synchronizer.cpp revision 7051:0420e825bb3c
1/* 2 * Copyright (c) 1998, 2014, 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 "memory/resourceArea.hpp" 28#include "oops/markOop.hpp" 29#include "oops/oop.inline.hpp" 30#include "runtime/atomic.inline.hpp" 31#include "runtime/biasedLocking.hpp" 32#include "runtime/handles.inline.hpp" 33#include "runtime/interfaceSupport.hpp" 34#include "runtime/mutexLocker.hpp" 35#include "runtime/objectMonitor.hpp" 36#include "runtime/objectMonitor.inline.hpp" 37#include "runtime/osThread.hpp" 38#include "runtime/stubRoutines.hpp" 39#include "runtime/synchronizer.hpp" 40#include "runtime/thread.inline.hpp" 41#include "utilities/dtrace.hpp" 42#include "utilities/events.hpp" 43#include "utilities/preserveException.hpp" 44 45#if defined(__GNUC__) && !defined(PPC64) 46// Need to inhibit inlining for older versions of GCC to avoid build-time failures 47 #define NOINLINE __attribute__((noinline)) 48#else 49 #define NOINLINE 50#endif 51 52PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 53 54// The "core" versions of monitor enter and exit reside in this file. 55// The interpreter and compilers contain specialized transliterated 56// variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 57// for instance. If you make changes here, make sure to modify the 58// interpreter, and both C1 and C2 fast-path inline locking code emission. 59// 60// ----------------------------------------------------------------------------- 61 62#ifdef DTRACE_ENABLED 63 64// Only bother with this argument setup if dtrace is available 65// TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 66 67#define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 68 char* bytes = NULL; \ 69 int len = 0; \ 70 jlong jtid = SharedRuntime::get_java_tid(thread); \ 71 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 72 if (klassname != NULL) { \ 73 bytes = (char*)klassname->bytes(); \ 74 len = klassname->utf8_length(); \ 75 } 76 77#define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 78 { \ 79 if (DTraceMonitorProbes) { \ 80 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 81 HOTSPOT_MONITOR_WAIT(jtid, \ 82 (uintptr_t)(monitor), bytes, len, (millis)); \ 83 } \ 84 } 85 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 105int 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 111static volatile intptr_t InflationLocks[NINFLATIONLOCKS]; 112 113ObjectMonitor * ObjectSynchronizer::gBlockList = NULL; 114ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 115ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 116int ObjectSynchronizer::gOmInUseCount = 0; 117static volatile intptr_t ListLock = 0; // protects global monitor free-list cache 118static volatile int MonitorFreeCount = 0; // # on gFreeList 119static volatile int MonitorPopulation = 0; // # Extant -- in circulation 120#define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 121 122// ----------------------------------------------------------------------------- 123// Fast Monitor Enter/Exit 124// This the fast monitor enter. The interpreter and compiler use 125// some assembly copies of this code. Make sure update those code 126// if the following function is changed. The implementation is 127// extremely sensitive to race condition. Be careful. 128 129void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 130 bool attempt_rebias, TRAPS) { 131 if (UseBiasedLocking) { 132 if (!SafepointSynchronize::is_at_safepoint()) { 133 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); 134 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { 135 return; 136 } 137 } else { 138 assert(!attempt_rebias, "can not rebias toward VM thread"); 139 BiasedLocking::revoke_at_safepoint(obj); 140 } 141 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 142 } 143 144 slow_enter(obj, lock, THREAD); 145} 146 147void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { 148 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); 149 // if displaced header is null, the previous enter is recursive enter, no-op 150 markOop dhw = lock->displaced_header(); 151 markOop mark; 152 if (dhw == NULL) { 153 // Recursive stack-lock. 154 // Diagnostics -- Could be: stack-locked, inflating, inflated. 155 mark = object->mark(); 156 assert(!mark->is_neutral(), "invariant"); 157 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { 158 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant"); 159 } 160 if (mark->has_monitor()) { 161 ObjectMonitor * m = mark->monitor(); 162 assert(((oop)(m->object()))->mark() == mark, "invariant"); 163 assert(m->is_entered(THREAD), "invariant"); 164 } 165 return; 166 } 167 168 mark = object->mark(); 169 170 // If the object is stack-locked by the current thread, try to 171 // swing the displaced header from the box back to the mark. 172 if (mark == (markOop) lock) { 173 assert(dhw->is_neutral(), "invariant"); 174 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { 175 TEVENT(fast_exit: release stacklock); 176 return; 177 } 178 } 179 180 ObjectSynchronizer::inflate(THREAD, object)->exit(true, THREAD); 181} 182 183// ----------------------------------------------------------------------------- 184// Interpreter/Compiler Slow Case 185// This routine is used to handle interpreter/compiler slow case 186// We don't need to use fast path here, because it must have been 187// failed in the interpreter/compiler code. 188void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { 189 markOop mark = obj->mark(); 190 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); 191 192 if (mark->is_neutral()) { 193 // Anticipate successful CAS -- the ST of the displaced mark must 194 // be visible <= the ST performed by the CAS. 195 lock->set_displaced_header(mark); 196 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { 197 TEVENT(slow_enter: release stacklock); 198 return; 199 } 200 // Fall through to inflate() ... 201 } else if (mark->has_locker() && 202 THREAD->is_lock_owned((address)mark->locker())) { 203 assert(lock != mark->locker(), "must not re-lock the same lock"); 204 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); 205 lock->set_displaced_header(NULL); 206 return; 207 } 208 209 // The object header will never be displaced to this lock, 210 // so it does not matter what the value is, except that it 211 // must be non-zero to avoid looking like a re-entrant lock, 212 // and must not look locked either. 213 lock->set_displaced_header(markOopDesc::unused_mark()); 214 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 215} 216 217// This routine is used to handle interpreter/compiler slow case 218// We don't need to use fast path here, because it must have 219// failed in the interpreter/compiler code. Simply use the heavy 220// weight monitor should be ok, unless someone find otherwise. 