synchronizer.cpp revision 7050:03835eaaab2d
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 63#ifdef DTRACE_ENABLED 64 65// Only bother with this argument setup if dtrace is available 66// TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 67 68#define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 69 char* bytes = NULL; \ 70 int len = 0; \ 71 jlong jtid = SharedRuntime::get_java_tid(thread); \ 72 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 73 if (klassname != NULL) { \ 74 bytes = (char*)klassname->bytes(); \ 75 len = klassname->utf8_length(); \ 76 } 77 78#define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 79 { \ 80 if (DTraceMonitorProbes) { \ 81 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 82 HOTSPOT_MONITOR_WAIT(jtid, \ 83 (uintptr_t)(monitor), bytes, len, (millis)); \ 84 } \ 85 } 86 87#define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 88 89#define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 90 { \ 91 if (DTraceMonitorProbes) { \ 92 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 93 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 94 (uintptr_t)(monitor), bytes, len); \ 95 } \ 96 } 97 98#else // ndef DTRACE_ENABLED 99 100#define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 101#define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 102 103#endif // ndef DTRACE_ENABLED 104 105// This exists only as a workaround of dtrace bug 6254741 106int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 107 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 108 return 0; 109} 110 111#define NINFLATIONLOCKS 256 112static volatile intptr_t InflationLocks[NINFLATIONLOCKS]; 113 114ObjectMonitor * ObjectSynchronizer::gBlockList = NULL; 115ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL; 116ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL; 117int ObjectSynchronizer::gOmInUseCount = 0; 118static volatile intptr_t ListLock = 0; // protects global monitor free-list cache 119static volatile int MonitorFreeCount = 0; // # on gFreeList 120static volatile int MonitorPopulation = 0; // # Extant -- in circulation 121#define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 122 123// ----------------------------------------------------------------------------- 124// Fast Monitor Enter/Exit 125// This the fast monitor enter. The interpreter and compiler use 126// some assembly copies of this code. Make sure update those code 127// if the following function is changed. The implementation is 128// extremely sensitive to race condition. Be careful. 129 130void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, 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 202 if (mark->has_locker() && 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) { // possible entry from jni enter 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// 514 515static inline intptr_t get_next_hash(Thread * Self, oop obj) { 516 intptr_t value = 0; 517 if (hashCode == 0) { 518 // This form uses an unguarded global Park-Miller RNG, 519 // so it's possible for two threads to race and generate the same RNG. 520 // On MP system we'll have lots of RW access to a global, so the 521 // mechanism induces lots of coherency traffic. 522 value = os::random(); 523 } else 524 if (hashCode == 1) { 525 // This variation has the property of being stable (idempotent) 526 // between STW operations. This can be useful in some of the 1-0 527 // synchronization schemes. 528 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3; 529 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom; 530 } else 531 if (hashCode == 2) { 532 value = 1; // for sensitivity testing 533 } else 534 if (hashCode == 3) { 535 value = ++GVars.hcSequence; 536 } else 537 if (hashCode == 4) { 538 value = cast_from_oop<intptr_t>(obj); 539 } else { 540 // Marsaglia's xor-shift scheme with thread-specific state 541 // This is probably the best overall implementation -- we'll 542 // likely make this the default in future releases. 