVars.hcSequence; 664 } else if (hashCode == 4) { 665 value = cast_from_oop<intptr_t>(obj); 666 } else { 667 // Marsaglia's xor-shift scheme with thread-specific state 668 // This is probably the best overall implementation -- we'll 669 // likely make this the default in future releases. 670 unsigned t = Self->_hashStateX; 671 t ^= (t << 11); 672 Self->_hashStateX = Self->_hashStateY; 673 Self->_hashStateY = Self->_hashStateZ; 674 Self->_hashStateZ = Self->_hashStateW; 675 unsigned v = Self->_hashStateW; 676 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 677 Self->_hashStateW = v; 678 value = v; 679 } 680 681 value &= markOopDesc::hash_mask; 682 if (value == 0) value = 0xBAD; 683 assert(value != markOopDesc::no_hash, "invariant"); 684 TEVENT(hashCode: GENERATE); 685 return value; 686 } 687 688 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 689 if (UseBiasedLocking) { 690 // NOTE: many places throughout the JVM do not expect a safepoint 691 // to be taken here, in particular most operations on perm gen 692 // objects. However, we only ever bias Java instances and all of 693 // the call sites of identity_hash that might revoke biases have 694 // been checked to make sure they can handle a safepoint. The 695 // added check of the bias pattern is to avoid useless calls to 696 // thread-local storage. 697 if (obj->mark()->has_bias_pattern()) { 698 // Handle for oop obj in case of STW safepoint 699 Handle hobj(Self, obj); 700 // Relaxing assertion for bug 6320749. 701 assert(Universe::verify_in_progress() || 702 !SafepointSynchronize::is_at_safepoint(), 703 "biases should not be seen by VM thread here"); 704 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 705 obj = hobj(); 706 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 707 } 708 } 709 710 // hashCode() is a heap mutator ... 711 // Relaxing assertion for bug 6320749. 712 assert(Universe::verify_in_progress() || DumpSharedSpaces || 713 !SafepointSynchronize::is_at_safepoint(), "invariant"); 714 assert(Universe::verify_in_progress() || DumpSharedSpaces || 715 Self->is_Java_thread() , "invariant"); 716 assert(Universe::verify_in_progress() || DumpSharedSpaces || 717 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 718 719 ObjectMonitor* monitor = NULL; 720 markOop temp, test; 721 intptr_t hash; 722 markOop mark = ReadStableMark(obj); 723 724 // object should remain ineligible for biased locking 725 assert(!mark->has_bias_pattern(), "invariant"); 726 727 if (mark->is_neutral()) { 728 hash = mark->hash(); // this is a normal header 729 if (hash) { // if it has hash, just return it 730 return hash; 731 } 732 hash = get_next_hash(Self, obj); // allocate a new hash code 733 temp = mark->copy_set_hash(hash); // merge the hash code into header 734 // use (machine word version) atomic operation to install the hash 735 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 736 if (test == mark) { 737 return hash; 738 } 739 // If atomic operation failed, we must inflate the header 740 // into heavy weight monitor. We could add more code here 741 // for fast path, but it does not worth the complexity. 742 } else if (mark->has_monitor()) { 743 monitor = mark->monitor(); 744 temp = monitor->header(); 745 assert(temp->is_neutral(), "invariant"); 746 hash = temp->hash(); 747 if (hash) { 748 return hash; 749 } 750 // Skip to the following code to reduce code size 751 } else if (Self->is_lock_owned((address)mark->locker())) { 752 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 753 assert(temp->is_neutral(), "invariant"); 754 hash = temp->hash(); // by current thread, check if the displaced 755 if (hash) { // header contains hash code 756 return hash; 757 } 758 // WARNING: 759 // The displaced header is strictly immutable. 760 // It can NOT be changed in ANY cases. So we have 761 // to inflate the header into heavyweight monitor 762 // even the current thread owns the lock. The reason 763 // is the BasicLock (stack slot) will be asynchronously 764 // read by other threads during the inflate() function. 765 // Any change to stack may not propagate to other threads 766 // correctly. 767 } 768 769 // Inflate the monitor to set hash code 770 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code); 771 // Load displaced header and check it has hash code 772 mark = monitor->header(); 773 assert(mark->is_neutral(), "invariant"); 774 hash = mark->hash(); 775 if (hash == 0) { 776 hash = get_next_hash(Self, obj); 777 temp = mark->copy_set_hash(hash); // merge hash code into header 778 assert(temp->is_neutral(), "invariant"); 779 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 780 if (test != mark) { 781 // The only update to the header in the monitor (outside GC) 782 // is install the hash code. If someone add new usage of 783 // displaced header, please update this code 784 hash = test->hash(); 785 assert(test->is_neutral(), "invariant"); 786 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 787 } 788 } 789 // We finally get the hash 790 return hash; 791 } 792 793 // Deprecated -- use FastHashCode() instead. 794 795 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 796 return FastHashCode(Thread::current(), obj()); 797 } 798 799 800 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 801 Handle h_obj) { 802 if (UseBiasedLocking) { 803 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 804 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 805 } 806 807 assert(thread == JavaThread::current(), "Can only be called on current thread"); 808 oop obj = h_obj(); 809 810 markOop mark = ReadStableMark(obj); 811 812 // Uncontended case, header points to stack 813 if (mark->has_locker()) { 814 return thread->is_lock_owned((address)mark->locker()); 815 } 816 // Contended case, header points to ObjectMonitor (tagged pointer) 817 if (mark->has_monitor()) { 818 ObjectMonitor* monitor = mark->monitor(); 819 return monitor->is_entered(thread) != 0; 820 } 821 // Unlocked case, header in place 822 assert(mark->is_neutral(), "sanity check"); 823 return false; 824 } 825 826 // Be aware of this method could revoke bias of the lock object. 