1 /* 2 * Copyright (c) 1998, 2017, 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.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/align.hpp" 48 #include "utilities/dtrace.hpp" 49 #include "utilities/events.hpp" 50 #include "utilities/preserveException.hpp" 51 52 // The "core" versions of monitor enter and exit reside in this file. 53 // The interpreter and compilers contain specialized transliterated 54 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 55 // for instance. If you make changes here, make sure to modify the 56 // interpreter, and both C1 and C2 fast-path inline locking code emission. 57 // 58 // ----------------------------------------------------------------------------- 59 60 #ifdef DTRACE_ENABLED 61 62 // Only bother with this argument setup if dtrace is available 63 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 64 65 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 66 char* bytes = NULL; \ 67 int len = 0; \ 68 jlong jtid = SharedRuntime::get_java_tid(thread); \ 69 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 70 if (klassname != NULL) { \ 71 bytes = (char*)klassname->bytes(); \ 72 len = klassname->utf8_length(); \ 73 } 74 75 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 76 { \ 77 if (DTraceMonitorProbes) { \ 78 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 79 HOTSPOT_MONITOR_WAIT(jtid, \ 80 (uintptr_t)(monitor), bytes, len, (millis)); \ 81 } \ 82 } 83 84 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 85 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 86 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 87 88 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 89 { \ 90 if (DTraceMonitorProbes) { \ 91 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 92 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 93 (uintptr_t)(monitor), bytes, len); \ 94 } \ 95 } 96 97 #else // ndef DTRACE_ENABLED 98 99 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 100 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 101 102 #endif // ndef DTRACE_ENABLED 103 104 // This exists only as a workaround of dtrace bug 6254741 105 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 106 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 107 return 0; 108 } 109 110 #define NINFLATIONLOCKS 256 111 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS]; 112 113 // global list of blocks of monitors 114 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL; 115 // global monitor free list 116 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 117 // global monitor in-use list, for moribund threads, 118 // monitors they inflated need to be scanned for deflation 119 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 120 // count of entries in gOmInUseList 121 int ObjectSynchronizer::gOmInUseCount = 0; 122 123 static volatile intptr_t gListLock = 0; // protects global monitor lists 124 static volatile int gMonitorFreeCount = 0; // # on gFreeList 125 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation 126 127 static void post_monitor_inflate_event(EventJavaMonitorInflate&, 128 const oop, 129 const ObjectSynchronizer::InflateCause); 130 131 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 132 133 134 // =====================> Quick functions 135 136 // The quick_* forms are special fast-path variants used to improve 137 // performance. In the simplest case, a "quick_*" implementation could 138 // simply return false, in which case the caller will perform the necessary 139 // state transitions and call the slow-path form. 140 // The fast-path is designed to handle frequently arising cases in an efficient 141 // manner and is just a degenerate "optimistic" variant of the slow-path. 142 // returns true -- to indicate the call was satisfied. 143 // returns false -- to indicate the call needs the services of the slow-path. 144 // A no-loitering ordinance is in effect for code in the quick_* family 145 // operators: safepoints or indefinite blocking (blocking that might span a 146 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 147 // entry. 148 // 149 // Consider: An interesting optimization is to have the JIT recognize the 150 // following common idiom: 151 // synchronized (someobj) { .... ; notify(); } 152 // That is, we find a notify() or notifyAll() call that immediately precedes 153 // the monitorexit operation. In that case the JIT could fuse the operations 154 // into a single notifyAndExit() runtime primitive. 155 156 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) { 157 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 158 assert(self->is_Java_thread(), "invariant"); 159 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 160 NoSafepointVerifier nsv; 161 if (obj == NULL) return false; // slow-path for invalid obj 162 const markOop mark = obj->mark(); 163 164 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) { 165 // Degenerate notify 166 // stack-locked by caller so by definition the implied waitset is empty. 167 return true; 168 } 169 170 if (mark->has_monitor()) { 171 ObjectMonitor * const mon = mark->monitor(); 172 assert(mon->object() == obj, "invariant"); 173 if (mon->owner() != self) return false; // slow-path for IMS exception 174 175 if (mon->first_waiter() != NULL) { 176 // We have one or more waiters. Since this is an inflated monitor 177 // that we own, we can transfer one or more threads from the waitset 178 // to the entrylist here and now, avoiding the slow-path. 179 if (all) { 180 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 181 } else { 182 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 183 } 184 int tally = 0; 185 do { 186 mon->INotify(self); 187 ++tally; 188 } while (mon->first_waiter() != NULL && all); 189 OM_PERFDATA_OP(Notifications, inc(tally)); 190 } 191 return true; 192 } 193 194 // biased locking and any other IMS exception states take the slow-path 195 return false; 196 } 197 198 199 // The LockNode emitted directly at the synchronization site would have 200 // been too big if it were to have included support for the cases of inflated 201 // recursive enter and exit, so they go here instead. 202 // Note that we can't safely call AsyncPrintJavaStack() from within 203 // quick_enter() as our thread state remains _in_Java. 204 205 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self, 206 BasicLock * lock) { 207 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 208 assert(Self->is_Java_thread(), "invariant"); 209 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant"); 210 NoSafepointVerifier nsv; 211 if (obj == NULL) return false; // Need to throw NPE 212 const markOop mark = obj->mark(); 213 214 if (mark->has_monitor()) { 215 ObjectMonitor * const m = mark->monitor(); 216 assert(m->object() == obj, "invariant"); 217 Thread * const owner = (Thread *) m->_owner; 218 219 // Lock contention and Transactional Lock Elision (TLE) diagnostics 220 // and observability 221 // Case: light contention possibly amenable to TLE 222 // Case: TLE inimical operations such as nested/recursive synchronization 223 224 if (owner == Self) { 225 m->_recursions++; 226 return true; 227 } 228 229 // This Java Monitor is inflated so obj's header will never be 230 // displaced to this thread's BasicLock. Make the displaced header 231 // non-NULL so this BasicLock is not seen as recursive nor as 232 // being locked. We do this unconditionally so that this thread's 233 // BasicLock cannot be mis-interpreted by any stack walkers. For 234 // performance reasons, stack walkers generally first check for 235 // Biased Locking in the object's header, the second check is for 236 // stack-locking in the object's header, the third check is for 237 // recursive stack-locking in the displaced header in the BasicLock, 238 // and last are the inflated Java Monitor (ObjectMonitor) checks. 239 lock->set_displaced_header(markOopDesc::unused_mark()); 240 241 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) { 242 assert(m->_recursions == 0, "invariant"); 243 assert(m->_owner == Self, "invariant"); 244 return true; 245 } 246 } 247 248 // Note that we could inflate in quick_enter. 249 // This is likely a useful optimization 250 // Critically, in quick_enter() we must not: 251 // -- perform bias revocation, or 252 // -- block indefinitely, or 253 // -- reach a safepoint 254 255 return false; // revert to slow-path 256 } 257 258 // ----------------------------------------------------------------------------- 259 // Fast Monitor Enter/Exit 260 // This the fast monitor enter. The interpreter and compiler use 261 // some assembly copies of this code. Make sure update those code 262 // if the following function is changed. The implementation is 263 // extremely sensitive to race condition. Be careful. 264 265 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 266 bool attempt_rebias, TRAPS) { 267 if (UseBiasedLocking) { 268 if (!SafepointSynchronize::is_at_safepoint()) { 269 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 270 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 271 return; 272 } 273 } else { 274 assert(!attempt_rebias, "can not rebias toward VM thread"); 275 BiasedLocking::revoke_at_safepoint(obj); 276 } 277 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 278 } 279 280 slow_enter(obj, lock, THREAD); 281 } 282 283 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 284 markOop mark = object->mark(); 285 // We cannot check for Biased Locking if we are racing an inflation. 