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 && 242 Atomic::cmpxchg(Self, &(m->_owner), (void*)NULL) == NULL) { 243 assert(m->_recursions == 0, "invariant"); 244 assert(m->_owner == Self, "invariant"); 245 return true; 246 } 247 } 248 249 // Note that we could inflate in quick_enter. 250 // This is likely a useful optimization 251 // Critically, in quick_enter() we must not: 252 // -- perform bias revocation, or 253 // -- block indefinitely, or 254 // -- reach a safepoint 255 256 return false; // revert to slow-path 257 } 258 259 // ----------------------------------------------------------------------------- 260 // Fast Monitor Enter/Exit 261 // This the fast monitor enter. The interpreter and compiler use 262 // some assembly copies of this code. Make sure update those code 263 // if the following function is changed. The implementation is 264 // extremely sensitive to race condition. Be careful. 265 266 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 267 bool attempt_rebias, TRAPS) { 268 if (UseBiasedLocking) { 269 if (!SafepointSynchronize::is_at_safepoint()) { 270 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 271 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 272 return; 273 } 274 } else { 275 assert(!attempt_rebias, "can not rebias toward VM thread"); 276 BiasedLocking::revoke_at_safepoint(obj); 277 } 278 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 279 } 280 281 slow_enter(obj, lock, THREAD); 282 } 283 284 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 285 markOop mark = object->mark(); 286 // We cannot check for Biased Locking if we are racing an inflation. 287 assert(mark == markOopDesc::INFLATING() || 288 !mark->has_bias_pattern(), "should not see bias pattern here"); 289 290 markOop dhw = lock->displaced_header(); 291 if (dhw == NULL) { 292 // If the displaced header is NULL, then this exit matches up with 293 // a recursive enter. No real work to do here except for diagnostics. 294 #ifndef PRODUCT 295 if (mark != markOopDesc::INFLATING()) { 296 // Only do diagnostics if we are not racing an inflation. Simply 297 // exiting a recursive enter of a Java Monitor that is being 298 // inflated is safe; see the has_monitor() comment below. 299 assert(!mark->is_neutral(), "invariant"); 300 assert(!mark->has_locker() || 301 THREAD->is_lock_owned((address)mark->locker()), "invariant"); 302 if (mark->has_monitor()) { 303 // The BasicLock's displaced_header is marked as a recursive 304 // enter and we have an inflated Java Monitor (ObjectMonitor). 305 // This is a special case where the Java Monitor was inflated 306 // after this thread entered the stack-lock recursively. When a 307 // Java Monitor is inflated, we cannot safely walk the Java 308 // Monitor owner's stack and update the BasicLocks because a 309 // Java Monitor can be asynchronously inflated by a thread that 310 // does not own the Java Monitor. 311 ObjectMonitor * m = mark->monitor(); 312 assert(((oop)(m->object()))->mark() == mark, "invariant"); 313 assert(m->is_entered(THREAD), "invariant"); 314 } 315 } 316 #endif 317 return; 318 } 319 320 if (mark == (markOop) lock) { 321 // If the object is stack-locked by the current thread, try to 322 // swing the displaced header from the BasicLock back to the mark. 323 assert(dhw->is_neutral(), "invariant"); 324 if (object->cas_set_mark(dhw, mark) == mark) { 325 TEVENT(fast_exit: release stack-lock); 326 return; 327 } 328 } 329 330 // We have to take the slow-path of possible inflation and then exit. 331 ObjectSynchronizer::inflate(THREAD, 332 object, 333 inflate_cause_vm_internal)->exit(true, THREAD); 334 } 335 336 // ----------------------------------------------------------------------------- 337 // Interpreter/Compiler Slow Case 338 // This routine is used to handle interpreter/compiler slow case 339 // We don't need to use fast path here, because it must have been 340 // failed in the interpreter/compiler code. 341 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 342 markOop mark = obj->mark(); 343 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 344 345 if (mark->is_neutral()) { 346 // Anticipate successful CAS -- the ST of the displaced mark must 347 // be visible <= the ST performed by the CAS. 348 lock->set_displaced_header(mark); 349 if (mark == obj()->cas_set_mark((markOop) lock, mark)) { 350 TEVENT(slow_enter: release stacklock); 351 return; 352 } 353 // Fall through to inflate() ... 354 } else if (mark->has_locker() && 355 THREAD->is_lock_owned((address)mark->locker())) { 356 assert(lock != mark->locker(), "must not re-lock the same lock"); 357 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 358 lock->set_displaced_header(NULL); 359 return; 360 } 361 362 // The object header will never be displaced to this lock, 363 // so it does not matter what the value is, except that it 364 // must be non-zero to avoid looking like a re-entrant lock, 365 // and must not look locked either. 366 lock->set_displaced_header(markOopDesc::unused_mark()); 367 ObjectSynchronizer::inflate(THREAD, 368 obj(), 369 inflate_cause_monitor_enter)->enter(THREAD); 370 } 371 372 // This routine is used to handle interpreter/compiler slow case 373 // We don't need to use fast path here, because it must have 374 // failed in the interpreter/compiler code. Simply use the heavy 375 // weight monitor should be ok, unless someone find otherwise. 376 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 377 fast_exit(object, lock, THREAD); 378 } 379 380 // ----------------------------------------------------------------------------- 381 // Class Loader support to workaround deadlocks on the class loader lock objects 382 // Also used by GC 383 // complete_exit()/reenter() are used to wait on a nested lock 384 // i.e. to give up an outer lock completely and then re-enter 385 // Used when holding nested locks - lock acquisition order: lock1 then lock2 386 // 1) complete_exit lock1 - saving recursion count 387 // 2) wait on lock2 388 // 3) when notified on lock2, unlock lock2 389 // 4) reenter lock1 with original recursion count 390 // 5) lock lock2 391 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 392 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 393 TEVENT(complete_exit); 394 if (UseBiasedLocking) { 395 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 396 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 397 } 398 399 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 400 obj(), 401 inflate_cause_vm_internal); 402 403 return monitor->complete_exit(THREAD); 404 } 405 406 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 407 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 408 TEVENT(reenter); 409 if (UseBiasedLocking) { 410 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 411 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 412 } 413 414 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 415 obj(), 416 inflate_cause_vm_internal); 417 418 monitor->reenter(recursion, THREAD); 419 } 420 // ----------------------------------------------------------------------------- 421 // JNI locks on java objects 422 // NOTE: must use heavy weight monitor to handle jni monitor enter 423 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 424 // the current locking is from JNI instead of Java code 425 TEVENT(jni_enter); 426 if (UseBiasedLocking) { 427 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 428 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 429 } 430 THREAD->set_current_pending_monitor_is_from_java(false); 431 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD); 432 THREAD->set_current_pending_monitor_is_from_java(true); 433 } 434 435 // NOTE: must use heavy weight monitor to handle jni monitor exit 436 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 437 TEVENT(jni_exit); 438 if (UseBiasedLocking) { 439 Handle h_obj(THREAD, obj); 440 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 441 obj = h_obj(); 442 } 443 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 444 445 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 446 obj, 447 inflate_cause_jni_exit); 448 // If this thread has locked the object, exit the monitor. Note: can't use 449 // monitor->check(CHECK); must exit even if an exception is pending. 