543 unsigned t = Self->_hashStateX; 544 t ^= (t << 11); 545 Self->_hashStateX = Self->_hashStateY; 546 Self->_hashStateY = Self->_hashStateZ; 547 Self->_hashStateZ = Self->_hashStateW; 548 unsigned v = Self->_hashStateW; 549 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 550 Self->_hashStateW = v; 551 value = v; 552 } 553 554 value &= markOopDesc::hash_mask; 555 if (value == 0) value = 0xBAD; 556 assert(value != markOopDesc::no_hash, "invariant"); 557 TEVENT(hashCode: GENERATE); 558 return value; 559} 560// 561intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { 562 if (UseBiasedLocking) { 563 // NOTE: many places throughout the JVM do not expect a safepoint 564 // to be taken here, in particular most operations on perm gen 565 // objects. However, we only ever bias Java instances and all of 566 // the call sites of identity_hash that might revoke biases have 567 // been checked to make sure they can handle a safepoint. The 568 // added check of the bias pattern is to avoid useless calls to 569 // thread-local storage. 570 if (obj->mark()->has_bias_pattern()) { 571 // Handle for oop obj in case of STW safepoint 572 Handle hobj(Self, obj); 573 // Relaxing assertion for bug 6320749. 574 assert(Universe::verify_in_progress() || 575 !SafepointSynchronize::is_at_safepoint(), 576 "biases should not be seen by VM thread here"); 577 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); 578 obj = hobj(); 579 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 580 } 581 } 582 583 // hashCode() is a heap mutator ... 584 // Relaxing assertion for bug 6320749. 585 assert(Universe::verify_in_progress() || 586 !SafepointSynchronize::is_at_safepoint(), "invariant"); 587 assert(Universe::verify_in_progress() || 588 Self->is_Java_thread() , "invariant"); 589 assert(Universe::verify_in_progress() || 590 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 591 592 ObjectMonitor* monitor = NULL; 593 markOop temp, test; 594 intptr_t hash; 595 markOop mark = ReadStableMark (obj); 596 597 // object should remain ineligible for biased locking 598 assert(!mark->has_bias_pattern(), "invariant"); 599 600 if (mark->is_neutral()) { 601 hash = mark->hash(); // this is a normal header 602 if (hash) { // if it has hash, just return it 603 return hash; 604 } 605 hash = get_next_hash(Self, obj); // allocate a new hash code 606 temp = mark->copy_set_hash(hash); // merge the hash code into header 607 // use (machine word version) atomic operation to install the hash 608 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); 609 if (test == mark) { 610 return hash; 611 } 612 // If atomic operation failed, we must inflate the header 613 // into heavy weight monitor. We could add more code here 614 // for fast path, but it does not worth the complexity. 615 } else if (mark->has_monitor()) { 616 monitor = mark->monitor(); 617 temp = monitor->header(); 618 assert(temp->is_neutral(), "invariant"); 619 hash = temp->hash(); 620 if (hash) { 621 return hash; 622 } 623 // Skip to the following code to reduce code size 624 } else if (Self->is_lock_owned((address)mark->locker())) { 625 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned 626 assert(temp->is_neutral(), "invariant"); 627 hash = temp->hash(); // by current thread, check if the displaced 628 if (hash) { // header contains hash code 629 return hash; 630 } 631 // WARNING: 632 // The displaced header is strictly immutable. 633 // It can NOT be changed in ANY cases. So we have 634 // to inflate the header into heavyweight monitor 635 // even the current thread owns the lock. The reason 636 // is the BasicLock (stack slot) will be asynchronously 637 // read by other threads during the inflate() function. 638 // Any change to stack may not propagate to other threads 639 // correctly. 640 } 641 642 // Inflate the monitor to set hash code 643 monitor = ObjectSynchronizer::inflate(Self, obj); 644 // Load displaced header and check it has hash code 645 mark = monitor->header(); 646 assert(mark->is_neutral(), "invariant"); 647 hash = mark->hash(); 648 if (hash == 0) { 649 hash = get_next_hash(Self, obj); 650 temp = mark->copy_set_hash(hash); // merge hash code into header 651 assert(temp->is_neutral(), "invariant"); 652 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); 653 if (test != mark) { 654 // The only update to the header in the monitor (outside GC) 655 // is install the hash code. If someone add new usage of 656 // displaced header, please update this code 657 hash = test->hash(); 658 assert(test->is_neutral(), "invariant"); 659 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 660 } 661 } 662 // We finally get the hash 663 return hash; 664} 665 666// Deprecated -- use FastHashCode() instead. 