286 assert(mark == markOopDesc::INFLATING() || 287 !mark->has_bias_pattern(), "should not see bias pattern here"); 288 289 markOop dhw = lock->displaced_header(); 290 if (dhw == NULL) { 291 // If the displaced header is NULL, then this exit matches up with 292 // a recursive enter. No real work to do here except for diagnostics. 293 #ifndef PRODUCT 294 if (mark != markOopDesc::INFLATING()) { 295 // Only do diagnostics if we are not racing an inflation. Simply 296 // exiting a recursive enter of a Java Monitor that is being 297 // inflated is safe; see the has_monitor() comment below. 298 assert(!mark->is_neutral(), "invariant"); 299 assert(!mark->has_locker() || 300 THREAD->is_lock_owned((address)mark->locker()), "invariant"); 301 if (mark->has_monitor()) { 302 // The BasicLock's displaced_header is marked as a recursive 303 // enter and we have an inflated Java Monitor (ObjectMonitor). 304 // This is a special case where the Java Monitor was inflated 305 // after this thread entered the stack-lock recursively. When a 306 // Java Monitor is inflated, we cannot safely walk the Java 307 // Monitor owner's stack and update the BasicLocks because a 308 // Java Monitor can be asynchronously inflated by a thread that 309 // does not own the Java Monitor. 310 ObjectMonitor * m = mark->monitor(); 311 assert(((oop)(m->object()))->mark() == mark, "invariant"); 312 assert(m->is_entered(THREAD), "invariant"); 313 } 314 } 315 #endif 316 return; 317 } 318 319 if (mark == (markOop) lock) { 320 // If the object is stack-locked by the current thread, try to 321 // swing the displaced header from the BasicLock back to the mark. 322 assert(dhw->is_neutral(), "invariant"); 323 if (object->cas_set_mark(dhw, mark) == mark) { 324 TEVENT(fast_exit: release stack-lock); 325 return; 326 } 327 } 328 329 // We have to take the slow-path of possible inflation and then exit. 330 ObjectSynchronizer::inflate(THREAD, 331 object, 332 inflate_cause_vm_internal)->exit(true, THREAD); 333 } 334 335 // ----------------------------------------------------------------------------- 336 // Interpreter/Compiler Slow Case 337 // This routine is used to handle interpreter/compiler slow case 338 // We don't need to use fast path here, because it must have been 339 // failed in the interpreter/compiler code. 340 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 341 markOop mark = obj->mark(); 342 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 343 344 if (mark->is_neutral()) { 345 // Anticipate successful CAS -- the ST of the displaced mark must 346 // be visible <= the ST performed by the CAS. 347 lock->set_displaced_header(mark); 348 if (mark == obj()->cas_set_mark((markOop) lock, mark)) { 349 TEVENT(slow_enter: release stacklock); 350 return; 351 } 352 // Fall through to inflate() ... 353 } else if (mark->has_locker() && 354 THREAD->is_lock_owned((address)mark->locker())) { 355 assert(lock != mark->locker(), "must not re-lock the same lock"); 356 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 357 lock->set_displaced_header(NULL); 358 return; 359 } 360 361 // The object header will never be displaced to this lock, 362 // so it does not matter what the value is, except that it 363 // must be non-zero to avoid looking like a re-entrant lock, 364 // and must not look locked either. 365 lock->set_displaced_header(markOopDesc::unused_mark()); 366 ObjectSynchronizer::inflate(THREAD, 367 obj(), 368 inflate_cause_monitor_enter)->enter(THREAD); 369 } 370 371 // This routine is used to handle interpreter/compiler slow case 372 // We don't need to use fast path here, because it must have 373 // failed in the interpreter/compiler code. Simply use the heavy 374 // weight monitor should be ok, unless someone find otherwise. 375 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 376 fast_exit(object, lock, THREAD); 377 } 378 379 // ----------------------------------------------------------------------------- 380 // Class Loader support to workaround deadlocks on the class loader lock objects 381 // Also used by GC 382 // complete_exit()/reenter() are used to wait on a nested lock 383 // i.e. to give up an outer lock completely and then re-enter 384 // Used when holding nested locks - lock acquisition order: lock1 then lock2 385 // 1) complete_exit lock1 - saving recursion count 386 // 2) wait on lock2 387 // 3) when notified on lock2, unlock lock2 388 // 4) reenter lock1 with original recursion count 389 // 5) lock lock2 390 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 391 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 392 TEVENT(complete_exit); 393 if (UseBiasedLocking) { 394 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 395 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 396 } 397 398 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 399 obj(), 400 inflate_cause_vm_internal); 401 402 return monitor->complete_exit(THREAD); 403 } 404 405 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 406 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 407 TEVENT(reenter); 408 if (UseBiasedLocking) { 409 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 410 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 411 } 412 413 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 414 obj(), 415 inflate_cause_vm_internal); 416 417 monitor->reenter(recursion, THREAD); 418 } 419 // ----------------------------------------------------------------------------- 420 // JNI locks on java objects 421 // NOTE: must use heavy weight monitor to handle jni monitor enter 422 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 423 // the current locking is from JNI instead of Java code 424 TEVENT(jni_enter); 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 THREAD->set_current_pending_monitor_is_from_java(false); 430 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD); 431 THREAD->set_current_pending_monitor_is_from_java(true); 432 } 433 434 // NOTE: must use heavy weight monitor to handle jni monitor exit 435 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 436 TEVENT(jni_exit); 437 if (UseBiasedLocking) { 438 Handle h_obj(THREAD, obj); 439 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 440 obj = h_obj(); 441 } 442 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 443 444 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 445 obj, 446 inflate_cause_jni_exit); 447 // If this thread has locked the object, exit the monitor. Note: can't use 448 // monitor->check(CHECK); must exit even if an exception is pending. 449 if (monitor->check(THREAD)) { 450 monitor->exit(true, THREAD); 451 } 452 } 453 454 // ----------------------------------------------------------------------------- 455 // Internal VM locks on java objects 456 // standard constructor, allows locking failures 457 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 458 _dolock = doLock; 459 _thread = thread; 460 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 461 _obj = obj; 462 463 if (_dolock) { 464 TEVENT(ObjectLocker); 465 466 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 467 } 468 } 469 470 ObjectLocker::~ObjectLocker() { 471 if (_dolock) { 472 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 473 } 474 } 475 476 477 // ----------------------------------------------------------------------------- 478 // Wait/Notify/NotifyAll 479 // NOTE: must use heavy weight monitor to handle wait() 480 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 481 if (UseBiasedLocking) { 482 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 483 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 484 } 485 if (millis < 0) { 486 TEVENT(wait - throw IAX); 487 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 488 } 489 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 490 obj(), 491 inflate_cause_wait); 492 493 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 494 monitor->wait(millis, true, THREAD); 495 496 // This dummy call is in place to get around dtrace bug 6254741. Once 497 // that's fixed we can uncomment the following line, remove the call 498 // and change this function back into a "void" func. 499 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 500 return dtrace_waited_probe(monitor, obj, THREAD); 501 } 502 503 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 504 if (UseBiasedLocking) { 505 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 506 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 507 } 508 if (millis < 0) { 509 TEVENT(wait - throw IAX); 510 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 511 } 512 ObjectSynchronizer::inflate(THREAD, 513 obj(), 514 inflate_cause_wait)->wait(millis, false, THREAD); 515 } 516 517 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 518 if (UseBiasedLocking) { 519 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 520 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 521 } 522 523 markOop mark = obj->mark(); 524 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 525 return; 526 } 527 ObjectSynchronizer::inflate(THREAD, 528 obj(), 529 inflate_cause_notify)->notify(THREAD); 530 } 531 532 // NOTE: see comment of notify() 533 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 534 if (UseBiasedLocking) { 535 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 536 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 537 } 538 539 markOop mark = obj->mark(); 540 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 541 return; 542 } 543 ObjectSynchronizer::inflate(THREAD, 544 obj(), 545 inflate_cause_notify)->notifyAll(THREAD); 546 } 547 548 // ----------------------------------------------------------------------------- 549 // Hash Code handling 550 // 551 // Performance concern: 552 // OrderAccess::storestore() calls release() which at one time stored 0 553 // into the global volatile OrderAccess::dummy variable. This store was 554 // unnecessary for correctness. Many threads storing into a common location 555 // causes considerable cache migration or "sloshing" on large SMP systems. 