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 // gBlockList is really PaddedEnd<ObjectMonitor> *, but we don't 115 // want to expose the PaddedEnd template more than necessary. 116 ObjectMonitor * volatile ObjectSynchronizer::gBlockList = NULL; 117 // global monitor free list 118 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 119 // global monitor in-use list, for moribund threads, 120 // monitors they inflated need to be scanned for deflation 121 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 122 // count of entries in gOmInUseList 123 int ObjectSynchronizer::gOmInUseCount = 0; 124 125 static volatile intptr_t gListLock = 0; // protects global monitor lists 126 static volatile int gMonitorFreeCount = 0; // # on gFreeList 127 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation 128 129 static void post_monitor_inflate_event(EventJavaMonitorInflate&, 130 const oop, 131 const ObjectSynchronizer::InflateCause); 132 133 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 134 135 136 // =====================> Quick functions 137 138 // The quick_* forms are special fast-path variants used to improve 139 // performance. In the simplest case, a "quick_*" implementation could 140 // simply return false, in which case the caller will perform the necessary 141 // state transitions and call the slow-path form. 142 // The fast-path is designed to handle frequently arising cases in an efficient 143 // manner and is just a degenerate "optimistic" variant of the slow-path. 144 // returns true -- to indicate the call was satisfied. 145 // returns false -- to indicate the call needs the services of the slow-path. 146 // A no-loitering ordinance is in effect for code in the quick_* family 147 // operators: safepoints or indefinite blocking (blocking that might span a 148 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 149 // entry. 150 // 151 // Consider: An interesting optimization is to have the JIT recognize the 152 // following common idiom: 153 // synchronized (someobj) { .... ; notify(); } 154 // That is, we find a notify() or notifyAll() call that immediately precedes 155 // the monitorexit operation. In that case the JIT could fuse the operations 156 // into a single notifyAndExit() runtime primitive. 157 158 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) { 159 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 160 assert(self->is_Java_thread(), "invariant"); 161 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 162 NoSafepointVerifier nsv; 163 if (obj == NULL) return false; // slow-path for invalid obj 164 const markOop mark = obj->mark(); 165 166 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) { 167 // Degenerate notify 168 // stack-locked by caller so by definition the implied waitset is empty. 169 return true; 170 } 171 172 if (mark->has_monitor()) { 173 ObjectMonitor * const mon = mark->monitor(); 174 assert(mon->object() == obj, "invariant"); 175 if (mon->owner() != self) return false; // slow-path for IMS exception 176 177 if (mon->first_waiter() != NULL) { 178 // We have one or more waiters. Since this is an inflated monitor 179 // that we own, we can transfer one or more threads from the waitset 180 // to the entrylist here and now, avoiding the slow-path. 181 if (all) { 182 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 183 } else { 184 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 185 } 186 int tally = 0; 187 do { 188 mon->INotify(self); 189 ++tally; 190 } while (mon->first_waiter() != NULL && all); 191 OM_PERFDATA_OP(Notifications, inc(tally)); 192 } 193 return true; 194 } 195 196 // biased locking and any other IMS exception states take the slow-path 197 return false; 198 } 199 200 201 // The LockNode emitted directly at the synchronization site would have 202 // been too big if it were to have included support for the cases of inflated 203 // recursive enter and exit, so they go here instead. 204 // Note that we can't safely call AsyncPrintJavaStack() from within 205 // quick_enter() as our thread state remains _in_Java. 206 207 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self, 208 BasicLock * lock) { 209 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 210 assert(Self->is_Java_thread(), "invariant"); 211 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant"); 212 NoSafepointVerifier nsv; 213 if (obj == NULL) return false; // Need to throw NPE 214 const markOop mark = obj->mark(); 215 216 if (mark->has_monitor()) { 217 ObjectMonitor * const m = mark->monitor(); 218 assert(m->object() == obj, "invariant"); 219 Thread * const owner = (Thread *) m->_owner; 220 221 // Lock contention and Transactional Lock Elision (TLE) diagnostics 222 // and observability 223 // Case: light contention possibly amenable to TLE 224 // Case: TLE inimical operations such as nested/recursive synchronization 225 226 if (owner == Self) { 227 m->_recursions++; 228 return true; 229 } 230 231 // This Java Monitor is inflated so obj's header will never be 232 // displaced to this thread's BasicLock. Make the displaced header 233 // non-NULL so this BasicLock is not seen as recursive nor as 234 // being locked. We do this unconditionally so that this thread's 235 // BasicLock cannot be mis-interpreted by any stack walkers. For 236 // performance reasons, stack walkers generally first check for 237 // Biased Locking in the object's header, the second check is for 238 // stack-locking in the object's header, the third check is for 239 // recursive stack-locking in the displaced header in the BasicLock, 240 // and last are the inflated Java Monitor (ObjectMonitor) checks. 