450 if (monitor->check(THREAD)) { 451 monitor->exit(true, THREAD); 452 } 453 } 454 455 // ----------------------------------------------------------------------------- 456 // Internal VM locks on java objects 457 // standard constructor, allows locking failures 458 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 459 _dolock = doLock; 460 _thread = thread; 461 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 462 _obj = obj; 463 464 if (_dolock) { 465 TEVENT(ObjectLocker); 466 467 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 468 } 469 } 470 471 ObjectLocker::~ObjectLocker() { 472 if (_dolock) { 473 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 474 } 475 } 476 477 478 // ----------------------------------------------------------------------------- 479 // Wait/Notify/NotifyAll 480 // NOTE: must use heavy weight monitor to handle wait() 481 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 482 if (UseBiasedLocking) { 483 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 484 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 485 } 486 if (millis < 0) { 487 TEVENT(wait - throw IAX); 488 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 489 } 490 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 491 obj(), 492 inflate_cause_wait); 493 494 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 495 monitor->wait(millis, true, THREAD); 496 497 // This dummy call is in place to get around dtrace bug 6254741. Once 498 // that's fixed we can uncomment the following line, remove the call 499 // and change this function back into a "void" func. 500 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 501 return dtrace_waited_probe(monitor, obj, THREAD); 502 } 503 504 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 505 if (UseBiasedLocking) { 506 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 507 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 508 } 509 if (millis < 0) { 510 TEVENT(wait - throw IAX); 511 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 512 } 513 ObjectSynchronizer::inflate(THREAD, 514 obj(), 515 inflate_cause_wait)->wait(millis, false, THREAD); 516 } 517 518 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 519 if (UseBiasedLocking) { 520 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 521 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 522 } 523 524 markOop mark = obj->mark(); 525 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 526 return; 527 } 528 ObjectSynchronizer::inflate(THREAD, 529 obj(), 530 inflate_cause_notify)->notify(THREAD); 531 } 532 533 // NOTE: see comment of notify() 534 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 535 if (UseBiasedLocking) { 536 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 537 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 538 } 539 540 markOop mark = obj->mark(); 541 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 542 return; 543 } 544 ObjectSynchronizer::inflate(THREAD, 545 obj(), 546 inflate_cause_notify)->notifyAll(THREAD); 547 } 548 549 // ----------------------------------------------------------------------------- 550 // Hash Code handling 551 // 552 // Performance concern: 553 // OrderAccess::storestore() calls release() which at one time stored 0 554 // into the global volatile OrderAccess::dummy variable. This store was 555 // unnecessary for correctness. Many threads storing into a common location 556 // causes considerable cache migration or "sloshing" on large SMP systems. 557 // As such, I avoided using OrderAccess::storestore(). In some cases 558 // OrderAccess::fence() -- which incurs local latency on the executing 559 // processor -- is a better choice as it scales on SMP systems. 560 // 561 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 562 // a discussion of coherency costs. Note that all our current reference 563 // platforms provide strong ST-ST order, so the issue is moot on IA32, 564 // x64, and SPARC. 565 // 566 // As a general policy we use "volatile" to control compiler-based reordering 567 // and explicit fences (barriers) to control for architectural reordering 568 // performed by the CPU(s) or platform. 569 570 struct SharedGlobals { 571 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE]; 572 // These are highly shared mostly-read variables. 573 // To avoid false-sharing they need to be the sole occupants of a cache line. 574 volatile int stwRandom; 575 volatile int stwCycle; 576 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 577 // Hot RW variable -- Sequester to avoid false-sharing 578 volatile int hcSequence; 579 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int)); 580 }; 581 582 static SharedGlobals GVars; 583 static int MonitorScavengeThreshold = 1000000; 584 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 585 586 static markOop ReadStableMark(oop obj) { 587 markOop mark = obj->mark(); 588 if (!mark->is_being_inflated()) { 589 return mark; // normal fast-path return 590 } 591 592 int its = 0; 593 for (;;) { 594 markOop mark = obj->mark(); 595 if (!mark->is_being_inflated()) { 596 return mark; // normal fast-path return 597 } 598 599 // The object is being inflated by some other thread. 600 // The caller of ReadStableMark() must wait for inflation to complete. 601 // Avoid live-lock 602 // TODO: consider calling SafepointSynchronize::do_call_back() while 603 // spinning to see if there's a safepoint pending. If so, immediately 604 // yielding or blocking would be appropriate. Avoid spinning while 605 // there is a safepoint pending. 606 // TODO: add inflation contention performance counters. 607 // TODO: restrict the aggregate number of spinners. 608 609 ++its; 610 if (its > 10000 || !os::is_MP()) { 611 if (its & 1) { 612 os::naked_yield(); 613 TEVENT(Inflate: INFLATING - yield); 614 } else { 615 // Note that the following code attenuates the livelock problem but is not 616 // a complete remedy. A more complete solution would require that the inflating 617 // thread hold the associated inflation lock. The following code simply restricts 618 // the number of spinners to at most one. We'll have N-2 threads blocked 619 // on the inflationlock, 1 thread holding the inflation lock and using 620 // a yield/park strategy, and 1 thread in the midst of inflation. 621 // A more refined approach would be to change the encoding of INFLATING 622 // to allow encapsulation of a native thread pointer. Threads waiting for 623 // inflation to complete would use CAS to push themselves onto a singly linked 624 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 625 // and calling park(). When inflation was complete the thread that accomplished inflation 626 // would detach the list and set the markword to inflated with a single CAS and 627 // then for each thread on the list, set the flag and unpark() the thread. 628 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 629 // wakes at most one thread whereas we need to wake the entire list. 630 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 631 int YieldThenBlock = 0; 632 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 633 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 634 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 635 while (obj->mark() == markOopDesc::INFLATING()) { 636 // Beware: NakedYield() is advisory and has almost no effect on some platforms 637 // so we periodically call Self->_ParkEvent->park(1). 638 // We use a mixed spin/yield/block mechanism. 639 if ((YieldThenBlock++) >= 16) { 640 Thread::current()->_ParkEvent->park(1); 641 } else { 642 os::naked_yield(); 643 } 644 } 645 Thread::muxRelease(gInflationLocks + ix); 646 TEVENT(Inflate: INFLATING - yield/park); 647 } 648 } else { 649 SpinPause(); // SMP-polite spinning 650 } 651 } 652 } 653 654 // hashCode() generation : 655 // 656 // Possibilities: 657 // * MD5Digest of {obj,stwRandom} 658 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 659 // * A DES- or AES-style SBox[] mechanism 660 // * One of the Phi-based schemes, such as: 661 // 2654435761 = 2^32 * Phi (golden ratio) 662 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 663 // * A variation of Marsaglia's shift-xor RNG scheme. 664 // * (obj ^ stwRandom) is appealing, but can result 665 // in undesirable regularity in the hashCode values of adjacent objects 666 // (objects allocated back-to-back, in particular). This could potentially 667 // result in hashtable collisions and reduced hashtable efficiency. 