556 // As such, I avoided using OrderAccess::storestore(). In some cases 557 // OrderAccess::fence() -- which incurs local latency on the executing 558 // processor -- is a better choice as it scales on SMP systems. 559 // 560 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 561 // a discussion of coherency costs. Note that all our current reference 562 // platforms provide strong ST-ST order, so the issue is moot on IA32, 563 // x64, and SPARC. 564 // 565 // As a general policy we use "volatile" to control compiler-based reordering 566 // and explicit fences (barriers) to control for architectural reordering 567 // performed by the CPU(s) or platform. 568 569 struct SharedGlobals { 570 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE]; 571 // These are highly shared mostly-read variables. 572 // To avoid false-sharing they need to be the sole occupants of a cache line. 573 volatile int stwRandom; 574 volatile int stwCycle; 575 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 576 // Hot RW variable -- Sequester to avoid false-sharing 577 volatile int hcSequence; 578 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int)); 579 }; 580 581 static SharedGlobals GVars; 582 static int MonitorScavengeThreshold = 1000000; 583 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 584 585 static markOop ReadStableMark(oop obj) { 586 markOop mark = obj->mark(); 587 if (!mark->is_being_inflated()) { 588 return mark; // normal fast-path return 589 } 590 591 int its = 0; 592 for (;;) { 593 markOop mark = obj->mark(); 594 if (!mark->is_being_inflated()) { 595 return mark; // normal fast-path return 596 } 597 598 // The object is being inflated by some other thread. 599 // The caller of ReadStableMark() must wait for inflation to complete. 600 // Avoid live-lock 601 // TODO: consider calling SafepointSynchronize::do_call_back() while 602 // spinning to see if there's a safepoint pending. If so, immediately 603 // yielding or blocking would be appropriate. Avoid spinning while 604 // there is a safepoint pending. 605 // TODO: add inflation contention performance counters. 606 // TODO: restrict the aggregate number of spinners. 607 608 ++its; 609 if (its > 10000 || !os::is_MP()) { 610 if (its & 1) { 611 os::naked_yield(); 612 TEVENT(Inflate: INFLATING - yield); 613 } else { 614 // Note that the following code attenuates the livelock problem but is not 615 // a complete remedy. A more complete solution would require that the inflating 616 // thread hold the associated inflation lock. The following code simply restricts 617 // the number of spinners to at most one. We'll have N-2 threads blocked 618 // on the inflationlock, 1 thread holding the inflation lock and using 619 // a yield/park strategy, and 1 thread in the midst of inflation. 620 // A more refined approach would be to change the encoding of INFLATING 621 // to allow encapsulation of a native thread pointer. Threads waiting for 622 // inflation to complete would use CAS to push themselves onto a singly linked 623 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 624 // and calling park(). When inflation was complete the thread that accomplished inflation 625 // would detach the list and set the markword to inflated with a single CAS and 626 // then for each thread on the list, set the flag and unpark() the thread. 627 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 628 // wakes at most one thread whereas we need to wake the entire list. 629 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 630 int YieldThenBlock = 0; 631 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 632 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 633 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 634 while (obj->mark() == markOopDesc::INFLATING()) { 635 // Beware: NakedYield() is advisory and has almost no effect on some platforms 636 // so we periodically call Self->_ParkEvent->park(1). 637 // We use a mixed spin/yield/block mechanism. 638 if ((YieldThenBlock++) >= 16) { 639 Thread::current()->_ParkEvent->park(1); 640 } else { 641 os::naked_yield(); 642 } 643 } 644 Thread::muxRelease(gInflationLocks + ix); 645 TEVENT(Inflate: INFLATING - yield/park); 646 } 647 } else { 648 SpinPause(); // SMP-polite spinning 649 } 650 } 651 } 652 653 // hashCode() generation : 654 // 655 // Possibilities: 656 // * MD5Digest of {obj,stwRandom} 657 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 658 // * A DES- or AES-style SBox[] mechanism 659 // * One of the Phi-based schemes, such as: 660 // 2654435761 = 2^32 * Phi (golden ratio) 661 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 662 // * A variation of Marsaglia's shift-xor RNG scheme. 663 // * (obj ^ stwRandom) is appealing, but can result 664 // in undesirable regularity in the hashCode values of adjacent objects 665 // (objects allocated back-to-back, in particular). This could potentially 666 // result in hashtable collisions and reduced hashtable efficiency. 667 // There are simple ways to "diffuse" the middle address bits over the 668 // generated hashCode values: 669 670 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 671 intptr_t value = 0; 672 if (hashCode == 0) { 673 // This form uses global Park-Miller RNG. 674 // On MP system we'll have lots of RW access to a global, so the 675 // mechanism induces lots of coherency traffic. 676 value = os::random(); 677 } else if (hashCode == 1) { 678 // This variation has the property of being stable (idempotent) 679 // between STW operations. This can be useful in some of the 1-0 680 // synchronization schemes. 681 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 682 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 683 } else if (hashCode == 2) { 684 value = 1; // for sensitivity testing 685 } else if (hashCode == 3) { 686 value = ++GVars.hcSequence; 687 } else if (hashCode == 4) { 688 value = cast_from_oop<intptr_t>(obj); 689 } else { 690 // Marsaglia's xor-shift scheme with thread-specific state 691 // This is probably the best overall implementation -- we'll 692 // likely make this the default in future releases. 693 unsigned t = Self->_hashStateX; 694 t ^= (t << 11); 695 Self->_hashStateX = Self->_hashStateY; 696 Self->_hashStateY = Self->_hashStateZ; 697 Self->_hashStateZ = Self->_hashStateW; 698 unsigned v = Self->_hashStateW; 699 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 700 Self->_hashStateW = v; 701 value = v; 702 } 703 704 value &= markOopDesc::hash_mask; 705 if (value == 0) value = 0xBAD; 706 assert(value != markOopDesc::no_hash, "invariant"); 707 TEVENT(hashCode: GENERATE); 708 return value; 709 } 710 711 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 712 if (UseBiasedLocking) { 713 // NOTE: many places throughout the JVM do not expect a safepoint 714 // to be taken here, in particular most operations on perm gen 715 // objects. However, we only ever bias Java instances and all of 716 // the call sites of identity_hash that might revoke biases have 717 // been checked to make sure they can handle a safepoint. The 718 // added check of the bias pattern is to avoid useless calls to 719 // thread-local storage. 720 if (obj->mark()->has_bias_pattern()) { 721 // Handle for oop obj in case of STW safepoint 722 Handle hobj(Self, obj); 723 // Relaxing assertion for bug 6320749. 724 assert(Universe::verify_in_progress() || 725 !SafepointSynchronize::is_at_safepoint(), 726 "biases should not be seen by VM thread here"); 727 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 728 obj = hobj(); 729 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 730 } 731 } 732 733 // hashCode() is a heap mutator ... 734 // Relaxing assertion for bug 6320749. 735 assert(Universe::verify_in_progress() || DumpSharedSpaces || 736 !SafepointSynchronize::is_at_safepoint(), "invariant"); 737 assert(Universe::verify_in_progress() || DumpSharedSpaces || 738 Self->is_Java_thread() , "invariant"); 739 assert(Universe::verify_in_progress() || DumpSharedSpaces || 740 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 741 742 ObjectMonitor* monitor = NULL; 743 markOop temp, test; 744 intptr_t hash; 745 markOop mark = ReadStableMark(obj); 746 747 // object should remain ineligible for biased locking 748 assert(!mark->has_bias_pattern(), "invariant"); 749 750 if (mark->is_neutral()) { 751 hash = mark->hash(); // this is a normal header 752 if (hash) { // if it has hash, just return it 753 return hash; 754 } 755 hash = get_next_hash(Self, obj); // allocate a new hash code 756 temp = mark->copy_set_hash(hash); // merge the hash code into header 757 // use (machine word version) atomic operation to install the hash 758 test = obj->cas_set_mark(temp, mark); 759 if (test == mark) { 760 return hash; 761 } 762 // If atomic operation failed, we must inflate the header 763 // into heavy weight monitor. We could add more code here 764 // for fast path, but it does not worth the complexity. 