241 lock->set_displaced_header(markOopDesc::unused_mark()); 242 243 if (owner == NULL && 244 Atomic::cmpxchg_ptr(Self, &(m->_owner), NULL) == NULL) { 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 ((markOop) Atomic::cmpxchg_ptr(dhw, object->mark_addr(), 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 == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), 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 = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), 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 = (markOop) Atomic::cmpxchg_ptr(temp, monitor, 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(Handle h_obj, bool doLock) { 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(owner, doLock); 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 = 943 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 944 while (block != NULL) { 945 assert(block->object() == CHAINMARKER, "must be a block header"); 946 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 947 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 948 oop object = (oop)mid->object(); 949 if (object != NULL) { 950 closure->do_monitor(mid); 951 } 952 } 953 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 954 } 955 } 956 957 // Get the next block in the block list. 958 static inline ObjectMonitor* next(ObjectMonitor* block) { 959 assert(block->object() == CHAINMARKER, "must be a block header"); 960 block = block->FreeNext; 961 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 962 return block; 963 } 964 965 static bool monitors_used_above_threshold() { 966 if (gMonitorPopulation == 0) { 967 return false; 968 } 969 int monitors_used = gMonitorPopulation - gMonitorFreeCount; 970 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation; 971 return monitor_usage > MonitorUsedDeflationThreshold; 972 } 973 974 bool ObjectSynchronizer::is_cleanup_needed() { 975 if (MonitorUsedDeflationThreshold > 0) { 976 return monitors_used_above_threshold(); 977 } 978 return false; 979 } 980 981 void ObjectSynchronizer::oops_do(OopClosure* f) { 982 if (MonitorInUseLists) { 983 // When using thread local monitor lists, we only scan the 984 // global used list here (for moribund threads), and 985 // the thread-local monitors in Thread::oops_do(). 986 global_used_oops_do(f); 987 } else { 988 global_oops_do(f); 989 } 990 } 991 992 void ObjectSynchronizer::global_oops_do(OopClosure* f) { 993 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 994 PaddedEnd<ObjectMonitor> * block = 995 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 996 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) { 997 assert(block->object() == CHAINMARKER, "must be a block header"); 998 for (int i = 1; i < _BLOCKSIZE; i++) { 999 ObjectMonitor* mid = (ObjectMonitor *)&block[i]; 1000 if (mid->object() != NULL) { 1001 f->do_oop((oop*)mid->object_addr()); 1002 } 1003 } 1004 } 1005 } 1006 1007 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1008 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1009 list_oops_do(gOmInUseList, f); 1010 } 1011 1012 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1013 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1014 list_oops_do(thread->omInUseList, f); 1015 } 1016 1017 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) { 1018 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1019 ObjectMonitor* mid; 1020 for (mid = list; mid != NULL; mid = mid->FreeNext) { 1021 if (mid->object() != NULL) { 1022 f->do_oop((oop*)mid->object_addr()); 1023 } 1024 } 1025 } 1026 1027 1028 // ----------------------------------------------------------------------------- 1029 // ObjectMonitor Lifecycle 1030 // ----------------------- 1031 // Inflation unlinks monitors from the global gFreeList and 1032 // associates them with objects. Deflation -- which occurs at 1033 // STW-time -- disassociates idle monitors from objects. Such 1034 // scavenged monitors are returned to the gFreeList. 1035 // 1036 // The global list is protected by gListLock. All the critical sections 1037 // are short and operate in constant-time. 1038 // 1039 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1040 // 1041 // Lifecycle: 1042 // -- unassigned and on the global free list 1043 // -- unassigned and on a thread's private omFreeList 1044 // -- assigned to an object. The object is inflated and the mark refers 1045 // to the objectmonitor. 1046 1047 1048 // Constraining monitor pool growth via MonitorBound ... 1049 // 1050 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1051 // the rate of scavenging is driven primarily by GC. As such, we can find 1052 // an inordinate number of monitors in circulation. 1053 // To avoid that scenario we can artificially induce a STW safepoint 1054 // if the pool appears to be growing past some reasonable bound. 1055 // Generally we favor time in space-time tradeoffs, but as there's no 1056 // natural back-pressure on the # of extant monitors we need to impose some 1057 // type of limit. Beware that if MonitorBound is set to too low a value 1058 // we could just loop. In addition, if MonitorBound is set to a low value 1059 // we'll incur more safepoints, which are harmful to performance. 1060 // See also: GuaranteedSafepointInterval 1061 // 1062 // The current implementation uses asynchronous VM operations. 1063 1064 static void InduceScavenge(Thread * Self, const char * Whence) { 1065 // Induce STW safepoint to trim monitors 1066 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1067 // More precisely, trigger an asynchronous STW safepoint as the number 1068 // of active monitors passes the specified threshold. 1069 // TODO: assert thread state is reasonable 1070 1071 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 1072 if (ObjectMonitor::Knob_Verbose) { 1073 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)", 1074 Whence, ForceMonitorScavenge) ; 1075 tty->flush(); 1076 } 1077 // Induce a 'null' safepoint to scavenge monitors 1078 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 1079 // to the VMthread and have a lifespan longer than that of this activation record. 1080 // The VMThread will delete the op when completed. 