668 // There are simple ways to "diffuse" the middle address bits over the 669 // generated hashCode values: 670 671 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 672 intptr_t value = 0; 673 if (hashCode == 0) { 674 // This form uses global Park-Miller RNG. 675 // On MP system we'll have lots of RW access to a global, so the 676 // mechanism induces lots of coherency traffic. 677 value = os::random(); 678 } else if (hashCode == 1) { 679 // This variation has the property of being stable (idempotent) 680 // between STW operations. This can be useful in some of the 1-0 681 // synchronization schemes. 682 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 683 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 684 } else if (hashCode == 2) { 685 value = 1; // for sensitivity testing 686 } else if (hashCode == 3) { 687 value = ++GVars.hcSequence; 688 } else if (hashCode == 4) { 689 value = cast_from_oop<intptr_t>(obj); 690 } else { 691 // Marsaglia's xor-shift scheme with thread-specific state 692 // This is probably the best overall implementation -- we'll 693 // likely make this the default in future releases. 694 unsigned t = Self->_hashStateX; 695 t ^= (t << 11); 696 Self->_hashStateX = Self->_hashStateY; 697 Self->_hashStateY = Self->_hashStateZ; 698 Self->_hashStateZ = Self->_hashStateW; 699 unsigned v = Self->_hashStateW; 700 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 701 Self->_hashStateW = v; 702 value = v; 703 } 704 705 value &= markOopDesc::hash_mask; 706 if (value == 0) value = 0xBAD; 707 assert(value != markOopDesc::no_hash, "invariant"); 708 TEVENT(hashCode: GENERATE); 709 return value; 710 } 711 712 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 713 if (UseBiasedLocking) { 714 // NOTE: many places throughout the JVM do not expect a safepoint 715 // to be taken here, in particular most operations on perm gen 716 // objects. However, we only ever bias Java instances and all of 717 // the call sites of identity_hash that might revoke biases have 718 // been checked to make sure they can handle a safepoint. The 719 // added check of the bias pattern is to avoid useless calls to 720 // thread-local storage. 721 if (obj->mark()->has_bias_pattern()) { 722 // Handle for oop obj in case of STW safepoint 723 Handle hobj(Self, obj); 724 // Relaxing assertion for bug 6320749. 725 assert(Universe::verify_in_progress() || 726 !SafepointSynchronize::is_at_safepoint(), 727 "biases should not be seen by VM thread here"); 728 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 729 obj = hobj(); 730 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 731 } 732 } 733 734 // hashCode() is a heap mutator ... 735 // Relaxing assertion for bug 6320749. 736 assert(Universe::verify_in_progress() || DumpSharedSpaces || 737 !SafepointSynchronize::is_at_safepoint(), "invariant"); 738 assert(Universe::verify_in_progress() || DumpSharedSpaces || 739 Self->is_Java_thread() , "invariant"); 740 assert(Universe::verify_in_progress() || DumpSharedSpaces || 741 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 742 743 ObjectMonitor* monitor = NULL; 744 markOop temp, test; 745 intptr_t hash; 746 markOop mark = ReadStableMark(obj); 747 748 // object should remain ineligible for biased locking 749 assert(!mark->has_bias_pattern(), "invariant"); 750 751 if (mark->is_neutral()) { 752 hash = mark->hash(); // this is a normal header 753 if (hash) { // if it has hash, just return it 754 return hash; 755 } 756 hash = get_next_hash(Self, obj); // allocate a new hash code 757 temp = mark->copy_set_hash(hash); // merge the hash code into header 758 // use (machine word version) atomic operation to install the hash 759 test = obj->cas_set_mark(temp, mark); 760 if (test == mark) { 761 return hash; 762 } 763 // If atomic operation failed, we must inflate the header 764 // into heavy weight monitor. We could add more code here 765 // for fast path, but it does not worth the complexity. 766 } else if (mark->has_monitor()) { 767 monitor = mark->monitor(); 768 temp = monitor->header(); 769 assert(temp->is_neutral(), "invariant"); 770 hash = temp->hash(); 771 if (hash) { 772 return hash; 773 } 774 // Skip to the following code to reduce code size 775 } else if (Self->is_lock_owned((address)mark->locker())) { 776 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 777 assert(temp->is_neutral(), "invariant"); 778 hash = temp->hash(); // by current thread, check if the displaced 779 if (hash) { // header contains hash code 780 return hash; 781 } 782 // WARNING: 783 // The displaced header is strictly immutable. 784 // It can NOT be changed in ANY cases. So we have 785 // to inflate the header into heavyweight monitor 786 // even the current thread owns the lock. The reason 787 // is the BasicLock (stack slot) will be asynchronously 788 // read by other threads during the inflate() function. 789 // Any change to stack may not propagate to other threads 790 // correctly. 791 } 792 793 // Inflate the monitor to set hash code 794 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code); 795 // Load displaced header and check it has hash code 796 mark = monitor->header(); 797 assert(mark->is_neutral(), "invariant"); 798 hash = mark->hash(); 799 if (hash == 0) { 800 hash = get_next_hash(Self, obj); 801 temp = mark->copy_set_hash(hash); // merge hash code into header 802 assert(temp->is_neutral(), "invariant"); 803 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark); 804 if (test != mark) { 805 // The only update to the header in the monitor (outside GC) 806 // is install the hash code. If someone add new usage of 807 // displaced header, please update this code 808 hash = test->hash(); 809 assert(test->is_neutral(), "invariant"); 810 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 811 } 812 } 813 // We finally get the hash 814 return hash; 815 } 816 817 // Deprecated -- use FastHashCode() instead. 818 819 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 820 return FastHashCode(Thread::current(), obj()); 821 } 822 823 824 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 825 Handle h_obj) { 826 if (UseBiasedLocking) { 827 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 828 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 829 } 830 831 assert(thread == JavaThread::current(), "Can only be called on current thread"); 832 oop obj = h_obj(); 833 834 markOop mark = ReadStableMark(obj); 835 836 // Uncontended case, header points to stack 837 if (mark->has_locker()) { 838 return thread->is_lock_owned((address)mark->locker()); 839 } 840 // Contended case, header points to ObjectMonitor (tagged pointer) 841 if (mark->has_monitor()) { 842 ObjectMonitor* monitor = mark->monitor(); 843 return monitor->is_entered(thread) != 0; 844 } 845 // Unlocked case, header in place 846 assert(mark->is_neutral(), "sanity check"); 847 return false; 848 } 849 850 // Be aware of this method could revoke bias of the lock object. 851 // This method queries the ownership of the lock handle specified by 'h_obj'. 852 // If the current thread owns the lock, it returns owner_self. If no 853 // thread owns the lock, it returns owner_none. Otherwise, it will return 854 // owner_other. 855 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 856 (JavaThread *self, Handle h_obj) { 857 // The caller must beware this method can revoke bias, and 858 // revocation can result in a safepoint. 859 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 860 assert(self->thread_state() != _thread_blocked, "invariant"); 861 862 // Possible mark states: neutral, biased, stack-locked, inflated 863 864 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 865 // CASE: biased 866 BiasedLocking::revoke_and_rebias(h_obj, false, self); 867 assert(!h_obj->mark()->has_bias_pattern(), 868 "biases should be revoked by now"); 869 } 870 871 assert(self == JavaThread::current(), "Can only be called on current thread"); 872 oop obj = h_obj(); 873 markOop mark = ReadStableMark(obj); 874 875 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 876 if (mark->has_locker()) { 877 return self->is_lock_owned((address)mark->locker()) ? 878 owner_self : owner_other; 879 } 880 881 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 882 // The Object:ObjectMonitor relationship is stable as long as we're 883 // not at a safepoint. 