827 // This method queries the ownership of the lock handle specified by 'h_obj'. 828 // If the current thread owns the lock, it returns owner_self. If no 829 // thread owns the lock, it returns owner_none. Otherwise, it will return 830 // owner_other. 831 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 832 (JavaThread *self, Handle h_obj) { 833 // The caller must beware this method can revoke bias, and 834 // revocation can result in a safepoint. 835 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 836 assert(self->thread_state() != _thread_blocked, "invariant"); 837 838 // Possible mark states: neutral, biased, stack-locked, inflated 839 840 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 841 // CASE: biased 842 BiasedLocking::revoke_and_rebias(h_obj, false, self); 843 assert(!h_obj->mark()->has_bias_pattern(), 844 "biases should be revoked by now"); 845 } 846 847 assert(self == JavaThread::current(), "Can only be called on current thread"); 848 oop obj = h_obj(); 849 markOop mark = ReadStableMark(obj); 850 851 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 852 if (mark->has_locker()) { 853 return self->is_lock_owned((address)mark->locker()) ? 854 owner_self : owner_other; 855 } 856 857 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 858 // The Object:ObjectMonitor relationship is stable as long as we're 859 // not at a safepoint. 860 if (mark->has_monitor()) { 861 void * owner = mark->monitor()->_owner; 862 if (owner == NULL) return owner_none; 863 return (owner == self || 864 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 865 } 866 867 // CASE: neutral 868 assert(mark->is_neutral(), "sanity check"); 869 return owner_none; // it's unlocked 870 } 871 872 // FIXME: jvmti should call this 873 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 874 if (UseBiasedLocking) { 875 if (SafepointSynchronize::is_at_safepoint()) { 876 BiasedLocking::revoke_at_safepoint(h_obj); 877 } else { 878 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 879 } 880 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 881 } 882 883 oop obj = h_obj(); 884 address owner = NULL; 885 886 markOop mark = ReadStableMark(obj); 887 888 // Uncontended case, header points to stack 889 if (mark->has_locker()) { 890 owner = (address) mark->locker(); 891 } 892 893 // Contended case, header points to ObjectMonitor (tagged pointer) 894 if (mark->has_monitor()) { 895 ObjectMonitor* monitor = mark->monitor(); 896 assert(monitor != NULL, "monitor should be non-null"); 897 owner = (address) monitor->owner(); 898 } 899 900 if (owner != NULL) { 901 // owning_thread_from_monitor_owner() may also return NULL here 902 return Threads::owning_thread_from_monitor_owner(owner, doLock); 903 } 904 905 // Unlocked case, header in place 906 // Cannot have assertion since this object may have been 907 // locked by another thread when reaching here. 908 // assert(mark->is_neutral(), "sanity check"); 909 910 return NULL; 911 } 912 913 // Visitors ... 914 915 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 916 PaddedEnd<ObjectMonitor> * block = 917 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 918 while (block != NULL) { 919 assert(block->object() == CHAINMARKER, "must be a block header"); 920 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 921 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 922 oop object = (oop)mid->object(); 923 if (object != NULL) { 924 closure->do_monitor(mid); 925 } 926 } 927 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 928 } 929 } 930 931 // Get the next block in the block list. 932 static inline ObjectMonitor* next(ObjectMonitor* block) { 933 assert(block->object() == CHAINMARKER, "must be a block header"); 934 block = block->FreeNext; 935 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 936 return block; 937 } 938 939 940 void ObjectSynchronizer::oops_do(OopClosure* f) { 941 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 942 PaddedEnd<ObjectMonitor> * block = 943 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 944 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) { 945 assert(block->object() == CHAINMARKER, "must be a block header"); 946 for (int i = 1; i < _BLOCKSIZE; i++) { 947 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 948 if (mid->object() != NULL) { 949 f->do_oop((oop*)mid->object_addr()); 950 } 951 } 952 } 953 } 954 955 956 // ----------------------------------------------------------------------------- 957 // ObjectMonitor Lifecycle 958 // ----------------------- 959 // Inflation unlinks monitors from the global gFreeList and 960 // associates them with objects. Deflation -- which occurs at 961 // STW-time -- disassociates idle monitors from objects. Such 962 // scavenged monitors are returned to the gFreeList. 963 // 964 // The global list is protected by gListLock. All the critical sections 965 // are short and operate in constant-time. 966 // 967 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 968 // 969 // Lifecycle: 970 // -- unassigned and on the global free list 971 // -- unassigned and on a thread's private omFreeList 972 // -- assigned to an object. The object is inflated and the mark refers 973 // to the objectmonitor. 974 975 976 // Constraining monitor pool growth via MonitorBound ... 977 // 978 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 979 // the rate of scavenging is driven primarily by GC. As such, we can find 980 // an inordinate number of monitors in circulation. 981 // To avoid that scenario we can artificially induce a STW safepoint 982 // if the pool appears to be growing past some reasonable bound. 