1081 VMThread::execute(new VM_ScavengeMonitors()); 1082 1083 if (ObjectMonitor::Knob_Verbose) { 1084 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)", 1085 Whence, ForceMonitorScavenge) ; 1086 tty->flush(); 1087 } 1088 } 1089 } 1090 1091 void ObjectSynchronizer::verifyInUse(Thread *Self) { 1092 ObjectMonitor* mid; 1093 int in_use_tally = 0; 1094 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 1095 in_use_tally++; 1096 } 1097 assert(in_use_tally == Self->omInUseCount, "in-use count off"); 1098 1099 int free_tally = 0; 1100 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 1101 free_tally++; 1102 } 1103 assert(free_tally == Self->omFreeCount, "free count off"); 1104 } 1105 1106 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) { 1107 // A large MAXPRIVATE value reduces both list lock contention 1108 // and list coherency traffic, but also tends to increase the 1109 // number of objectMonitors in circulation as well as the STW 1110 // scavenge costs. As usual, we lean toward time in space-time 1111 // tradeoffs. 1112 const int MAXPRIVATE = 1024; 1113 for (;;) { 1114 ObjectMonitor * m; 1115 1116 // 1: try to allocate from the thread's local omFreeList. 1117 // Threads will attempt to allocate first from their local list, then 1118 // from the global list, and only after those attempts fail will the thread 1119 // attempt to instantiate new monitors. Thread-local free lists take 1120 // heat off the gListLock and improve allocation latency, as well as reducing 1121 // coherency traffic on the shared global list. 1122 m = Self->omFreeList; 1123 if (m != NULL) { 1124 Self->omFreeList = m->FreeNext; 1125 Self->omFreeCount--; 1126 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 1127 guarantee(m->object() == NULL, "invariant"); 1128 if (MonitorInUseLists) { 1129 m->FreeNext = Self->omInUseList; 1130 Self->omInUseList = m; 1131 Self->omInUseCount++; 1132 if (ObjectMonitor::Knob_VerifyInUse) { 1133 verifyInUse(Self); 1134 } 1135 } else { 1136 m->FreeNext = NULL; 1137 } 1138 return m; 1139 } 1140 1141 // 2: try to allocate from the global gFreeList 1142 // CONSIDER: use muxTry() instead of muxAcquire(). 1143 // If the muxTry() fails then drop immediately into case 3. 1144 // If we're using thread-local free lists then try 1145 // to reprovision the caller's free list. 1146 if (gFreeList != NULL) { 1147 // Reprovision the thread's omFreeList. 1148 // Use bulk transfers to reduce the allocation rate and heat 1149 // on various locks. 1150 Thread::muxAcquire(&gListLock, "omAlloc"); 1151 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 1152 gMonitorFreeCount--; 1153 ObjectMonitor * take = gFreeList; 1154 gFreeList = take->FreeNext; 1155 guarantee(take->object() == NULL, "invariant"); 1156 guarantee(!take->is_busy(), "invariant"); 1157 take->Recycle(); 1158 omRelease(Self, take, false); 1159 } 1160 Thread::muxRelease(&gListLock); 1161 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 1162 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 1163 TEVENT(omFirst - reprovision); 1164 1165 const int mx = MonitorBound; 1166 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) { 1167 // We can't safely induce a STW safepoint from omAlloc() as our thread 1168 // state may not be appropriate for such activities and callers may hold 1169 // naked oops, so instead we defer the action. 1170 InduceScavenge(Self, "omAlloc"); 1171 } 1172 continue; 1173 } 1174 1175 // 3: allocate a block of new ObjectMonitors 1176 // Both the local and global free lists are empty -- resort to malloc(). 1177 // In the current implementation objectMonitors are TSM - immortal. 1178 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1179 // each ObjectMonitor to start at the beginning of a cache line, 1180 // so we use align_up(). 1181 // A better solution would be to use C++ placement-new. 1182 // BEWARE: As it stands currently, we don't run the ctors! 1183 assert(_BLOCKSIZE > 1, "invariant"); 1184 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE; 1185 PaddedEnd<ObjectMonitor> * temp; 1186 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1187 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size, 1188 mtInternal); 1189 temp = (PaddedEnd<ObjectMonitor> *) 1190 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE); 1191 1192 // NOTE: (almost) no way to recover if allocation failed. 1193 // We might be able to induce a STW safepoint and scavenge enough 1194 // objectMonitors to permit progress. 1195 if (temp == NULL) { 1196 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR, 1197 "Allocate ObjectMonitors"); 1198 } 1199 (void)memset((void *) temp, 0, neededsize); 1200 1201 // Format the block. 1202 // initialize the linked list, each monitor points to its next 1203 // forming the single linked free list, the very first monitor 1204 // will points to next block, which forms the block list. 1205 // The trick of using the 1st element in the block as gBlockList 1206 // linkage should be reconsidered. A better implementation would 1207 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1208 1209 for (int i = 1; i < _BLOCKSIZE; i++) { 1210 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1]; 1211 } 1212 1213 // terminate the last monitor as the end of list 1214 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1215 1216 // Element [0] is reserved for global list linkage 1217 temp[0].set_object(CHAINMARKER); 1218 1219 // Consider carving out this thread's current request from the 1220 // block in hand. This avoids some lock traffic and redundant 1221 // list activity. 1222 1223 // Acquire the gListLock to manipulate gBlockList and gFreeList. 1224 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1225 Thread::muxAcquire(&gListLock, "omAlloc [2]"); 1226 gMonitorPopulation += _BLOCKSIZE-1; 1227 gMonitorFreeCount += _BLOCKSIZE-1; 1228 1229 // Add the new block to the list of extant blocks (gBlockList). 1230 // The very first objectMonitor in a block is reserved and dedicated. 1231 // It serves as blocklist "next" linkage. 