1 /* 2 * Copyright (c) 1998, 2018, 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/allocation.inline.hpp" 29 #include "memory/metaspaceShared.hpp" 30 #include "memory/padded.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/markOop.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "runtime/atomic.hpp" 35 #include "runtime/biasedLocking.hpp" 36 #include "runtime/handles.inline.hpp" 37 #include "runtime/interfaceSupport.inline.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "runtime/objectMonitor.hpp" 40 #include "runtime/objectMonitor.inline.hpp" 41 #include "runtime/osThread.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 #include "runtime/stubRoutines.hpp" 44 #include "runtime/synchronizer.hpp" 45 #include "runtime/thread.inline.hpp" 46 #include "runtime/vframe.hpp" 47 #include "runtime/vmThread.hpp" 48 #include "trace/traceMacros.hpp" 49 #include "trace/tracing.hpp" 50 #include "utilities/align.hpp" 51 #include "utilities/dtrace.hpp" 52 #include "utilities/events.hpp" 53 #include "utilities/preserveException.hpp" 54 55 // The "core" versions of monitor enter and exit reside in this file. 56 // The interpreter and compilers contain specialized transliterated 57 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 58 // for instance. If you make changes here, make sure to modify the 59 // interpreter, and both C1 and C2 fast-path inline locking code emission. 60 // 61 // ----------------------------------------------------------------------------- 62 63 #ifdef DTRACE_ENABLED 64 65 // Only bother with this argument setup if dtrace is available 66 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 67 68 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 69 char* bytes = NULL; \ 70 int len = 0; \ 71 jlong jtid = SharedRuntime::get_java_tid(thread); \ 72 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 73 if (klassname != NULL) { \ 74 bytes = (char*)klassname->bytes(); \ 75 len = klassname->utf8_length(); \ 76 } 77 78 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 79 { \ 80 if (DTraceMonitorProbes) { \ 81 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 82 HOTSPOT_MONITOR_WAIT(jtid, \ 83 (uintptr_t)(monitor), bytes, len, (millis)); \ 84 } \ 85 } 86 87 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 88 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 89 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 90 91 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 92 { \ 93 if (DTraceMonitorProbes) { \ 94 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 95 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 96 (uintptr_t)(monitor), bytes, len); \ 97 } \ 98 } 99 100 #else // ndef DTRACE_ENABLED 101 102 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 103 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 104 105 #endif // ndef DTRACE_ENABLED 106 107 // This exists only as a workaround of dtrace bug 6254741 108 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 109 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 110 return 0; 111 } 112 113 #define NINFLATIONLOCKS 256 114 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS]; 115 116 // global list of blocks of monitors 117 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL; 118 // global monitor free list 119 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 120 // global monitor in-use list, for moribund threads, 121 // monitors they inflated need to be scanned for deflation 122 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 123 // count of entries in gOmInUseList 124 int ObjectSynchronizer::gOmInUseCount = 0; 125 126 static volatile intptr_t gListLock = 0; // protects global monitor lists 127 static volatile int gMonitorFreeCount = 0; // # on gFreeList 128 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation 129 130 static void post_monitor_inflate_event(EventJavaMonitorInflate&, 131 const oop, 132 const ObjectSynchronizer::InflateCause); 133 134 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 135 136 137 // =====================> Quick functions 138 139 // The quick_* forms are special fast-path variants used to improve 140 // performance. In the simplest case, a "quick_*" implementation could 141 // simply return false, in which case the caller will perform the necessary 142 // state transitions and call the slow-path form. 143 // The fast-path is designed to handle frequently arising cases in an efficient 144 // manner and is just a degenerate "optimistic" variant of the slow-path. 145 // returns true -- to indicate the call was satisfied. 146 // returns false -- to indicate the call needs the services of the slow-path. 147 // A no-loitering ordinance is in effect for code in the quick_* family 148 // operators: safepoints or indefinite blocking (blocking that might span a 149 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 150 // entry. 151 // 152 // Consider: An interesting optimization is to have the JIT recognize the 153 // following common idiom: 154 // synchronized (someobj) { .... ; notify(); } 155 // That is, we find a notify() or notifyAll() call that immediately precedes 156 // the monitorexit operation. In that case the JIT could fuse the operations 157 // into a single notifyAndExit() runtime primitive. 158 159 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) { 160 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 161 assert(self->is_Java_thread(), "invariant"); 162 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 163 NoSafepointVerifier nsv; 164 if (obj == NULL) return false; // slow-path for invalid obj 165 const markOop mark = obj->mark(); 166 167 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) { 168 // Degenerate notify 169 // stack-locked by caller so by definition the implied waitset is empty. 170 return true; 171 } 172 173 if (mark->has_monitor()) { 174 ObjectMonitor * const mon = mark->monitor(); 175 assert(mon->object() == obj, "invariant"); 176 if (mon->owner() != self) return false; // slow-path for IMS exception 177 178 if (mon->first_waiter() != NULL) { 179 // We have one or more waiters. Since this is an inflated monitor 180 // that we own, we can transfer one or more threads from the waitset 181 // to the entrylist here and now, avoiding the slow-path. 182 if (all) { 183 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 184 } else { 185 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 186 } 187 int tally = 0; 188 do { 189 mon->INotify(self); 190 ++tally; 191 } while (mon->first_waiter() != NULL && all); 192 OM_PERFDATA_OP(Notifications, inc(tally)); 193 } 194 return true; 195 } 196 197 // biased locking and any other IMS exception states take the slow-path 198 return false; 199 } 200 201 202 // The LockNode emitted directly at the synchronization site would have 203 // been too big if it were to have included support for the cases of inflated 204 // recursive enter and exit, so they go here instead. 205 // Note that we can't safely call AsyncPrintJavaStack() from within 206 // quick_enter() as our thread state remains _in_Java. 207 208 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self, 209 BasicLock * lock) { 210 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 211 assert(Self->is_Java_thread(), "invariant"); 212 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant"); 213 NoSafepointVerifier nsv; 214 if (obj == NULL) return false; // Need to throw NPE 215 const markOop mark = obj->mark(); 216 217 if (mark->has_monitor()) { 218 ObjectMonitor * const m = mark->monitor(); 219 assert(m->object() == obj, "invariant"); 220 Thread * const owner = (Thread *) m->_owner; 221 222 // Lock contention and Transactional Lock Elision (TLE) diagnostics 223 // and observability 224 // Case: light contention possibly amenable to TLE 225 // Case: TLE inimical operations such as nested/recursive synchronization 226 227 if (owner == Self) { 228 m->_recursions++; 229 return true; 230 } 231 232 // This Java Monitor is inflated so obj's header will never be 233 // displaced to this thread's BasicLock. Make the displaced header 234 // non-NULL so this BasicLock is not seen as recursive nor as 235 // being locked. We do this unconditionally so that this thread's 236 // BasicLock cannot be mis-interpreted by any stack walkers. For 237 // performance reasons, stack walkers generally first check for 238 // Biased Locking in the object's header, the second check is for 239 // stack-locking in the object's header, the third check is for 240 // recursive stack-locking in the displaced header in the BasicLock, 241 // and last are the inflated Java Monitor (ObjectMonitor) checks. 242 lock->set_displaced_header(markOopDesc::unused_mark()); 243 244 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) { 245 assert(m->_recursions == 0, "invariant"); 246 assert(m->_owner == Self, "invariant"); 247 return true; 248 } 249 } 250 251 // Note that we could inflate in quick_enter. 252 // This is likely a useful optimization 253 // Critically, in quick_enter() we must not: 254 // -- perform bias revocation, or 255 // -- block indefinitely, or 256 // -- reach a safepoint 257 258 return false; // revert to slow-path 259 } 260 261 // ----------------------------------------------------------------------------- 262 // Fast Monitor Enter/Exit 263 // This the fast monitor enter. The interpreter and compiler use 264 // some assembly copies of this code. Make sure update those code 265 // if the following function is changed. The implementation is 266 // extremely sensitive to race condition. Be careful. 