884 if (mark->has_monitor()) { 885 void * owner = mark->monitor()->_owner; 886 if (owner == NULL) return owner_none; 887 return (owner == self || 888 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 889 } 890 891 // CASE: neutral 892 assert(mark->is_neutral(), "sanity check"); 893 return owner_none; // it's unlocked 894 } 895 896 // FIXME: jvmti should call this 897 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 898 if (UseBiasedLocking) { 899 if (SafepointSynchronize::is_at_safepoint()) { 900 BiasedLocking::revoke_at_safepoint(h_obj); 901 } else { 902 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 903 } 904 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 905 } 906 907 oop obj = h_obj(); 908 address owner = NULL; 909 910 markOop mark = ReadStableMark(obj); 911 912 // Uncontended case, header points to stack 913 if (mark->has_locker()) { 914 owner = (address) mark->locker(); 915 } 916 917 // Contended case, header points to ObjectMonitor (tagged pointer) 918 if (mark->has_monitor()) { 919 ObjectMonitor* monitor = mark->monitor(); 920 assert(monitor != NULL, "monitor should be non-null"); 921 owner = (address) monitor->owner(); 922 } 923 924 if (owner != NULL) { 925 // owning_thread_from_monitor_owner() may also return NULL here 926 return Threads::owning_thread_from_monitor_owner(t_list, owner); 927 } 928 929 // Unlocked case, header in place 930 // Cannot have assertion since this object may have been 931 // locked by another thread when reaching here. 932 // assert(mark->is_neutral(), "sanity check"); 933 934 return NULL; 935 } 936 937 // Visitors ... 938 939 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 940 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 941 while (block != NULL) { 942 assert(block->object() == CHAINMARKER, "must be a block header"); 943 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 944 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 945 oop object = (oop)mid->object(); 946 if (object != NULL) { 947 closure->do_monitor(mid); 948 } 949 } 950 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 951 } 952 } 953 954 // Get the next block in the block list. 955 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) { 956 assert(block->object() == CHAINMARKER, "must be a block header"); 957 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext; 958 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 959 return block; 960 } 961 962 static bool monitors_used_above_threshold() { 963 if (gMonitorPopulation == 0) { 964 return false; 965 } 966 int monitors_used = gMonitorPopulation - gMonitorFreeCount; 967 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation; 968 return monitor_usage > MonitorUsedDeflationThreshold; 969 } 970 971 bool ObjectSynchronizer::is_cleanup_needed() { 972 if (MonitorUsedDeflationThreshold > 0) { 973 return monitors_used_above_threshold(); 974 } 975 return false; 976 } 977 978 void ObjectSynchronizer::oops_do(OopClosure* f) { 979 if (MonitorInUseLists) { 980 // When using thread local monitor lists, we only scan the 981 // global used list here (for moribund threads), and 982 // the thread-local monitors in Thread::oops_do(). 983 global_used_oops_do(f); 984 } else { 985 global_oops_do(f); 986 } 987 } 988 989 void ObjectSynchronizer::global_oops_do(OopClosure* f) { 990 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 991 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 992 for (; block != NULL; block = next(block)) { 993 assert(block->object() == CHAINMARKER, "must be a block header"); 994 for (int i = 1; i < _BLOCKSIZE; i++) { 995 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 996 if (mid->object() != NULL) { 997 f->do_oop((oop*)mid->object_addr()); 998 } 999 } 1000 } 1001 } 1002 1003 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1004 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1005 list_oops_do(gOmInUseList, f); 1006 } 1007 1008 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1009 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1010 list_oops_do(thread->omInUseList, f); 1011 } 1012 1013 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) { 1014 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1015 ObjectMonitor* mid; 1016 for (mid = list; mid != NULL; mid = mid->FreeNext) { 1017 if (mid->object() != NULL) { 1018 f->do_oop((oop*)mid->object_addr()); 1019 } 1020 } 1021 } 1022 1023 1024 // ----------------------------------------------------------------------------- 1025 // ObjectMonitor Lifecycle 1026 // ----------------------- 1027 // Inflation unlinks monitors from the global gFreeList and 1028 // associates them with objects. Deflation -- which occurs at 1029 // STW-time -- disassociates idle monitors from objects. Such 1030 // scavenged monitors are returned to the gFreeList. 1031 // 1032 // The global list is protected by gListLock. All the critical sections 1033 // are short and operate in constant-time. 1034 // 1035 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1036 // 1037 // Lifecycle: 1038 // -- unassigned and on the global free list 1039 // -- unassigned and on a thread's private omFreeList 1040 // -- assigned to an object. The object is inflated and the mark refers 1041 // to the objectmonitor. 1042 1043 1044 // Constraining monitor pool growth via MonitorBound ... 1045 // 1046 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1047 // the rate of scavenging is driven primarily by GC. As such, we can find 1048 // an inordinate number of monitors in circulation. 1049 // To avoid that scenario we can artificially induce a STW safepoint 1050 // if the pool appears to be growing past some reasonable bound. 1051 // Generally we favor time in space-time tradeoffs, but as there's no 1052 // natural back-pressure on the # of extant monitors we need to impose some 1053 // type of limit. Beware that if MonitorBound is set to too low a value 1054 // we could just loop. In addition, if MonitorBound is set to a low value 1055 // we'll incur more safepoints, which are harmful to performance. 1056 // See also: GuaranteedSafepointInterval 1057 // 1058 // The current implementation uses asynchronous VM operations. 1059 1060 static void InduceScavenge(Thread * Self, const char * Whence) { 1061 // Induce STW safepoint to trim monitors 1062 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1063 // More precisely, trigger an asynchronous STW safepoint as the number 1064 // of active monitors passes the specified threshold. 1065 // TODO: assert thread state is reasonable 1066 1067 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 1068 if (ObjectMonitor::Knob_Verbose) { 1069 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)", 1070 Whence, ForceMonitorScavenge) ; 1071 tty->flush(); 1072 } 1073 // Induce a 'null' safepoint to scavenge monitors 1074 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 1075 // to the VMthread and have a lifespan longer than that of this activation record. 1076 // The VMThread will delete the op when completed. 1077 VMThread::execute(new VM_ScavengeMonitors()); 1078 1079 if (ObjectMonitor::Knob_Verbose) { 1080 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)", 1081 Whence, ForceMonitorScavenge) ; 1082 tty->flush(); 1083 } 1084 } 1085 } 1086 1087 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1088 ObjectMonitor* mid; 1089 int in_use_tally = 0; 1090 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1091 in_use_tally++; 1092 } 1093 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1094 1095 int free_tally = 0; 1096 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1097 free_tally++; 1098 } 1099 assert(free_tally == Self->omFreeCount, "free count off"); 1100 } 1101 1102 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) { 1103 // A large MAXPRIVATE value reduces both list lock contention 1104 // and list coherency traffic, but also tends to increase the 1105 // number of objectMonitors in circulation as well as the STW 1106 // scavenge costs. As usual, we lean toward time in space-time 1107 // tradeoffs. 1108 const int MAXPRIVATE = 1024; 1109 for (;;) { 1110 ObjectMonitor * m; 1111 1112 // 1: try to allocate from the thread's local omFreeList. 1113 // Threads will attempt to allocate first from their local list, then 1114 // from the global list, and only after those attempts fail will the thread 1115 // attempt to instantiate new monitors. Thread-local free lists take 1116 // heat off the gListLock and improve allocation latency, as well as reducing 1117 // coherency traffic on the shared global list. 1118 m = Self->omFreeList; 1119 if (m != NULL) { 1120 Self->omFreeList = m->FreeNext; 1121 Self->omFreeCount--; 1122 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1123 guarantee(m->object() == NULL, "invariant"); 1124 if (MonitorInUseLists) { 1125 m->FreeNext = Self->omInUseList; 1126 Self->omInUseList = m; 1127 Self->omInUseCount++; 1128 if (ObjectMonitor::Knob_VerifyInUse) { 1129 verifyInUse(Self); 1130 } 1131 } else { 1132 m->FreeNext = NULL; 1133 } 1134 return m; 1135 } 1136 1137 // 2: try to allocate from the global gFreeList 1138 // CONSIDER: use muxTry() instead of muxAcquire(). 