765 } else if (mark->has_monitor()) { 766 monitor = mark->monitor(); 767 temp = monitor->header(); 768 assert(temp->is_neutral(), "invariant"); 769 hash = temp->hash(); 770 if (hash) { 771 return hash; 772 } 773 // Skip to the following code to reduce code size 774 } else if (Self->is_lock_owned((address)mark->locker())) { 775 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 776 assert(temp->is_neutral(), "invariant"); 777 hash = temp->hash(); // by current thread, check if the displaced 778 if (hash) { // header contains hash code 779 return hash; 780 } 781 // WARNING: 782 // The displaced header is strictly immutable. 783 // It can NOT be changed in ANY cases. So we have 784 // to inflate the header into heavyweight monitor 785 // even the current thread owns the lock. The reason 786 // is the BasicLock (stack slot) will be asynchronously 787 // read by other threads during the inflate() function. 788 // Any change to stack may not propagate to other threads 789 // correctly. 790 } 791 792 // Inflate the monitor to set hash code 793 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code); 794 // Load displaced header and check it has hash code 795 mark = monitor->header(); 796 assert(mark->is_neutral(), "invariant"); 797 hash = mark->hash(); 798 if (hash == 0) { 799 hash = get_next_hash(Self, obj); 800 temp = mark->copy_set_hash(hash); // merge hash code into header 801 assert(temp->is_neutral(), "invariant"); 802 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark); 803 if (test != mark) { 804 // The only update to the header in the monitor (outside GC) 805 // is install the hash code. If someone add new usage of 806 // displaced header, please update this code 807 hash = test->hash(); 808 assert(test->is_neutral(), "invariant"); 809 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 810 } 811 } 812 // We finally get the hash 813 return hash; 814 } 815 816 // Deprecated -- use FastHashCode() instead. 817 818 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 819 return FastHashCode(Thread::current(), obj()); 820 } 821 822 823 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 824 Handle h_obj) { 825 if (UseBiasedLocking) { 826 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 827 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 828 } 829 830 assert(thread == JavaThread::current(), "Can only be called on current thread"); 831 oop obj = h_obj(); 832 833 markOop mark = ReadStableMark(obj); 834 835 // Uncontended case, header points to stack 836 if (mark->has_locker()) { 837 return thread->is_lock_owned((address)mark->locker()); 838 } 839 // Contended case, header points to ObjectMonitor (tagged pointer) 840 if (mark->has_monitor()) { 841 ObjectMonitor* monitor = mark->monitor(); 842 return monitor->is_entered(thread) != 0; 843 } 844 // Unlocked case, header in place 845 assert(mark->is_neutral(), "sanity check"); 846 return false; 847 } 848 849 // Be aware of this method could revoke bias of the lock object. 850 // This method queries the ownership of the lock handle specified by 'h_obj'. 851 // If the current thread owns the lock, it returns owner_self. If no 852 // thread owns the lock, it returns owner_none. Otherwise, it will return 853 // owner_other. 854 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 855 (JavaThread *self, Handle h_obj) { 856 // The caller must beware this method can revoke bias, and 857 // revocation can result in a safepoint. 858 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 859 assert(self->thread_state() != _thread_blocked, "invariant"); 860 861 // Possible mark states: neutral, biased, stack-locked, inflated 862 863 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 864 // CASE: biased 865 BiasedLocking::revoke_and_rebias(h_obj, false, self); 866 assert(!h_obj->mark()->has_bias_pattern(), 867 "biases should be revoked by now"); 868 } 869 870 assert(self == JavaThread::current(), "Can only be called on current thread"); 871 oop obj = h_obj(); 872 markOop mark = ReadStableMark(obj); 873 874 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 875 if (mark->has_locker()) { 876 return self->is_lock_owned((address)mark->locker()) ? 877 owner_self : owner_other; 878 } 879 880 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 881 // The Object:ObjectMonitor relationship is stable as long as we're 882 // not at a safepoint. 883 if (mark->has_monitor()) { 884 void * owner = mark->monitor()->_owner; 885 if (owner == NULL) return owner_none; 886 return (owner == self || 887 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 888 } 889 890 // CASE: neutral 891 assert(mark->is_neutral(), "sanity check"); 892 return owner_none; // it's unlocked 893 } 894 895 // FIXME: jvmti should call this 896 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 897 if (UseBiasedLocking) { 898 if (SafepointSynchronize::is_at_safepoint()) { 899 BiasedLocking::revoke_at_safepoint(h_obj); 900 } else { 901 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 902 } 903 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 904 } 905 906 oop obj = h_obj(); 907 address owner = NULL; 908 909 markOop mark = ReadStableMark(obj); 910 911 // Uncontended case, header points to stack 912 if (mark->has_locker()) { 913 owner = (address) mark->locker(); 914 } 915 916 // Contended case, header points to ObjectMonitor (tagged pointer) 917 if (mark->has_monitor()) { 918 ObjectMonitor* monitor = mark->monitor(); 919 assert(monitor != NULL, "monitor should be non-null"); 920 owner = (address) monitor->owner(); 921 } 922 923 if (owner != NULL) { 924 // owning_thread_from_monitor_owner() may also return NULL here 925 return Threads::owning_thread_from_monitor_owner(t_list, owner); 926 } 927 928 // Unlocked case, header in place 929 // Cannot have assertion since this object may have been 930 // locked by another thread when reaching here. 931 // assert(mark->is_neutral(), "sanity check"); 932 933 return NULL; 934 } 935 936 // Visitors ... 937 938 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 939 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 940 while (block != NULL) { 941 assert(block->object() == CHAINMARKER, "must be a block header"); 942 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 943 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 944 oop object = (oop)mid->object(); 945 if (object != NULL) { 946 closure->do_monitor(mid); 947 } 948 } 949 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 950 } 951 } 952 953 // Get the next block in the block list. 954 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) { 955 assert(block->object() == CHAINMARKER, "must be a block header"); 956 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext; 957 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 958 return block; 959 } 960 961 static bool monitors_used_above_threshold() { 962 if (gMonitorPopulation == 0) { 963 return false; 964 } 965 int monitors_used = gMonitorPopulation - gMonitorFreeCount; 966 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation; 967 return monitor_usage > MonitorUsedDeflationThreshold; 968 } 969 970 bool ObjectSynchronizer::is_cleanup_needed() { 971 if (MonitorUsedDeflationThreshold > 0) { 972 return monitors_used_above_threshold(); 973 } 974 return false; 975 } 976 977 void ObjectSynchronizer::oops_do(OopClosure* f) { 978 if (MonitorInUseLists) { 979 // When using thread local monitor lists, we only scan the 980 // global used list here (for moribund threads), and 981 // the thread-local monitors in Thread::oops_do(). 982 global_used_oops_do(f); 983 } else { 984 global_oops_do(f); 985 } 986 } 987 988 void ObjectSynchronizer::global_oops_do(OopClosure* f) { 989 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 990 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 991 for (; block != NULL; block = next(block)) { 992 assert(block->object() == CHAINMARKER, "must be a block header"); 993 for (int i = 1; i < _BLOCKSIZE; i++) { 994 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 995 if (mid->object() != NULL) { 996 f->do_oop((oop*)mid->object_addr()); 997 } 998 } 999 } 1000 } 1001 1002 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1003 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1004 list_oops_do(gOmInUseList, f); 1005 } 1006 1007 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1008 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1009 list_oops_do(thread->omInUseList, f); 1010 } 1011 1012 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) { 1013 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1014 ObjectMonitor* mid; 1015 for (mid = list; mid != NULL; mid = mid->FreeNext) { 1016 if (mid->object() != NULL) { 1017 f->do_oop((oop*)mid->object_addr()); 1018 } 1019 } 1020 } 1021 1022 1023 // ----------------------------------------------------------------------------- 1024 // ObjectMonitor Lifecycle 1025 // ----------------------- 1026 // Inflation unlinks monitors from the global gFreeList and 1027 // associates them with objects. Deflation -- which occurs at 1028 // STW-time -- disassociates idle monitors from objects. Such 1029 // scavenged monitors are returned to the gFreeList. 