221void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { 222 fast_exit(object, lock, THREAD); 223} 224 225// ----------------------------------------------------------------------------- 226// Class Loader support to workaround deadlocks on the class loader lock objects 227// Also used by GC 228// complete_exit()/reenter() are used to wait on a nested lock 229// i.e. to give up an outer lock completely and then re-enter 230// Used when holding nested locks - lock acquisition order: lock1 then lock2 231// 1) complete_exit lock1 - saving recursion count 232// 2) wait on lock2 233// 3) when notified on lock2, unlock lock2 234// 4) reenter lock1 with original recursion count 235// 5) lock lock2 236// NOTE: must use heavy weight monitor to handle complete_exit/reenter() 237intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 238 TEVENT(complete_exit); 239 if (UseBiasedLocking) { 240 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 241 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 242 } 243 244 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 245 246 return monitor->complete_exit(THREAD); 247} 248 249// NOTE: must use heavy weight monitor to handle complete_exit/reenter() 250void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { 251 TEVENT(reenter); 252 if (UseBiasedLocking) { 253 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 254 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 255 } 256 257 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 258 259 monitor->reenter(recursion, THREAD); 260} 261// ----------------------------------------------------------------------------- 262// JNI locks on java objects 263// NOTE: must use heavy weight monitor to handle jni monitor enter 264void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 265 // the current locking is from JNI instead of Java code 266 TEVENT(jni_enter); 267 if (UseBiasedLocking) { 268 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 269 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 270 } 271 THREAD->set_current_pending_monitor_is_from_java(false); 272 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); 273 THREAD->set_current_pending_monitor_is_from_java(true); 274} 275 276// NOTE: must use heavy weight monitor to handle jni monitor enter 277bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) { 278 if (UseBiasedLocking) { 279 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 280 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 281 } 282 283 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj()); 284 return monitor->try_enter(THREAD); 285} 286 287 288// NOTE: must use heavy weight monitor to handle jni monitor exit 289void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 290 TEVENT(jni_exit); 291 if (UseBiasedLocking) { 292 Handle h_obj(THREAD, obj); 293 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); 294 obj = h_obj(); 295 } 296 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 297 298 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); 299 // If this thread has locked the object, exit the monitor. Note: can't use 300 // monitor->check(CHECK); must exit even if an exception is pending. 301 if (monitor->check(THREAD)) { 302 monitor->exit(true, THREAD); 303 } 304} 305 306// ----------------------------------------------------------------------------- 307// Internal VM locks on java objects 308// standard constructor, allows locking failures 309ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { 310 _dolock = doLock; 311 _thread = thread; 312 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) 313 _obj = obj; 314 315 if (_dolock) { 316 TEVENT(ObjectLocker); 317 318 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); 319 } 320} 321 322ObjectLocker::~ObjectLocker() { 323 if (_dolock) { 324 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); 325 } 326} 327 328 329// ----------------------------------------------------------------------------- 330// Wait/Notify/NotifyAll 331// NOTE: must use heavy weight monitor to handle wait() 332int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 333 if (UseBiasedLocking) { 334 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 335 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 336 } 337 if (millis < 0) { 338 TEVENT(wait - throw IAX); 339 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 340 } 341 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); 342 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 343 monitor->wait(millis, true, THREAD); 344 345 // This dummy call is in place to get around dtrace bug 6254741. Once 346 // that's fixed we can uncomment the following line, remove the call 347 // and change this function back into a "void" func. 348 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 349 return dtrace_waited_probe(monitor, obj, THREAD); 350} 351 352void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 353 if (UseBiasedLocking) { 354 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 355 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 356 } 357 if (millis < 0) { 358 TEVENT(wait - throw IAX); 359 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 360 } 361 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD); 362} 363 364void ObjectSynchronizer::notify(Handle obj, TRAPS) { 365 if (UseBiasedLocking) { 366 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 367 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 368 } 369 370 markOop mark = obj->mark(); 371 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 372 return; 373 } 374 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); 375} 376 377// NOTE: see comment of notify() 378void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 379 if (UseBiasedLocking) { 380 BiasedLocking::revoke_and_rebias(obj, false, THREAD); 381 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 382 } 383 384 markOop mark = obj->mark(); 385 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { 386 return; 387 } 388 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); 389} 390 391// ----------------------------------------------------------------------------- 392// Hash Code handling 393// 394// Performance concern: 395// OrderAccess::storestore() calls release() which at one time stored 0 396// into the global volatile OrderAccess::dummy variable. This store was 397// unnecessary for correctness. Many threads storing into a common location 398// causes considerable cache migration or "sloshing" on large SMP systems. 399// As such, I avoided using OrderAccess::storestore(). In some cases 400// OrderAccess::fence() -- which incurs local latency on the executing 401// processor -- is a better choice as it scales on SMP systems. 402// 403// See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 404// a discussion of coherency costs. Note that all our current reference 405// platforms provide strong ST-ST order, so the issue is moot on IA32, 406// x64, and SPARC. 407// 408// As a general policy we use "volatile" to control compiler-based reordering 409// and explicit fences (barriers) to control for architectural reordering 410// performed by the CPU(s) or platform. 411 412struct SharedGlobals { 413 // These are highly shared mostly-read variables. 414 // To avoid false-sharing they need to be the sole occupants of a $ line. 415 double padPrefix[8]; 416 volatile int stwRandom; 417 volatile int stwCycle; 418 419 // Hot RW variables -- Sequester to avoid false-sharing 420 double padSuffix[16]; 421 volatile int hcSequence; 422 double padFinal[8]; 423}; 424 425static SharedGlobals GVars; 426static int MonitorScavengeThreshold = 1000000; 427static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 428 429static markOop ReadStableMark(oop obj) { 430 markOop mark = obj->mark(); 431 if (!mark->is_being_inflated()) { 432 return mark; // normal fast-path return 433 } 434 435 int its = 0; 436 for (;;) { 437 markOop mark = obj->mark(); 438 if (!mark->is_being_inflated()) { 439 return mark; // normal fast-path return 440 } 441 442 // The object is being inflated by some other thread. 443 // The caller of ReadStableMark() must wait for inflation to complete. 444 // Avoid live-lock 445 // TODO: consider calling SafepointSynchronize::do_call_back() while 446 // spinning to see if there's a safepoint pending. If so, immediately 447 // yielding or blocking would be appropriate. Avoid spinning while 448 // there is a safepoint pending. 449 // TODO: add inflation contention performance counters. 450 // TODO: restrict the aggregate number of spinners. 451 452 ++its; 453 if (its > 10000 || !os::is_MP()) { 454 if (its & 1) { 455 os::naked_yield(); 456 TEVENT(Inflate: INFLATING - yield); 457 } else { 458 // Note that the following code attenuates the livelock problem but is not 459 // a complete remedy. A more complete solution would require that the inflating 460 // thread hold the associated inflation lock. The following code simply restricts 461 // the number of spinners to at most one. We'll have N-2 threads blocked 462 // on the inflationlock, 1 thread holding the inflation lock and using 463 // a yield/park strategy, and 1 thread in the midst of inflation. 