267 268 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 269 bool attempt_rebias, TRAPS) { 270 if (UseBiasedLocking) { 271 if (!SafepointSynchronize::is_at_safepoint()) { 272 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 273 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 274 return; 275 } 276 } else { 277 assert(!attempt_rebias, "can not rebias toward VM thread"); 278 BiasedLocking::revoke_at_safepoint(obj); 279 } 280 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 281 } 282 283 slow_enter(obj, lock, THREAD); 284 } 285 286 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 287 markOop mark = object->mark(); 288 // We cannot check for Biased Locking if we are racing an inflation. 289 assert(mark == markOopDesc::INFLATING() || 290 !mark->has_bias_pattern(), "should not see bias pattern here"); 291 292 markOop dhw = lock->displaced_header(); 293 if (dhw == NULL) { 294 // If the displaced header is NULL, then this exit matches up with 295 // a recursive enter. No real work to do here except for diagnostics. 296 #ifndef PRODUCT 297 if (mark != markOopDesc::INFLATING()) { 298 // Only do diagnostics if we are not racing an inflation. Simply 299 // exiting a recursive enter of a Java Monitor that is being 300 // inflated is safe; see the has_monitor() comment below. 301 assert(!mark->is_neutral(), "invariant"); 302 assert(!mark->has_locker() || 303 THREAD->is_lock_owned((address)mark->locker()), "invariant"); 304 if (mark->has_monitor()) { 305 // The BasicLock's displaced_header is marked as a recursive 306 // enter and we have an inflated Java Monitor (ObjectMonitor). 307 // This is a special case where the Java Monitor was inflated 308 // after this thread entered the stack-lock recursively. When a 309 // Java Monitor is inflated, we cannot safely walk the Java 310 // Monitor owner's stack and update the BasicLocks because a 311 // Java Monitor can be asynchronously inflated by a thread that 312 // does not own the Java Monitor. 313 ObjectMonitor * m = mark->monitor(); 314 assert(((oop)(m->object()))->mark() == mark, "invariant"); 315 assert(m->is_entered(THREAD), "invariant"); 316 } 317 } 318 #endif 319 return; 320 } 321 322 if (mark == (markOop) lock) { 323 // If the object is stack-locked by the current thread, try to 324 // swing the displaced header from the BasicLock back to the mark. 325 assert(dhw->is_neutral(), "invariant"); 326 if (object->cas_set_mark(dhw, mark) == mark) { 327 TEVENT(fast_exit: release stack-lock); 328 return; 329 } 330 } 331 332 // We have to take the slow-path of possible inflation and then exit. 333 ObjectSynchronizer::inflate(THREAD, 334 object, 335 inflate_cause_vm_internal)->exit(true, THREAD); 336 } 337 338 // ----------------------------------------------------------------------------- 339 // Interpreter/Compiler Slow Case 340 // This routine is used to handle interpreter/compiler slow case 341 // We don't need to use fast path here, because it must have been 342 // failed in the interpreter/compiler code. 343 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 344 markOop mark = obj->mark(); 345 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 346 347 if (mark->is_neutral()) { 348 // Anticipate successful CAS -- the ST of the displaced mark must 349 // be visible <= the ST performed by the CAS. 350 lock->set_displaced_header(mark); 351 if (mark == obj()->cas_set_mark((markOop) lock, mark)) { 352 TEVENT(slow_enter: release stacklock); 353 return; 354 } 355 // Fall through to inflate() ... 356 } else if (mark->has_locker() && 357 THREAD->is_lock_owned((address)mark->locker())) { 358 assert(lock != mark->locker(), "must not re-lock the same lock"); 359 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 360 lock->set_displaced_header(NULL); 361 return; 362 } 363 364 // The object header will never be displaced to this lock, 365 // so it does not matter what the value is, except that it 366 // must be non-zero to avoid looking like a re-entrant lock, 367 // and must not look locked either. 368 lock->set_displaced_header(markOopDesc::unused_mark()); 369 ObjectSynchronizer::inflate(THREAD, 370 obj(), 371 inflate_cause_monitor_enter)->enter(THREAD); 372 } 373 374 // This routine is used to handle interpreter/compiler slow case 375 // We don't need to use fast path here, because it must have 376 // failed in the interpreter/compiler code. Simply use the heavy 377 // weight monitor should be ok, unless someone find otherwise. 378 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 379 fast_exit(object, lock, THREAD); 380 } 381 382 // ----------------------------------------------------------------------------- 383 // Class Loader support to workaround deadlocks on the class loader lock objects 384 // Also used by GC 385 // complete_exit()/reenter() are used to wait on a nested lock 386 // i.e. to give up an outer lock completely and then re-enter 387 // Used when holding nested locks - lock acquisition order: lock1 then lock2 388 // 1) complete_exit lock1 - saving recursion count 389 // 2) wait on lock2 390 // 3) when notified on lock2, unlock lock2 391 // 4) reenter lock1 with original recursion count 392 // 5) lock lock2 393 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 394 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 395 TEVENT(complete_exit); 396 if (UseBiasedLocking) { 397 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 398 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 399 } 400 401 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 402 obj(), 403 inflate_cause_vm_internal); 404 405 return monitor->complete_exit(THREAD); 406 } 407 408 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 409 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 410 TEVENT(reenter); 411 if (UseBiasedLocking) { 412 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 413 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 414 } 415 416 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 417 obj(), 418 inflate_cause_vm_internal); 419 420 monitor->reenter(recursion, THREAD); 421 } 422 // ----------------------------------------------------------------------------- 423 // JNI locks on java objects 424 // NOTE: must use heavy weight monitor to handle jni monitor enter 425 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 426 // the current locking is from JNI instead of Java code 427 TEVENT(jni_enter); 428 if (UseBiasedLocking) { 429 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 430 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 431 } 432 THREAD->set_current_pending_monitor_is_from_java(false); 433 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD); 434 THREAD->set_current_pending_monitor_is_from_java(true); 435 } 436 437 // NOTE: must use heavy weight monitor to handle jni monitor exit 438 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 439 TEVENT(jni_exit); 440 if (UseBiasedLocking) { 441 Handle h_obj(THREAD, obj); 442 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 443 obj = h_obj(); 444 } 445 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 446 447 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 448 obj, 449 inflate_cause_jni_exit); 450 // If this thread has locked the object, exit the monitor. Note: can't use 451 // monitor->check(CHECK); must exit even if an exception is pending. 452 if (monitor->check(THREAD)) { 453 monitor->exit(true, THREAD); 454 } 455 } 456 457 // ----------------------------------------------------------------------------- 458 // Internal VM locks on java objects 459 // standard constructor, allows locking failures 460 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 461 _dolock = doLock; 462 _thread = thread; 463 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 464 _obj = obj; 465 466 if (_dolock) { 467 TEVENT(ObjectLocker); 468 469 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 470 } 471 } 472 473 ObjectLocker::~ObjectLocker() { 474 if (_dolock) { 475 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 476 } 477 } 478 479 480 // ----------------------------------------------------------------------------- 481 // Wait/Notify/NotifyAll 482 // NOTE: must use heavy weight monitor to handle wait() 483 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 484 if (UseBiasedLocking) { 485 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 486 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 487 } 488 if (millis < 0) { 489 TEVENT(wait - throw IAX); 490 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 491 } 492 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, 493 obj(), 494 inflate_cause_wait); 495 496 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 497 monitor->wait(millis, true, THREAD); 498 499 // This dummy call is in place to get around dtrace bug 6254741. Once 500 // that's fixed we can uncomment the following line, remove the call 501 // and change this function back into a "void" func. 502 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 503 return dtrace_waited_probe(monitor, obj, THREAD); 504 } 505 506 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 507 if (UseBiasedLocking) { 508 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 509 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 510 } 511 if (millis < 0) { 512 TEVENT(wait - throw IAX); 513 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 514 } 515 ObjectSynchronizer::inflate(THREAD, 516 obj(), 517 inflate_cause_wait)->wait(millis, false, THREAD); 518 } 519 520 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 521 if (UseBiasedLocking) { 522 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 523 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 524 } 525 526 markOop mark = obj->mark(); 527 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 528 return; 529 } 530 ObjectSynchronizer::inflate(THREAD, 531 obj(), 532 inflate_cause_notify)->notify(THREAD); 533 } 534 535 // NOTE: see comment of notify() 536 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 537 if (UseBiasedLocking) { 538 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 539 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 540 } 541 542 markOop mark = obj->mark(); 543 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 544 return; 545 } 546 ObjectSynchronizer::inflate(THREAD, 547 obj(), 548 inflate_cause_notify)->notifyAll(THREAD); 549 } 550 551 // ----------------------------------------------------------------------------- 552 // Hash Code handling 553 // 554 // Performance concern: 555 // OrderAccess::storestore() calls release() which at one time stored 0 556 // into the global volatile OrderAccess::dummy variable. This store was 557 // unnecessary for correctness. Many threads storing into a common location 558 // causes considerable cache migration or "sloshing" on large SMP systems. 