1139 // If the muxTry() fails then drop immediately into case 3. 1140 // If we're using thread-local free lists then try 1141 // to reprovision the caller's free list. 1142 if (gFreeList != NULL) { 1143 // Reprovision the thread's omFreeList. 1144 // Use bulk transfers to reduce the allocation rate and heat 1145 // on various locks. 1146 Thread::muxAcquire(&gListLock, "omAlloc"); 1147 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1148 gMonitorFreeCount--; 1149 ObjectMonitor * take = gFreeList; 1150 gFreeList = take->FreeNext; 1151 guarantee(take->object() == NULL, "invariant"); 1152 guarantee(!take->is_busy(), "invariant"); 1153 take->Recycle(); 1154 omRelease(Self, take, false); 1155 } 1156 Thread::muxRelease(&gListLock); 1157 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1158 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1159 TEVENT(omFirst - reprovision); 1160 1161 const int mx = MonitorBound; 1162 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1163 // We can't safely induce a STW safepoint from omAlloc() as our thread 1164 // state may not be appropriate for such activities and callers may hold 1165 // naked oops, so instead we defer the action. 1166 InduceScavenge(Self, "omAlloc"); 1167 } 1168 continue; 1169 } 1170 1171 // 3: allocate a block of new ObjectMonitors 1172 // Both the local and global free lists are empty -- resort to malloc(). 1173 // In the current implementation objectMonitors are TSM - immortal. 1174 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1175 // each ObjectMonitor to start at the beginning of a cache line, 1176 // so we use align_up(). 1177 // A better solution would be to use C++ placement-new. 1178 // BEWARE: As it stands currently, we don't run the ctors! 1179 assert(_BLOCKSIZE > 1, "invariant"); 1180 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1181 PaddedEnd<ObjectMonitor> * temp; 1182 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1183 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1184 mtInternal); 1185 temp = (PaddedEnd<ObjectMonitor> *) 1186 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE); 1187 1188 // NOTE: (almost) no way to recover if allocation failed. 1189 // We might be able to induce a STW safepoint and scavenge enough 1190 // objectMonitors to permit progress. 1191 if (temp == NULL) { 1192 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1193 "Allocate ObjectMonitors"); 1194 } 1195 (void)memset((void *) temp, 0, neededsize); 1196 1197 // Format the block. 1198 // initialize the linked list, each monitor points to its next 1199 // forming the single linked free list, the very first monitor 1200 // will points to next block, which forms the block list. 1201 // The trick of using the 1st element in the block as gBlockList 1202 // linkage should be reconsidered. A better implementation would 1203 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1204 1205 for (int i = 1; i < _BLOCKSIZE; i++) { 1206 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1207 } 1208 1209 // terminate the last monitor as the end of list 1210 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1211 1212 // Element [0] is reserved for global list linkage 1213 temp[0].set_object(CHAINMARKER); 1214 1215 // Consider carving out this thread's current request from the 1216 // block in hand. This avoids some lock traffic and redundant 1217 // list activity. 1218 1219 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1220 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1221 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1222 gMonitorPopulation += _BLOCKSIZE-1; 1223 gMonitorFreeCount += _BLOCKSIZE-1; 1224 1225 // Add the new block to the list of extant blocks (gBlockList). 1226 // The very first objectMonitor in a block is reserved and dedicated. 1227 // It serves as blocklist "next" linkage. 1228 temp[0].FreeNext = gBlockList; 1229 // There are lock-free uses of gBlockList so make sure that 1230 // the previous stores happen before we update gBlockList. 1231 OrderAccess::release_store(&gBlockList, temp); 1232 1233 // Add the new string of objectMonitors to the global free list 1234 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1235 gFreeList = temp + 1; 1236 Thread::muxRelease(&gListLock); 1237 TEVENT(Allocate block of monitors); 1238 } 1239 } 1240 1241 // Place "m" on the caller's private per-thread omFreeList. 1242 // In practice there's no need to clamp or limit the number of 1243 // monitors on a thread's omFreeList as the only time we'll call 1244 // omRelease is to return a monitor to the free list after a CAS 1245 // attempt failed. This doesn't allow unbounded #s of monitors to 1246 // accumulate on a thread's free list. 1247 // 1248 // Key constraint: all ObjectMonitors on a thread's free list and the global 1249 // free list must have their object field set to null. This prevents the 1250 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1251 1252 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1253 bool fromPerThreadAlloc) { 1254 guarantee(m->object() == NULL, "invariant"); 1255 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1256 // Remove from omInUseList 1257 if (MonitorInUseLists && fromPerThreadAlloc) { 1258 ObjectMonitor* cur_mid_in_use = NULL; 1259 bool extracted = false; 1260 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1261 if (m == mid) { 1262 // extract from per-thread in-use list 1263 if (mid == Self->omInUseList) { 1264 Self->omInUseList = mid->FreeNext; 1265 } else if (cur_mid_in_use != NULL) { 1266 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1267 } 1268 extracted = true; 1269 Self->omInUseCount--; 1270 if (ObjectMonitor::Knob_VerifyInUse) { 1271 verifyInUse(Self); 1272 } 1273 break; 1274 } 1275 } 1276 assert(extracted, "Should have extracted from in-use list"); 1277 } 1278 1279 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1280 m->FreeNext = Self->omFreeList; 1281 Self->omFreeList = m; 1282 Self->omFreeCount++; 1283 } 1284 1285 // Return the monitors of a moribund thread's local free list to 1286 // the global free list. Typically a thread calls omFlush() when 1287 // it's dying. We could also consider having the VM thread steal 1288 // monitors from threads that have not run java code over a few 1289 // consecutive STW safepoints. Relatedly, we might decay 1290 // omFreeProvision at STW safepoints. 1291 // 1292 // Also return the monitors of a moribund thread's omInUseList to 1293 // a global gOmInUseList under the global list lock so these 1294 // will continue to be scanned. 1295 // 1296 // We currently call omFlush() from Threads::remove() _before the thread 1297 // has been excised from the thread list and is no longer a mutator. 1298 // This means that omFlush() can not run concurrently with a safepoint and 1299 // interleave with the scavenge operator. In particular, this ensures that 1300 // the thread's monitors are scanned by a GC safepoint, either via 1301 // Thread::oops_do() (if safepoint happens before omFlush()) or via 1302 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's 1303 // monitors have been transferred to the global in-use list). 1304 1305 void ObjectSynchronizer::omFlush(Thread * Self) { 1306 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1307 Self->omFreeList = NULL; 1308 ObjectMonitor * tail = NULL; 1309 int tally = 0; 1310 if (list != NULL) { 1311 ObjectMonitor * s; 1312 // The thread is going away, the per-thread free monitors 1313 // are freed via set_owner(NULL) 1314 // Link them to tail, which will be linked into the global free list 1315 // gFreeList below, under the gListLock 1316 for (s = list; s != NULL; s = s->FreeNext) { 1317 tally++; 1318 tail = s; 1319 guarantee(s->object() == NULL, "invariant"); 1320 guarantee(!s->is_busy(), "invariant"); 1321 s->set_owner(NULL); // redundant but good hygiene 1322 TEVENT(omFlush - Move one); 1323 } 1324 guarantee(tail != NULL && list != NULL, "invariant"); 1325 } 1326 1327 ObjectMonitor * inUseList = Self->omInUseList; 1328 ObjectMonitor * inUseTail = NULL; 1329 int inUseTally = 0; 1330 if (inUseList != NULL) { 1331 Self->omInUseList = NULL; 1332 ObjectMonitor *cur_om; 1333 // The thread is going away, however the omInUseList inflated 1334 // monitors may still be in-use by other threads. 