464 // A more refined approach would be to change the encoding of INFLATING 465 // to allow encapsulation of a native thread pointer. Threads waiting for 466 // inflation to complete would use CAS to push themselves onto a singly linked 467 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 468 // and calling park(). When inflation was complete the thread that accomplished inflation 469 // would detach the list and set the markword to inflated with a single CAS and 470 // then for each thread on the list, set the flag and unpark() the thread. 471 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 472 // wakes at most one thread whereas we need to wake the entire list. 473 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 474 int YieldThenBlock = 0; 475 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 476 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 477 Thread::muxAcquire(InflationLocks + ix, "InflationLock"); 478 while (obj->mark() == markOopDesc::INFLATING()) { 479 // Beware: NakedYield() is advisory and has almost no effect on some platforms 480 // so we periodically call Self->_ParkEvent->park(1). 481 // We use a mixed spin/yield/block mechanism. 482 if ((YieldThenBlock++) >= 16) { 483 Thread::current()->_ParkEvent->park(1); 484 } else { 485 os::naked_yield(); 486 } 487 } 488 Thread::muxRelease(InflationLocks + ix); 489 TEVENT(Inflate: INFLATING - yield/park); 490 } 491 } else { 492 SpinPause(); // SMP-polite spinning 493 } 494 } 495} 496 497// hashCode() generation : 498// 499// Possibilities: 500// * MD5Digest of {obj,stwRandom} 501// * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. 502// * A DES- or AES-style SBox[] mechanism 503// * One of the Phi-based schemes, such as: 504// 2654435761 = 2^32 * Phi (golden ratio) 505// HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; 506// * A variation of Marsaglia's shift-xor RNG scheme. 507// * (obj ^ stwRandom) is appealing, but can result 508// in undesirable regularity in the hashCode values of adjacent objects 509// (objects allocated back-to-back, in particular). This could potentially 510// result in hashtable collisions and reduced hashtable efficiency. 511// There are simple ways to "diffuse" the middle address bits over the 512// generated hashCode values: 513 514static inline intptr_t get_next_hash(Thread * Self, oop obj) { 515 intptr_t value = 0; 516 if (hashCode == 0) { 517 // This form uses an unguarded global Park-Miller RNG, 518 // so it's possible for two threads to race and generate the same RNG. 519 // On MP system we'll have lots of RW access to a global, so the 520 // mechanism induces lots of coherency traffic. 521 value = os::random(); 522 } else if (hashCode == 1) { 523 // This variation has the property of being stable (idempotent) 524 // between STW operations. This can be useful in some of the 1-0 525 // synchronization schemes. 526 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 527 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 528 } else if (hashCode == 2) { 529 value = 1; // for sensitivity testing 530 } else if (hashCode == 3) { 531 value = ++GVars.hcSequence; 532 } else if (hashCode == 4) { 533 value = cast_from_oop<intptr_t>(obj); 534 } else { 535 // Marsaglia's xor-shift scheme with thread-specific state 536 // This is probably the best overall implementation -- we'll 537 // likely make this the default in future releases. 538 unsigned t = Self->_hashStateX; 539 t ^= (t << 11); 540 Self->_hashStateX = Self->_hashStateY; 541 Self->_hashStateY = Self->_hashStateZ; 542 Self->_hashStateZ = Self->_hashStateW; 543 unsigned v = Self->_hashStateW; 544 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 545 Self->_hashStateW = v; 546 value = v; 547 } 548 549 value &= markOopDesc::hash_mask; 550 if (value == 0) value = 0xBAD; 551 assert(value != markOopDesc::no_hash, "invariant"); 552 TEVENT(hashCode: GENERATE); 553 return value; 554} 555 556intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) { 557 if (UseBiasedLocking) { 558 // NOTE: many places throughout the JVM do not expect a safepoint 559 // to be taken here, in particular most operations on perm gen 560 // objects. However, we only ever bias Java instances and all of 561 // the call sites of identity_hash that might revoke biases have 562 // been checked to make sure they can handle a safepoint. The 563 // added check of the bias pattern is to avoid useless calls to 564 // thread-local storage. 565 if (obj->mark()->has_bias_pattern()) { 566 // Handle for oop obj in case of STW safepoint 567 Handle hobj(Self, obj); 568 // Relaxing assertion for bug 6320749. 569 assert(Universe::verify_in_progress() || 570 !SafepointSynchronize::is_at_safepoint(), 571 "biases should not be seen by VM thread here"); 572 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 573 obj = hobj(); 574 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 575 } 576 } 577 578 // hashCode() is a heap mutator ... 579 // Relaxing assertion for bug 6320749. 580 assert(Universe::verify_in_progress() || 581 !SafepointSynchronize::is_at_safepoint(), "invariant"); 582 assert(Universe::verify_in_progress() || 583 Self->is_Java_thread() , "invariant"); 584 assert(Universe::verify_in_progress() || 585 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 586 587 ObjectMonitor* monitor = NULL; 588 markOop temp, test; 589 intptr_t hash; 590 markOop mark = ReadStableMark(obj); 591 592 // object should remain ineligible for biased locking 593 assert(!mark->has_bias_pattern(), "invariant"); 594 595 if (mark->is_neutral()) { 596 hash = mark->hash(); // this is a normal header 597 if (hash) { // if it has hash, just return it 598 return hash; 599 } 600 hash = get_next_hash(Self, obj); // allocate a new hash code 601 temp = mark->copy_set_hash(hash); // merge the hash code into header 602 // use (machine word version) atomic operation to install the hash 603 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 604 if (test == mark) { 605 return hash; 606 } 607 // If atomic operation failed, we must inflate the header 608 // into heavy weight monitor. We could add more code here 609 // for fast path, but it does not worth the complexity. 610 } else if (mark->has_monitor()) { 611 monitor = mark->monitor(); 612 temp = monitor->header(); 613 assert(temp->is_neutral(), "invariant"); 614 hash = temp->hash(); 615 if (hash) { 616 return hash; 617 } 618 // Skip to the following code to reduce code size 619 } else if (Self->is_lock_owned((address)mark->locker())) { 620 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 621 assert(temp->is_neutral(), "invariant"); 622 hash = temp->hash(); // by current thread, check if the displaced 623 if (hash) { // header contains hash code 624 return hash; 625 } 626 // WARNING: 627 // The displaced header is strictly immutable. 628 // It can NOT be changed in ANY cases. So we have 629 // to inflate the header into heavyweight monitor 630 // even the current thread owns the lock. The reason 631 // is the BasicLock (stack slot) will be asynchronously 632 // read by other threads during the inflate() function. 633 // Any change to stack may not propagate to other threads 634 // correctly. 