667 668intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 669 return FastHashCode(Thread::current(), obj()); 670} 671 672 673bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 674 Handle h_obj) { 675 if (UseBiasedLocking) { 676 BiasedLocking::revoke_and_rebias(h_obj, false, thread); 677 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 678 } 679 680 assert(thread == JavaThread::current(), "Can only be called on current thread"); 681 oop obj = h_obj(); 682 683 markOop mark = ReadStableMark(obj); 684 685 // Uncontended case, header points to stack 686 if (mark->has_locker()) { 687 return thread->is_lock_owned((address)mark->locker()); 688 } 689 // Contended case, header points to ObjectMonitor (tagged pointer) 690 if (mark->has_monitor()) { 691 ObjectMonitor* monitor = mark->monitor(); 692 return monitor->is_entered(thread) != 0; 693 } 694 // Unlocked case, header in place 695 assert(mark->is_neutral(), "sanity check"); 696 return false; 697} 698 699// Be aware of this method could revoke bias of the lock object. 700// This method queries the ownership of the lock handle specified by 'h_obj'. 701// If the current thread owns the lock, it returns owner_self. If no 702// thread owns the lock, it returns owner_none. Otherwise, it will return 703// owner_other. 704ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 705(JavaThread *self, Handle h_obj) { 706 // The caller must beware this method can revoke bias, and 707 // revocation can result in a safepoint. 708 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 709 assert(self->thread_state() != _thread_blocked , "invariant"); 710 711 // Possible mark states: neutral, biased, stack-locked, inflated 712 713 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { 714 // CASE: biased 715 BiasedLocking::revoke_and_rebias(h_obj, false, self); 716 assert(!h_obj->mark()->has_bias_pattern(), 717 "biases should be revoked by now"); 718 } 719 720 assert(self == JavaThread::current(), "Can only be called on current thread"); 721 oop obj = h_obj(); 722 markOop mark = ReadStableMark(obj); 723 724 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 725 if (mark->has_locker()) { 726 return self->is_lock_owned((address)mark->locker()) ? 727 owner_self : owner_other; 728 } 729 730 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. 731 // The Object:ObjectMonitor relationship is stable as long as we're 732 // not at a safepoint. 733 if (mark->has_monitor()) { 734 void * owner = mark->monitor()->_owner; 735 if (owner == NULL) return owner_none; 736 return (owner == self || 737 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 738 } 739 740 // CASE: neutral 741 assert(mark->is_neutral(), "sanity check"); 742 return owner_none; // it's unlocked 743} 744 745// FIXME: jvmti should call this 746JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { 747 if (UseBiasedLocking) { 748 if (SafepointSynchronize::is_at_safepoint()) { 749 BiasedLocking::revoke_at_safepoint(h_obj); 750 } else { 751 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); 752 } 753 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); 754 } 755 756 oop obj = h_obj(); 757 address owner = NULL; 758 759 markOop mark = ReadStableMark(obj); 760 761 // Uncontended case, header points to stack 762 if (mark->has_locker()) { 763 owner = (address) mark->locker(); 764 } 765 766 // Contended case, header points to ObjectMonitor (tagged pointer) 767 if (mark->has_monitor()) { 768 ObjectMonitor* monitor = mark->monitor(); 769 assert(monitor != NULL, "monitor should be non-null"); 770 owner = (address) monitor->owner(); 771 } 772 773 if (owner != NULL) { 774 // owning_thread_from_monitor_owner() may also return NULL here 775 return Threads::owning_thread_from_monitor_owner(owner, doLock); 776 } 777 778 // Unlocked case, header in place 779 // Cannot have assertion since this object may have been 780 // locked by another thread when reaching here. 781 // assert(mark->is_neutral(), "sanity check"); 782 783 return NULL; 784} 785// Visitors ... 