1232 temp[0].FreeNext = gBlockList; 1233 // There are lock-free uses of gBlockList so make sure that 1234 // the previous stores happen before we update gBlockList. 1235 OrderAccess::release_store_ptr(&gBlockList, temp); 1236 1237 // Add the new string of objectMonitors to the global free list 1238 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1239 gFreeList = temp + 1; 1240 Thread::muxRelease(&gListLock); 1241 TEVENT(Allocate block of monitors); 1242 } 1243 } 1244 1245 // Place "m" on the caller's private per-thread omFreeList. 1246 // In practice there's no need to clamp or limit the number of 1247 // monitors on a thread's omFreeList as the only time we'll call 1248 // omRelease is to return a monitor to the free list after a CAS 1249 // attempt failed. This doesn't allow unbounded #s of monitors to 1250 // accumulate on a thread's free list. 1251 // 1252 // Key constraint: all ObjectMonitors on a thread's free list and the global 1253 // free list must have their object field set to null. This prevents the 1254 // scavenger -- deflate_idle_monitors -- from reclaiming them. 1255 1256 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1257 bool fromPerThreadAlloc) { 1258 guarantee(m->object() == NULL, "invariant"); 1259 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor"); 1260 // Remove from omInUseList 1261 if (MonitorInUseLists && fromPerThreadAlloc) { 1262 ObjectMonitor* cur_mid_in_use = NULL; 1263 bool extracted = false; 1264 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) { 1265 if (m == mid) { 1266 // extract from per-thread in-use list 1267 if (mid == Self->omInUseList) { 1268 Self->omInUseList = mid->FreeNext; 1269 } else if (cur_mid_in_use != NULL) { 1270 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1271 } 1272 extracted = true; 1273 Self->omInUseCount--; 1274 if (ObjectMonitor::Knob_VerifyInUse) { 1275 verifyInUse(Self); 1276 } 1277 break; 1278 } 1279 } 1280 assert(extracted, "Should have extracted from in-use list"); 1281 } 1282 1283 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1284 m->FreeNext = Self->omFreeList; 1285 Self->omFreeList = m; 1286 Self->omFreeCount++; 1287 } 1288 1289 // Return the monitors of a moribund thread's local free list to 1290 // the global free list. Typically a thread calls omFlush() when 1291 // it's dying. We could also consider having the VM thread steal 1292 // monitors from threads that have not run java code over a few 1293 // consecutive STW safepoints. Relatedly, we might decay 1294 // omFreeProvision at STW safepoints. 1295 // 1296 // Also return the monitors of a moribund thread's omInUseList to 1297 // a global gOmInUseList under the global list lock so these 1298 // will continue to be scanned. 1299 // 1300 // We currently call omFlush() from Threads::remove() _before the thread 1301 // has been excised from the thread list and is no longer a mutator. 1302 // This means that omFlush() can not run concurrently with a safepoint and 1303 // interleave with the scavenge operator. In particular, this ensures that 1304 // the thread's monitors are scanned by a GC safepoint, either via 1305 // Thread::oops_do() (if safepoint happens before omFlush()) or via 1306 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's 1307 // monitors have been transferred to the global in-use list). 1308 1309 void ObjectSynchronizer::omFlush(Thread * Self) { 1310 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL 1311 Self->omFreeList = NULL; 1312 ObjectMonitor * tail = NULL; 1313 int tally = 0; 1314 if (list != NULL) { 1315 ObjectMonitor * s; 1316 // The thread is going away, the per-thread free monitors 1317 // are freed via set_owner(NULL) 1318 // Link them to tail, which will be linked into the global free list 1319 // gFreeList below, under the gListLock 1320 for (s = list; s != NULL; s = s->FreeNext) { 1321 tally++; 1322 tail = s; 1323 guarantee(s->object() == NULL, "invariant"); 1324 guarantee(!s->is_busy(), "invariant"); 1325 s->set_owner(NULL); // redundant but good hygiene 1326 TEVENT(omFlush - Move one); 1327 } 1328 guarantee(tail != NULL && list != NULL, "invariant"); 1329 } 1330 1331 ObjectMonitor * inUseList = Self->omInUseList; 1332 ObjectMonitor * inUseTail = NULL; 1333 int inUseTally = 0; 1334 if (inUseList != NULL) { 1335 Self->omInUseList = NULL; 1336 ObjectMonitor *cur_om; 1337 // The thread is going away, however the omInUseList inflated 1338 // monitors may still be in-use by other threads. 1339 // Link them to inUseTail, which will be linked into the global in-use list 1340 // gOmInUseList below, under the gListLock 1341 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) { 1342 inUseTail = cur_om; 1343 inUseTally++; 1344 } 1345 assert(Self->omInUseCount == inUseTally, "in-use count off"); 1346 Self->omInUseCount = 0; 1347 guarantee(inUseTail != NULL && inUseList != NULL, "invariant"); 1348 } 1349 1350 Thread::muxAcquire(&gListLock, "omFlush"); 1351 if (tail != NULL) { 1352 tail->FreeNext = gFreeList; 1353 gFreeList = list; 1354 gMonitorFreeCount += tally; 1355 assert(Self->omFreeCount == tally, "free-count off"); 1356 Self->omFreeCount = 0; 1357 } 1358 1359 if (inUseTail != NULL) { 1360 inUseTail->FreeNext = gOmInUseList; 1361 gOmInUseList = inUseList; 1362 gOmInUseCount += inUseTally; 1363 } 1364 1365 Thread::muxRelease(&gListLock); 1366 TEVENT(omFlush); 1367 } 1368 1369 // Fast path code shared by multiple functions 1370 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1371 markOop mark = obj->mark(); 1372 if (mark->has_monitor()) { 1373 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1374 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1375 return mark->monitor(); 1376 } 1377 return ObjectSynchronizer::inflate(Thread::current(), 1378 obj, 1379 inflate_cause_vm_internal); 1380 } 1381 1382 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self, 1383 oop object, 1384 const InflateCause cause) { 1385 1386 // Inflate mutates the heap ... 1387 // Relaxing assertion for bug 6320749. 