1335 // Link them to inUseTail, which will be linked into the global in-use list 1336 // gOmInUseList below, under the gListLock 1337 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1338 inUseTail = cur_om; 1339 inUseTally++; 1340 } 1341 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1342 Self->omInUseCount = 0; 1343 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1344 } 1345 1346 Thread::muxAcquire(&gListLock, "omFlush"); 1347 if (tail != NULL) { 1348 tail->FreeNext = gFreeList; 1349 gFreeList = list; 1350 gMonitorFreeCount += tally; 1351 assert(Self->omFreeCount == tally, "free-count off"); 1352 Self->omFreeCount = 0; 1353 } 1354 1355 if (inUseTail != NULL) { 1356 inUseTail->FreeNext = gOmInUseList; 1357 gOmInUseList = inUseList; 1358 gOmInUseCount += inUseTally; 1359 } 1360 1361 Thread::muxRelease(&gListLock); 1362 TEVENT(omFlush); 1363 } 1364 1365 // Fast path code shared by multiple functions 1366 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1367 markOop mark = obj->mark(); 1368 if (mark->has_monitor()) { 1369 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1370 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1371 return mark->monitor(); 1372 } 1373 return ObjectSynchronizer::inflate(Thread::current(), 1374 obj, 1375 inflate_cause_vm_internal); 1376 } 1377 1378 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self, 1379 oop object, 1380 const InflateCause cause) { 1381 1382 // Inflate mutates the heap ... 1383 // Relaxing assertion for bug 6320749. 1384 assert(Universe::verify_in_progress() || 1385 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1386 1387 EventJavaMonitorInflate event; 1388 1389 for (;;) { 1390 const markOop mark = object->mark(); 1391 assert(!mark->has_bias_pattern(), "invariant"); 1392 1393 // The mark can be in one of the following states: 1394 // * Inflated - just return 1395 // * Stack-locked - coerce it to inflated 1396 // * INFLATING - busy wait for conversion to complete 1397 // * Neutral - aggressively inflate the object. 1398 // * BIASED - Illegal. We should never see this 1399 1400 // CASE: inflated 1401 if (mark->has_monitor()) { 1402 ObjectMonitor * inf = mark->monitor(); 1403 assert(inf->header()->is_neutral(), "invariant"); 1404 assert(inf->object() == object, "invariant"); 1405 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1406 return inf; 1407 } 1408 1409 // CASE: inflation in progress - inflating over a stack-lock. 1410 // Some other thread is converting from stack-locked to inflated. 1411 // Only that thread can complete inflation -- other threads must wait. 1412 // The INFLATING value is transient. 1413 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1414 // We could always eliminate polling by parking the thread on some auxiliary list. 1415 if (mark == markOopDesc::INFLATING()) { 1416 TEVENT(Inflate: spin while INFLATING); 1417 ReadStableMark(object); 1418 continue; 1419 } 1420 1421 // CASE: stack-locked 1422 // Could be stack-locked either by this thread or by some other thread. 1423 // 1424 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1425 // to install INFLATING into the mark word. We originally installed INFLATING, 1426 // allocated the objectmonitor, and then finally STed the address of the 1427 // objectmonitor into the mark. This was correct, but artificially lengthened 1428 // the interval in which INFLATED appeared in the mark, thus increasing 1429 // the odds of inflation contention. 1430 // 1431 // We now use per-thread private objectmonitor free lists. 1432 // These list are reprovisioned from the global free list outside the 1433 // critical INFLATING...ST interval. A thread can transfer 1434 // multiple objectmonitors en-mass from the global free list to its local free list. 1435 // This reduces coherency traffic and lock contention on the global free list. 1436 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1437 // before or after the CAS(INFLATING) operation. 1438 // See the comments in omAlloc(). 1439 1440 if (mark->has_locker()) { 1441 ObjectMonitor * m = omAlloc(Self); 1442 // Optimistically prepare the objectmonitor - anticipate successful CAS 1443 // We do this before the CAS in order to minimize the length of time 1444 // in which INFLATING appears in the mark. 1445 m->Recycle(); 1446 m->_Responsible = NULL; 1447 m->_recursions = 0; 1448 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1449 1450 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark); 1451 if (cmp != mark) { 1452 omRelease(Self, m, true); 1453 continue; // Interference -- just retry 1454 } 1455 1456 // We've successfully installed INFLATING (0) into the mark-word. 1457 // This is the only case where 0 will appear in a mark-word. 1458 // Only the singular thread that successfully swings the mark-word 1459 // to 0 can perform (or more precisely, complete) inflation. 1460 // 1461 // Why do we CAS a 0 into the mark-word instead of just CASing the 1462 // mark-word from the stack-locked value directly to the new inflated state? 1463 // Consider what happens when a thread unlocks a stack-locked object. 1464 // It attempts to use CAS to swing the displaced header value from the 1465 // on-stack basiclock back into the object header. Recall also that the 1466 // header value (hashcode, etc) can reside in (a) the object header, or 1467 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1468 // header in an objectMonitor. The inflate() routine must copy the header 1469 // value from the basiclock on the owner's stack to the objectMonitor, all 1470 // the while preserving the hashCode stability invariants. If the owner 1471 // decides to release the lock while the value is 0, the unlock will fail 1472 // and control will eventually pass from slow_exit() to inflate. The owner 1473 // will then spin, waiting for the 0 value to disappear. Put another way, 1474 // the 0 causes the owner to stall if the owner happens to try to 1475 // drop the lock (restoring the header from the basiclock to the object) 1476 // while inflation is in-progress. This protocol avoids races that might 1477 // would otherwise permit hashCode values to change or "flicker" for an object. 1478 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1479 // 0 serves as a "BUSY" inflate-in-progress indicator. 1480 1481 1482 // fetch the displaced mark from the owner's stack. 1483 // The owner can't die or unwind past the lock while our INFLATING 1484 // object is in the mark. Furthermore the owner can't complete 1485 // an unlock on the object, either. 1486 markOop dmw = mark->displaced_mark_helper(); 1487 assert(dmw->is_neutral(), "invariant"); 1488 1489 // Setup monitor fields to proper values -- prepare the monitor 1490 m->set_header(dmw); 1491 1492 // Optimization: if the mark->locker stack address is associated 1493 // with this thread we could simply set m->_owner = Self. 1494 // Note that a thread can inflate an object 1495 // that it has stack-locked -- as might happen in wait() -- directly 1496 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1497 m->set_owner(mark->locker()); 1498 m->set_object(object); 1499 // TODO-FIXME: assert BasicLock->dhw != 0. 1500 1501 // Must preserve store ordering. The monitor state must 1502 // be stable at the time of publishing the monitor address. 1503 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1504 object->release_set_mark(markOopDesc::encode(m)); 1505 1506 // Hopefully the performance counters are allocated on distinct cache lines 1507 // to avoid false sharing on MP systems ... 1508 OM_PERFDATA_OP(Inflations, inc()); 1509 TEVENT(Inflate: overwrite stacklock); 1510 if (log_is_enabled(Debug, monitorinflation)) { 1511 if (object->is_instance()) { 1512 ResourceMark rm; 1513 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1514 p2i(object), p2i(object->mark()), 1515 object->klass()->external_name()); 1516 } 1517 } 1518 if (event.should_commit()) { 1519 post_monitor_inflate_event(event, object, cause); 1520 } 1521 return m; 1522 } 1523 1524 // CASE: neutral 1525 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1526 // If we know we're inflating for entry it's better to inflate by swinging a 1527 // pre-locked objectMonitor pointer into the object header. A successful 1528 // CAS inflates the object *and* confers ownership to the inflating thread. 