786 787void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 788 ObjectMonitor* block = gBlockList; 789 ObjectMonitor* mid; 790 while (block) { 791 assert(block->object() == CHAINMARKER, "must be a block header"); 792 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 793 mid = block + i; 794 oop object = (oop) mid->object(); 795 if (object != NULL) { 796 closure->do_monitor(mid); 797 } 798 } 799 block = (ObjectMonitor*) block->FreeNext; 800 } 801} 802 803// Get the next block in the block list. 804static inline ObjectMonitor* next(ObjectMonitor* block) { 805 assert(block->object() == CHAINMARKER, "must be a block header"); 806 block = block->FreeNext; 807 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); 808 return block; 809} 810 811 812void ObjectSynchronizer::oops_do(OopClosure* f) { 813 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 814 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { 815 assert(block->object() == CHAINMARKER, "must be a block header"); 816 for (int i = 1; i < _BLOCKSIZE; i++) { 817 ObjectMonitor* mid = &block[i]; 818 if (mid->object() != NULL) { 819 f->do_oop((oop*)mid->object_addr()); 820 } 821 } 822 } 823} 824 825 826// ----------------------------------------------------------------------------- 827// ObjectMonitor Lifecycle 828// ----------------------- 829// Inflation unlinks monitors from the global gFreeList and 830// associates them with objects. Deflation -- which occurs at 831// STW-time -- disassociates idle monitors from objects. Such 832// scavenged monitors are returned to the gFreeList. 833// 834// The global list is protected by ListLock. All the critical sections 835// are short and operate in constant-time. 836// 837// ObjectMonitors reside in type-stable memory (TSM) and are immortal. 838// 839// Lifecycle: 840// -- unassigned and on the global free list 841// -- unassigned and on a thread's private omFreeList 842// -- assigned to an object. The object is inflated and the mark refers 843// to the objectmonitor. 844// 845 846 847// Constraining monitor pool growth via MonitorBound ... 848// 849// The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 850// the rate of scavenging is driven primarily by GC. As such, we can find 851// an inordinate number of monitors in circulation. 852// To avoid that scenario we can artificially induce a STW safepoint 853// if the pool appears to be growing past some reasonable bound. 854// Generally we favor time in space-time tradeoffs, but as there's no 855// natural back-pressure on the # of extant monitors we need to impose some 856// type of limit. Beware that if MonitorBound is set to too low a value 857// we could just loop. In addition, if MonitorBound is set to a low value 858// we'll incur more safepoints, which are harmful to performance. 859// See also: GuaranteedSafepointInterval 860// 861// The current implementation uses asynchronous VM operations. 862// 863 864static void InduceScavenge (Thread * Self, const char * Whence) { 865 // Induce STW safepoint to trim monitors 866 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 867 // More precisely, trigger an asynchronous STW safepoint as the number 868 // of active monitors passes the specified threshold. 869 // TODO: assert thread state is reasonable 870 871 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 872 if (ObjectMonitor::Knob_Verbose) { 873 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; 874 ::fflush(stdout); 875 } 876 // Induce a 'null' safepoint to scavenge monitors 877 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 878 // to the VMthread and have a lifespan longer than that of this activation record. 879 // The VMThread will delete the op when completed. 880 VMThread::execute(new VM_ForceAsyncSafepoint()); 881 882 if (ObjectMonitor::Knob_Verbose) { 883 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; 884 ::fflush(stdout); 885 } 886 } 887} 888 889void ObjectSynchronizer::verifyInUse (Thread *Self) { 890 ObjectMonitor* mid; 891 int inusetally = 0; 892 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { 893 inusetally++; 894 } 895 assert(inusetally == Self->omInUseCount, "inuse count off"); 896 897 int freetally = 0; 898 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { 899 freetally++; 900 } 901 assert(freetally == Self->omFreeCount, "free count off"); 902} 903 904ObjectMonitor * NOINLINE ObjectSynchronizer::omAlloc (Thread * Self) { 905 // A large MAXPRIVATE value reduces both list lock contention 906 // and list coherency traffic, but also tends to increase the 907 // number of objectMonitors in circulation as well as the STW 908 // scavenge costs. As usual, we lean toward time in space-time 909 // tradeoffs. 