635 } 636 637 // Inflate the monitor to set hash code 638 monitor = ObjectSynchronizer::inflate(Self, obj); 639 // Load displaced header and check it has hash code 640 mark = monitor->header(); 641 assert(mark->is_neutral(), "invariant"); 642 hash = mark->hash(); 643 if (hash == 0) { 644 hash = get_next_hash(Self, obj); 645 temp = mark->copy_set_hash(hash); // merge hash code into header 646 assert(temp->is_neutral(), "invariant"); 647 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 648 if (test != mark) { 649 // The only update to the header in the monitor (outside GC) 650 // is install the hash code. If someone add new usage of 651 // displaced header, please update this code 652 hash = test->hash(); 653 assert(test->is_neutral(), "invariant"); 654 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 655 } 656 } 657 // We finally get the hash 658 return hash; 659} 660 661// Deprecated -- use FastHashCode() instead. 662 663intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 664 return FastHashCode(Thread::current(), obj()); 665} 666 667 668bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 669 Handle h_obj) { 670 if (UseBiasedLocking) { 671 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 672 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 673 } 674 675 assert(thread == JavaThread::current(), "Can only be called on current thread"); 676 oop obj = h_obj(); 677 678 markOop mark = ReadStableMark(obj); 679 680 // Uncontended case, header points to stack 681 if (mark->has_locker()) { 682 return thread->is_lock_owned((address)mark->locker()); 683 } 684 // Contended case, header points to ObjectMonitor (tagged pointer) 685 if (mark->has_monitor()) { 686 ObjectMonitor* monitor = mark->monitor(); 687 return monitor->is_entered(thread) != 0; 688 } 689 // Unlocked case, header in place 690 assert(mark->is_neutral(), "sanity check"); 691 return false; 692} 693 694// Be aware of this method could revoke bias of the lock object. 695// This method queries the ownership of the lock handle specified by 'h_obj'. 696// If the current thread owns the lock, it returns owner_self. If no 697// thread owns the lock, it returns owner_none. Otherwise, it will return 698// owner_other. 699ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 700(JavaThread *self, Handle h_obj) { 701 // The caller must beware this method can revoke bias, and 702 // revocation can result in a safepoint. 703 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 704 assert(self->thread_state() != _thread_blocked, "invariant"); 705 706 // Possible mark states: neutral, biased, stack-locked, inflated 707 708 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 709 // CASE: biased 710 BiasedLocking::revoke_and_rebias(h_obj, false, self); 711 assert(!h_obj->mark()->has_bias_pattern(), 712 "biases should be revoked by now"); 713 } 714 715 assert(self == JavaThread::current(), "Can only be called on current thread"); 716 oop obj = h_obj(); 717 markOop mark = ReadStableMark(obj); 718 719 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 720 if (mark->has_locker()) { 721 return self->is_lock_owned((address)mark->locker()) ? 722 owner_self : owner_other; 723 } 724 725 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 726 // The Object:ObjectMonitor relationship is stable as long as we're 727 // not at a safepoint. 728 if (mark->has_monitor()) { 729 void * owner = mark->monitor()->_owner; 730 if (owner == NULL) return owner_none; 731 return (owner == self || 732 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 733 } 734 735 // CASE: neutral 736 assert(mark->is_neutral(), "sanity check"); 737 return owner_none; // it's unlocked 738} 739 740// FIXME: jvmti should call this 741JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 742 if (UseBiasedLocking) { 743 if (SafepointSynchronize::is_at_safepoint()) { 744 BiasedLocking::revoke_at_safepoint(h_obj); 745 } else { 746 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 747 } 748 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 749 } 750 751 oop obj = h_obj(); 752 address owner = NULL; 753 754 markOop mark = ReadStableMark(obj); 755 756 // Uncontended case, header points to stack 757 if (mark->has_locker()) { 758 owner = (address) mark->locker(); 759 } 760 761 // Contended case, header points to ObjectMonitor (tagged pointer) 762 if (mark->has_monitor()) { 763 ObjectMonitor* monitor = mark->monitor(); 764 assert(monitor != NULL, "monitor should be non-null"); 765 owner = (address) monitor->owner(); 766 } 767 768 if (owner != NULL) { 769 // owning_thread_from_monitor_owner() may also return NULL here 770 return Threads::owning_thread_from_monitor_owner(owner, doLock); 771 } 772 773 // Unlocked case, header in place 774 // Cannot have assertion since this object may have been 775 // locked by another thread when reaching here. 776 // assert(mark->is_neutral(), "sanity check"); 777 778 return NULL; 779} 780// Visitors ... 781 782void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 783 ObjectMonitor* block = gBlockList; 784 ObjectMonitor* mid; 785 while (block) { 786 assert(block->object() == CHAINMARKER, "must be a block header"); 787 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 788 mid = block + i; 789 oop object = (oop) mid->object(); 790 if (object != NULL) { 791 closure->do_monitor(mid); 792 } 793 } 794 block = (ObjectMonitor*) block->FreeNext; 795 } 796} 797 798// Get the next block in the block list. 799static inline ObjectMonitor* next(ObjectMonitor* block) { 800 assert(block->object() == CHAINMARKER, "must be a block header"); 801 block = block->FreeNext; 802 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 803 return block; 804} 805 806 807void ObjectSynchronizer::oops_do(OopClosure* f) { 808 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 809 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { 810 assert(block->object() == CHAINMARKER, "must be a block header"); 811 for (int i = 1; i < _BLOCKSIZE; i++) { 812 ObjectMonitor* mid = &block[i]; 813 if (mid->object() != NULL) { 814 f->do_oop((oop*)mid->object_addr()); 815 } 816 } 817 } 818} 819 820 821// ----------------------------------------------------------------------------- 822// ObjectMonitor Lifecycle 823// ----------------------- 824// Inflation unlinks monitors from the global gFreeList and 825// associates them with objects. Deflation -- which occurs at 826// STW-time -- disassociates idle monitors from objects. Such 827// scavenged monitors are returned to the gFreeList. 828// 829// The global list is protected by ListLock. All the critical sections 830// are short and operate in constant-time. 831// 832// ObjectMonitors reside in type-stable memory (TSM) and are immortal. 833// 834// Lifecycle: 835// -- unassigned and on the global free list 836// -- unassigned and on a thread's private omFreeList 837// -- assigned to an object. The object is inflated and the mark refers 838// to the objectmonitor. 839 840 841// Constraining monitor pool growth via MonitorBound ... 842// 843// The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 844// the rate of scavenging is driven primarily by GC. As such, we can find 845// an inordinate number of monitors in circulation. 846// To avoid that scenario we can artificially induce a STW safepoint 847// if the pool appears to be growing past some reasonable bound. 848// Generally we favor time in space-time tradeoffs, but as there's no 849// natural back-pressure on the # of extant monitors we need to impose some 850// type of limit. Beware that if MonitorBound is set to too low a value 851// we could just loop. In addition, if MonitorBound is set to a low value 852// we'll incur more safepoints, which are harmful to performance. 853// See also: GuaranteedSafepointInterval 854// 855// The current implementation uses asynchronous VM operations. 856 857static void InduceScavenge(Thread * Self, const char * Whence) { 858 // Induce STW safepoint to trim monitors 859 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 860 // More precisely, trigger an asynchronous STW safepoint as the number 861 // of active monitors passes the specified threshold. 