1529 // In the current implementation we use a 2-step mechanism where we CAS() 1530 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1531 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1532 // would be useful. 1533 1534 assert(mark->is_neutral(), "invariant"); 1535 ObjectMonitor * m = omAlloc(Self); 1536 // prepare m for installation - set monitor to initial state 1537 m->Recycle(); 1538 m->set_header(mark); 1539 m->set_owner(NULL); 1540 m->set_object(object); 1541 m->_recursions = 0; 1542 m->_Responsible = NULL; 1543 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1544 1545 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) { 1546 m->set_object(NULL); 1547 m->set_owner(NULL); 1548 m->Recycle(); 1549 omRelease(Self, m, true); 1550 m = NULL; 1551 continue; 1552 // interference - the markword changed - just retry. 1553 // The state-transitions are one-way, so there's no chance of 1554 // live-lock -- "Inflated" is an absorbing state. 1555 } 1556 1557 // Hopefully the performance counters are allocated on distinct 1558 // cache lines to avoid false sharing on MP systems ... 1559 OM_PERFDATA_OP(Inflations, inc()); 1560 TEVENT(Inflate: overwrite neutral); 1561 if (log_is_enabled(Debug, monitorinflation)) { 1562 if (object->is_instance()) { 1563 ResourceMark rm; 1564 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1565 p2i(object), p2i(object->mark()), 1566 object->klass()->external_name()); 1567 } 1568 } 1569 if (event.should_commit()) { 1570 post_monitor_inflate_event(event, object, cause); 1571 } 1572 return m; 1573 } 1574 } 1575 1576 1577 // Deflate_idle_monitors() is called at all safepoints, immediately 1578 // after all mutators are stopped, but before any objects have moved. 1579 // It traverses the list of known monitors, deflating where possible. 1580 // The scavenged monitor are returned to the monitor free list. 1581 // 1582 // Beware that we scavenge at *every* stop-the-world point. 1583 // Having a large number of monitors in-circulation negatively 1584 // impacts the performance of some applications (e.g., PointBase). 1585 // Broadly, we want to minimize the # of monitors in circulation. 1586 // 1587 // We have added a flag, MonitorInUseLists, which creates a list 1588 // of active monitors for each thread. deflate_idle_monitors() 1589 // only scans the per-thread in-use lists. omAlloc() puts all 1590 // assigned monitors on the per-thread list. deflate_idle_monitors() 1591 // returns the non-busy monitors to the global free list. 1592 // When a thread dies, omFlush() adds the list of active monitors for 1593 // that thread to a global gOmInUseList acquiring the 1594 // global list lock. deflate_idle_monitors() acquires the global 1595 // list lock to scan for non-busy monitors to the global free list. 1596 // An alternative could have used a single global in-use list. The 1597 // downside would have been the additional cost of acquiring the global list lock 1598 // for every omAlloc(). 1599 // 1600 // Perversely, the heap size -- and thus the STW safepoint rate -- 1601 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1602 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1603 // This is an unfortunate aspect of this design. 1604 1605 enum ManifestConstants { 1606 ClearResponsibleAtSTW = 0 1607 }; 1608 1609 // Deflate a single monitor if not in-use 1610 // Return true if deflated, false if in-use 1611 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1612 ObjectMonitor** freeHeadp, 1613 ObjectMonitor** freeTailp) { 1614 bool deflated; 1615 // Normal case ... The monitor is associated with obj. 1616 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1617 guarantee(mid == obj->mark()->monitor(), "invariant"); 1618 guarantee(mid->header()->is_neutral(), "invariant"); 1619 1620 if (mid->is_busy()) { 1621 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1622 deflated = false; 1623 } else { 1624 // Deflate the monitor if it is no longer being used 1625 // It's idle - scavenge and return to the global free list 1626 // plain old deflation ... 1627 TEVENT(deflate_idle_monitors - scavenge1); 1628 if (log_is_enabled(Debug, monitorinflation)) { 1629 if (obj->is_instance()) { 1630 ResourceMark rm; 1631 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , " 1632 "mark " INTPTR_FORMAT " , type %s", 1633 p2i(obj), p2i(obj->mark()), 1634 obj->klass()->external_name()); 1635 } 1636 } 1637 1638 // Restore the header back to obj 1639 obj->release_set_mark(mid->header()); 1640 mid->clear(); 1641 1642 assert(mid->object() == NULL, "invariant"); 1643 1644 // Move the object to the working free list defined by freeHeadp, freeTailp 1645 if (*freeHeadp == NULL) *freeHeadp = mid; 1646 if (*freeTailp != NULL) { 1647 ObjectMonitor * prevtail = *freeTailp; 1648 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1649 prevtail->FreeNext = mid; 1650 } 1651 *freeTailp = mid; 1652 deflated = true; 1653 } 1654 return deflated; 1655 } 1656 1657 // Walk a given monitor list, and deflate idle monitors 1658 // The given list could be a per-thread list or a global list 1659 // Caller acquires gListLock. 1660 // 1661 // In the case of parallel processing of thread local monitor lists, 1662 // work is done by Threads::parallel_threads_do() which ensures that 1663 // each Java thread is processed by exactly one worker thread, and 1664 // thus avoid conflicts that would arise when worker threads would 1665 // process the same monitor lists concurrently. 1666 // 1667 // See also ParallelSPCleanupTask and 1668 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and 1669 // Threads::parallel_java_threads_do() in thread.cpp. 1670 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1671 ObjectMonitor** freeHeadp, 1672 ObjectMonitor** freeTailp) { 1673 ObjectMonitor* mid; 1674 ObjectMonitor* next; 1675 ObjectMonitor* cur_mid_in_use = NULL; 1676 int deflated_count = 0; 1677 1678 for (mid = *listHeadp; mid != NULL;) { 1679 oop obj = (oop) mid->object(); 1680 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1681 // if deflate_monitor succeeded, 1682 // extract from per-thread in-use list 1683 if (mid == *listHeadp) { 1684 *listHeadp = mid->FreeNext; 1685 } else if (cur_mid_in_use != NULL) { 1686 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1687 } 1688 next = mid->FreeNext; 1689 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1690 mid = next; 1691 deflated_count++; 1692 } else { 1693 cur_mid_in_use = mid; 1694 mid = mid->FreeNext; 1695 } 1696 } 1697 return deflated_count; 1698 } 1699 1700 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1701 counters->nInuse = 0; // currently associated with objects 1702 counters->nInCirculation = 0; // extant 1703 counters->nScavenged = 0; // reclaimed 1704 } 1705 1706 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 1707 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1708 bool deflated = false; 1709 1710 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1711 ObjectMonitor * freeTailp = NULL; 1712 1713 TEVENT(deflate_idle_monitors); 1714 // Prevent omFlush from changing mids in Thread dtor's during deflation 1715 // And in case the vm thread is acquiring a lock during a safepoint 1716 // See e.g. 6320749 1717 Thread::muxAcquire(&gListLock, "scavenge - return"); 1718 1719 if (MonitorInUseLists) { 1720 // Note: the thread-local monitors lists get deflated in 1721 // a separate pass. See deflate_thread_local_monitors(). 1722 1723 // For moribund threads, scan gOmInUseList 1724 if (gOmInUseList) { 1725 counters->nInCirculation += gOmInUseCount; 1726 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1727 gOmInUseCount -= deflated_count; 1728 counters->nScavenged += deflated_count; 1729 counters->nInuse += gOmInUseCount; 1730 } 1731 1732 } else { 1733 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 1734 for (; block != NULL; block = next(block)) { 1735 // Iterate over all extant monitors - Scavenge all idle monitors. 1736 assert(block->object() == CHAINMARKER, "must be a block header"); 1737 counters->nInCirculation += _BLOCKSIZE; 1738 for (int i = 1; i < _BLOCKSIZE; i++) { 1739 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1740 oop obj = (oop)mid->object(); 1741 1742 if (obj == NULL) { 1743 // The monitor is not associated with an object. 