910 const int MAXPRIVATE = 1024; 911 for (;;) { 912 ObjectMonitor * m; 913 914 // 1: try to allocate from the thread's local omFreeList. 915 // Threads will attempt to allocate first from their local list, then 916 // from the global list, and only after those attempts fail will the thread 917 // attempt to instantiate new monitors. Thread-local free lists take 918 // heat off the ListLock and improve allocation latency, as well as reducing 919 // coherency traffic on the shared global list. 920 m = Self->omFreeList; 921 if (m != NULL) { 922 Self->omFreeList = m->FreeNext; 923 Self->omFreeCount--; 924 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene 925 guarantee(m->object() == NULL, "invariant"); 926 if (MonitorInUseLists) { 927 m->FreeNext = Self->omInUseList; 928 Self->omInUseList = m; 929 Self->omInUseCount++; 930 if (ObjectMonitor::Knob_VerifyInUse) { 931 verifyInUse(Self); 932 } 933 } else { 934 m->FreeNext = NULL; 935 } 936 return m; 937 } 938 939 // 2: try to allocate from the global gFreeList 940 // CONSIDER: use muxTry() instead of muxAcquire(). 941 // If the muxTry() fails then drop immediately into case 3. 942 // If we're using thread-local free lists then try 943 // to reprovision the caller's free list. 944 if (gFreeList != NULL) { 945 // Reprovision the thread's omFreeList. 946 // Use bulk transfers to reduce the allocation rate and heat 947 // on various locks. 948 Thread::muxAcquire(&ListLock, "omAlloc"); 949 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) { 950 MonitorFreeCount--; 951 ObjectMonitor * take = gFreeList; 952 gFreeList = take->FreeNext; 953 guarantee(take->object() == NULL, "invariant"); 954 guarantee(!take->is_busy(), "invariant"); 955 take->Recycle(); 956 omRelease(Self, take, false); 957 } 958 Thread::muxRelease(&ListLock); 959 Self->omFreeProvision += 1 + (Self->omFreeProvision/2); 960 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE; 961 TEVENT(omFirst - reprovision); 962 963 const int mx = MonitorBound; 964 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) { 965 // We can't safely induce a STW safepoint from omAlloc() as our thread 966 // state may not be appropriate for such activities and callers may hold 967 // naked oops, so instead we defer the action. 968 InduceScavenge(Self, "omAlloc"); 969 } 970 continue; 971 } 972 973 // 3: allocate a block of new ObjectMonitors 974 // Both the local and global free lists are empty -- resort to malloc(). 975 // In the current implementation objectMonitors are TSM - immortal. 976 assert(_BLOCKSIZE > 1, "invariant"); 977 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; 978 979 // NOTE: (almost) no way to recover if allocation failed. 980 // We might be able to induce a STW safepoint and scavenge enough 981 // objectMonitors to permit progress. 982 if (temp == NULL) { 983 vm_exit_out_of_memory(sizeof (ObjectMonitor[_BLOCKSIZE]), OOM_MALLOC_ERROR, 984 "Allocate ObjectMonitors"); 985 } 986 987 // Format the block. 988 // initialize the linked list, each monitor points to its next 989 // forming the single linked free list, the very first monitor 990 // will points to next block, which forms the block list. 991 // The trick of using the 1st element in the block as gBlockList 992 // linkage should be reconsidered. A better implementation would 993 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 994 995 for (int i = 1; i < _BLOCKSIZE; i++) { 996 temp[i].FreeNext = &temp[i+1]; 997 } 998 999 // terminate the last monitor as the end of list 1000 temp[_BLOCKSIZE - 1].FreeNext = NULL; 1001 1002 // Element [0] is reserved for global list linkage 1003 temp[0].set_object(CHAINMARKER); 1004 1005 // Consider carving out this thread's current request from the 1006 // block in hand. This avoids some lock traffic and redundant 1007 // list activity. 1008 1009 // Acquire the ListLock to manipulate BlockList and FreeList. 1010 // An Oyama-Taura-Yonezawa scheme might be more efficient. 