983 // Generally we favor time in space-time tradeoffs, but as there's no 984 // natural back-pressure on the # of extant monitors we need to impose some 985 // type of limit. Beware that if MonitorBound is set to too low a value 986 // we could just loop. In addition, if MonitorBound is set to a low value 987 // we'll incur more safepoints, which are harmful to performance. 988 // See also: GuaranteedSafepointInterval 989 // 990 // The current implementation uses asynchronous VM operations. 991 992 static void InduceScavenge(Thread * Self, const char * Whence) { 993 // Induce STW safepoint to trim monitors 994 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 995 // More precisely, trigger an asynchronous STW safepoint as the number 996 // of active monitors passes the specified threshold. 997 // TODO: assert thread state is reasonable 998 999 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 1000 if (ObjectMonitor::Knob_Verbose) { 1001 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)", 1002 Whence, ForceMonitorScavenge) ; 1003 tty->flush(); 1004 } 1005 // Induce a 'null' safepoint to scavenge monitors 1006 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 1007 // to the VMthread and have a lifespan longer than that of this activation record. 1008 // The VMThread will delete the op when completed. 1009 VMThread::execute(new VM_ForceAsyncSafepoint()); 1010 1011 if (ObjectMonitor::Knob_Verbose) { 1012 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)", 1013 Whence, ForceMonitorScavenge) ; 1014 tty->flush(); 1015 } 1016 } 1017 } 1018 1019 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1020 ObjectMonitor* mid; 1021 int in_use_tally = 0; 1022 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1023 in_use_tally++; 1024 } 1025 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1026 1027 int free_tally = 0; 1028 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1029 free_tally++; 1030 } 1031 assert(free_tally == Self->omFreeCount, "free count off"); 1032 } 1033 1034 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) { 1035 // A large MAXPRIVATE value reduces both list lock contention 1036 // and list coherency traffic, but also tends to increase the 1037 // number of objectMonitors in circulation as well as the STW 1038 // scavenge costs. As usual, we lean toward time in space-time 1039 // tradeoffs. 1040 const int MAXPRIVATE = 1024; 1041 for (;;) { 1042 ObjectMonitor * m; 1043 1044 // 1: try to allocate from the thread's local omFreeList. 1045 // Threads will attempt to allocate first from their local list, then 1046 // from the global list, and only after those attempts fail will the thread 1047 // attempt to instantiate new monitors. Thread-local free lists take 1048 // heat off the gListLock and improve allocation latency, as well as reducing 1049 // coherency traffic on the shared global list. 1050 m = Self->omFreeList; 1051 if (m != NULL) { 1052 Self->omFreeList = m->FreeNext; 1053 Self->omFreeCount--; 1054 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1055 guarantee(m->object() == NULL, "invariant"); 1056 if (MonitorInUseLists) { 1057 m->FreeNext = Self->omInUseList; 1058 Self->omInUseList = m; 1059 Self->omInUseCount++; 1060 if (ObjectMonitor::Knob_VerifyInUse) { 1061 verifyInUse(Self); 1062 } 1063 } else { 1064 m->FreeNext = NULL; 1065 } 1066 return m; 1067 } 1068 1069 // 2: try to allocate from the global gFreeList 1070 // CONSIDER: use muxTry() instead of muxAcquire(). 1071 // If the muxTry() fails then drop immediately into case 3. 1072 // If we're using thread-local free lists then try 1073 // to reprovision the caller's free list. 1074 if (gFreeList != NULL) { 1075 // Reprovision the thread's omFreeList. 1076 // Use bulk transfers to reduce the allocation rate and heat 1077 // on various locks. 1078 Thread::muxAcquire(&gListLock, "omAlloc"); 1079 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1080 gMonitorFreeCount--; 1081 ObjectMonitor * take = gFreeList; 1082 gFreeList = take->FreeNext; 1083 guarantee(take->object() == NULL, "invariant"); 1084 guarantee(!take->is_busy(), "invariant"); 1085 take->Recycle(); 1086 omRelease(Self, take, false); 1087 } 1088 Thread::muxRelease(&gListLock); 1089 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1090 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1091 TEVENT(omFirst - reprovision); 1092 1093 const int mx = MonitorBound; 1094 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1095 // We can't safely induce a STW safepoint from omAlloc() as our thread 1096 // state may not be appropriate for such activities and callers may hold 1097 // naked oops, so instead we defer the action. 1098 InduceScavenge(Self, "omAlloc"); 1099 } 1100 continue; 1101 } 1102 1103 // 3: allocate a block of new ObjectMonitors 1104 // Both the local and global free lists are empty -- resort to malloc(). 1105 // In the current implementation objectMonitors are TSM - immortal. 1106 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1107 // each ObjectMonitor to start at the beginning of a cache line, 1108 // so we use align_size_up(). 1109 // A better solution would be to use C++ placement-new. 1110 // BEWARE: As it stands currently, we don't run the ctors! 1111 assert(_BLOCKSIZE > 1, "invariant"); 1112 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1113 PaddedEnd<ObjectMonitor> * temp; 1114 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1115 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1116 mtInternal); 1117 temp = (PaddedEnd<ObjectMonitor> *) 1118 align_size_up((intptr_t)real_malloc_addr, 1119 DEFAULT_CACHE_LINE_SIZE); 1120 1121 // NOTE: (almost) no way to recover if allocation failed. 1122 // We might be able to induce a STW safepoint and scavenge enough 1123 // objectMonitors to permit progress. 1124 if (temp == NULL) { 1125 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1126 "Allocate ObjectMonitors"); 1127 } 1128 (void)memset((void *) temp, 0, neededsize); 1129 1130 // Format the block. 