559 // As such, I avoided using OrderAccess::storestore(). In some cases 560 // OrderAccess::fence() -- which incurs local latency on the executing 561 // processor -- is a better choice as it scales on SMP systems. 562 // 563 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 564 // a discussion of coherency costs. Note that all our current reference 565 // platforms provide strong ST-ST order, so the issue is moot on IA32, 566 // x64, and SPARC. 567 // 568 // As a general policy we use "volatile" to control compiler-based reordering 569 // and explicit fences (barriers) to control for architectural reordering 570 // performed by the CPU(s) or platform. 571 572 struct SharedGlobals { 573 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE]; 574 // These are highly shared mostly-read variables. 575 // To avoid false-sharing they need to be the sole occupants of a cache line. 576 volatile int stwRandom; 577 volatile int stwCycle; 578 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 579 // Hot RW variable -- Sequester to avoid false-sharing 580 volatile int hcSequence; 581 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int)); 582 }; 583 584 static SharedGlobals GVars; 585 static int MonitorScavengeThreshold = 1000000; 586 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 587 588 static markOop ReadStableMark(oop obj) { 589 markOop mark = obj->mark(); 590 if (!mark->is_being_inflated()) { 591 return mark; // normal fast-path return 592 } 593 594 int its = 0; 595 for (;;) { 596 markOop mark = obj->mark(); 597 if (!mark->is_being_inflated()) { 598 return mark; // normal fast-path return 599 } 600 601 // The object is being inflated by some other thread. 602 // The caller of ReadStableMark() must wait for inflation to complete. 603 // Avoid live-lock 604 // TODO: consider calling SafepointSynchronize::do_call_back() while 605 // spinning to see if there's a safepoint pending. If so, immediately 606 // yielding or blocking would be appropriate. Avoid spinning while 607 // there is a safepoint pending. 608 // TODO: add inflation contention performance counters. 609 // TODO: restrict the aggregate number of spinners. 610 611 ++its; 612 if (its > 10000 || !os::is_MP()) { 613 if (its & 1) { 614 os::naked_yield(); 615 TEVENT(Inflate: INFLATING - yield); 616 } else { 617 // Note that the following code attenuates the livelock problem but is not 618 // a complete remedy. A more complete solution would require that the inflating 619 // thread hold the associated inflation lock. The following code simply restricts 620 // the number of spinners to at most one. We'll have N-2 threads blocked 621 // on the inflationlock, 1 thread holding the inflation lock and using 622 // a yield/park strategy, and 1 thread in the midst of inflation. 623 // A more refined approach would be to change the encoding of INFLATING 624 // to allow encapsulation of a native thread pointer. Threads waiting for 625 // inflation to complete would use CAS to push themselves onto a singly linked 626 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 627 // and calling park(). When inflation was complete the thread that accomplished inflation 628 // would detach the list and set the markword to inflated with a single CAS and 629 // then for each thread on the list, set the flag and unpark() the thread. 630 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 631 // wakes at most one thread whereas we need to wake the entire list. 632 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 633 int YieldThenBlock = 0; 634 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 635 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 636 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 637 while (obj->mark() == markOopDesc::INFLATING()) { 638 // Beware: NakedYield() is advisory and has almost no effect on some platforms 639 // so we periodically call Self->_ParkEvent->park(1). 640 // We use a mixed spin/yield/block mechanism. 641 if ((YieldThenBlock++) >= 16) { 642 Thread::current()->_ParkEvent->park(1); 643 } else { 644 os::naked_yield(); 645 } 646 } 647 Thread::muxRelease(gInflationLocks + ix); 648 TEVENT(Inflate: INFLATING - yield/park); 649 } 650 } else { 651 SpinPause(); // SMP-polite spinning 652 } 653 } 654 } 655 656 // hashCode() generation : 657 // 658 // Possibilities: 659 // * MD5Digest of {obj,stwRandom} 660 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 661 // * A DES- or AES-style SBox[] mechanism 662 // * One of the Phi-based schemes, such as: 663 // 2654435761 = 2^32 * Phi (golden ratio) 664 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 665 // * A variation of Marsaglia's shift-xor RNG scheme. 666 // * (obj ^ stwRandom) is appealing, but can result 667 // in undesirable regularity in the hashCode values of adjacent objects 668 // (objects allocated back-to-back, in particular). This could potentially 669 // result in hashtable collisions and reduced hashtable efficiency. 670 // There are simple ways to "diffuse" the middle address bits over the 671 // generated hashCode values: 672 673 static inline intptr_t get_next_hash(Thread * Self, oop obj) { 674 intptr_t value = 0; 675 if (hashCode == 0) { 676 // This form uses global Park-Miller RNG. 677 // On MP system we'll have lots of RW access to a global, so the 678 // mechanism induces lots of coherency traffic. 679 value = os::random(); 680 } else if (hashCode == 1) { 681 // This variation has the property of being stable (idempotent) 682 // between STW operations. This can be useful in some of the 1-0 683 // synchronization schemes. 684 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 685 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 686 } else if (hashCode == 2) { 687 value = 1; // for sensitivity testing 688 } else if (hashCode == 3) { 689 value = ++GVars.hcSequence; 690 } else if (hashCode == 4) { 691 value = cast_from_oop<intptr_t>(obj); 692 } else { 693 // Marsaglia's xor-shift scheme with thread-specific state 694 // This is probably the best overall implementation -- we'll 695 // likely make this the default in future releases. 696 unsigned t = Self->_hashStateX; 697 t ^= (t << 11); 698 Self->_hashStateX = Self->_hashStateY; 699 Self->_hashStateY = Self->_hashStateZ; 700 Self->_hashStateZ = Self->_hashStateW; 701 unsigned v = Self->_hashStateW; 702 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 703 Self->_hashStateW = v; 704 value = v; 705 } 706 707 value &= markOopDesc::hash_mask; 708 if (value == 0) value = 0xBAD; 709 assert(value != markOopDesc::no_hash, "invariant"); 710 TEVENT(hashCode: GENERATE); 711 return value; 712 } 713 714 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 715 if (UseBiasedLocking) { 716 // NOTE: many places throughout the JVM do not expect a safepoint 717 // to be taken here, in particular most operations on perm gen 718 // objects. However, we only ever bias Java instances and all of 719 // the call sites of identity_hash that might revoke biases have 720 // been checked to make sure they can handle a safepoint. The 721 // added check of the bias pattern is to avoid useless calls to 722 // thread-local storage. 723 if (obj->mark()->has_bias_pattern()) { 724 // Handle for oop obj in case of STW safepoint 725 Handle hobj(Self, obj); 726 // Relaxing assertion for bug 6320749. 727 assert(Universe::verify_in_progress() || 728 !SafepointSynchronize::is_at_safepoint(), 729 "biases should not be seen by VM thread here"); 730 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 731 obj = hobj(); 732 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 733 } 734 } 735 736 // hashCode() is a heap mutator ... 737 // Relaxing assertion for bug 6320749. 