1030 // 1031 // The global list is protected by gListLock. All the critical sections 1032 // are short and operate in constant-time. 1033 // 1034 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1035 // 1036 // Lifecycle: 1037 // -- unassigned and on the global free list 1038 // -- unassigned and on a thread's private omFreeList 1039 // -- assigned to an object. The object is inflated and the mark refers 1040 // to the objectmonitor. 1041 1042 1043 // Constraining monitor pool growth via MonitorBound ... 1044 // 1045 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1046 // the rate of scavenging is driven primarily by GC. As such, we can find 1047 // an inordinate number of monitors in circulation. 1048 // To avoid that scenario we can artificially induce a STW safepoint 1049 // if the pool appears to be growing past some reasonable bound. 1050 // Generally we favor time in space-time tradeoffs, but as there's no 1051 // natural back-pressure on the # of extant monitors we need to impose some 1052 // type of limit. Beware that if MonitorBound is set to too low a value 1053 // we could just loop. In addition, if MonitorBound is set to a low value 1054 // we'll incur more safepoints, which are harmful to performance. 1055 // See also: GuaranteedSafepointInterval 1056 // 1057 // The current implementation uses asynchronous VM operations. 1058 1059 static void InduceScavenge(Thread * Self, const char * Whence) { 1060 // Induce STW safepoint to trim monitors 1061 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1062 // More precisely, trigger an asynchronous STW safepoint as the number 1063 // of active monitors passes the specified threshold. 1064 // TODO: assert thread state is reasonable 1065 1066 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 1067 if (ObjectMonitor::Knob_Verbose) { 1068 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)", 1069 Whence, ForceMonitorScavenge) ; 1070 tty->flush(); 1071 } 1072 // Induce a 'null' safepoint to scavenge monitors 1073 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 1074 // to the VMthread and have a lifespan longer than that of this activation record. 1075 // The VMThread will delete the op when completed. 1076 VMThread::execute(new VM_ScavengeMonitors()); 1077 1078 if (ObjectMonitor::Knob_Verbose) { 1079 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)", 1080 Whence, ForceMonitorScavenge) ; 1081 tty->flush(); 1082 } 1083 } 1084 } 1085 1086 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1087 ObjectMonitor* mid; 1088 int in_use_tally = 0; 1089 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1090 in_use_tally++; 1091 } 1092 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1093 1094 int free_tally = 0; 1095 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1096 free_tally++; 1097 } 1098 assert(free_tally == Self->omFreeCount, "free count off"); 1099 } 1100 1101 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) { 1102 // A large MAXPRIVATE value reduces both list lock contention 1103 // and list coherency traffic, but also tends to increase the 1104 // number of objectMonitors in circulation as well as the STW 1105 // scavenge costs. As usual, we lean toward time in space-time 1106 // tradeoffs. 1107 const int MAXPRIVATE = 1024; 1108 for (;;) { 1109 ObjectMonitor * m; 1110 1111 // 1: try to allocate from the thread's local omFreeList. 1112 // Threads will attempt to allocate first from their local list, then 1113 // from the global list, and only after those attempts fail will the thread 1114 // attempt to instantiate new monitors. Thread-local free lists take 1115 // heat off the gListLock and improve allocation latency, as well as reducing 1116 // coherency traffic on the shared global list. 1117 m = Self->omFreeList; 1118 if (m != NULL) { 1119 Self->omFreeList = m->FreeNext; 1120 Self->omFreeCount--; 1121 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1122 guarantee(m->object() == NULL, "invariant"); 1123 if (MonitorInUseLists) { 1124 m->FreeNext = Self->omInUseList; 1125 Self->omInUseList = m; 1126 Self->omInUseCount++; 1127 if (ObjectMonitor::Knob_VerifyInUse) { 1128 verifyInUse(Self); 1129 } 1130 } else { 1131 m->FreeNext = NULL; 1132 } 1133 return m; 1134 } 1135 1136 // 2: try to allocate from the global gFreeList 1137 // CONSIDER: use muxTry() instead of muxAcquire(). 1138 // If the muxTry() fails then drop immediately into case 3. 1139 // If we're using thread-local free lists then try 1140 // to reprovision the caller's free list. 1141 if (gFreeList != NULL) { 1142 // Reprovision the thread's omFreeList. 1143 // Use bulk transfers to reduce the allocation rate and heat 1144 // on various locks. 1145 Thread::muxAcquire(&gListLock, "omAlloc"); 1146 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1147 gMonitorFreeCount--; 1148 ObjectMonitor * take = gFreeList; 1149 gFreeList = take->FreeNext; 1150 guarantee(take->object() == NULL, "invariant"); 1151 guarantee(!take->is_busy(), "invariant"); 1152 take->Recycle(); 1153 omRelease(Self, take, false); 1154 } 1155 Thread::muxRelease(&gListLock); 1156 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1157 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1158 TEVENT(omFirst - reprovision); 1159 1160 const int mx = MonitorBound; 1161 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1162 // We can't safely induce a STW safepoint from omAlloc() as our thread 1163 // state may not be appropriate for such activities and callers may hold 1164 // naked oops, so instead we defer the action. 1165 InduceScavenge(Self, "omAlloc"); 1166 } 1167 continue; 1168 } 1169 1170 // 3: allocate a block of new ObjectMonitors 1171 // Both the local and global free lists are empty -- resort to malloc(). 1172 // In the current implementation objectMonitors are TSM - immortal. 1173 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1174 // each ObjectMonitor to start at the beginning of a cache line, 1175 // so we use align_up(). 1176 // A better solution would be to use C++ placement-new. 1177 // BEWARE: As it stands currently, we don't run the ctors! 1178 assert(_BLOCKSIZE > 1, "invariant"); 1179 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1180 PaddedEnd<ObjectMonitor> * temp; 1181 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1182 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1183 mtInternal); 1184 temp = (PaddedEnd<ObjectMonitor> *) 1185 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE); 1186 1187 // NOTE: (almost) no way to recover if allocation failed. 1188 // We might be able to induce a STW safepoint and scavenge enough 1189 // objectMonitors to permit progress. 1190 if (temp == NULL) { 1191 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1192 "Allocate ObjectMonitors"); 1193 } 1194 (void)memset((void *) temp, 0, neededsize); 1195 1196 // Format the block. 1197 // initialize the linked list, each monitor points to its next 1198 // forming the single linked free list, the very first monitor 1199 // will points to next block, which forms the block list. 1200 // The trick of using the 1st element in the block as gBlockList 1201 // linkage should be reconsidered. A better implementation would 1202 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1203 1204 for (int i = 1; i < _BLOCKSIZE; i++) { 1205 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1206 } 1207 1208 // terminate the last monitor as the end of list 1209 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1210 1211 // Element [0] is reserved for global list linkage 1212 temp[0].set_object(CHAINMARKER); 1213 1214 // Consider carving out this thread's current request from the 1215 // block in hand. This avoids some lock traffic and redundant 1216 // list activity. 1217 1218 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1219 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1220 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1221 gMonitorPopulation += _BLOCKSIZE-1; 1222 gMonitorFreeCount += _BLOCKSIZE-1; 1223 1224 // Add the new block to the list of extant blocks (gBlockList). 1225 // The very first objectMonitor in a block is reserved and dedicated. 1226 // It serves as blocklist "next" linkage. 1227 temp[0].FreeNext = gBlockList; 1228 // There are lock-free uses of gBlockList so make sure that 1229 // the previous stores happen before we update gBlockList. 1230 OrderAccess::release_store(&gBlockList, temp); 1231 1232 // Add the new string of objectMonitors to the global free list 1233 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1234 gFreeList = temp + 1; 1235 Thread::muxRelease(&gListLock); 1236 TEVENT(Allocate block of monitors); 1237 } 1238 } 1239 1240 // Place "m" on the caller's private per-thread omFreeList. 