862 // TODO: assert thread state is reasonable 863 864 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 865 if (ObjectMonitor::Knob_Verbose) { 866 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; 867 ::fflush(stdout); 868 } 869 // Induce a 'null' safepoint to scavenge monitors 870 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 871 // to the VMthread and have a lifespan longer than that of this activation record. 872 // The VMThread will delete the op when completed. 873 VMThread::execute(new VM_ForceAsyncSafepoint()); 874 875 if (ObjectMonitor::Knob_Verbose) { 876 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; 877 ::fflush(stdout); 878 } 879 } 880} 881 882void ObjectSynchronizer::verifyInUse(Thread *Self) { 883 ObjectMonitor* mid; 884 int inusetally = 0; 885 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 886 inusetally++; 887 } 888 assert(inusetally == Self->omInUseCount, "inuse count off"); 889 890 int freetally = 0; 891 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 892 freetally++; 893 } 894 assert(freetally == Self->omFreeCount, "free count off"); 895} 896 897ObjectMonitor * NOINLINE ObjectSynchronizer::omAlloc(Thread * Self) { 898 // A large MAXPRIVATE value reduces both list lock contention 899 // and list coherency traffic, but also tends to increase the 900 // number of objectMonitors in circulation as well as the STW 901 // scavenge costs. As usual, we lean toward time in space-time 902 // tradeoffs. 903 const int MAXPRIVATE = 1024; 904 for (;;) { 905 ObjectMonitor * m; 906 907 // 1: try to allocate from the thread's local omFreeList. 908 // Threads will attempt to allocate first from their local list, then 909 // from the global list, and only after those attempts fail will the thread 910 // attempt to instantiate new monitors. Thread-local free lists take 911 // heat off the ListLock and improve allocation latency, as well as reducing 912 // coherency traffic on the shared global list. 913 m = Self->omFreeList; 914 if (m != NULL) { 915 Self->omFreeList = m->FreeNext; 916 Self->omFreeCount--; 917 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 918 guarantee(m->object() == NULL, "invariant"); 919 if (MonitorInUseLists) { 920 m->FreeNext = Self->omInUseList; 921 Self->omInUseList = m; 922 Self->omInUseCount++; 923 if (ObjectMonitor::Knob_VerifyInUse) { 924 verifyInUse(Self); 925 } 926 } else { 927 m->FreeNext = NULL; 928 } 929 return m; 930 } 931 932 // 2: try to allocate from the global gFreeList 933 // CONSIDER: use muxTry() instead of muxAcquire(). 934 // If the muxTry() fails then drop immediately into case 3. 935 // If we're using thread-local free lists then try 936 // to reprovision the caller's free list. 937 if (gFreeList != NULL) { 938 // Reprovision the thread's omFreeList. 939 // Use bulk transfers to reduce the allocation rate and heat 940 // on various locks. 941 Thread::muxAcquire(&ListLock, "omAlloc"); 942 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 943 MonitorFreeCount--; 944 ObjectMonitor * take = gFreeList; 945 gFreeList = take->FreeNext; 946 guarantee(take->object() == NULL, "invariant"); 947 guarantee(!take->is_busy(), "invariant"); 948 take->Recycle(); 949 omRelease(Self, take, false); 950 } 951 Thread::muxRelease(&ListLock); 952 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 953 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 954 TEVENT(omFirst - reprovision); 955 956 const int mx = MonitorBound; 957 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) { 958 // We can't safely induce a STW safepoint from omAlloc() as our thread 959 // state may not be appropriate for such activities and callers may hold 960 // naked oops, so instead we defer the action. 961 InduceScavenge(Self, "omAlloc"); 962 } 963 continue; 964 } 965 966 // 3: allocate a block of new ObjectMonitors 967 // Both the local and global free lists are empty -- resort to malloc(). 968 // In the current implementation objectMonitors are TSM - immortal. 969 assert(_BLOCKSIZE > 1, "invariant"); 970 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; 971 972 // NOTE: (almost) no way to recover if allocation failed. 973 // We might be able to induce a STW safepoint and scavenge enough 974 // objectMonitors to permit progress. 975 if (temp == NULL) { 976 vm_exit_out_of_memory(sizeof (ObjectMonitor[_BLOCKSIZE]), OOM_MALLOC_ERROR, 977 "Allocate ObjectMonitors"); 978 } 979 980 // Format the block. 981 // initialize the linked list, each monitor points to its next 982 // forming the single linked free list, the very first monitor 983 // will points to next block, which forms the block list. 984 // The trick of using the 1st element in the block as gBlockList 985 // linkage should be reconsidered. A better implementation would 986 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 987 988 for (int i = 1; i < _BLOCKSIZE; i++) { 989 temp[i].FreeNext = &temp[i+1]; 990 } 991 992 // terminate the last monitor as the end of list 993 temp[_BLOCKSIZE - 1].FreeNext = NULL; 994 995 // Element [0] is reserved for global list linkage 996 temp[0].set_object(CHAINMARKER); 997 998 // Consider carving out this thread's current request from the 999 // block in hand. This avoids some lock traffic and redundant 1000 // list activity. 1001 1002 // Acquire the ListLock to manipulate BlockList and FreeList. 1003 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1004 Thread::muxAcquire(&ListLock, "omAlloc [2]"); 1005 MonitorPopulation += _BLOCKSIZE-1; 1006 MonitorFreeCount += _BLOCKSIZE-1; 1007 1008 // Add the new block to the list of extant blocks (gBlockList). 1009 // The very first objectMonitor in a block is reserved and dedicated. 1010 // It serves as blocklist "next" linkage. 1011 temp[0].FreeNext = gBlockList; 1012 gBlockList = temp; 1013 1014 // Add the new string of objectMonitors to the global free list 1015 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1016 gFreeList = temp + 1; 1017 Thread::muxRelease(&ListLock); 1018 TEVENT(Allocate block of monitors); 1019 } 1020} 1021 1022// Place "m" on the caller's private per-thread omFreeList. 1023// In practice there's no need to clamp or limit the number of 1024// monitors on a thread's omFreeList as the only time we'll call 1025// omRelease is to return a monitor to the free list after a CAS 1026// attempt failed. This doesn't allow unbounded #s of monitors to 1027// accumulate on a thread's free list. 1028 1029void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m, 1030 bool fromPerThreadAlloc) { 1031 guarantee(m->object() == NULL, "invariant"); 1032 1033 // Remove from omInUseList 1034 if (MonitorInUseLists && fromPerThreadAlloc) { 1035 ObjectMonitor* curmidinuse = NULL; 1036 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL;) { 1037 if (m == mid) { 1038 // extract from per-thread in-use-list 1039 if (mid == Self->omInUseList) { 1040 Self->omInUseList = mid->FreeNext; 1041 } else if (curmidinuse != NULL) { 1042 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist 1043 } 1044 Self->omInUseCount--; 1045 if (ObjectMonitor::Knob_VerifyInUse) { 1046 verifyInUse(Self); 1047 } 1048 break; 1049 } else { 1050 curmidinuse = mid; 1051 mid = mid->FreeNext; 1052 } 1053 } 1054 } 1055 1056 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1057 m->FreeNext = Self->omFreeList; 1058 Self->omFreeList = m; 1059 Self->omFreeCount++; 1060} 1061 1062// Return the monitors of a moribund thread's local free list to 1063// the global free list. Typically a thread calls omFlush() when 1064// it's dying. We could also consider having the VM thread steal 1065// monitors from threads that have not run java code over a few 1066// consecutive STW safepoints. Relatedly, we might decay 1067// omFreeProvision at STW safepoints. 1068// 1069// Also return the monitors of a moribund thread's omInUseList to 1070// a global gOmInUseList under the global list lock so these 1071// will continue to be scanned. 