1388 assert(Universe::verify_in_progress() || 1389 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1390 1391 EventJavaMonitorInflate event; 1392 1393 for (;;) { 1394 const markOop mark = object->mark(); 1395 assert(!mark->has_bias_pattern(), "invariant"); 1396 1397 // The mark can be in one of the following states: 1398 // * Inflated - just return 1399 // * Stack-locked - coerce it to inflated 1400 // * INFLATING - busy wait for conversion to complete 1401 // * Neutral - aggressively inflate the object. 1402 // * BIASED - Illegal. We should never see this 1403 1404 // CASE: inflated 1405 if (mark->has_monitor()) { 1406 ObjectMonitor * inf = mark->monitor(); 1407 assert(inf->header()->is_neutral(), "invariant"); 1408 assert(inf->object() == object, "invariant"); 1409 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1410 event.cancel(); // let's not post an inflation event, unless we did the deed ourselves 1411 return inf; 1412 } 1413 1414 // CASE: inflation in progress - inflating over a stack-lock. 1415 // Some other thread is converting from stack-locked to inflated. 1416 // Only that thread can complete inflation -- other threads must wait. 1417 // The INFLATING value is transient. 1418 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1419 // We could always eliminate polling by parking the thread on some auxiliary list. 1420 if (mark == markOopDesc::INFLATING()) { 1421 TEVENT(Inflate: spin while INFLATING); 1422 ReadStableMark(object); 1423 continue; 1424 } 1425 1426 // CASE: stack-locked 1427 // Could be stack-locked either by this thread or by some other thread. 1428 // 1429 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1430 // to install INFLATING into the mark word. We originally installed INFLATING, 1431 // allocated the objectmonitor, and then finally STed the address of the 1432 // objectmonitor into the mark. This was correct, but artificially lengthened 1433 // the interval in which INFLATED appeared in the mark, thus increasing 1434 // the odds of inflation contention. 1435 // 1436 // We now use per-thread private objectmonitor free lists. 1437 // These list are reprovisioned from the global free list outside the 1438 // critical INFLATING...ST interval. A thread can transfer 1439 // multiple objectmonitors en-mass from the global free list to its local free list. 1440 // This reduces coherency traffic and lock contention on the global free list. 1441 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1442 // before or after the CAS(INFLATING) operation. 1443 // See the comments in omAlloc(). 1444 1445 if (mark->has_locker()) { 1446 ObjectMonitor * m = omAlloc(Self); 1447 // Optimistically prepare the objectmonitor - anticipate successful CAS 1448 // We do this before the CAS in order to minimize the length of time 1449 // in which INFLATING appears in the mark. 1450 m->Recycle(); 1451 m->_Responsible = NULL; 1452 m->_recursions = 0; 1453 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1454 1455 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark); 1456 if (cmp != mark) { 1457 omRelease(Self, m, true); 1458 continue; // Interference -- just retry 1459 } 1460 1461 // We've successfully installed INFLATING (0) into the mark-word. 1462 // This is the only case where 0 will appear in a mark-word. 1463 // Only the singular thread that successfully swings the mark-word 1464 // to 0 can perform (or more precisely, complete) inflation. 1465 // 1466 // Why do we CAS a 0 into the mark-word instead of just CASing the 1467 // mark-word from the stack-locked value directly to the new inflated state? 1468 // Consider what happens when a thread unlocks a stack-locked object. 1469 // It attempts to use CAS to swing the displaced header value from the 1470 // on-stack basiclock back into the object header. Recall also that the 1471 // header value (hashcode, etc) can reside in (a) the object header, or 1472 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1473 // header in an objectMonitor. The inflate() routine must copy the header 1474 // value from the basiclock on the owner's stack to the objectMonitor, all 1475 // the while preserving the hashCode stability invariants. If the owner 1476 // decides to release the lock while the value is 0, the unlock will fail 1477 // and control will eventually pass from slow_exit() to inflate. The owner 1478 // will then spin, waiting for the 0 value to disappear. Put another way, 1479 // the 0 causes the owner to stall if the owner happens to try to 1480 // drop the lock (restoring the header from the basiclock to the object) 1481 // while inflation is in-progress. This protocol avoids races that might 1482 // would otherwise permit hashCode values to change or "flicker" for an object. 1483 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1484 // 0 serves as a "BUSY" inflate-in-progress indicator. 1485 1486 1487 // fetch the displaced mark from the owner's stack. 1488 // The owner can't die or unwind past the lock while our INFLATING 1489 // object is in the mark. Furthermore the owner can't complete 1490 // an unlock on the object, either. 1491 markOop dmw = mark->displaced_mark_helper(); 1492 assert(dmw->is_neutral(), "invariant"); 1493 1494 // Setup monitor fields to proper values -- prepare the monitor 1495 m->set_header(dmw); 1496 1497 // Optimization: if the mark->locker stack address is associated 1498 // with this thread we could simply set m->_owner = Self. 1499 // Note that a thread can inflate an object 1500 // that it has stack-locked -- as might happen in wait() -- directly 1501 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1502 m->set_owner(mark->locker()); 1503 m->set_object(object); 1504 // TODO-FIXME: assert BasicLock->dhw != 0. 1505 1506 // Must preserve store ordering. The monitor state must 1507 // be stable at the time of publishing the monitor address. 