1131 // initialize the linked list, each monitor points to its next 1132 // forming the single linked free list, the very first monitor 1133 // will points to next block, which forms the block list. 1134 // The trick of using the 1st element in the block as gBlockList 1135 // linkage should be reconsidered. A better implementation would 1136 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1137 1138 for (int i = 1; i < _BLOCKSIZE; i++) { 1139 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1140 } 1141 1142 // terminate the last monitor as the end of list 1143 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1144 1145 // Element [0] is reserved for global list linkage 1146 temp[0].set_object(CHAINMARKER); 1147 1148 // Consider carving out this thread's current request from the 1149 // block in hand. This avoids some lock traffic and redundant 1150 // list activity. 1151 1152 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1153 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1154 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1155 gMonitorPopulation += _BLOCKSIZE-1; 1156 gMonitorFreeCount += _BLOCKSIZE-1; 1157 1158 // Add the new block to the list of extant blocks (gBlockList). 1159 // The very first objectMonitor in a block is reserved and dedicated. 1160 // It serves as blocklist "next" linkage. 1161 temp[0].FreeNext = gBlockList; 1162 // There are lock-free uses of gBlockList so make sure that 1163 // the previous stores happen before we update gBlockList. 1164 OrderAccess::release_store_ptr(&gBlockList, temp); 1165 1166 // Add the new string of objectMonitors to the global free list 1167 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1168 gFreeList = temp + 1; 1169 Thread::muxRelease(&gListLock); 1170 TEVENT(Allocate block of monitors); 1171 } 1172 } 1173 1174 // Place "m" on the caller's private per-thread omFreeList. 1175 // In practice there's no need to clamp or limit the number of 1176 // monitors on a thread's omFreeList as the only time we'll call 1177 // omRelease is to return a monitor to the free list after a CAS 1178 // attempt failed. This doesn't allow unbounded #s of monitors to 1179 // accumulate on a thread's free list. 1180 // 1181 // Key constraint: all ObjectMonitors on a thread's free list and the global 1182 // free list must have their object field set to null. This prevents the 1183 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1184 1185 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1186 bool fromPerThreadAlloc) { 1187 guarantee(m->object() == NULL, "invariant"); 1188 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1189 // Remove from omInUseList 1190 if (MonitorInUseLists && fromPerThreadAlloc) { 1191 ObjectMonitor* cur_mid_in_use = NULL; 1192 bool extracted = false; 1193 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1194 if (m == mid) { 1195 // extract from per-thread in-use list 1196 if (mid == Self->omInUseList) { 1197 Self->omInUseList = mid->FreeNext; 1198 } else if (cur_mid_in_use != NULL) { 1199 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1200 } 1201 extracted = true; 1202 Self->omInUseCount--; 1203 if (ObjectMonitor::Knob_VerifyInUse) { 1204 verifyInUse(Self); 1205 } 1206 break; 1207 } 1208 } 1209 assert(extracted, "Should have extracted from in-use list"); 1210 } 1211 1212 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1213 m->FreeNext = Self->omFreeList; 1214 Self->omFreeList = m; 1215 Self->omFreeCount++; 1216 } 1217 1218 // Return the monitors of a moribund thread's local free list to 1219 // the global free list. Typically a thread calls omFlush() when 1220 // it's dying. We could also consider having the VM thread steal 1221 // monitors from threads that have not run java code over a few 1222 // consecutive STW safepoints. Relatedly, we might decay 1223 // omFreeProvision at STW safepoints. 1224 // 1225 // Also return the monitors of a moribund thread's omInUseList to 1226 // a global gOmInUseList under the global list lock so these 1227 // will continue to be scanned. 1228 // 1229 // We currently call omFlush() from the Thread:: dtor _after the thread 1230 // has been excised from the thread list and is no longer a mutator. 1231 // That means that omFlush() can run concurrently with a safepoint and 1232 // the scavenge operator. Calling omFlush() from JavaThread::exit() might 1233 // be a better choice as we could safely reason that that the JVM is 1234 // not at a safepoint at the time of the call, and thus there could 1235 // be not inopportune interleavings between omFlush() and the scavenge 1236 // operator. 1237 1238 void ObjectSynchronizer::omFlush(Thread * Self) { 1239 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1240 Self->omFreeList = NULL; 1241 ObjectMonitor * tail = NULL; 1242 int tally = 0; 1243 if (list != NULL) { 1244 ObjectMonitor * s; 1245 // The thread is going away, the per-thread free monitors 1246 // are freed via set_owner(NULL) 1247 // Link them to tail, which will be linked into the global free list 1248 // gFreeList below, under the gListLock 1249 for (s = list; s != NULL; s = s->FreeNext) { 1250 tally++; 1251 tail = s; 1252 guarantee(s->object() == NULL, "invariant"); 1253 guarantee(!s->is_busy(), "invariant"); 1254 s->set_owner(NULL); // redundant but good hygiene 1255 TEVENT(omFlush - Move one); 1256 } 1257 guarantee(tail != NULL && list != NULL, "invariant"); 1258 } 1259 1260 ObjectMonitor * inUseList = Self->omInUseList; 1261 ObjectMonitor * inUseTail = NULL; 1262 int inUseTally = 0; 1263 if (inUseList != NULL) { 1264 Self->omInUseList = NULL; 1265 ObjectMonitor *cur_om; 1266 // The thread is going away, however the omInUseList inflated 1267 // monitors may still be in-use by other threads. 