1072// 1073// We currently call omFlush() from the Thread:: dtor _after the thread 1074// has been excised from the thread list and is no longer a mutator. 1075// That means that omFlush() can run concurrently with a safepoint and 1076// the scavenge operator. Calling omFlush() from JavaThread::exit() might 1077// be a better choice as we could safely reason that that the JVM is 1078// not at a safepoint at the time of the call, and thus there could 1079// be not inopportune interleavings between omFlush() and the scavenge 1080// operator. 1081 1082void ObjectSynchronizer::omFlush(Thread * Self) { 1083 ObjectMonitor * List = Self->omFreeList; // Null-terminated SLL 1084 Self->omFreeList = NULL; 1085 ObjectMonitor * Tail = NULL; 1086 int Tally = 0; 1087 if (List != NULL) { 1088 ObjectMonitor * s; 1089 for (s = List; s != NULL; s = s->FreeNext) { 1090 Tally++; 1091 Tail = s; 1092 guarantee(s->object() == NULL, "invariant"); 1093 guarantee(!s->is_busy(), "invariant"); 1094 s->set_owner(NULL); // redundant but good hygiene 1095 TEVENT(omFlush - Move one); 1096 } 1097 guarantee(Tail != NULL && List != NULL, "invariant"); 1098 } 1099 1100 ObjectMonitor * InUseList = Self->omInUseList; 1101 ObjectMonitor * InUseTail = NULL; 1102 int InUseTally = 0; 1103 if (InUseList != NULL) { 1104 Self->omInUseList = NULL; 1105 ObjectMonitor *curom; 1106 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) { 1107 InUseTail = curom; 1108 InUseTally++; 1109 } 1110 assert(Self->omInUseCount == InUseTally, "inuse count off"); 1111 Self->omInUseCount = 0; 1112 guarantee(InUseTail != NULL && InUseList != NULL, "invariant"); 1113 } 1114 1115 Thread::muxAcquire(&ListLock, "omFlush"); 1116 if (Tail != NULL) { 1117 Tail->FreeNext = gFreeList; 1118 gFreeList = List; 1119 MonitorFreeCount += Tally; 1120 } 1121 1122 if (InUseTail != NULL) { 1123 InUseTail->FreeNext = gOmInUseList; 1124 gOmInUseList = InUseList; 1125 gOmInUseCount += InUseTally; 1126 } 1127 1128 Thread::muxRelease(&ListLock); 1129 TEVENT(omFlush); 1130} 1131 1132// Fast path code shared by multiple functions 1133ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1134 markOop mark = obj->mark(); 1135 if (mark->has_monitor()) { 1136 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1137 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1138 return mark->monitor(); 1139 } 1140 return ObjectSynchronizer::inflate(Thread::current(), obj); 1141} 1142 1143 1144// Note that we could encounter some performance loss through false-sharing as 1145// multiple locks occupy the same $ line. Padding might be appropriate. 1146 1147 1148ObjectMonitor * NOINLINE ObjectSynchronizer::inflate(Thread * Self, 1149 oop object) { 1150 // Inflate mutates the heap ... 1151 // Relaxing assertion for bug 6320749. 1152 assert(Universe::verify_in_progress() || 1153 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1154 1155 for (;;) { 1156 const markOop mark = object->mark(); 1157 assert(!mark->has_bias_pattern(), "invariant"); 1158 1159 // The mark can be in one of the following states: 1160 // * Inflated - just return 1161 // * Stack-locked - coerce it to inflated 1162 // * INFLATING - busy wait for conversion to complete 1163 // * Neutral - aggressively inflate the object. 1164 // * BIASED - Illegal. We should never see this 1165 1166 // CASE: inflated 1167 if (mark->has_monitor()) { 1168 ObjectMonitor * inf = mark->monitor(); 1169 assert(inf->header()->is_neutral(), "invariant"); 1170 assert(inf->object() == object, "invariant"); 1171 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1172 return inf; 1173 } 1174 1175 // CASE: inflation in progress - inflating over a stack-lock. 1176 // Some other thread is converting from stack-locked to inflated. 1177 // Only that thread can complete inflation -- other threads must wait. 1178 // The INFLATING value is transient. 1179 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1180 // We could always eliminate polling by parking the thread on some auxiliary list. 1181 if (mark == markOopDesc::INFLATING()) { 1182 TEVENT(Inflate: spin while INFLATING); 1183 ReadStableMark(object); 1184 continue; 1185 } 1186 1187 // CASE: stack-locked 1188 // Could be stack-locked either by this thread or by some other thread. 1189 // 1190 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1191 // to install INFLATING into the mark word. We originally installed INFLATING, 1192 // allocated the objectmonitor, and then finally STed the address of the 1193 // objectmonitor into the mark. This was correct, but artificially lengthened 1194 // the interval in which INFLATED appeared in the mark, thus increasing 1195 // the odds of inflation contention. 1196 // 1197 // We now use per-thread private objectmonitor free lists. 1198 // These list are reprovisioned from the global free list outside the 1199 // critical INFLATING...ST interval. A thread can transfer 1200 // multiple objectmonitors en-mass from the global free list to its local free list. 1201 // This reduces coherency traffic and lock contention on the global free list. 1202 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1203 // before or after the CAS(INFLATING) operation. 1204 // See the comments in omAlloc(). 1205 1206 if (mark->has_locker()) { 1207 ObjectMonitor * m = omAlloc(Self); 1208 // Optimistically prepare the objectmonitor - anticipate successful CAS 1209 // We do this before the CAS in order to minimize the length of time 1210 // in which INFLATING appears in the mark. 1211 m->Recycle(); 1212 m->_Responsible = NULL; 1213 m->OwnerIsThread = 0; 1214 m->_recursions = 0; 1215 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1216 1217 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark); 1218 if (cmp != mark) { 1219 omRelease(Self, m, true); 1220 continue; // Interference -- just retry 1221 } 1222 1223 // We've successfully installed INFLATING (0) into the mark-word. 1224 // This is the only case where 0 will appear in a mark-work. 1225 // Only the singular thread that successfully swings the mark-word 1226 // to 0 can perform (or more precisely, complete) inflation. 1227 // 1228 // Why do we CAS a 0 into the mark-word instead of just CASing the 1229 // mark-word from the stack-locked value directly to the new inflated state? 1230 // Consider what happens when a thread unlocks a stack-locked object. 1231 // It attempts to use CAS to swing the displaced header value from the 1232 // on-stack basiclock back into the object header. Recall also that the 1233 // header value (hashcode, etc) can reside in (a) the object header, or 1234 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1235 // header in an objectMonitor. The inflate() routine must copy the header 1236 // value from the basiclock on the owner's stack to the objectMonitor, all 1237 // the while preserving the hashCode stability invariants. If the owner 1238 // decides to release the lock while the value is 0, the unlock will fail 1239 // and control will eventually pass from slow_exit() to inflate. The owner 1240 // will then spin, waiting for the 0 value to disappear. Put another way, 1241 // the 0 causes the owner to stall if the owner happens to try to 1242 // drop the lock (restoring the header from the basiclock to the object) 1243 // while inflation is in-progress. This protocol avoids races that might 1244 // would otherwise permit hashCode values to change or "flicker" for an object. 1245 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1246 // 0 serves as a "BUSY" inflate-in-progress indicator. 1247 1248 1249 // fetch the displaced mark from the owner's stack. 1250 // The owner can't die or unwind past the lock while our INFLATING 1251 // object is in the mark. Furthermore the owner can't complete 1252 // an unlock on the object, either. 