738 assert(Universe::verify_in_progress() || DumpSharedSpaces || 739 !SafepointSynchronize::is_at_safepoint(), "invariant"); 740 assert(Universe::verify_in_progress() || DumpSharedSpaces || 741 Self->is_Java_thread() , "invariant"); 742 assert(Universe::verify_in_progress() || DumpSharedSpaces || 743 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 744 745 ObjectMonitor* monitor = NULL; 746 markOop temp, test; 747 intptr_t hash; 748 markOop mark = ReadStableMark(obj); 749 750 // object should remain ineligible for biased locking 751 assert(!mark->has_bias_pattern(), "invariant"); 752 753 if (mark->is_neutral()) { 754 hash = mark->hash(); // this is a normal header 755 if (hash) { // if it has hash, just return it 756 return hash; 757 } 758 hash = get_next_hash(Self, obj); // allocate a new hash code 759 temp = mark->copy_set_hash(hash); // merge the hash code into header 760 // use (machine word version) atomic operation to install the hash 761 test = obj->cas_set_mark(temp, mark); 762 if (test == mark) { 763 return hash; 764 } 765 // If atomic operation failed, we must inflate the header 766 // into heavy weight monitor. We could add more code here 767 // for fast path, but it does not worth the complexity. 768 } else if (mark->has_monitor()) { 769 monitor = mark->monitor(); 770 temp = monitor->header(); 771 assert(temp->is_neutral(), "invariant"); 772 hash = temp->hash(); 773 if (hash) { 774 return hash; 775 } 776 // Skip to the following code to reduce code size 777 } else if (Self->is_lock_owned((address)mark->locker())) { 778 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 779 assert(temp->is_neutral(), "invariant"); 780 hash = temp->hash(); // by current thread, check if the displaced 781 if (hash) { // header contains hash code 782 return hash; 783 } 784 // WARNING: 785 // The displaced header is strictly immutable. 786 // It can NOT be changed in ANY cases. So we have 787 // to inflate the header into heavyweight monitor 788 // even the current thread owns the lock. The reason 789 // is the BasicLock (stack slot) will be asynchronously 790 // read by other threads during the inflate() function. 791 // Any change to stack may not propagate to other threads 792 // correctly. 793 } 794 795 // Inflate the monitor to set hash code 796 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code); 797 // Load displaced header and check it has hash code 798 mark = monitor->header(); 799 assert(mark->is_neutral(), "invariant"); 800 hash = mark->hash(); 801 if (hash == 0) { 802 hash = get_next_hash(Self, obj); 803 temp = mark->copy_set_hash(hash); // merge hash code into header 804 assert(temp->is_neutral(), "invariant"); 805 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark); 806 if (test != mark) { 807 // The only update to the header in the monitor (outside GC) 808 // is install the hash code. If someone add new usage of 809 // displaced header, please update this code 810 hash = test->hash(); 811 assert(test->is_neutral(), "invariant"); 812 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 813 } 814 } 815 // We finally get the hash 816 return hash; 817 } 818 819 // Deprecated -- use FastHashCode() instead. 820 821 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 822 return FastHashCode(Thread::current(), obj()); 823 } 824 825 826 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 827 Handle h_obj) { 828 if (UseBiasedLocking) { 829 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 830 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 831 } 832 833 assert(thread == JavaThread::current(), "Can only be called on current thread"); 834 oop obj = h_obj(); 835 836 markOop mark = ReadStableMark(obj); 837 838 // Uncontended case, header points to stack 839 if (mark->has_locker()) { 840 return thread->is_lock_owned((address)mark->locker()); 841 } 842 // Contended case, header points to ObjectMonitor (tagged pointer) 843 if (mark->has_monitor()) { 844 ObjectMonitor* monitor = mark->monitor(); 845 return monitor->is_entered(thread) != 0; 846 } 847 // Unlocked case, header in place 848 assert(mark->is_neutral(), "sanity check"); 849 return false; 850 } 851 852 // Be aware of this method could revoke bias of the lock object. 853 // This method queries the ownership of the lock handle specified by 'h_obj'. 854 // If the current thread owns the lock, it returns owner_self. If no 855 // thread owns the lock, it returns owner_none. Otherwise, it will return 856 // owner_other. 857 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 858 (JavaThread *self, Handle h_obj) { 859 // The caller must beware this method can revoke bias, and 860 // revocation can result in a safepoint. 861 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 862 assert(self->thread_state() != _thread_blocked, "invariant"); 863 864 // Possible mark states: neutral, biased, stack-locked, inflated 865 866 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 867 // CASE: biased 868 BiasedLocking::revoke_and_rebias(h_obj, false, self); 869 assert(!h_obj->mark()->has_bias_pattern(), 870 "biases should be revoked by now"); 871 } 872 873 assert(self == JavaThread::current(), "Can only be called on current thread"); 874 oop obj = h_obj(); 875 markOop mark = ReadStableMark(obj); 876 877 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 878 if (mark->has_locker()) { 879 return self->is_lock_owned((address)mark->locker()) ? 880 owner_self : owner_other; 881 } 882 883 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 884 // The Object:ObjectMonitor relationship is stable as long as we're 885 // not at a safepoint. 886 if (mark->has_monitor()) { 887 void * owner = mark->monitor()->_owner; 888 if (owner == NULL) return owner_none; 889 return (owner == self || 890 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 891 } 892 893 // CASE: neutral 894 assert(mark->is_neutral(), "sanity check"); 895 return owner_none; // it's unlocked 896 } 897 898 // FIXME: jvmti should call this 899 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 900 if (UseBiasedLocking) { 901 if (SafepointSynchronize::is_at_safepoint()) { 902 BiasedLocking::revoke_at_safepoint(h_obj); 903 } else { 904 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 905 } 906 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 907 } 908 909 oop obj = h_obj(); 910 address owner = NULL; 911 912 markOop mark = ReadStableMark(obj); 913 914 // Uncontended case, header points to stack 915 if (mark->has_locker()) { 916 owner = (address) mark->locker(); 917 } 918 919 // Contended case, header points to ObjectMonitor (tagged pointer) 920 if (mark->has_monitor()) { 921 ObjectMonitor* monitor = mark->monitor(); 922 assert(monitor != NULL, "monitor should be non-null"); 923 owner = (address) monitor->owner(); 924 } 925 926 if (owner != NULL) { 927 // owning_thread_from_monitor_owner() may also return NULL here 928 return Threads::owning_thread_from_monitor_owner(t_list, owner); 929 } 930 931 // Unlocked case, header in place 932 // Cannot have assertion since this object may have been 933 // locked by another thread when reaching here. 934 // assert(mark->is_neutral(), "sanity check"); 935 936 return NULL; 937 } 938 939 // Visitors ... 940 941 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 942 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 943 while (block != NULL) { 944 assert(block->object() == CHAINMARKER, "must be a block header"); 945 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 946 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 947 oop object = (oop)mid->object(); 948 if (object != NULL) { 949 closure->do_monitor(mid); 950 } 951 } 952 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 953 } 954 } 955 956 // Get the next block in the block list. 957 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) { 958 assert(block->object() == CHAINMARKER, "must be a block header"); 959 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext; 960 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 961 return block; 962 } 963 964 static bool monitors_used_above_threshold() { 965 if (gMonitorPopulation == 0) { 966 return false; 967 } 968 int monitors_used = gMonitorPopulation - gMonitorFreeCount; 969 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation; 970 return monitor_usage > MonitorUsedDeflationThreshold; 971 } 972 973 bool ObjectSynchronizer::is_cleanup_needed() { 974 if (MonitorUsedDeflationThreshold > 0) { 975 return monitors_used_above_threshold(); 976 } 977 return false; 978 } 979 980 void ObjectSynchronizer::oops_do(OopClosure* f) { 981 if (MonitorInUseLists) { 982 // When using thread local monitor lists, we only scan the 983 // global used list here (for moribund threads), and 984 // the thread-local monitors in Thread::oops_do(). 985 global_used_oops_do(f); 986 } else { 987 global_oops_do(f); 988 } 989 } 990 991 void ObjectSynchronizer::global_oops_do(OopClosure* f) { 992 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 993 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 994 for (; block != NULL; block = next(block)) { 995 assert(block->object() == CHAINMARKER, "must be a block header"); 996 for (int i = 1; i < _BLOCKSIZE; i++) { 997 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 998 if (mid->object() != NULL) { 999 f->do_oop((oop*)mid->object_addr()); 1000 } 1001 } 1002 } 1003 } 1004 1005 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1006 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1007 list_oops_do(gOmInUseList, f); 1008 } 1009 1010 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1011 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1012 list_oops_do(thread->omInUseList, f); 1013 } 1014 1015 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) { 1016 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1017 ObjectMonitor* mid; 1018 for (mid = list; mid != NULL; mid = mid->FreeNext) { 1019 if (mid->object() != NULL) { 1020 f->do_oop((oop*)mid->object_addr()); 1021 } 1022 } 1023 } 1024 1025 1026 // ----------------------------------------------------------------------------- 1027 // ObjectMonitor Lifecycle 1028 // ----------------------- 1029 // Inflation unlinks monitors from the global gFreeList and 1030 // associates them with objects. Deflation -- which occurs at 1031 // STW-time -- disassociates idle monitors from objects. Such 1032 // scavenged monitors are returned to the gFreeList. 