1268 // Link them to inUseTail, which will be linked into the global in-use list 1269 // gOmInUseList below, under the gListLock 1270 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1271 inUseTail = cur_om; 1272 inUseTally++; 1273 } 1274 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1275 Self->omInUseCount = 0; 1276 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1277 } 1278 1279 Thread::muxAcquire(&gListLock, "omFlush"); 1280 if (tail != NULL) { 1281 tail->FreeNext = gFreeList; 1282 gFreeList = list; 1283 gMonitorFreeCount += tally; 1284 } 1285 1286 if (inUseTail != NULL) { 1287 inUseTail->FreeNext = gOmInUseList; 1288 gOmInUseList = inUseList; 1289 gOmInUseCount += inUseTally; 1290 } 1291 1292 Thread::muxRelease(&gListLock); 1293 TEVENT(omFlush); 1294 } 1295 1296 // Fast path code shared by multiple functions 1297 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1298 markOop mark = obj->mark(); 1299 if (mark->has_monitor()) { 1300 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1301 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1302 return mark->monitor(); 1303 } 1304 return ObjectSynchronizer::inflate(Thread::current(), 1305 obj, 1306 inflate_cause_vm_internal); 1307 } 1308 1309 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self, 1310 oop object, 1311 const InflateCause cause) { 1312 1313 // Inflate mutates the heap ... 1314 // Relaxing assertion for bug 6320749. 1315 assert(Universe::verify_in_progress() || 1316 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1317 1318 EventJavaMonitorInflate event; 1319 1320 for (;;) { 1321 const markOop mark = object->mark(); 1322 assert(!mark->has_bias_pattern(), "invariant"); 1323 1324 // The mark can be in one of the following states: 1325 // * Inflated - just return 1326 // * Stack-locked - coerce it to inflated 1327 // * INFLATING - busy wait for conversion to complete 1328 // * Neutral - aggressively inflate the object. 1329 // * BIASED - Illegal. We should never see this 1330 1331 // CASE: inflated 1332 if (mark->has_monitor()) { 1333 ObjectMonitor * inf = mark->monitor(); 1334 assert(inf->header()->is_neutral(), "invariant"); 1335 assert(inf->object() == object, "invariant"); 1336 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1337 event.cancel(); // let's not post an inflation event, unless we did the deed ourselves 1338 return inf; 1339 } 1340 1341 // CASE: inflation in progress - inflating over a stack-lock. 1342 // Some other thread is converting from stack-locked to inflated. 1343 // Only that thread can complete inflation -- other threads must wait. 1344 // The INFLATING value is transient. 1345 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1346 // We could always eliminate polling by parking the thread on some auxiliary list. 1347 if (mark == markOopDesc::INFLATING()) { 1348 TEVENT(Inflate: spin while INFLATING); 1349 ReadStableMark(object); 1350 continue; 1351 } 1352 1353 // CASE: stack-locked 1354 // Could be stack-locked either by this thread or by some other thread. 1355 // 1356 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1357 // to install INFLATING into the mark word. We originally installed INFLATING, 1358 // allocated the objectmonitor, and then finally STed the address of the 1359 // objectmonitor into the mark. This was correct, but artificially lengthened 1360 // the interval in which INFLATED appeared in the mark, thus increasing 1361 // the odds of inflation contention. 1362 // 1363 // We now use per-thread private objectmonitor free lists. 1364 // These list are reprovisioned from the global free list outside the 1365 // critical INFLATING...ST interval. A thread can transfer 1366 // multiple objectmonitors en-mass from the global free list to its local free list. 1367 // This reduces coherency traffic and lock contention on the global free list. 1368 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1369 // before or after the CAS(INFLATING) operation. 1370 // See the comments in omAlloc(). 1371 1372 if (mark->has_locker()) { 1373 ObjectMonitor * m = omAlloc(Self); 1374 // Optimistically prepare the objectmonitor - anticipate successful CAS 1375 // We do this before the CAS in order to minimize the length of time 1376 // in which INFLATING appears in the mark. 1377 m->Recycle(); 1378 m->_Responsible = NULL; 1379 m->_recursions = 0; 1380 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1381 1382 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark); 1383 if (cmp != mark) { 1384 omRelease(Self, m, true); 1385 continue; // Interference -- just retry 1386 } 1387 1388 // We've successfully installed INFLATING (0) into the mark-word. 1389 // This is the only case where 0 will appear in a mark-word. 1390 // Only the singular thread that successfully swings the mark-word 1391 // to 0 can perform (or more precisely, complete) inflation. 1392 // 1393 // Why do we CAS a 0 into the mark-word instead of just CASing the 1394 // mark-word from the stack-locked value directly to the new inflated state? 1395 // Consider what happens when a thread unlocks a stack-locked object. 1396 // It attempts to use CAS to swing the displaced header value from the 1397 // on-stack basiclock back into the object header. Recall also that the 1398 // header value (hashcode, etc) can reside in (a) the object header, or 1399 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1400 // header in an objectMonitor. The inflate() routine must copy the header 1401 // value from the basiclock on the owner's stack to the objectMonitor, all 1402 // the while preserving the hashCode stability invariants. If the owner 1403 // decides to release the lock while the value is 0, the unlock will fail 1404 // and control will eventually pass from slow_exit() to inflate. The owner 1405 // will then spin, waiting for the 0 value to disappear. Put another way, 1406 // the 0 causes the owner to stall if the owner happens to try to 1407 // drop the lock (restoring the header from the basiclock to the object) 1408 // while inflation is in-progress. This protocol avoids races that might 1409 // would otherwise permit hashCode values to change or "flicker" for an object. 1410 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1411 // 0 serves as a "BUSY" inflate-in-progress indicator. 