1508 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1509 object->release_set_mark(markOopDesc::encode(m)); 1510 1511 // Hopefully the performance counters are allocated on distinct cache lines 1512 // to avoid false sharing on MP systems ... 1513 OM_PERFDATA_OP(Inflations, inc()); 1514 TEVENT(Inflate: overwrite stacklock); 1515 if (log_is_enabled(Debug, monitorinflation)) { 1516 if (object->is_instance()) { 1517 ResourceMark rm; 1518 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1519 p2i(object), p2i(object->mark()), 1520 object->klass()->external_name()); 1521 } 1522 } 1523 if (event.should_commit()) { 1524 post_monitor_inflate_event(event, object, cause); 1525 } 1526 return m; 1527 } 1528 1529 // CASE: neutral 1530 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1531 // If we know we're inflating for entry it's better to inflate by swinging a 1532 // pre-locked objectMonitor pointer into the object header. A successful 1533 // CAS inflates the object *and* confers ownership to the inflating thread. 1534 // In the current implementation we use a 2-step mechanism where we CAS() 1535 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1536 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1537 // would be useful. 1538 1539 assert(mark->is_neutral(), "invariant"); 1540 ObjectMonitor * m = omAlloc(Self); 1541 // prepare m for installation - set monitor to initial state 1542 m->Recycle(); 1543 m->set_header(mark); 1544 m->set_owner(NULL); 1545 m->set_object(object); 1546 m->_recursions = 0; 1547 m->_Responsible = NULL; 1548 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1549 1550 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1551 m->set_object(NULL); 1552 m->set_owner(NULL); 1553 m->Recycle(); 1554 omRelease(Self, m, true); 1555 m = NULL; 1556 continue; 1557 // interference - the markword changed - just retry. 1558 // The state-transitions are one-way, so there's no chance of 1559 // live-lock -- "Inflated" is an absorbing state. 1560 } 1561 1562 // Hopefully the performance counters are allocated on distinct 1563 // cache lines to avoid false sharing on MP systems ... 1564 OM_PERFDATA_OP(Inflations, inc()); 1565 TEVENT(Inflate: overwrite neutral); 1566 if (log_is_enabled(Debug, monitorinflation)) { 1567 if (object->is_instance()) { 1568 ResourceMark rm; 1569 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1570 p2i(object), p2i(object->mark()), 1571 object->klass()->external_name()); 1572 } 1573 } 1574 if (event.should_commit()) { 1575 post_monitor_inflate_event(event, object, cause); 1576 } 1577 return m; 1578 } 1579 } 1580 1581 1582 // Deflate_idle_monitors() is called at all safepoints, immediately 1583 // after all mutators are stopped, but before any objects have moved. 1584 // It traverses the list of known monitors, deflating where possible. 1585 // The scavenged monitor are returned to the monitor free list. 1586 // 1587 // Beware that we scavenge at *every* stop-the-world point. 1588 // Having a large number of monitors in-circulation negatively 1589 // impacts the performance of some applications (e.g., PointBase). 1590 // Broadly, we want to minimize the # of monitors in circulation. 1591 // 1592 // We have added a flag, MonitorInUseLists, which creates a list 1593 // of active monitors for each thread. deflate_idle_monitors() 1594 // only scans the per-thread in-use lists. omAlloc() puts all 1595 // assigned monitors on the per-thread list. deflate_idle_monitors() 1596 // returns the non-busy monitors to the global free list. 1597 // When a thread dies, omFlush() adds the list of active monitors for 1598 // that thread to a global gOmInUseList acquiring the 1599 // global list lock. deflate_idle_monitors() acquires the global 1600 // list lock to scan for non-busy monitors to the global free list. 1601 // An alternative could have used a single global in-use list. The 1602 // downside would have been the additional cost of acquiring the global list lock 1603 // for every omAlloc(). 1604 // 1605 // Perversely, the heap size -- and thus the STW safepoint rate -- 1606 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1607 // which in turn can mean large(r) numbers of objectmonitors in circulation. 1608 // This is an unfortunate aspect of this design. 1609 1610 enum ManifestConstants { 1611 ClearResponsibleAtSTW = 0 1612 }; 1613 1614 // Deflate a single monitor if not in-use 1615 // Return true if deflated, false if in-use 1616 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1617 ObjectMonitor** freeHeadp, 1618 ObjectMonitor** freeTailp) { 1619 bool deflated; 1620 // Normal case ... The monitor is associated with obj. 1621 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1622 guarantee(mid == obj->mark()->monitor(), "invariant"); 1623 guarantee(mid->header()->is_neutral(), "invariant"); 1624 1625 if (mid->is_busy()) { 1626 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1627 deflated = false; 1628 } else { 1629 // Deflate the monitor if it is no longer being used 1630 // It's idle - scavenge and return to the global free list 1631 // plain old deflation ... 1632 TEVENT(deflate_idle_monitors - scavenge1); 1633 if (log_is_enabled(Debug, monitorinflation)) { 1634 if (obj->is_instance()) { 1635 ResourceMark rm; 1636 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , " 1637 "mark " INTPTR_FORMAT " , type %s", 1638 p2i(obj), p2i(obj->mark()), 1639 obj->klass()->external_name()); 1640 } 1641 } 1642 1643 // Restore the header back to obj 1644 obj->release_set_mark(mid->header()); 1645 mid->clear(); 1646 1647 assert(mid->object() == NULL, "invariant"); 1648 1649 // Move the object to the working free list defined by freeHeadp, freeTailp 1650 if (*freeHeadp == NULL) *freeHeadp = mid; 1651 if (*freeTailp != NULL) { 1652 ObjectMonitor * prevtail = *freeTailp; 1653 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); 1654 prevtail->FreeNext = mid; 1655 } 1656 *freeTailp = mid; 1657 deflated = true; 1658 } 1659 return deflated; 1660 } 1661 1662 // Walk a given monitor list, and deflate idle monitors 1663 // The given list could be a per-thread list or a global list 1664 // Caller acquires gListLock. 