1033 // 1034 // The global list is protected by gListLock. All the critical sections 1035 // are short and operate in constant-time. 1036 // 1037 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1038 // 1039 // Lifecycle: 1040 // -- unassigned and on the global free list 1041 // -- unassigned and on a thread's private omFreeList 1042 // -- assigned to an object. The object is inflated and the mark refers 1043 // to the objectmonitor. 1044 1045 1046 // Constraining monitor pool growth via MonitorBound ... 1047 // 1048 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1049 // the rate of scavenging is driven primarily by GC. As such, we can find 1050 // an inordinate number of monitors in circulation. 1051 // To avoid that scenario we can artificially induce a STW safepoint 1052 // if the pool appears to be growing past some reasonable bound. 1053 // Generally we favor time in space-time tradeoffs, but as there's no 1054 // natural back-pressure on the # of extant monitors we need to impose some 1055 // type of limit. Beware that if MonitorBound is set to too low a value 1056 // we could just loop. In addition, if MonitorBound is set to a low value 1057 // we'll incur more safepoints, which are harmful to performance. 1058 // See also: GuaranteedSafepointInterval 1059 // 1060 // The current implementation uses asynchronous VM operations. 1061 1062 static void InduceScavenge(Thread * Self, const char * Whence) { 1063 // Induce STW safepoint to trim monitors 1064 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1065 // More precisely, trigger an asynchronous STW safepoint as the number 1066 // of active monitors passes the specified threshold. 1067 // TODO: assert thread state is reasonable 1068 1069 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 1070 if (ObjectMonitor::Knob_Verbose) { 1071 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)", 1072 Whence, ForceMonitorScavenge) ; 1073 tty->flush(); 1074 } 1075 // Induce a 'null' safepoint to scavenge monitors 1076 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 1077 // to the VMthread and have a lifespan longer than that of this activation record. 1078 // The VMThread will delete the op when completed. 1079 VMThread::execute(new VM_ScavengeMonitors()); 1080 1081 if (ObjectMonitor::Knob_Verbose) { 1082 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)", 1083 Whence, ForceMonitorScavenge) ; 1084 tty->flush(); 1085 } 1086 } 1087 } 1088 1089 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1090 ObjectMonitor* mid; 1091 int in_use_tally = 0; 1092 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1093 in_use_tally++; 1094 } 1095 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1096 1097 int free_tally = 0; 1098 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1099 free_tally++; 1100 } 1101 assert(free_tally == Self->omFreeCount, "free count off"); 1102 } 1103 1104 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) { 1105 // A large MAXPRIVATE value reduces both list lock contention 1106 // and list coherency traffic, but also tends to increase the 1107 // number of objectMonitors in circulation as well as the STW 1108 // scavenge costs. As usual, we lean toward time in space-time 1109 // tradeoffs. 1110 const int MAXPRIVATE = 1024; 1111 for (;;) { 1112 ObjectMonitor * m; 1113 1114 // 1: try to allocate from the thread's local omFreeList. 1115 // Threads will attempt to allocate first from their local list, then 1116 // from the global list, and only after those attempts fail will the thread 1117 // attempt to instantiate new monitors. Thread-local free lists take 1118 // heat off the gListLock and improve allocation latency, as well as reducing 1119 // coherency traffic on the shared global list. 1120 m = Self->omFreeList; 1121 if (m != NULL) { 1122 Self->omFreeList = m->FreeNext; 1123 Self->omFreeCount--; 1124 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1125 guarantee(m->object() == NULL, "invariant"); 1126 if (MonitorInUseLists) { 1127 m->FreeNext = Self->omInUseList; 1128 Self->omInUseList = m; 1129 Self->omInUseCount++; 1130 if (ObjectMonitor::Knob_VerifyInUse) { 1131 verifyInUse(Self); 1132 } 1133 } else { 1134 m->FreeNext = NULL; 1135 } 1136 return m; 1137 } 1138 1139 // 2: try to allocate from the global gFreeList 1140 // CONSIDER: use muxTry() instead of muxAcquire(). 1141 // If the muxTry() fails then drop immediately into case 3. 1142 // If we're using thread-local free lists then try 1143 // to reprovision the caller's free list. 1144 if (gFreeList != NULL) { 1145 // Reprovision the thread's omFreeList. 1146 // Use bulk transfers to reduce the allocation rate and heat 1147 // on various locks. 1148 Thread::muxAcquire(&gListLock, "omAlloc"); 1149 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1150 gMonitorFreeCount--; 1151 ObjectMonitor * take = gFreeList; 1152 gFreeList = take->FreeNext; 1153 guarantee(take->object() == NULL, "invariant"); 1154 guarantee(!take->is_busy(), "invariant"); 1155 take->Recycle(); 1156 omRelease(Self, take, false); 1157 } 1158 Thread::muxRelease(&gListLock); 1159 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1160 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1161 TEVENT(omFirst - reprovision); 1162 1163 const int mx = MonitorBound; 1164 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1165 // We can't safely induce a STW safepoint from omAlloc() as our thread 1166 // state may not be appropriate for such activities and callers may hold 1167 // naked oops, so instead we defer the action. 1168 InduceScavenge(Self, "omAlloc"); 1169 } 1170 continue; 1171 } 1172 1173 // 3: allocate a block of new ObjectMonitors 1174 // Both the local and global free lists are empty -- resort to malloc(). 1175 // In the current implementation objectMonitors are TSM - immortal. 1176 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1177 // each ObjectMonitor to start at the beginning of a cache line, 1178 // so we use align_up(). 1179 // A better solution would be to use C++ placement-new. 1180 // BEWARE: As it stands currently, we don't run the ctors! 1181 assert(_BLOCKSIZE > 1, "invariant"); 1182 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1183 PaddedEnd<ObjectMonitor> * temp; 1184 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1185 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1186 mtInternal); 1187 temp = (PaddedEnd<ObjectMonitor> *) 1188 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE); 1189 1190 // NOTE: (almost) no way to recover if allocation failed. 1191 // We might be able to induce a STW safepoint and scavenge enough 1192 // objectMonitors to permit progress. 1193 if (temp == NULL) { 1194 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1195 "Allocate ObjectMonitors"); 1196 } 1197 (void)memset((void *) temp, 0, neededsize); 1198 1199 // Format the block. 1200 // initialize the linked list, each monitor points to its next 1201 // forming the single linked free list, the very first monitor 1202 // will points to next block, which forms the block list. 1203 // The trick of using the 1st element in the block as gBlockList 1204 // linkage should be reconsidered. A better implementation would 1205 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1206 1207 for (int i = 1; i < _BLOCKSIZE; i++) { 1208 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1209 } 1210 1211 // terminate the last monitor as the end of list 1212 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1213 1214 // Element [0] is reserved for global list linkage 1215 temp[0].set_object(CHAINMARKER); 1216 1217 // Consider carving out this thread's current request from the 1218 // block in hand. This avoids some lock traffic and redundant 1219 // list activity. 1220 1221 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1222 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1223 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1224 gMonitorPopulation += _BLOCKSIZE-1; 1225 gMonitorFreeCount += _BLOCKSIZE-1; 1226 1227 // Add the new block to the list of extant blocks (gBlockList). 1228 // The very first objectMonitor in a block is reserved and dedicated. 1229 // It serves as blocklist "next" linkage. 1230 temp[0].FreeNext = gBlockList; 1231 // There are lock-free uses of gBlockList so make sure that 1232 // the previous stores happen before we update gBlockList. 1233 OrderAccess::release_store(&gBlockList, temp); 1234 1235 // Add the new string of objectMonitors to the global free list 1236 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1237 gFreeList = temp + 1; 1238 Thread::muxRelease(&gListLock); 1239 TEVENT(Allocate block of monitors); 1240 } 1241 } 1242 1243 // Place "m" on the caller's private per-thread omFreeList. 