1412 1413 1414 // fetch the displaced mark from the owner's stack. 1415 // The owner can't die or unwind past the lock while our INFLATING 1416 // object is in the mark. Furthermore the owner can't complete 1417 // an unlock on the object, either. 1418 markOop dmw = mark->displaced_mark_helper(); 1419 assert(dmw->is_neutral(), "invariant"); 1420 1421 // Setup monitor fields to proper values -- prepare the monitor 1422 m->set_header(dmw); 1423 1424 // Optimization: if the mark->locker stack address is associated 1425 // with this thread we could simply set m->_owner = Self. 1426 // Note that a thread can inflate an object 1427 // that it has stack-locked -- as might happen in wait() -- directly 1428 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1429 m->set_owner(mark->locker()); 1430 m->set_object(object); 1431 // TODO-FIXME: assert BasicLock->dhw != 0. 1432 1433 // Must preserve store ordering. The monitor state must 1434 // be stable at the time of publishing the monitor address. 1435 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1436 object->release_set_mark(markOopDesc::encode(m)); 1437 1438 // Hopefully the performance counters are allocated on distinct cache lines 1439 // to avoid false sharing on MP systems ... 1440 OM_PERFDATA_OP(Inflations, inc()); 1441 TEVENT(Inflate: overwrite stacklock); 1442 if (log_is_enabled(Debug, monitorinflation)) { 1443 if (object->is_instance()) { 1444 ResourceMark rm; 1445 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1446 p2i(object), p2i(object->mark()), 1447 object->klass()->external_name()); 1448 } 1449 } 1450 if (event.should_commit()) { 1451 post_monitor_inflate_event(event, object, cause); 1452 } 1453 return m; 1454 } 1455 1456 // CASE: neutral 1457 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1458 // If we know we're inflating for entry it's better to inflate by swinging a 1459 // pre-locked objectMonitor pointer into the object header. A successful 1460 // CAS inflates the object *and* confers ownership to the inflating thread. 1461 // In the current implementation we use a 2-step mechanism where we CAS() 1462 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1463 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1464 // would be useful. 1465 1466 assert(mark->is_neutral(), "invariant"); 1467 ObjectMonitor * m = omAlloc(Self); 1468 // prepare m for installation - set monitor to initial state 1469 m->Recycle(); 1470 m->set_header(mark); 1471 m->set_owner(NULL); 1472 m->set_object(object); 1473 m->_recursions = 0; 1474 m->_Responsible = NULL; 1475 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1476 1477 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1478 m->set_object(NULL); 1479 m->set_owner(NULL); 1480 m->Recycle(); 1481 omRelease(Self, m, true); 1482 m = NULL; 1483 continue; 1484 // interference - the markword changed - just retry. 1485 // The state-transitions are one-way, so there's no chance of 1486 // live-lock -- "Inflated" is an absorbing state. 1487 } 1488 1489 // Hopefully the performance counters are allocated on distinct 1490 // cache lines to avoid false sharing on MP systems ... 1491 OM_PERFDATA_OP(Inflations, inc()); 1492 TEVENT(Inflate: overwrite neutral); 1493 if (log_is_enabled(Debug, monitorinflation)) { 1494 if (object->is_instance()) { 1495 ResourceMark rm; 1496 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1497 p2i(object), p2i(object->mark()), 1498 object->klass()->external_name()); 1499 } 1500 } 1501 if (event.should_commit()) { 1502 post_monitor_inflate_event(event, object, cause); 1503 } 1504 return m; 1505 } 1506 } 1507 1508 1509 // Deflate_idle_monitors() is called at all safepoints, immediately 1510 // after all mutators are stopped, but before any objects have moved. 1511 // It traverses the list of known monitors, deflating where possible. 1512 // The scavenged monitor are returned to the monitor free list. 1513 // 1514 // Beware that we scavenge at *every* stop-the-world point. 1515 // Having a large number of monitors in-circulation negatively 1516 // impacts the performance of some applications (e.g., PointBase). 1517 // Broadly, we want to minimize the # of monitors in circulation. 1518 // 1519 // We have added a flag, MonitorInUseLists, which creates a list 1520 // of active monitors for each thread. deflate_idle_monitors() 1521 // only scans the per-thread in-use lists. omAlloc() puts all 1522 // assigned monitors on the per-thread list. deflate_idle_monitors() 1523 // returns the non-busy monitors to the global free list. 1524 // When a thread dies, omFlush() adds the list of active monitors for 1525 // that thread to a global gOmInUseList acquiring the 1526 // global list lock. deflate_idle_monitors() acquires the global 1527 // list lock to scan for non-busy monitors to the global free list. 1528 // An alternative could have used a single global in-use list. The 1529 // downside would have been the additional cost of acquiring the global list lock 1530 // for every omAlloc(). 1531 // 1532 // Perversely, the heap size -- and thus the STW safepoint rate -- 1533 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1534 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1535 // This is an unfortunate aspect of this design. 1536 1537 enum ManifestConstants { 1538 ClearResponsibleAtSTW = 0 1539 }; 1540 1541 // Deflate a single monitor if not in-use 1542 // Return true if deflated, false if in-use 1543 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1544 ObjectMonitor** freeHeadp, 1545 ObjectMonitor** freeTailp) { 1546 bool deflated; 1547 // Normal case ... The monitor is associated with obj. 1548 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1549 guarantee(mid == obj->mark()->monitor(), "invariant"); 1550 guarantee(mid->header()->is_neutral(), "invariant"); 1551 1552 if (mid->is_busy()) { 1553 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1554 deflated = false; 1555 } else { 1556 // Deflate the monitor if it is no longer being used 1557 // It's idle - scavenge and return to the global free list 1558 // plain old deflation ... 