1665 // 1666 // In the case of parallel processing of thread local monitor lists, 1667 // work is done by Threads::parallel_threads_do() which ensures that 1668 // each Java thread is processed by exactly one worker thread, and 1669 // thus avoid conflicts that would arise when worker threads would 1670 // process the same monitor lists concurrently. 1671 // 1672 // See also ParallelSPCleanupTask and 1673 // SafepointSynchronizer::do_cleanup_tasks() in safepoint.cpp and 1674 // Threads::parallel_java_threads_do() in thread.cpp. 1675 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp, 1676 ObjectMonitor** freeHeadp, 1677 ObjectMonitor** freeTailp) { 1678 ObjectMonitor* mid; 1679 ObjectMonitor* next; 1680 ObjectMonitor* cur_mid_in_use = NULL; 1681 int deflated_count = 0; 1682 1683 for (mid = *listHeadp; mid != NULL;) { 1684 oop obj = (oop) mid->object(); 1685 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) { 1686 // if deflate_monitor succeeded, 1687 // extract from per-thread in-use list 1688 if (mid == *listHeadp) { 1689 *listHeadp = mid->FreeNext; 1690 } else if (cur_mid_in_use != NULL) { 1691 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list 1692 } 1693 next = mid->FreeNext; 1694 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list 1695 mid = next; 1696 deflated_count++; 1697 } else { 1698 cur_mid_in_use = mid; 1699 mid = mid->FreeNext; 1700 } 1701 } 1702 return deflated_count; 1703 } 1704 1705 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1706 counters->nInuse = 0; // currently associated with objects 1707 counters->nInCirculation = 0; // extant 1708 counters->nScavenged = 0; // reclaimed 1709 } 1710 1711 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 1712 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1713 bool deflated = false; 1714 1715 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1716 ObjectMonitor * freeTailp = NULL; 1717 1718 TEVENT(deflate_idle_monitors); 1719 // Prevent omFlush from changing mids in Thread dtor's during deflation 1720 // And in case the vm thread is acquiring a lock during a safepoint 1721 // See e.g. 6320749 1722 Thread::muxAcquire(&gListLock, "scavenge - return"); 1723 1724 if (MonitorInUseLists) { 1725 // Note: the thread-local monitors lists get deflated in 1726 // a separate pass. See deflate_thread_local_monitors(). 1727 1728 // For moribund threads, scan gOmInUseList 1729 if (gOmInUseList) { 1730 counters->nInCirculation += gOmInUseCount; 1731 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp); 1732 gOmInUseCount -= deflated_count; 1733 counters->nScavenged += deflated_count; 1734 counters->nInuse += gOmInUseCount; 1735 } 1736 1737 } else { 1738 PaddedEnd<ObjectMonitor> * block = 1739 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 1740 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) { 1741 // Iterate over all extant monitors - Scavenge all idle monitors. 1742 assert(block->object() == CHAINMARKER, "must be a block header"); 1743 counters->nInCirculation += _BLOCKSIZE; 1744 for (int i = 1; i < _BLOCKSIZE; i++) { 1745 ObjectMonitor* mid = (ObjectMonitor*)&block[i]; 1746 oop obj = (oop)mid->object(); 1747 1748 if (obj == NULL) { 1749 // The monitor is not associated with an object. 1750 // The monitor should either be a thread-specific private 1751 // free list or the global free list. 1752 // obj == NULL IMPLIES mid->is_busy() == 0 1753 guarantee(!mid->is_busy(), "invariant"); 1754 continue; 1755 } 1756 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp); 1757 1758 if (deflated) { 1759 mid->FreeNext = NULL; 1760 counters->nScavenged++; 1761 } else { 1762 counters->nInuse++; 1763 } 1764 } 1765 } 1766 } 1767 1768 // Move the scavenged monitors back to the global free list. 1769 if (freeHeadp != NULL) { 1770 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant"); 1771 assert(freeTailp->FreeNext == NULL, "invariant"); 1772 // constant-time list splice - prepend scavenged segment to gFreeList 1773 freeTailp->FreeNext = gFreeList; 1774 gFreeList = freeHeadp; 1775 } 1776 Thread::muxRelease(&gListLock); 1777 1778 } 1779 1780 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1781 gMonitorFreeCount += counters->nScavenged; 1782 1783 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree. 1784 1785 if (ObjectMonitor::Knob_Verbose) { 1786 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d " 1787 "ForceMonitorScavenge=%d : pop=%d free=%d", 1788 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge, 1789 gMonitorPopulation, gMonitorFreeCount); 1790 tty->flush(); 1791 } 1792 1793 ForceMonitorScavenge = 0; // Reset 1794 1795 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged)); 1796 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation)); 1797 1798 // TODO: Add objectMonitor leak detection. 1799 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1800 GVars.stwRandom = os::random(); 1801 GVars.stwCycle++; 1802 } 1803 1804 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 1805 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1806 if (!MonitorInUseLists) return; 1807 1808 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors 1809 ObjectMonitor * freeTailp = NULL; 1810 1811 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp); 1812 1813 Thread::muxAcquire(&gListLock, "scavenge - return"); 1814 1815 // Adjust counters 1816 counters->nInCirculation += thread->omInUseCount; 1817 thread->omInUseCount -= deflated_count; 1818 if (ObjectMonitor::Knob_VerifyInUse) { 1819 verifyInUse(thread); 1820 } 1821 counters->nScavenged += deflated_count; 1822 counters->nInuse += thread->omInUseCount; 1823 1824 // Move the scavenged monitors back to the global free list. 1825 if (freeHeadp != NULL) { 1826 guarantee(freeTailp != NULL && deflated_count > 0, "invariant"); 1827 assert(freeTailp->FreeNext == NULL, "invariant"); 1828 1829 // constant-time list splice - prepend scavenged segment to gFreeList 1830 freeTailp->FreeNext = gFreeList; 1831 gFreeList = freeHeadp; 1832 } 1833 Thread::muxRelease(&gListLock); 1834 } 1835 1836 // Monitor cleanup on JavaThread::exit 1837 1838 // Iterate through monitor cache and attempt to release thread's monitors 1839 // Gives up on a particular monitor if an exception occurs, but continues 1840 // the overall iteration, swallowing the exception. 