1241 // In practice there's no need to clamp or limit the number of 1242 // monitors on a thread's omFreeList as the only time we'll call 1243 // omRelease is to return a monitor to the free list after a CAS 1244 // attempt failed. This doesn't allow unbounded #s of monitors to 1245 // accumulate on a thread's free list. 1246 // 1247 // Key constraint: all ObjectMonitors on a thread's free list and the global 1248 // free list must have their object field set to null. This prevents the 1249 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1250 1251 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1252 bool fromPerThreadAlloc) { 1253 guarantee(m->object() == NULL, "invariant"); 1254 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1255 // Remove from omInUseList 1256 if (MonitorInUseLists && fromPerThreadAlloc) { 1257 ObjectMonitor* cur_mid_in_use = NULL; 1258 bool extracted = false; 1259 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1260 if (m == mid) { 1261 // extract from per-thread in-use list 1262 if (mid == Self->omInUseList) { 1263 Self->omInUseList = mid->FreeNext; 1264 } else if (cur_mid_in_use != NULL) { 1265 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1266 } 1267 extracted = true; 1268 Self->omInUseCount--; 1269 if (ObjectMonitor::Knob_VerifyInUse) { 1270 verifyInUse(Self); 1271 } 1272 break; 1273 } 1274 } 1275 assert(extracted, "Should have extracted from in-use list"); 1276 } 1277 1278 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1279 m->FreeNext = Self->omFreeList; 1280 Self->omFreeList = m; 1281 Self->omFreeCount++; 1282 } 1283 1284 // Return the monitors of a moribund thread's local free list to 1285 // the global free list. Typically a thread calls omFlush() when 1286 // it's dying. We could also consider having the VM thread steal 1287 // monitors from threads that have not run java code over a few 1288 // consecutive STW safepoints. Relatedly, we might decay 1289 // omFreeProvision at STW safepoints. 1290 // 1291 // Also return the monitors of a moribund thread's omInUseList to 1292 // a global gOmInUseList under the global list lock so these 1293 // will continue to be scanned. 1294 // 1295 // We currently call omFlush() from Threads::remove() _before the thread 1296 // has been excised from the thread list and is no longer a mutator. 1297 // This means that omFlush() can not run concurrently with a safepoint and 1298 // interleave with the scavenge operator. In particular, this ensures that 1299 // the thread's monitors are scanned by a GC safepoint, either via 1300 // Thread::oops_do() (if safepoint happens before omFlush()) or via 1301 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's 1302 // monitors have been transferred to the global in-use list). 1303 1304 void ObjectSynchronizer::omFlush(Thread * Self) { 1305 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1306 Self->omFreeList = NULL; 1307 ObjectMonitor * tail = NULL; 1308 int tally = 0; 1309 if (list != NULL) { 1310 ObjectMonitor * s; 1311 // The thread is going away, the per-thread free monitors 1312 // are freed via set_owner(NULL) 1313 // Link them to tail, which will be linked into the global free list 1314 // gFreeList below, under the gListLock 1315 for (s = list; s != NULL; s = s->FreeNext) { 1316 tally++; 1317 tail = s; 1318 guarantee(s->object() == NULL, "invariant"); 1319 guarantee(!s->is_busy(), "invariant"); 1320 s->set_owner(NULL); // redundant but good hygiene 1321 TEVENT(omFlush - Move one); 1322 } 1323 guarantee(tail != NULL && list != NULL, "invariant"); 1324 } 1325 1326 ObjectMonitor * inUseList = Self->omInUseList; 1327 ObjectMonitor * inUseTail = NULL; 1328 int inUseTally = 0; 1329 if (inUseList != NULL) { 1330 Self->omInUseList = NULL; 1331 ObjectMonitor *cur_om; 1332 // The thread is going away, however the omInUseList inflated 1333 // monitors may still be in-use by other threads. 1334 // Link them to inUseTail, which will be linked into the global in-use list 1335 // gOmInUseList below, under the gListLock 1336 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1337 inUseTail = cur_om; 1338 inUseTally++; 1339 } 1340 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1341 Self->omInUseCount = 0; 1342 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1343 } 1344 1345 Thread::muxAcquire(&gListLock, "omFlush"); 1346 if (tail != NULL) { 1347 tail->FreeNext = gFreeList; 1348 gFreeList = list; 1349 gMonitorFreeCount += tally; 1350 assert(Self->omFreeCount == tally, "free-count off"); 1351 Self->omFreeCount = 0; 1352 } 1353 1354 if (inUseTail != NULL) { 1355 inUseTail->FreeNext = gOmInUseList; 1356 gOmInUseList = inUseList; 1357 gOmInUseCount += inUseTally; 1358 } 1359 1360 Thread::muxRelease(&gListLock); 1361 TEVENT(omFlush); 1362 } 1363 1364 // Fast path code shared by multiple functions 1365 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1366 markOop mark = obj->mark(); 1367 if (mark->has_monitor()) { 1368 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1369 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1370 return mark->monitor(); 1371 } 1372 return ObjectSynchronizer::inflate(Thread::current(), 1373 obj, 1374 inflate_cause_vm_internal); 1375 } 1376 1377 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self, 1378 oop object, 1379 const InflateCause cause) { 1380 1381 // Inflate mutates the heap ... 1382 // Relaxing assertion for bug 6320749. 1383 assert(Universe::verify_in_progress() || 1384 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1385 1386 EventJavaMonitorInflate event; 1387 1388 for (;;) { 1389 const markOop mark = object->mark(); 1390 assert(!mark->has_bias_pattern(), "invariant"); 1391 1392 // The mark can be in one of the following states: 1393 // * Inflated - just return 1394 // * Stack-locked - coerce it to inflated 1395 // * INFLATING - busy wait for conversion to complete 1396 // * Neutral - aggressively inflate the object. 1397 // * BIASED - Illegal. We should never see this 1398 1399 // CASE: inflated 1400 if (mark->has_monitor()) { 1401 ObjectMonitor * inf = mark->monitor(); 1402 assert(inf->header()->is_neutral(), "invariant"); 1403 assert(inf->object() == object, "invariant"); 1404 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1405 return inf; 1406 } 1407 1408 // CASE: inflation in progress - inflating over a stack-lock. 1409 // Some other thread is converting from stack-locked to inflated. 1410 // Only that thread can complete inflation -- other threads must wait. 1411 // The INFLATING value is transient. 1412 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1413 // We could always eliminate polling by parking the thread on some auxiliary list. 1414 if (mark == markOopDesc::INFLATING()) { 1415 TEVENT(Inflate: spin while INFLATING); 1416 ReadStableMark(object); 1417 continue; 1418 } 1419 1420 // CASE: stack-locked 1421 // Could be stack-locked either by this thread or by some other thread. 1422 // 1423 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1424 // to install INFLATING into the mark word. We originally installed INFLATING, 1425 // allocated the objectmonitor, and then finally STed the address of the 1426 // objectmonitor into the mark. This was correct, but artificially lengthened 1427 // the interval in which INFLATED appeared in the mark, thus increasing 1428 // the odds of inflation contention. 1429 // 1430 // We now use per-thread private objectmonitor free lists. 1431 // These list are reprovisioned from the global free list outside the 1432 // critical INFLATING...ST interval. A thread can transfer 1433 // multiple objectmonitors en-mass from the global free list to its local free list. 1434 // This reduces coherency traffic and lock contention on the global free list. 1435 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1436 // before or after the CAS(INFLATING) operation. 1437 // See the comments in omAlloc(). 1438 1439 if (mark->has_locker()) { 1440 ObjectMonitor * m = omAlloc(Self); 1441 // Optimistically prepare the objectmonitor - anticipate successful CAS 1442 // We do this before the CAS in order to minimize the length of time 1443 // in which INFLATING appears in the mark. 1444 m->Recycle(); 1445 m->_Responsible = NULL; 1446 m->_recursions = 0; 1447 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1448 1449 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark); 1450 if (cmp != mark) { 1451 omRelease(Self, m, true); 1452 continue; // Interference -- just retry 1453 } 1454 1455 // We've successfully installed INFLATING (0) into the mark-word. 1456 // This is the only case where 0 will appear in a mark-word. 1457 // Only the singular thread that successfully swings the mark-word 1458 // to 0 can perform (or more precisely, complete) inflation. 1459 // 1460 // Why do we CAS a 0 into the mark-word instead of just CASing the 1461 // mark-word from the stack-locked value directly to the new inflated state? 1462 // Consider what happens when a thread unlocks a stack-locked object. 1463 // It attempts to use CAS to swing the displaced header value from the 1464 // on-stack basiclock back into the object header. Recall also that the 1465 // header value (hashcode, etc) can reside in (a) the object header, or 1466 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1467 // header in an objectMonitor. The inflate() routine must copy the header 1468 // value from the basiclock on the owner's stack to the objectMonitor, all 1469 // the while preserving the hashCode stability invariants. If the owner 1470 // decides to release the lock while the value is 0, the unlock will fail 1471 // and control will eventually pass from slow_exit() to inflate. The owner 1472 // will then spin, waiting for the 0 value to disappear. Put another way, 1473 // the 0 causes the owner to stall if the owner happens to try to 1474 // drop the lock (restoring the header from the basiclock to the object) 1475 // while inflation is in-progress. This protocol avoids races that might 1476 // would otherwise permit hashCode values to change or "flicker" for an object. 