1253 markOop dmw = mark->displaced_mark_helper(); 1254 assert(dmw->is_neutral(), "invariant"); 1255 1256 // Setup monitor fields to proper values -- prepare the monitor 1257 m->set_header(dmw); 1258 1259 // Optimization: if the mark->locker stack address is associated 1260 // with this thread we could simply set m->_owner = Self and 1261 // m->OwnerIsThread = 1. Note that a thread can inflate an object 1262 // that it has stack-locked -- as might happen in wait() -- directly 1263 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1264 m->set_owner(mark->locker()); 1265 m->set_object(object); 1266 // TODO-FIXME: assert BasicLock->dhw != 0. 1267 1268 // Must preserve store ordering. The monitor state must 1269 // be stable at the time of publishing the monitor address. 1270 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1271 object->release_set_mark(markOopDesc::encode(m)); 1272 1273 // Hopefully the performance counters are allocated on distinct cache lines 1274 // to avoid false sharing on MP systems ... 1275 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc(); 1276 TEVENT(Inflate: overwrite stacklock); 1277 if (TraceMonitorInflation) { 1278 if (object->is_instance()) { 1279 ResourceMark rm; 1280 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1281 (void *) object, (intptr_t) object->mark(), 1282 object->klass()->external_name()); 1283 } 1284 } 1285 return m; 1286 } 1287 1288 // CASE: neutral 1289 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1290 // If we know we're inflating for entry it's better to inflate by swinging a 1291 // pre-locked objectMonitor pointer into the object header. A successful 1292 // CAS inflates the object *and* confers ownership to the inflating thread. 1293 // In the current implementation we use a 2-step mechanism where we CAS() 1294 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1295 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1296 // would be useful. 1297 1298 assert(mark->is_neutral(), "invariant"); 1299 ObjectMonitor * m = omAlloc(Self); 1300 // prepare m for installation - set monitor to initial state 1301 m->Recycle(); 1302 m->set_header(mark); 1303 m->set_owner(NULL); 1304 m->set_object(object); 1305 m->OwnerIsThread = 1; 1306 m->_recursions = 0; 1307 m->_Responsible = NULL; 1308 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1309 1310 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1311 m->set_object(NULL); 1312 m->set_owner(NULL); 1313 m->OwnerIsThread = 0; 1314 m->Recycle(); 1315 omRelease(Self, m, true); 1316 m = NULL; 1317 continue; 1318 // interference - the markword changed - just retry. 1319 // The state-transitions are one-way, so there's no chance of 1320 // live-lock -- "Inflated" is an absorbing state. 1321 } 1322 1323 // Hopefully the performance counters are allocated on distinct 1324 // cache lines to avoid false sharing on MP systems ... 1325 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc(); 1326 TEVENT(Inflate: overwrite neutral); 1327 if (TraceMonitorInflation) { 1328 if (object->is_instance()) { 1329 ResourceMark rm; 1330 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1331 (void *) object, (intptr_t) object->mark(), 1332 object->klass()->external_name()); 1333 } 1334 } 1335 return m; 1336 } 1337} 1338 1339// Note that we could encounter some performance loss through false-sharing as 1340// multiple locks occupy the same $ line. Padding might be appropriate. 1341 1342 1343// Deflate_idle_monitors() is called at all safepoints, immediately 1344// after all mutators are stopped, but before any objects have moved. 1345// It traverses the list of known monitors, deflating where possible. 1346// The scavenged monitor are returned to the monitor free list. 1347// 1348// Beware that we scavenge at *every* stop-the-world point. 1349// Having a large number of monitors in-circulation negatively 1350// impacts the performance of some applications (e.g., PointBase). 1351// Broadly, we want to minimize the # of monitors in circulation. 1352// 1353// We have added a flag, MonitorInUseLists, which creates a list 1354// of active monitors for each thread. deflate_idle_monitors() 1355// only scans the per-thread inuse lists. omAlloc() puts all 1356// assigned monitors on the per-thread list. deflate_idle_monitors() 1357// returns the non-busy monitors to the global free list. 1358// When a thread dies, omFlush() adds the list of active monitors for 1359// that thread to a global gOmInUseList acquiring the 1360// global list lock. deflate_idle_monitors() acquires the global 1361// list lock to scan for non-busy monitors to the global free list. 1362// An alternative could have used a single global inuse list. The 1363// downside would have been the additional cost of acquiring the global list lock 1364// for every omAlloc(). 1365// 1366// Perversely, the heap size -- and thus the STW safepoint rate -- 1367// typically drives the scavenge rate. Large heaps can mean infrequent GC, 1368// which in turn can mean large(r) numbers of objectmonitors in circulation. 1369// This is an unfortunate aspect of this design. 1370 1371enum ManifestConstants { 1372 ClearResponsibleAtSTW = 0, 1373 MaximumRecheckInterval = 1000 1374}; 1375 1376// Deflate a single monitor if not in use 1377// Return true if deflated, false if in use 1378bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1379 ObjectMonitor** freeHeadp, 1380 ObjectMonitor** freeTailp) { 1381 bool deflated; 1382 // Normal case ... The monitor is associated with obj. 1383 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1384 guarantee(mid == obj->mark()->monitor(), "invariant"); 1385 guarantee(mid->header()->is_neutral(), "invariant"); 1386 1387 if (mid->is_busy()) { 1388 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1389 deflated = false; 1390 } else { 1391 // Deflate the monitor if it is no longer being used 1392 // It's idle - scavenge and return to the global free list 1393 // plain old deflation ... 1394 TEVENT(deflate_idle_monitors - scavenge1); 1395 if (TraceMonitorInflation) { 1396 if (obj->is_instance()) { 1397 ResourceMark rm; 1398 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1399 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name()); 1400 } 1401 } 1402 1403 // Restore the header back to obj 1404 obj->release_set_mark(mid->header()); 1405 mid->clear(); 1406 1407 assert(mid->object() == NULL, "invariant"); 1408 1409 // Move the object to the working free list defined by FreeHead,FreeTail. 1410 if (*freeHeadp == NULL) *freeHeadp = mid; 1411 if (*freeTailp != NULL) { 1412 ObjectMonitor * prevtail = *freeTailp; 1413 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK 1414 prevtail->FreeNext = mid; 1415 } 1416 *freeTailp = mid; 1417 deflated = true; 1418 } 1419 return deflated; 1420} 1421 1422// Caller acquires ListLock 1423int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp, 1424 ObjectMonitor** freeHeadp, 1425 ObjectMonitor** freeTailp) { 1426 ObjectMonitor* mid; 1427 ObjectMonitor* next; 1428 ObjectMonitor* curmidinuse = NULL; 1429 int deflatedcount = 0; 1430 1431 for (mid = *listheadp; mid != NULL;) { 1432 oop obj = (oop) mid->object(); 1433 bool deflated = false; 1434 if (obj != NULL) { 1435 deflated = deflate_monitor(mid, obj, freeHeadp, freeTailp); 1436 } 1437 if (deflated) { 1438 // extract from per-thread in-use-list 1439 if (mid == *listheadp) { 1440 *listheadp = mid->FreeNext; 1441 } else if (curmidinuse != NULL) { 1442 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist 1443 } 1444 next = mid->FreeNext; 1445 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list 1446 mid = next; 1447 deflatedcount++; 1448 } else { 1449 curmidinuse = mid; 1450 mid = mid->FreeNext; 1451 } 1452 } 1453 return deflatedcount; 1454} 1455 1456void ObjectSynchronizer::deflate_idle_monitors() { 1457 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1458 int nInuse = 0; // currently associated with objects 1459 int nInCirculation = 0; // extant 