1559 TEVENT(deflate_idle_monitors - scavenge1); 1560 if (log_is_enabled(Debug, monitorinflation)) { 1561 if (obj->is_instance()) { 1562 ResourceMark rm; 1563 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , " 1564 "mark " INTPTR_FORMAT " , type %s", 1565 p2i(obj), p2i(obj->mark()), 1566 obj->klass()->external_name()); 1567 } 1568 } 1569 1570 // Restore the header back to obj 1571 obj->release_set_mark(mid->header()); 1572 mid->clear(); 1573 1574 assert(mid->object() == NULL, "invariant"); 1575 1576 // Move the object to the working free list defined by freeHeadp, freeTailp 1577 if (*freeHeadp == NULL) *freeHeadp = mid; 1578 if (*freeTailp != NULL) { 1579 ObjectMonitor * prevtail = *freeTailp; 1580 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1581 prevtail->FreeNext = mid; 1582 } 1583 *freeTailp = mid; 1584 deflated = true; 1585 } 1586 return deflated; 1587 } 1588 1589 // Walk a given monitor list, and deflate idle monitors 1590 // The given list could be a per-thread list or a global list 1591 // Caller acquires gListLock 1592 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1593 ObjectMonitor** freeHeadp, 1594 ObjectMonitor** freeTailp) { 1595 ObjectMonitor* mid; 1596 ObjectMonitor* next; 1597 ObjectMonitor* cur_mid_in_use = NULL; 1598 int deflated_count = 0; 1599 1600 for (mid = *listHeadp; mid != NULL;) { 1601 oop obj = (oop) mid->object(); 1602 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1603 // if deflate_monitor succeeded, 1604 // extract from per-thread in-use list 1605 if (mid == *listHeadp) { 1606 *listHeadp = mid->FreeNext; 1607 } else if (cur_mid_in_use != NULL) { 1608 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1609 } 1610 next = mid->FreeNext; 1611 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1612 mid = next; 1613 deflated_count++; 1614 } else { 1615 cur_mid_in_use = mid; 1616 mid = mid->FreeNext; 1617 } 1618 } 1619 return deflated_count; 1620 } 1621 1622 void ObjectSynchronizer::deflate_idle_monitors() { 1623 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1624 int nInuse = 0; // currently associated with objects 1625 int nInCirculation = 0; // extant 1626 int nScavenged = 0; // reclaimed 1627 bool deflated = false; 1628 1629 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1630 ObjectMonitor * freeTailp = NULL; 1631 1632 TEVENT(deflate_idle_monitors); 1633 // Prevent omFlush from changing mids in Thread dtor's during deflation 1634 // And in case the vm thread is acquiring a lock during a safepoint 1635 // See e.g. 6320749 1636 Thread::muxAcquire(&gListLock, "scavenge - return"); 1637 1638 if (MonitorInUseLists) { 1639 int inUse = 0; 1640 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { 1641 nInCirculation+= cur->omInUseCount; 1642 int deflated_count = deflate_monitor_list(cur->omInUseList_addr(), &freeHeadp, &freeTailp); 1643 cur->omInUseCount-= deflated_count; 1644 if (ObjectMonitor::Knob_VerifyInUse) { 1645 verifyInUse(cur); 1646 } 1647 nScavenged += deflated_count; 1648 nInuse += cur->omInUseCount; 1649 } 1650 1651 // For moribund threads, scan gOmInUseList 1652 if (gOmInUseList) { 1653 nInCirculation += gOmInUseCount; 1654 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1655 gOmInUseCount-= deflated_count; 1656 nScavenged += deflated_count; 1657 nInuse += gOmInUseCount; 1658 } 1659 1660 } else { 1661 PaddedEnd<ObjectMonitor> * block = 1662 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 1663 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) { 1664 // Iterate over all extant monitors - Scavenge all idle monitors. 1665 assert(block->object() == CHAINMARKER, "must be a block header"); 1666 nInCirculation += _BLOCKSIZE; 1667 for (int i = 1; i < _BLOCKSIZE; i++) { 1668 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1669 oop obj = (oop)mid->object(); 1670 1671 if (obj == NULL) { 1672 // The monitor is not associated with an object. 1673 // The monitor should either be a thread-specific private 1674 // free list or the global free list. 1675 // obj == NULL IMPLIES mid->is_busy() == 0 1676 guarantee(!mid->is_busy(), "invariant"); 1677 continue; 1678 } 1679 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1680 1681 if (deflated) { 1682 mid->FreeNext = NULL; 1683 nScavenged++; 1684 } else { 1685 nInuse++; 1686 } 1687 } 1688 } 1689 } 1690 1691 gMonitorFreeCount += nScavenged; 1692 1693 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1694 1695 if (ObjectMonitor::Knob_Verbose) { 1696 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d " 1697 "ForceMonitorScavenge=%d : pop=%d free=%d", 1698 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, 1699 gMonitorPopulation, gMonitorFreeCount); 1700 tty->flush(); 1701 } 1702 1703 ForceMonitorScavenge = 0; // Reset 1704 1705 // Move the scavenged monitors back to the global free list. 1706 if (freeHeadp != NULL) { 1707 guarantee(freeTailp != NULL && nScavenged > 0, "invariant"); 1708 assert(freeTailp->FreeNext == NULL, "invariant"); 1709 // constant-time list splice - prepend scavenged segment to gFreeList 1710 freeTailp->FreeNext = gFreeList; 1711 gFreeList = freeHeadp; 1712 } 1713 Thread::muxRelease(&gListLock); 1714 1715 OM_PERFDATA_OP(Deflations, inc(nScavenged)); 1716 OM_PERFDATA_OP(MonExtant, set_value(nInCirculation)); 1717 1718 // TODO: Add objectMonitor leak detection. 1719 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1720 GVars.stwRandom = os::random(); 1721 GVars.stwCycle++; 1722 } 1723 1724 // Monitor cleanup on JavaThread::exit 1725 1726 // Iterate through monitor cache and attempt to release thread's monitors 1727 // Gives up on a particular monitor if an exception occurs, but continues 1728 // the overall iteration, swallowing the exception. 