1477 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1478 // 0 serves as a "BUSY" inflate-in-progress indicator. 1479 1480 1481 // fetch the displaced mark from the owner's stack. 1482 // The owner can't die or unwind past the lock while our INFLATING 1483 // object is in the mark. Furthermore the owner can't complete 1484 // an unlock on the object, either. 1485 markOop dmw = mark->displaced_mark_helper(); 1486 assert(dmw->is_neutral(), "invariant"); 1487 1488 // Setup monitor fields to proper values -- prepare the monitor 1489 m->set_header(dmw); 1490 1491 // Optimization: if the mark->locker stack address is associated 1492 // with this thread we could simply set m->_owner = Self. 1493 // Note that a thread can inflate an object 1494 // that it has stack-locked -- as might happen in wait() -- directly 1495 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1496 m->set_owner(mark->locker()); 1497 m->set_object(object); 1498 // TODO-FIXME: assert BasicLock->dhw != 0. 1499 1500 // Must preserve store ordering. The monitor state must 1501 // be stable at the time of publishing the monitor address. 1502 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1503 object->release_set_mark(markOopDesc::encode(m)); 1504 1505 // Hopefully the performance counters are allocated on distinct cache lines 1506 // to avoid false sharing on MP systems ... 1507 OM_PERFDATA_OP(Inflations, inc()); 1508 TEVENT(Inflate: overwrite stacklock); 1509 if (log_is_enabled(Debug, monitorinflation)) { 1510 if (object->is_instance()) { 1511 ResourceMark rm; 1512 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1513 p2i(object), p2i(object->mark()), 1514 object->klass()->external_name()); 1515 } 1516 } 1517 if (event.should_commit()) { 1518 post_monitor_inflate_event(event, object, cause); 1519 } 1520 return m; 1521 } 1522 1523 // CASE: neutral 1524 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1525 // If we know we're inflating for entry it's better to inflate by swinging a 1526 // pre-locked objectMonitor pointer into the object header. A successful 1527 // CAS inflates the object *and* confers ownership to the inflating thread. 1528 // In the current implementation we use a 2-step mechanism where we CAS() 1529 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1530 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1531 // would be useful. 1532 1533 assert(mark->is_neutral(), "invariant"); 1534 ObjectMonitor * m = omAlloc(Self); 1535 // prepare m for installation - set monitor to initial state 1536 m->Recycle(); 1537 m->set_header(mark); 1538 m->set_owner(NULL); 1539 m->set_object(object); 1540 m->_recursions = 0; 1541 m->_Responsible = NULL; 1542 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1543 1544 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) { 1545 m->set_object(NULL); 1546 m->set_owner(NULL); 1547 m->Recycle(); 1548 omRelease(Self, m, true); 1549 m = NULL; 1550 continue; 1551 // interference - the markword changed - just retry. 1552 // The state-transitions are one-way, so there's no chance of 1553 // live-lock -- "Inflated" is an absorbing state. 1554 } 1555 1556 // Hopefully the performance counters are allocated on distinct 1557 // cache lines to avoid false sharing on MP systems ... 1558 OM_PERFDATA_OP(Inflations, inc()); 1559 TEVENT(Inflate: overwrite neutral); 1560 if (log_is_enabled(Debug, monitorinflation)) { 1561 if (object->is_instance()) { 1562 ResourceMark rm; 1563 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1564 p2i(object), p2i(object->mark()), 1565 object->klass()->external_name()); 1566 } 1567 } 1568 if (event.should_commit()) { 1569 post_monitor_inflate_event(event, object, cause); 1570 } 1571 return m; 1572 } 1573 } 1574 1575 1576 // Deflate_idle_monitors() is called at all safepoints, immediately 1577 // after all mutators are stopped, but before any objects have moved. 1578 // It traverses the list of known monitors, deflating where possible. 1579 // The scavenged monitor are returned to the monitor free list. 1580 // 1581 // Beware that we scavenge at *every* stop-the-world point. 1582 // Having a large number of monitors in-circulation negatively 1583 // impacts the performance of some applications (e.g., PointBase). 1584 // Broadly, we want to minimize the # of monitors in circulation. 1585 // 1586 // We have added a flag, MonitorInUseLists, which creates a list 1587 // of active monitors for each thread. deflate_idle_monitors() 1588 // only scans the per-thread in-use lists. omAlloc() puts all 1589 // assigned monitors on the per-thread list. deflate_idle_monitors() 1590 // returns the non-busy monitors to the global free list. 1591 // When a thread dies, omFlush() adds the list of active monitors for 1592 // that thread to a global gOmInUseList acquiring the 1593 // global list lock. deflate_idle_monitors() acquires the global 1594 // list lock to scan for non-busy monitors to the global free list. 1595 // An alternative could have used a single global in-use list. The 1596 // downside would have been the additional cost of acquiring the global list lock 1597 // for every omAlloc(). 1598 // 1599 // Perversely, the heap size -- and thus the STW safepoint rate -- 1600 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1601 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1602 // This is an unfortunate aspect of this design. 1603 1604 enum ManifestConstants { 1605 ClearResponsibleAtSTW = 0 1606 }; 1607 1608 // Deflate a single monitor if not in-use 1609 // Return true if deflated, false if in-use 1610 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1611 ObjectMonitor** freeHeadp, 1612 ObjectMonitor** freeTailp) { 1613 bool deflated; 1614 // Normal case ... The monitor is associated with obj. 1615 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1616 guarantee(mid == obj->mark()->monitor(), "invariant"); 1617 guarantee(mid->header()->is_neutral(), "invariant"); 1618 1619 if (mid->is_busy()) { 1620 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1621 deflated = false; 1622 } else { 1623 // Deflate the monitor if it is no longer being used 1624 // It's idle - scavenge and return to the global free list 1625 // plain old deflation ... 1626 TEVENT(deflate_idle_monitors - scavenge1); 1627 if (log_is_enabled(Debug, monitorinflation)) { 1628 if (obj->is_instance()) { 1629 ResourceMark rm; 1630 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , " 1631 "mark " INTPTR_FORMAT " , type %s", 1632 p2i(obj), p2i(obj->mark()), 1633 obj->klass()->external_name()); 1634 } 1635 } 1636 1637 // Restore the header back to obj 1638 obj->release_set_mark(mid->header()); 1639 mid->clear(); 1640 1641 assert(mid->object() == NULL, "invariant"); 1642 1643 // Move the object to the working free list defined by freeHeadp, freeTailp 1644 if (*freeHeadp == NULL) *freeHeadp = mid; 1645 if (*freeTailp != NULL) { 1646 ObjectMonitor * prevtail = *freeTailp; 1647 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1648 prevtail->FreeNext = mid; 1649 } 1650 *freeTailp = mid; 1651 deflated = true; 1652 } 1653 return deflated; 1654 } 1655 1656 // Walk a given monitor list, and deflate idle monitors 1657 // The given list could be a per-thread list or a global list 1658 // Caller acquires gListLock. 1659 // 1660 // In the case of parallel processing of thread local monitor lists, 1661 // work is done by Threads::parallel_threads_do() which ensures that 1662 // each Java thread is processed by exactly one worker thread, and 1663 // thus avoid conflicts that would arise when worker threads would 1664 // process the same monitor lists concurrently. 1665 // 1666 // See also ParallelSPCleanupTask and 1667 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and 1668 // Threads::parallel_java_threads_do() in thread.cpp. 1669 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1670 ObjectMonitor** freeHeadp, 1671 ObjectMonitor** freeTailp) { 1672 ObjectMonitor* mid; 1673 ObjectMonitor* next; 1674 ObjectMonitor* cur_mid_in_use = NULL; 1675 int deflated_count = 0; 1676 1677 for (mid = *listHeadp; mid != NULL;) { 1678 oop obj = (oop) mid->object(); 1679 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1680 // if deflate_monitor succeeded, 1681 // extract from per-thread in-use list 1682 if (mid == *listHeadp) { 1683 *listHeadp = mid->FreeNext; 1684 } else if (cur_mid_in_use != NULL) { 1685 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1686 } 1687 next = mid->FreeNext; 1688 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1689 mid = next; 1690 deflated_count++; 1691 } else { 1692 cur_mid_in_use = mid; 1693 mid = mid->FreeNext; 1694 } 1695 } 1696 return deflated_count; 1697 } 1698 1699 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1700 counters->nInuse = 0; // currently associated with objects 1701 counters->nInCirculation = 0; // extant 1702 counters->nScavenged = 0; // reclaimed 1703 } 1704 1705 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 1706 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1707 bool deflated = false; 1708 1709 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1710 ObjectMonitor * freeTailp = NULL; 1711 1712 TEVENT(deflate_idle_monitors); 1713 // Prevent omFlush from changing mids in Thread dtor's during deflation 1714 // And in case the vm thread is acquiring a lock during a safepoint 1715 // See e.g. 6320749 1716 Thread::muxAcquire(&gListLock, "scavenge - return"); 1717 1718 if (MonitorInUseLists) { 1719 // Note: the thread-local monitors lists get deflated in 1720 // a separate pass. See deflate_thread_local_monitors(). 