1841 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1842 private: 1843 TRAPS; 1844 1845 public: 1846 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1847 void do_monitor(ObjectMonitor* mid) { 1848 if (mid->owner() == THREAD) { 1849 if (ObjectMonitor::Knob_VerifyMatch != 0) { 1850 ResourceMark rm; 1851 Handle obj(THREAD, (oop) mid->object()); 1852 tty->print("INFO: unexpected locked object:"); 1853 javaVFrame::print_locked_object_class_name(tty, obj, "locked"); 1854 fatal("exiting JavaThread=" INTPTR_FORMAT 1855 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT, 1856 p2i(THREAD), p2i(mid)); 1857 } 1858 (void)mid->complete_exit(CHECK); 1859 } 1860 } 1861 }; 1862 1863 // Release all inflated monitors owned by THREAD. Lightweight monitors are 1864 // ignored. This is meant to be called during JNI thread detach which assumes 1865 // all remaining monitors are heavyweight. All exceptions are swallowed. 1866 // Scanning the extant monitor list can be time consuming. 1867 // A simple optimization is to add a per-thread flag that indicates a thread 1868 // called jni_monitorenter() during its lifetime. 1869 // 1870 // Instead of No_Savepoint_Verifier it might be cheaper to 1871 // use an idiom of the form: 1872 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1873 // <code that must not run at safepoint> 1874 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1875 // Since the tests are extremely cheap we could leave them enabled 1876 // for normal product builds. 1877 1878 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1879 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1880 NoSafepointVerifier nsv; 1881 ReleaseJavaMonitorsClosure rjmc(THREAD); 1882 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread"); 1883 ObjectSynchronizer::monitors_iterate(&rjmc); 1884 Thread::muxRelease(&gListLock); 1885 THREAD->clear_pending_exception(); 1886 } 1887 1888 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1889 switch (cause) { 1890 case inflate_cause_vm_internal: return "VM Internal"; 1891 case inflate_cause_monitor_enter: return "Monitor Enter"; 1892 case inflate_cause_wait: return "Monitor Wait"; 1893 case inflate_cause_notify: return "Monitor Notify"; 1894 case inflate_cause_hash_code: return "Monitor Hash Code"; 1895 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1896 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1897 default: 1898 ShouldNotReachHere(); 1899 } 1900 return "Unknown"; 1901 } 1902 1903 static void post_monitor_inflate_event(EventJavaMonitorInflate& event, 1904 const oop obj, 1905 const ObjectSynchronizer::InflateCause cause) { 1906 #if INCLUDE_TRACE 1907 assert(event.should_commit(), "check outside"); 1908 event.set_monitorClass(obj->klass()); 1909 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj); 1910 event.set_cause((u1)cause); 1911 event.commit(); 1912 #endif 1913 } 1914 1915 //------------------------------------------------------------------------------ 1916 // Debugging code 1917 1918 void ObjectSynchronizer::sanity_checks(const bool verbose, 1919 const uint cache_line_size, 1920 int *error_cnt_ptr, 1921 int *warning_cnt_ptr) { 1922 u_char *addr_begin = (u_char*)&GVars; 1923 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1924 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1925 1926 if (verbose) { 1927 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1928 sizeof(SharedGlobals)); 1929 } 1930 1931 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1932 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1933 1934 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1935 if (verbose) { 1936 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1937 } 1938 1939 if (cache_line_size != 0) { 1940 // We were able to determine the L1 data cache line size so 1941 // do some cache line specific sanity checks 1942 1943 if (offset_stwRandom < cache_line_size) { 1944 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1945 "to the struct beginning than a cache line which permits " 1946 "false sharing."); 1947 (*warning_cnt_ptr)++; 1948 } 1949 1950 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1951 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1952 "SharedGlobals.hcSequence fields are closer than a cache " 1953 "line which permits false sharing."); 1954 (*warning_cnt_ptr)++; 1955 } 1956 1957 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1958 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1959 "to the struct end than a cache line which permits false " 1960 "sharing."); 1961 (*warning_cnt_ptr)++; 1962 } 1963 } 1964 } 1965 1966 #ifndef PRODUCT 1967 1968 // Check if monitor belongs to the monitor cache 1969 // The list is grow-only so it's *relatively* safe to traverse 1970 // the list of extant blocks without taking a lock. 1971 1972 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1973 PaddedEnd<ObjectMonitor> * block = 1974 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList); 1975 while (block != NULL) { 1976 assert(block->object() == CHAINMARKER, "must be a block header"); 1977 if (monitor > (ObjectMonitor *)&block[0] && 1978 monitor < (ObjectMonitor *)&block[_BLOCKSIZE]) { 1979 address mon = (address)monitor; 1980 address blk = (address)block; 1981 size_t diff = mon - blk; 1982 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned"); 1983 return 1; 1984 } 1985 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext; 1986 } 1987 return 0; 1988 } 1989 1990 #endif