1729 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1730 private: 1731 TRAPS; 1732 1733 public: 1734 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1735 void do_monitor(ObjectMonitor* mid) { 1736 if (mid->owner() == THREAD) { 1737 if (ObjectMonitor::Knob_VerifyMatch != 0) { 1738 Handle obj((oop) mid->object()); 1739 tty->print("INFO: unexpected locked object:"); 1740 javaVFrame::print_locked_object_class_name(tty, obj, "locked"); 1741 fatal("exiting JavaThread=" INTPTR_FORMAT 1742 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT, 1743 p2i(THREAD), p2i(mid)); 1744 } 1745 (void)mid->complete_exit(CHECK); 1746 } 1747 } 1748 }; 1749 1750 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1751 // ignored. This is meant to be called during JNI thread detach which assumes 1752 // all remaining monitors are heavyweight. All exceptions are swallowed. 1753 // Scanning the extant monitor list can be time consuming. 1754 // A simple optimization is to add a per-thread flag that indicates a thread 1755 // called jni_monitorenter() during its lifetime. 1756 // 1757 // Instead of No_Savepoint_Verifier it might be cheaper to 1758 // use an idiom of the form: 1759 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1760 // <code that must not run at safepoint> 1761 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1762 // Since the tests are extremely cheap we could leave them enabled 1763 // for normal product builds. 1764 1765 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1766 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1767 NoSafepointVerifier nsv; 1768 ReleaseJavaMonitorsClosure rjmc(THREAD); 1769 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1770 ObjectSynchronizer::monitors_iterate(&rjmc); 1771 Thread::muxRelease(&gListLock); 1772 THREAD->clear_pending_exception(); 1773 } 1774 1775 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1776 switch (cause) { 1777 case inflate_cause_vm_internal: return "VM Internal"; 1778 case inflate_cause_monitor_enter: return "Monitor Enter"; 1779 case inflate_cause_wait: return "Monitor Wait"; 1780 case inflate_cause_notify: return "Monitor Notify"; 1781 case inflate_cause_hash_code: return "Monitor Hash Code"; 1782 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1783 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1784 default: 1785 ShouldNotReachHere(); 1786 } 1787 return "Unknown"; 1788 } 1789 1790 static void post_monitor_inflate_event(EventJavaMonitorInflate& event, 1791 const oop obj, 1792 const ObjectSynchronizer::InflateCause cause) { 1793 #if INCLUDE_TRACE 1794 assert(event.should_commit(), "check outside"); 1795 event.set_klass(obj->klass()); 1796 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj); 1797 event.set_cause((u1)cause); 1798 event.commit(); 1799 #endif 1800 } 1801 1802 //------------------------------------------------------------------------------ 1803 // Debugging code 1804 1805 void ObjectSynchronizer::sanity_checks(const bool verbose, 1806 const uint cache_line_size, 1807 int *error_cnt_ptr, 1808 int *warning_cnt_ptr) { 1809 u_char *addr_begin = (u_char*)&GVars; 1810 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1811 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1812 1813 if (verbose) { 1814 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1815 sizeof(SharedGlobals)); 1816 } 1817 1818 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1819 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1820 1821 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1822 if (verbose) { 1823 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1824 } 1825 1826 if (cache_line_size != 0) { 1827 // We were able to determine the L1 data cache line size so 1828 // do some cache line specific sanity checks 1829 1830 if (offset_stwRandom < cache_line_size) { 1831 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1832 "to the struct beginning than a cache line which permits " 1833 "false sharing."); 1834 (*warning_cnt_ptr)++; 1835 } 1836 1837 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1838 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1839 "SharedGlobals.hcSequence fields are closer than a cache " 1840 "line which permits false sharing."); 1841 (*warning_cnt_ptr)++; 1842 } 1843 1844 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1845 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1846 "to the struct end than a cache line which permits false " 1847 "sharing."); 1848 (*warning_cnt_ptr)++; 1849 } 1850 } 1851 } 1852 1853 #ifndef PRODUCT 1854 1855 // Verify all monitors in the monitor cache, the verification is weak. 1856 void ObjectSynchronizer::verify() { 1857 PaddedEnd<ObjectMonitor> * block = 1858 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 1859 while (block != NULL) { 1860 assert(block->object() == CHAINMARKER, "must be a block header"); 1861 for (int i = 1; i < _BLOCKSIZE; i++) { 1862 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 1863 oop object = (oop)mid->object(); 1864 if (object != NULL) { 1865 mid->verify(); 1866 } 1867 } 1868 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 1869 } 1870 } 1871 1872 // Check if monitor belongs to the monitor cache 1873 // The list is grow-only so it's *relatively* safe to traverse 1874 // the list of extant blocks without taking a lock. 1875 1876 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1877 PaddedEnd<ObjectMonitor> * block = 1878 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 1879 while (block != NULL) { 1880 assert(block->object() == CHAINMARKER, "must be a block header"); 1881 if (monitor > (ObjectMonitor *)&block[0] && 1882 monitor < (ObjectMonitor *)&block[_BLOCKSIZE]) { 1883 address mon = (address)monitor; 1884 address blk = (address)block; 1885 size_t diff = mon - blk; 1886 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned"); 1887 return 1; 1888 } 1889 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 1890 } 1891 return 0; 1892 } 1893 1894 #endif