1744 // The monitor should either be a thread-specific private 1745 // free list or the global free list. 1746 // obj == NULL IMPLIES mid->is_busy() == 0 1747 guarantee(!mid->is_busy(), "invariant"); 1748 continue; 1749 } 1750 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1751 1752 if (deflated) { 1753 mid->FreeNext = NULL; 1754 counters->nScavenged++; 1755 } else { 1756 counters->nInuse++; 1757 } 1758 } 1759 } 1760 } 1761 1762 // Move the scavenged monitors back to the global free list. 1763 if (freeHeadp != NULL) { 1764 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant"); 1765 assert(freeTailp->FreeNext == NULL, "invariant"); 1766 // constant-time list splice - prepend scavenged segment to gFreeList 1767 freeTailp->FreeNext = gFreeList; 1768 gFreeList = freeHeadp; 1769 } 1770 Thread::muxRelease(&gListLock); 1771 1772 } 1773 1774 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1775 gMonitorFreeCount += counters->nScavenged; 1776 1777 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1778 1779 if (ObjectMonitor::Knob_Verbose) { 1780 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d " 1781 "ForceMonitorScavenge=%d : pop=%d free=%d", 1782 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge, 1783 gMonitorPopulation, gMonitorFreeCount); 1784 tty->flush(); 1785 } 1786 1787 ForceMonitorScavenge = 0; // Reset 1788 1789 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged)); 1790 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation)); 1791 1792 // TODO: Add objectMonitor leak detection. 1793 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1794 GVars.stwRandom = os::random(); 1795 GVars.stwCycle++; 1796 } 1797 1798 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 1799 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1800 if (!MonitorInUseLists) return; 1801 1802 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1803 ObjectMonitor * freeTailp = NULL; 1804 1805 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp); 1806 1807 Thread::muxAcquire(&gListLock, "scavenge - return"); 1808 1809 // Adjust counters 1810 counters->nInCirculation += thread->omInUseCount; 1811 thread->omInUseCount -= deflated_count; 1812 if (ObjectMonitor::Knob_VerifyInUse) { 1813 verifyInUse(thread); 1814 } 1815 counters->nScavenged += deflated_count; 1816 counters->nInuse += thread->omInUseCount; 1817 1818 // Move the scavenged monitors back to the global free list. 1819 if (freeHeadp != NULL) { 1820 guarantee(freeTailp != NULL && deflated_count > 0, "invariant"); 1821 assert(freeTailp->FreeNext == NULL, "invariant"); 1822 1823 // constant-time list splice - prepend scavenged segment to gFreeList 1824 freeTailp->FreeNext = gFreeList; 1825 gFreeList = freeHeadp; 1826 } 1827 Thread::muxRelease(&gListLock); 1828 } 1829 1830 // Monitor cleanup on JavaThread::exit 1831 1832 // Iterate through monitor cache and attempt to release thread's monitors 1833 // Gives up on a particular monitor if an exception occurs, but continues 1834 // the overall iteration, swallowing the exception. 1835 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1836 private: 1837 TRAPS; 1838 1839 public: 1840 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1841 void do_monitor(ObjectMonitor* mid) { 1842 if (mid->owner() == THREAD) { 1843 if (ObjectMonitor::Knob_VerifyMatch != 0) { 1844 ResourceMark rm; 1845 Handle obj(THREAD, (oop) mid->object()); 1846 tty->print("INFO: unexpected locked object:"); 1847 javaVFrame::print_locked_object_class_name(tty, obj, "locked"); 1848 fatal("exiting JavaThread=" INTPTR_FORMAT 1849 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT, 1850 p2i(THREAD), p2i(mid)); 1851 } 1852 (void)mid->complete_exit(CHECK); 1853 } 1854 } 1855 }; 1856 1857 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1858 // ignored. This is meant to be called during JNI thread detach which assumes 1859 // all remaining monitors are heavyweight. All exceptions are swallowed. 1860 // Scanning the extant monitor list can be time consuming. 1861 // A simple optimization is to add a per-thread flag that indicates a thread 1862 // called jni_monitorenter() during its lifetime. 1863 // 1864 // Instead of No_Savepoint_Verifier it might be cheaper to 1865 // use an idiom of the form: 1866 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1867 // <code that must not run at safepoint> 1868 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1869 // Since the tests are extremely cheap we could leave them enabled 1870 // for normal product builds. 1871 1872 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1873 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1874 NoSafepointVerifier nsv; 1875 ReleaseJavaMonitorsClosure rjmc(THREAD); 1876 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1877 ObjectSynchronizer::monitors_iterate(&rjmc); 1878 Thread::muxRelease(&gListLock); 1879 THREAD->clear_pending_exception(); 1880 } 1881 1882 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1883 switch (cause) { 1884 case inflate_cause_vm_internal: return "VM Internal"; 1885 case inflate_cause_monitor_enter: return "Monitor Enter"; 1886 case inflate_cause_wait: return "Monitor Wait"; 1887 case inflate_cause_notify: return "Monitor Notify"; 1888 case inflate_cause_hash_code: return "Monitor Hash Code"; 1889 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1890 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1891 default: 1892 ShouldNotReachHere(); 1893 } 1894 return "Unknown"; 1895 } 1896 1897 static void post_monitor_inflate_event(EventJavaMonitorInflate& event, 1898 const oop obj, 1899 const ObjectSynchronizer::InflateCause cause) { 1900 #if INCLUDE_TRACE 1901 assert(event.should_commit(), "check outside"); 1902 event.set_monitorClass(obj->klass()); 1903 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj); 1904 event.set_cause((u1)cause); 1905 event.commit(); 1906 #endif 1907 } 1908 1909 //------------------------------------------------------------------------------ 1910 // Debugging code 1911 1912 void ObjectSynchronizer::sanity_checks(const bool verbose, 1913 const uint cache_line_size, 1914 int *error_cnt_ptr, 1915 int *warning_cnt_ptr) { 1916 u_char *addr_begin = (u_char*)&GVars; 1917 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1918 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1919 1920 if (verbose) { 1921 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1922 sizeof(SharedGlobals)); 1923 } 1924 1925 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1926 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1927 1928 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1929 if (verbose) { 1930 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1931 } 1932 1933 if (cache_line_size != 0) { 1934 // We were able to determine the L1 data cache line size so 1935 // do some cache line specific sanity checks 1936 1937 if (offset_stwRandom < cache_line_size) { 1938 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1939 "to the struct beginning than a cache line which permits " 1940 "false sharing."); 1941 (*warning_cnt_ptr)++; 1942 } 1943 1944 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1945 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1946 "SharedGlobals.hcSequence fields are closer than a cache " 1947 "line which permits false sharing."); 1948 (*warning_cnt_ptr)++; 1949 } 1950 1951 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1952 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1953 "to the struct end than a cache line which permits false " 1954 "sharing."); 1955 (*warning_cnt_ptr)++; 1956 } 1957 } 1958 } 1959 1960 #ifndef PRODUCT 1961 1962 // Check if monitor belongs to the monitor cache 1963 // The list is grow-only so it's *relatively* safe to traverse 1964 // the list of extant blocks without taking a lock. 1965 1966 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1967 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 1968 while (block != NULL) { 1969 assert(block->object() == CHAINMARKER, "must be a block header"); 1970 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 1971 address mon = (address)monitor; 1972 address blk = (address)block; 1973 size_t diff = mon - blk; 1974 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned"); 1975 return 1; 1976 } 1977 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 1978 } 1979 return 0; 1980 } 1981 1982 #endif