1721 1722 // For moribund threads, scan gOmInUseList 1723 if (gOmInUseList) { 1724 counters->nInCirculation += gOmInUseCount; 1725 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1726 gOmInUseCount -= deflated_count; 1727 counters->nScavenged += deflated_count; 1728 counters->nInuse += gOmInUseCount; 1729 } 1730 1731 } else { 1732 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 1733 for (; block != NULL; block = next(block)) { 1734 // Iterate over all extant monitors - Scavenge all idle monitors. 1735 assert(block->object() == CHAINMARKER, "must be a block header"); 1736 counters->nInCirculation += _BLOCKSIZE; 1737 for (int i = 1; i < _BLOCKSIZE; i++) { 1738 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1739 oop obj = (oop)mid->object(); 1740 1741 if (obj == NULL) { 1742 // The monitor is not associated with an object. 1743 // The monitor should either be a thread-specific private 1744 // free list or the global free list. 1745 // obj == NULL IMPLIES mid->is_busy() == 0 1746 guarantee(!mid->is_busy(), "invariant"); 1747 continue; 1748 } 1749 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1750 1751 if (deflated) { 1752 mid->FreeNext = NULL; 1753 counters->nScavenged++; 1754 } else { 1755 counters->nInuse++; 1756 } 1757 } 1758 } 1759 } 1760 1761 // Move the scavenged monitors back to the global free list. 1762 if (freeHeadp != NULL) { 1763 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant"); 1764 assert(freeTailp->FreeNext == NULL, "invariant"); 1765 // constant-time list splice - prepend scavenged segment to gFreeList 1766 freeTailp->FreeNext = gFreeList; 1767 gFreeList = freeHeadp; 1768 } 1769 Thread::muxRelease(&gListLock); 1770 1771 } 1772 1773 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1774 gMonitorFreeCount += counters->nScavenged; 1775 1776 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1777 1778 if (ObjectMonitor::Knob_Verbose) { 1779 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d " 1780 "ForceMonitorScavenge=%d : pop=%d free=%d", 1781 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge, 1782 gMonitorPopulation, gMonitorFreeCount); 1783 tty->flush(); 1784 } 1785 1786 ForceMonitorScavenge = 0; // Reset 1787 1788 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged)); 1789 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation)); 1790 1791 // TODO: Add objectMonitor leak detection. 1792 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1793 GVars.stwRandom = os::random(); 1794 GVars.stwCycle++; 1795 } 1796 1797 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 1798 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1799 if (!MonitorInUseLists) return; 1800 1801 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1802 ObjectMonitor * freeTailp = NULL; 1803 1804 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp); 1805 1806 Thread::muxAcquire(&gListLock, "scavenge - return"); 1807 1808 // Adjust counters 1809 counters->nInCirculation += thread->omInUseCount; 1810 thread->omInUseCount -= deflated_count; 1811 if (ObjectMonitor::Knob_VerifyInUse) { 1812 verifyInUse(thread); 1813 } 1814 counters->nScavenged += deflated_count; 1815 counters->nInuse += thread->omInUseCount; 1816 1817 // Move the scavenged monitors back to the global free list. 1818 if (freeHeadp != NULL) { 1819 guarantee(freeTailp != NULL && deflated_count > 0, "invariant"); 1820 assert(freeTailp->FreeNext == NULL, "invariant"); 1821 1822 // constant-time list splice - prepend scavenged segment to gFreeList 1823 freeTailp->FreeNext = gFreeList; 1824 gFreeList = freeHeadp; 1825 } 1826 Thread::muxRelease(&gListLock); 1827 } 1828 1829 // Monitor cleanup on JavaThread::exit 1830 1831 // Iterate through monitor cache and attempt to release thread's monitors 1832 // Gives up on a particular monitor if an exception occurs, but continues 1833 // the overall iteration, swallowing the exception. 1834 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1835 private: 1836 TRAPS; 1837 1838 public: 1839 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1840 void do_monitor(ObjectMonitor* mid) { 1841 if (mid->owner() == THREAD) { 1842 if (ObjectMonitor::Knob_VerifyMatch != 0) { 1843 ResourceMark rm; 1844 Handle obj(THREAD, (oop) mid->object()); 1845 tty->print("INFO: unexpected locked object:"); 1846 javaVFrame::print_locked_object_class_name(tty, obj, "locked"); 1847 fatal("exiting JavaThread=" INTPTR_FORMAT 1848 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT, 1849 p2i(THREAD), p2i(mid)); 1850 } 1851 (void)mid->complete_exit(CHECK); 1852 } 1853 } 1854 }; 1855 1856 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1857 // ignored. This is meant to be called during JNI thread detach which assumes 1858 // all remaining monitors are heavyweight. All exceptions are swallowed. 1859 // Scanning the extant monitor list can be time consuming. 1860 // A simple optimization is to add a per-thread flag that indicates a thread 1861 // called jni_monitorenter() during its lifetime. 1862 // 1863 // Instead of No_Savepoint_Verifier it might be cheaper to 1864 // use an idiom of the form: 1865 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1866 // <code that must not run at safepoint> 1867 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1868 // Since the tests are extremely cheap we could leave them enabled 1869 // for normal product builds. 1870 1871 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1872 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1873 NoSafepointVerifier nsv; 1874 ReleaseJavaMonitorsClosure rjmc(THREAD); 1875 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1876 ObjectSynchronizer::monitors_iterate(&rjmc); 1877 Thread::muxRelease(&gListLock); 1878 THREAD->clear_pending_exception(); 1879 } 1880 1881 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1882 switch (cause) { 1883 case inflate_cause_vm_internal: return "VM Internal"; 1884 case inflate_cause_monitor_enter: return "Monitor Enter"; 1885 case inflate_cause_wait: return "Monitor Wait"; 1886 case inflate_cause_notify: return "Monitor Notify"; 1887 case inflate_cause_hash_code: return "Monitor Hash Code"; 1888 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1889 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1890 default: 1891 ShouldNotReachHere(); 1892 } 1893 return "Unknown"; 1894 } 1895 1896 static void post_monitor_inflate_event(EventJavaMonitorInflate& event, 1897 const oop obj, 1898 const ObjectSynchronizer::InflateCause cause) { 1899 #if INCLUDE_TRACE 1900 assert(event.should_commit(), "check outside"); 1901 event.set_monitorClass(obj->klass()); 1902 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj); 1903 event.set_cause((u1)cause); 1904 event.commit(); 1905 #endif 1906 } 1907 1908 //------------------------------------------------------------------------------ 1909 // Debugging code 1910 1911 void ObjectSynchronizer::sanity_checks(const bool verbose, 1912 const uint cache_line_size, 1913 int *error_cnt_ptr, 1914 int *warning_cnt_ptr) { 1915 u_char *addr_begin = (u_char*)&GVars; 1916 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1917 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1918 1919 if (verbose) { 1920 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1921 sizeof(SharedGlobals)); 1922 } 1923 1924 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1925 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1926 1927 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1928 if (verbose) { 1929 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1930 } 1931 1932 if (cache_line_size != 0) { 1933 // We were able to determine the L1 data cache line size so 1934 // do some cache line specific sanity checks 1935 1936 if (offset_stwRandom < cache_line_size) { 1937 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1938 "to the struct beginning than a cache line which permits " 1939 "false sharing."); 1940 (*warning_cnt_ptr)++; 1941 } 1942 1943 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1944 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1945 "SharedGlobals.hcSequence fields are closer than a cache " 1946 "line which permits false sharing."); 1947 (*warning_cnt_ptr)++; 1948 } 1949 1950 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1951 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1952 "to the struct end than a cache line which permits false " 1953 "sharing."); 1954 (*warning_cnt_ptr)++; 1955 } 1956 } 1957 } 1958 1959 #ifndef PRODUCT 1960 1961 // Check if monitor belongs to the monitor cache 1962 // The list is grow-only so it's *relatively* safe to traverse 1963 // the list of extant blocks without taking a lock. 1964 1965 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1966 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 1967 while (block != NULL) { 1968 assert(block->object() == CHAINMARKER, "must be a block header"); 1969 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 1970 address mon = (address)monitor; 1971 address blk = (address)block; 1972 size_t diff = mon - blk; 1973 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned"); 1974 return 1; 1975 } 1976 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 1977 } 1978 return 0; 1979 } 1980 1981 #endif