1244 // In practice there's no need to clamp or limit the number of 1245 // monitors on a thread's omFreeList as the only time we'll call 1246 // omRelease is to return a monitor to the free list after a CAS 1247 // attempt failed. This doesn't allow unbounded #s of monitors to 1248 // accumulate on a thread's free list. 1249 // 1250 // Key constraint: all ObjectMonitors on a thread's free list and the global 1251 // free list must have their object field set to null. This prevents the 1252 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1253 1254 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1255 bool fromPerThreadAlloc) { 1256 guarantee(m->object() == NULL, "invariant"); 1257 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1258 // Remove from omInUseList 1259 if (MonitorInUseLists && fromPerThreadAlloc) { 1260 ObjectMonitor* cur_mid_in_use = NULL; 1261 bool extracted = false; 1262 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1263 if (m == mid) { 1264 // extract from per-thread in-use list 1265 if (mid == Self->omInUseList) { 1266 Self->omInUseList = mid->FreeNext; 1267 } else if (cur_mid_in_use != NULL) { 1268 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1269 } 1270 extracted = true; 1271 Self->omInUseCount--; 1272 if (ObjectMonitor::Knob_VerifyInUse) { 1273 verifyInUse(Self); 1274 } 1275 break; 1276 } 1277 } 1278 assert(extracted, "Should have extracted from in-use list"); 1279 } 1280 1281 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1282 m->FreeNext = Self->omFreeList; 1283 Self->omFreeList = m; 1284 Self->omFreeCount++; 1285 } 1286 1287 // Return the monitors of a moribund thread's local free list to 1288 // the global free list. Typically a thread calls omFlush() when 1289 // it's dying. We could also consider having the VM thread steal 1290 // monitors from threads that have not run java code over a few 1291 // consecutive STW safepoints. Relatedly, we might decay 1292 // omFreeProvision at STW safepoints. 1293 // 1294 // Also return the monitors of a moribund thread's omInUseList to 1295 // a global gOmInUseList under the global list lock so these 1296 // will continue to be scanned. 1297 // 1298 // We currently call omFlush() from Threads::remove() _before the thread 1299 // has been excised from the thread list and is no longer a mutator. 1300 // This means that omFlush() can not run concurrently with a safepoint and 1301 // interleave with the scavenge operator. In particular, this ensures that 1302 // the thread's monitors are scanned by a GC safepoint, either via 1303 // Thread::oops_do() (if safepoint happens before omFlush()) or via 1304 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's 1305 // monitors have been transferred to the global in-use list). 1306 1307 void ObjectSynchronizer::omFlush(Thread * Self) { 1308 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1309 Self->omFreeList = NULL; 1310 ObjectMonitor * tail = NULL; 1311 int tally = 0; 1312 if (list != NULL) { 1313 ObjectMonitor * s; 1314 // The thread is going away, the per-thread free monitors 1315 // are freed via set_owner(NULL) 1316 // Link them to tail, which will be linked into the global free list 1317 // gFreeList below, under the gListLock 1318 for (s = list; s != NULL; s = s->FreeNext) { 1319 tally++; 1320 tail = s; 1321 guarantee(s->object() == NULL, "invariant"); 1322 guarantee(!s->is_busy(), "invariant"); 1323 s->set_owner(NULL); // redundant but good hygiene 1324 TEVENT(omFlush - Move one); 1325 } 1326 guarantee(tail != NULL && list != NULL, "invariant"); 1327 } 1328 1329 ObjectMonitor * inUseList = Self->omInUseList; 1330 ObjectMonitor * inUseTail = NULL; 1331 int inUseTally = 0; 1332 if (inUseList != NULL) { 1333 Self->omInUseList = NULL; 1334 ObjectMonitor *cur_om; 1335 // The thread is going away, however the omInUseList inflated 1336 // monitors may still be in-use by other threads. 1337 // Link them to inUseTail, which will be linked into the global in-use list 1338 // gOmInUseList below, under the gListLock 1339 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1340 inUseTail = cur_om; 1341 inUseTally++; 1342 } 1343 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1344 Self->omInUseCount = 0; 1345 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1346 } 1347 1348 Thread::muxAcquire(&gListLock, "omFlush"); 1349 if (tail != NULL) { 1350 tail->FreeNext = gFreeList; 1351 gFreeList = list; 1352 gMonitorFreeCount += tally; 1353 assert(Self->omFreeCount == tally, "free-count off"); 1354 Self->omFreeCount = 0; 1355 } 1356 1357 if (inUseTail != NULL) { 1358 inUseTail->FreeNext = gOmInUseList; 1359 gOmInUseList = inUseList; 1360 gOmInUseCount += inUseTally; 1361 } 1362 1363 Thread::muxRelease(&gListLock); 1364 TEVENT(omFlush); 1365 } 1366 1367 // Fast path code shared by multiple functions 1368 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1369 markOop mark = obj->mark(); 1370 if (mark->has_monitor()) { 1371 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1372 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1373 return mark->monitor(); 1374 } 1375 return ObjectSynchronizer::inflate(Thread::current(), 1376 obj, 1377 inflate_cause_vm_internal); 1378 } 1379 1380 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self, 1381 oop object, 1382 const InflateCause cause) { 1383 1384 // Inflate mutates the heap ... 1385 // Relaxing assertion for bug 6320749. 1386 assert(Universe::verify_in_progress() || 1387 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1388 1389 EventJavaMonitorInflate event; 1390 1391 for (;;) { 1392 const markOop mark = object->mark(); 1393 assert(!mark->has_bias_pattern(), "invariant"); 1394 1395 // The mark can be in one of the following states: 1396 // * Inflated - just return 1397 // * Stack-locked - coerce it to inflated 1398 // * INFLATING - busy wait for conversion to complete 1399 // * Neutral - aggressively inflate the object. 1400 // * BIASED - Illegal. We should never see this 1401 1402 // CASE: inflated 1403 if (mark->has_monitor()) { 1404 ObjectMonitor * inf = mark->monitor(); 1405 assert(inf->header()->is_neutral(), "invariant"); 1406 assert(inf->object() == object, "invariant"); 1407 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1408 return inf; 1409 } 1410 1411 // CASE: inflation in progress - inflating over a stack-lock. 1412 // Some other thread is converting from stack-locked to inflated. 1413 // Only that thread can complete inflation -- other threads must wait. 1414 // The INFLATING value is transient. 1415 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1416 // We could always eliminate polling by parking the thread on some auxiliary list. 1417 if (mark == markOopDesc::INFLATING()) { 1418 TEVENT(Inflate: spin while INFLATING); 1419 ReadStableMark(object); 1420 continue; 1421 } 1422 1423 // CASE: stack-locked 1424 // Could be stack-locked either by this thread or by some other thread. 1425 // 1426 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1427 // to install INFLATING into the mark word. We originally installed INFLATING, 1428 // allocated the objectmonitor, and then finally STed the address of the 1429 // objectmonitor into the mark. This was correct, but artificially lengthened 1430 // the interval in which INFLATED appeared in the mark, thus increasing 1431 // the odds of inflation contention. 1432 // 1433 // We now use per-thread private objectmonitor free lists. 1434 // These list are reprovisioned from the global free list outside the 1435 // critical INFLATING...ST interval. A thread can transfer 1436 // multiple objectmonitors en-mass from the global free list to its local free list. 1437 // This reduces coherency traffic and lock contention on the global free list. 1438 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1439 // before or after the CAS(INFLATING) operation. 1440 // See the comments in omAlloc(). 1441 1442 if (mark->has_locker()) { 1443 ObjectMonitor * m = omAlloc(Self); 1444 // Optimistically prepare the objectmonitor - anticipate successful CAS 1445 // We do this before the CAS in order to minimize the length of time 1446 // in which INFLATING appears in the mark. 1447 m->Recycle(); 1448 m->_Responsible = NULL; 1449 m->_recursions = 0; 1450 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1451 1452 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark); 1453 if (cmp != mark) { 1454 omRelease(Self, m, true); 1455 continue; // Interference -- just retry 1456 } 1457 1458 // We've successfully installed INFLATING (0) into the mark-word. 1459 // This is the only case where 0 will appear in a mark-word. 1460 // Only the singular thread that successfully swings the mark-word 1461 // to 0 can perform (or more precisely, complete) inflation. 1462 // 1463 // Why do we CAS a 0 into the mark-word instead of just CASing the 1464 // mark-word from the stack-locked value directly to the new inflated state? 1465 // Consider what happens when a thread unlocks a stack-locked object. 1466 // It attempts to use CAS to swing the displaced header value from the 1467 // on-stack basiclock back into the object header. Recall also that the 1468 // header value (hashcode, etc) can reside in (a) the object header, or 1469 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1470 // header in an objectMonitor. The inflate() routine must copy the header 1471 // value from the basiclock on the owner's stack to the objectMonitor, all 1472 // the while preserving the hashCode stability invariants. If the owner 1473 // decides to release the lock while the value is 0, the unlock will fail 1474 // and control will eventually pass from slow_exit() to inflate. The owner 1475 // will then spin, waiting for the 0 value to disappear. Put another way, 1476 // the 0 causes the owner to stall if the owner happens to try to 1477 // drop the lock (restoring the header from the basiclock to the object) 1478 // while inflation is in-progress. This protocol avoids races that might 1479 // would otherwise permit hashCode values to change or "flicker" for an object. 