1460 int nScavenged = 0; // reclaimed 1461 bool deflated = false; 1462 1463 ObjectMonitor * FreeHead = NULL; // Local SLL of scavenged monitors 1464 ObjectMonitor * FreeTail = NULL; 1465 1466 TEVENT(deflate_idle_monitors); 1467 // Prevent omFlush from changing mids in Thread dtor's during deflation 1468 // And in case the vm thread is acquiring a lock during a safepoint 1469 // See e.g. 6320749 1470 Thread::muxAcquire(&ListLock, "scavenge - return"); 1471 1472 if (MonitorInUseLists) { 1473 int inUse = 0; 1474 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { 1475 nInCirculation+= cur->omInUseCount; 1476 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail); 1477 cur->omInUseCount-= deflatedcount; 1478 if (ObjectMonitor::Knob_VerifyInUse) { 1479 verifyInUse(cur); 1480 } 1481 nScavenged += deflatedcount; 1482 nInuse += cur->omInUseCount; 1483 } 1484 1485 // For moribund threads, scan gOmInUseList 1486 if (gOmInUseList) { 1487 nInCirculation += gOmInUseCount; 1488 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail); 1489 gOmInUseCount-= deflatedcount; 1490 nScavenged += deflatedcount; 1491 nInuse += gOmInUseCount; 1492 } 1493 1494 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { 1495 // Iterate over all extant monitors - Scavenge all idle monitors. 1496 assert(block->object() == CHAINMARKER, "must be a block header"); 1497 nInCirculation += _BLOCKSIZE; 1498 for (int i = 1; i < _BLOCKSIZE; i++) { 1499 ObjectMonitor* mid = &block[i]; 1500 oop obj = (oop) mid->object(); 1501 1502 if (obj == NULL) { 1503 // The monitor is not associated with an object. 1504 // The monitor should either be a thread-specific private 1505 // free list or the global free list. 1506 // obj == NULL IMPLIES mid->is_busy() == 0 1507 guarantee(!mid->is_busy(), "invariant"); 1508 continue; 1509 } 1510 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail); 1511 1512 if (deflated) { 1513 mid->FreeNext = NULL; 1514 nScavenged++; 1515 } else { 1516 nInuse++; 1517 } 1518 } 1519 } 1520 1521 MonitorFreeCount += nScavenged; 1522 1523 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree. 1524 1525 if (ObjectMonitor::Knob_Verbose) { 1526 ::printf("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n", 1527 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, 1528 MonitorPopulation, MonitorFreeCount); 1529 ::fflush(stdout); 1530 } 1531 1532 ForceMonitorScavenge = 0; // Reset 1533 1534 // Move the scavenged monitors back to the global free list. 1535 if (FreeHead != NULL) { 1536 guarantee(FreeTail != NULL && nScavenged > 0, "invariant"); 1537 assert(FreeTail->FreeNext == NULL, "invariant"); 1538 // constant-time list splice - prepend scavenged segment to gFreeList 1539 FreeTail->FreeNext = gFreeList; 1540 gFreeList = FreeHead; 1541 } 1542 Thread::muxRelease(&ListLock); 1543 1544 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged); 1545 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation); 1546 1547 // TODO: Add objectMonitor leak detection. 1548 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1549 GVars.stwRandom = os::random(); 1550 GVars.stwCycle++; 1551} 1552 1553// Monitor cleanup on JavaThread::exit 1554 1555// Iterate through monitor cache and attempt to release thread's monitors 1556// Gives up on a particular monitor if an exception occurs, but continues 1557// the overall iteration, swallowing the exception. 1558class ReleaseJavaMonitorsClosure: public MonitorClosure { 1559 private: 1560 TRAPS; 1561 1562 public: 1563 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1564 void do_monitor(ObjectMonitor* mid) { 1565 if (mid->owner() == THREAD) { 1566 (void)mid->complete_exit(CHECK); 1567 } 1568 } 1569}; 1570 1571// Release all inflated monitors owned by THREAD. Lightweight monitors are 1572// ignored. This is meant to be called during JNI thread detach which assumes 1573// all remaining monitors are heavyweight. All exceptions are swallowed. 1574// Scanning the extant monitor list can be time consuming. 1575// A simple optimization is to add a per-thread flag that indicates a thread 1576// called jni_monitorenter() during its lifetime. 1577// 1578// Instead of No_Savepoint_Verifier it might be cheaper to 1579// use an idiom of the form: 1580// auto int tmp = SafepointSynchronize::_safepoint_counter ; 1581// <code that must not run at safepoint> 1582// guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1583// Since the tests are extremely cheap we could leave them enabled 1584// for normal product builds. 1585 1586void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1587 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1588 No_Safepoint_Verifier nsv; 1589 ReleaseJavaMonitorsClosure rjmc(THREAD); 1590 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); 1591 ObjectSynchronizer::monitors_iterate(&rjmc); 1592 Thread::muxRelease(&ListLock); 1593 THREAD->clear_pending_exception(); 1594} 1595 1596//------------------------------------------------------------------------------ 1597// Debugging code 1598 1599void ObjectSynchronizer::sanity_checks(const bool verbose, 1600 const uint cache_line_size, 1601 int *error_cnt_ptr, 1602 int *warning_cnt_ptr) { 1603 u_char *addr_begin = (u_char*)&GVars; 1604 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1605 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1606 1607 if (verbose) { 1608 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1609 sizeof(SharedGlobals)); 1610 } 1611 1612 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1613 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1614 1615 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1616 if (verbose) { 1617 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1618 } 1619 1620 if (cache_line_size != 0) { 1621 // We were able to determine the L1 data cache line size so 1622 // do some cache line specific sanity checks 1623 1624 if (offset_stwRandom < cache_line_size) { 1625 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1626 "to the struct beginning than a cache line which permits " 1627 "false sharing."); 1628 (*warning_cnt_ptr)++; 1629 } 1630 1631 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1632 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1633 "SharedGlobals.hcSequence fields are closer than a cache " 1634 "line which permits false sharing."); 1635 (*warning_cnt_ptr)++; 1636 } 1637 1638 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1639 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1640 "to the struct end than a cache line which permits false " 1641 "sharing."); 1642 (*warning_cnt_ptr)++; 1643 } 1644 } 1645} 1646 1647#ifndef PRODUCT 1648 1649// Verify all monitors in the monitor cache, the verification is weak. 1650void ObjectSynchronizer::verify() { 1651 ObjectMonitor* block = gBlockList; 1652 ObjectMonitor* mid; 1653 while (block) { 1654 assert(block->object() == CHAINMARKER, "must be a block header"); 1655 for (int i = 1; i < _BLOCKSIZE; i++) { 1656 mid = block + i; 1657 oop object = (oop) mid->object(); 1658 if (object != NULL) { 1659 mid->verify(); 1660 } 1661 } 1662 block = (ObjectMonitor*) block->FreeNext; 1663 } 1664} 1665 1666// Check if monitor belongs to the monitor cache 1667// The list is grow-only so it's *relatively* safe to traverse 1668// the list of extant blocks without taking a lock. 1669 1670int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1671 ObjectMonitor* block = gBlockList; 1672 1673 while (block) { 1674 assert(block->object() == CHAINMARKER, "must be a block header"); 1675 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 1676 address mon = (address) monitor; 1677 address blk = (address) block; 1678 size_t diff = mon - blk; 1679 assert((diff % sizeof(ObjectMonitor)) == 0, "check"); 1680 return 1; 1681 } 1682 block = (ObjectMonitor*) block->FreeNext; 1683 } 1684 return 0; 1685} 1686 1687#endif 1688