1480 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1481 // 0 serves as a "BUSY" inflate-in-progress indicator. 1482 1483 1484 // fetch the displaced mark from the owner's stack. 1485 // The owner can't die or unwind past the lock while our INFLATING 1486 // object is in the mark. Furthermore the owner can't complete 1487 // an unlock on the object, either. 1488 markOop dmw = mark->displaced_mark_helper(); 1489 assert(dmw->is_neutral(), "invariant"); 1490 1491 // Setup monitor fields to proper values -- prepare the monitor 1492 m->set_header(dmw); 1493 1494 // Optimization: if the mark->locker stack address is associated 1495 // with this thread we could simply set m->_owner = Self. 1496 // Note that a thread can inflate an object 1497 // that it has stack-locked -- as might happen in wait() -- directly 1498 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1499 m->set_owner(mark->locker()); 1500 m->set_object(object); 1501 // TODO-FIXME: assert BasicLock->dhw != 0. 1502 1503 // Must preserve store ordering. The monitor state must 1504 // be stable at the time of publishing the monitor address. 1505 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1506 object->release_set_mark(markOopDesc::encode(m)); 1507 1508 // Hopefully the performance counters are allocated on distinct cache lines 1509 // to avoid false sharing on MP systems ... 1510 OM_PERFDATA_OP(Inflations, inc()); 1511 TEVENT(Inflate: overwrite stacklock); 1512 if (log_is_enabled(Debug, monitorinflation)) { 1513 if (object->is_instance()) { 1514 ResourceMark rm; 1515 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1516 p2i(object), p2i(object->mark()), 1517 object->klass()->external_name()); 1518 } 1519 } 1520 if (event.should_commit()) { 1521 post_monitor_inflate_event(event, object, cause); 1522 } 1523 return m; 1524 } 1525 1526 // CASE: neutral 1527 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1528 // If we know we're inflating for entry it's better to inflate by swinging a 1529 // pre-locked objectMonitor pointer into the object header. A successful 1530 // CAS inflates the object *and* confers ownership to the inflating thread. 1531 // In the current implementation we use a 2-step mechanism where we CAS() 1532 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1533 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1534 // would be useful. 1535 1536 assert(mark->is_neutral(), "invariant"); 1537 ObjectMonitor * m = omAlloc(Self); 1538 // prepare m for installation - set monitor to initial state 1539 m->Recycle(); 1540 m->set_header(mark); 1541 m->set_owner(NULL); 1542 m->set_object(object); 1543 m->_recursions = 0; 1544 m->_Responsible = NULL; 1545 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1546 1547 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) { 1548 m->set_object(NULL); 1549 m->set_owner(NULL); 1550 m->Recycle(); 1551 omRelease(Self, m, true); 1552 m = NULL; 1553 continue; 1554 // interference - the markword changed - just retry. 1555 // The state-transitions are one-way, so there's no chance of 1556 // live-lock -- "Inflated" is an absorbing state. 1557 } 1558 1559 // Hopefully the performance counters are allocated on distinct 1560 // cache lines to avoid false sharing on MP systems ... 1561 OM_PERFDATA_OP(Inflations, inc()); 1562 TEVENT(Inflate: overwrite neutral); 1563 if (log_is_enabled(Debug, monitorinflation)) { 1564 if (object->is_instance()) { 1565 ResourceMark rm; 1566 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1567 p2i(object), p2i(object->mark()), 1568 object->klass()->external_name()); 1569 } 1570 } 1571 if (event.should_commit()) { 1572 post_monitor_inflate_event(event, object, cause); 1573 } 1574 return m; 1575 } 1576 } 1577 1578 1579 // Deflate_idle_monitors() is called at all safepoints, immediately 1580 // after all mutators are stopped, but before any objects have moved. 1581 // It traverses the list of known monitors, deflating where possible. 1582 // The scavenged monitor are returned to the monitor free list. 1583 // 1584 // Beware that we scavenge at *every* stop-the-world point. 1585 // Having a large number of monitors in-circulation negatively 1586 // impacts the performance of some applications (e.g., PointBase). 1587 // Broadly, we want to minimize the # of monitors in circulation. 1588 // 1589 // We have added a flag, MonitorInUseLists, which creates a list 1590 // of active monitors for each thread. deflate_idle_monitors() 1591 // only scans the per-thread in-use lists. omAlloc() puts all 1592 // assigned monitors on the per-thread list. deflate_idle_monitors() 1593 // returns the non-busy monitors to the global free list. 1594 // When a thread dies, omFlush() adds the list of active monitors for 1595 // that thread to a global gOmInUseList acquiring the 1596 // global list lock. deflate_idle_monitors() acquires the global 1597 // list lock to scan for non-busy monitors to the global free list. 1598 // An alternative could have used a single global in-use list. The 1599 // downside would have been the additional cost of acquiring the global list lock 1600 // for every omAlloc(). 1601 // 1602 // Perversely, the heap size -- and thus the STW safepoint rate -- 1603 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1604 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1605 // This is an unfortunate aspect of this design. 1606 1607 enum ManifestConstants { 1608 ClearResponsibleAtSTW = 0 1609 }; 1610 1611 // Deflate a single monitor if not in-use 1612 // Return true if deflated, false if in-use 1613 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1614 ObjectMonitor** freeHeadp, 1615 ObjectMonitor** freeTailp) { 1616 bool deflated; 1617 // Normal case ... The monitor is associated with obj. 1618 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1619 guarantee(mid == obj->mark()->monitor(), "invariant"); 1620 guarantee(mid->header()->is_neutral(), "invariant"); 1621 1622 if (mid->is_busy()) { 1623 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1624 deflated = false; 1625 } else { 1626 // Deflate the monitor if it is no longer being used 1627 // It's idle - scavenge and return to the global free list 1628 // plain old deflation ... 1629 TEVENT(deflate_idle_monitors - scavenge1); 1630 if (log_is_enabled(Debug, monitorinflation)) { 1631 if (obj->is_instance()) { 1632 ResourceMark rm; 1633 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , " 1634 "mark " INTPTR_FORMAT " , type %s", 1635 p2i(obj), p2i(obj->mark()), 1636 obj->klass()->external_name()); 1637 } 1638 } 1639 1640 // Restore the header back to obj 1641 obj->release_set_mark(mid->header()); 1642 mid->clear(); 1643 1644 assert(mid->object() == NULL, "invariant"); 1645 1646 // Move the object to the working free list defined by freeHeadp, freeTailp 1647 if (*freeHeadp == NULL) *freeHeadp = mid; 1648 if (*freeTailp != NULL) { 1649 ObjectMonitor * prevtail = *freeTailp; 1650 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1651 prevtail->FreeNext = mid; 1652 } 1653 *freeTailp = mid; 1654 deflated = true; 1655 } 1656 return deflated; 1657 } 1658 1659 // Walk a given monitor list, and deflate idle monitors 1660 // The given list could be a per-thread list or a global list 1661 // Caller acquires gListLock. 1662 // 1663 // In the case of parallel processing of thread local monitor lists, 1664 // work is done by Threads::parallel_threads_do() which ensures that 1665 // each Java thread is processed by exactly one worker thread, and 1666 // thus avoid conflicts that would arise when worker threads would 1667 // process the same monitor lists concurrently. 1668 // 1669 // See also ParallelSPCleanupTask and 1670 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and 1671 // Threads::parallel_java_threads_do() in thread.cpp. 1672 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1673 ObjectMonitor** freeHeadp, 1674 ObjectMonitor** freeTailp) { 1675 ObjectMonitor* mid; 1676 ObjectMonitor* next; 1677 ObjectMonitor* cur_mid_in_use = NULL; 1678 int deflated_count = 0; 1679 1680 for (mid = *listHeadp; mid != NULL;) { 1681 oop obj = (oop) mid->object(); 1682 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1683 // if deflate_monitor succeeded, 1684 // extract from per-thread in-use list 1685 if (mid == *listHeadp) { 1686 *listHeadp = mid->FreeNext; 1687 } else if (cur_mid_in_use != NULL) { 1688 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1689 } 1690 next = mid->FreeNext; 1691 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1692 mid = next; 1693 deflated_count++; 1694 } else { 1695 cur_mid_in_use = mid; 1696 mid = mid->FreeNext; 1697 } 1698 } 1699 return deflated_count; 1700 } 1701 1702 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1703 counters->nInuse = 0; // currently associated with objects 1704 counters->nInCirculation = 0; // extant 1705 counters->nScavenged = 0; // reclaimed 1706 } 1707 1708 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 1709 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1710 bool deflated = false; 1711 1712 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1713 ObjectMonitor * freeTailp = NULL; 1714 1715 TEVENT(deflate_idle_monitors); 1716 // Prevent omFlush from changing mids in Thread dtor's during deflation 1717 // And in case the vm thread is acquiring a lock during a safepoint 1718 // See e.g. 6320749 1719 Thread::muxAcquire(&gListLock, "scavenge - return"); 1720 1721 if (MonitorInUseLists) { 1722 // Note: the thread-local monitors lists get deflated in 1723 // a separate pass. See deflate_thread_local_monitors(). 1724 1725 // For moribund threads, scan gOmInUseList 1726 if (gOmInUseList) { 1727 counters->nInCirculation += gOmInUseCount; 1728 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1729 gOmInUseCount -= deflated_count; 1730 counters->nScavenged += deflated_count; 1731 counters->nInuse += gOmInUseCount; 1732 } 1733 1734 } else { 1735 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 1736 for (; block != NULL; block = next(block)) { 1737 // Iterate over all extant monitors - Scavenge all idle monitors. 1738 assert(block->object() == CHAINMARKER, "must be a block header"); 1739 counters->nInCirculation += _BLOCKSIZE; 1740 for (int i = 1; i < _BLOCKSIZE; i++) { 1741 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1742 oop obj = (oop)mid->object(); 1743 1744 if (obj == NULL) { 1745 // The monitor is not associated with an object. 1746 // The monitor should either be a thread-specific private 1747 // free list or the global free list. 1748 // obj == NULL IMPLIES mid->is_busy() == 0 1749 guarantee(!mid->is_busy(), "invariant"); 1750 continue; 1751 } 1752 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1753 1754 if (deflated) { 1755 mid->FreeNext = NULL; 1756 counters->nScavenged++; 1757 } else { 1758 counters->nInuse++; 1759 } 1760 } 1761 } 1762 } 1763 1764 // Move the scavenged monitors back to the global free list. 1765 if (freeHeadp != NULL) { 1766 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant"); 1767 assert(freeTailp->FreeNext == NULL, "invariant"); 1768 // constant-time list splice - prepend scavenged segment to gFreeList 1769 freeTailp->FreeNext = gFreeList; 1770 gFreeList = freeHeadp; 1771 } 1772 Thread::muxRelease(&gListLock); 1773 1774 } 1775 1776 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1777 gMonitorFreeCount += counters->nScavenged; 1778 1779 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1780 1781 if (ObjectMonitor::Knob_Verbose) { 1782 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d " 1783 "ForceMonitorScavenge=%d : pop=%d free=%d", 1784 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge, 1785 gMonitorPopulation, gMonitorFreeCount); 1786 tty->flush(); 1787 } 1788 1789 ForceMonitorScavenge = 0; // Reset 1790 1791 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged)); 1792 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation)); 1793 1794 // TODO: Add objectMonitor leak detection. 1795 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1796 GVars.stwRandom = os::random(); 1797 GVars.stwCycle++; 1798 } 1799 1800 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 1801 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1802 if (!MonitorInUseLists) return; 1803 1804 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1805 ObjectMonitor * freeTailp = NULL; 1806 1807 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp); 1808 1809 Thread::muxAcquire(&gListLock, "scavenge - return"); 1810 1811 // Adjust counters 1812 counters->nInCirculation += thread->omInUseCount; 1813 thread->omInUseCount -= deflated_count; 1814 if (ObjectMonitor::Knob_VerifyInUse) { 1815 verifyInUse(thread); 1816 } 1817 counters->nScavenged += deflated_count; 1818 counters->nInuse += thread->omInUseCount; 1819 1820 // Move the scavenged monitors back to the global free list. 1821 if (freeHeadp != NULL) { 1822 guarantee(freeTailp != NULL && deflated_count > 0, "invariant"); 1823 assert(freeTailp->FreeNext == NULL, "invariant"); 1824 1825 // constant-time list splice - prepend scavenged segment to gFreeList 1826 freeTailp->FreeNext = gFreeList; 1827 gFreeList = freeHeadp; 1828 } 1829 Thread::muxRelease(&gListLock); 1830 } 1831 1832 // Monitor cleanup on JavaThread::exit 1833 1834 // Iterate through monitor cache and attempt to release thread's monitors 1835 // Gives up on a particular monitor if an exception occurs, but continues 1836 // the overall iteration, swallowing the exception. 1837 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1838 private: 1839 TRAPS; 1840 1841 public: 1842 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1843 void do_monitor(ObjectMonitor* mid) { 1844 if (mid->owner() == THREAD) { 1845 if (ObjectMonitor::Knob_VerifyMatch != 0) { 1846 ResourceMark rm; 1847 Handle obj(THREAD, (oop) mid->object()); 1848 tty->print("INFO: unexpected locked object:"); 1849 javaVFrame::print_locked_object_class_name(tty, obj, "locked"); 1850 fatal("exiting JavaThread=" INTPTR_FORMAT 1851 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT, 1852 p2i(THREAD), p2i(mid)); 1853 } 1854 (void)mid->complete_exit(CHECK); 1855 } 1856 } 1857 }; 1858 1859 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1860 // ignored. This is meant to be called during JNI thread detach which assumes 1861 // all remaining monitors are heavyweight. All exceptions are swallowed. 1862 // Scanning the extant monitor list can be time consuming. 1863 // A simple optimization is to add a per-thread flag that indicates a thread 1864 // called jni_monitorenter() during its lifetime. 1865 // 1866 // Instead of No_Savepoint_Verifier it might be cheaper to 1867 // use an idiom of the form: 1868 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1869 // <code that must not run at safepoint> 1870 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1871 // Since the tests are extremely cheap we could leave them enabled 1872 // for normal product builds. 1873 1874 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1875 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1876 NoSafepointVerifier nsv; 1877 ReleaseJavaMonitorsClosure rjmc(THREAD); 1878 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1879 ObjectSynchronizer::monitors_iterate(&rjmc); 1880 Thread::muxRelease(&gListLock); 1881 THREAD->clear_pending_exception(); 1882 } 1883 1884 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1885 switch (cause) { 1886 case inflate_cause_vm_internal: return "VM Internal"; 1887 case inflate_cause_monitor_enter: return "Monitor Enter"; 1888 case inflate_cause_wait: return "Monitor Wait"; 1889 case inflate_cause_notify: return "Monitor Notify"; 1890 case inflate_cause_hash_code: return "Monitor Hash Code"; 1891 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1892 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1893 default: 1894 ShouldNotReachHere(); 1895 } 1896 return "Unknown"; 1897 } 1898 1899 static void post_monitor_inflate_event(EventJavaMonitorInflate& event, 1900 const oop obj, 1901 const ObjectSynchronizer::InflateCause cause) { 1902 #if INCLUDE_TRACE 1903 assert(event.should_commit(), "check outside"); 1904 event.set_monitorClass(obj->klass()); 1905 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj); 1906 event.set_cause((u1)cause); 1907 event.commit(); 1908 #endif 1909 } 1910 1911 //------------------------------------------------------------------------------ 1912 // Debugging code 1913 1914 void ObjectSynchronizer::sanity_checks(const bool verbose, 1915 const uint cache_line_size, 1916 int *error_cnt_ptr, 1917 int *warning_cnt_ptr) { 1918 u_char *addr_begin = (u_char*)&GVars; 1919 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1920 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1921 1922 if (verbose) { 1923 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1924 sizeof(SharedGlobals)); 1925 } 1926 1927 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1928 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1929 1930 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1931 if (verbose) { 1932 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1933 } 1934 1935 if (cache_line_size != 0) { 1936 // We were able to determine the L1 data cache line size so 1937 // do some cache line specific sanity checks 1938 1939 if (offset_stwRandom < cache_line_size) { 1940 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1941 "to the struct beginning than a cache line which permits " 1942 "false sharing."); 1943 (*warning_cnt_ptr)++; 1944 } 1945 1946 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1947 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1948 "SharedGlobals.hcSequence fields are closer than a cache " 1949 "line which permits false sharing."); 1950 (*warning_cnt_ptr)++; 1951 } 1952 1953 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1954 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1955 "to the struct end than a cache line which permits false " 1956 "sharing."); 1957 (*warning_cnt_ptr)++; 1958 } 1959 } 1960 } 1961 1962 #ifndef PRODUCT 1963 1964 // Check if monitor belongs to the monitor cache 1965 // The list is grow-only so it's *relatively* safe to traverse 1966 // the list of extant blocks without taking a lock. 1967 1968 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1969 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList); 1970 while (block != NULL) { 1971 assert(block->object() == CHAINMARKER, "must be a block header"); 1972 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 1973 address mon = (address)monitor; 1974 address blk = (address)block; 1975 size_t diff = mon - blk; 1976 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned"); 1977 return 1; 1978 } 1979 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 1980 } 1981 return 0; 1982 } 1983 1984 #endif