1011 Thread::muxAcquire(&ListLock, "omAlloc [2]"); 1012 MonitorPopulation += _BLOCKSIZE-1; 1013 MonitorFreeCount += _BLOCKSIZE-1; 1014 1015 // Add the new block to the list of extant blocks (gBlockList). 1016 // The very first objectMonitor in a block is reserved and dedicated. 1017 // It serves as blocklist "next" linkage. 1018 temp[0].FreeNext = gBlockList; 1019 gBlockList = temp; 1020 1021 // Add the new string of objectMonitors to the global free list 1022 temp[_BLOCKSIZE - 1].FreeNext = gFreeList; 1023 gFreeList = temp + 1; 1024 Thread::muxRelease(&ListLock); 1025 TEVENT(Allocate block of monitors); 1026 } 1027} 1028 1029// Place "m" on the caller's private per-thread omFreeList. 1030// In practice there's no need to clamp or limit the number of 1031// monitors on a thread's omFreeList as the only time we'll call 1032// omRelease is to return a monitor to the free list after a CAS 1033// attempt failed. This doesn't allow unbounded #s of monitors to 1034// accumulate on a thread's free list. 1035// 1036 1037void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) { 1038 guarantee(m->object() == NULL, "invariant"); 1039 1040 // Remove from omInUseList 1041 if (MonitorInUseLists && fromPerThreadAlloc) { 1042 ObjectMonitor* curmidinuse = NULL; 1043 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL;) { 1044 if (m == mid) { 1045 // extract from per-thread in-use-list 1046 if (mid == Self->omInUseList) { 1047 Self->omInUseList = mid->FreeNext; 1048 } else if (curmidinuse != NULL) { 1049 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist 1050 } 1051 Self->omInUseCount--; 1052 if (ObjectMonitor::Knob_VerifyInUse) { 1053 verifyInUse(Self); 1054 } 1055 break; 1056 } else { 1057 curmidinuse = mid; 1058 mid = mid->FreeNext; 1059 } 1060 } 1061 } 1062 1063 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new 1064 m->FreeNext = Self->omFreeList; 1065 Self->omFreeList = m; 1066 Self->omFreeCount++; 1067} 1068 1069// Return the monitors of a moribund thread's local free list to 1070// the global free list. Typically a thread calls omFlush() when 1071// it's dying. We could also consider having the VM thread steal 1072// monitors from threads that have not run java code over a few 1073// consecutive STW safepoints. Relatedly, we might decay 1074// omFreeProvision at STW safepoints. 1075// 1076// Also return the monitors of a moribund thread's omInUseList to 1077// a global gOmInUseList under the global list lock so these 1078// will continue to be scanned. 1079// 1080// We currently call omFlush() from the Thread:: dtor _after the thread 1081// has been excised from the thread list and is no longer a mutator. 1082// That means that omFlush() can run concurrently with a safepoint and 1083// the scavenge operator. Calling omFlush() from JavaThread::exit() might 1084// be a better choice as we could safely reason that that the JVM is 1085// not at a safepoint at the time of the call, and thus there could 1086// be not inopportune interleavings between omFlush() and the scavenge 1087// operator. 1088 1089void ObjectSynchronizer::omFlush (Thread * Self) { 1090 ObjectMonitor * List = Self->omFreeList; // Null-terminated SLL 1091 Self->omFreeList = NULL; 1092 ObjectMonitor * Tail = NULL; 1093 int Tally = 0; 1094 if (List != NULL) { 1095 ObjectMonitor * s; 1096 for (s = List; s != NULL; s = s->FreeNext) { 1097 Tally++; 1098 Tail = s; 1099 guarantee(s->object() == NULL, "invariant"); 1100 guarantee(!s->is_busy(), "invariant"); 1101 s->set_owner(NULL); // redundant but good hygiene 1102 TEVENT(omFlush - Move one); 1103 } 1104 guarantee(Tail != NULL && List != NULL, "invariant"); 1105 } 1106 1107 ObjectMonitor * InUseList = Self->omInUseList; 1108 ObjectMonitor * InUseTail = NULL; 1109 int InUseTally = 0; 1110 if (InUseList != NULL) { 1111 Self->omInUseList = NULL; 1112 ObjectMonitor *curom; 1113 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) { 1114 InUseTail = curom; 1115 InUseTally++; 1116 } 1117 assert(Self->omInUseCount == InUseTally, "inuse count off"); 1118 Self->omInUseCount = 0; 1119 guarantee(InUseTail != NULL && InUseList != NULL, "invariant"); 1120 } 1121 1122 Thread::muxAcquire(&ListLock, "omFlush"); 1123 if (Tail != NULL) { 1124 Tail->FreeNext = gFreeList; 1125 gFreeList = List; 1126 MonitorFreeCount += Tally; 1127 } 1128 1129 if (InUseTail != NULL) { 1130 InUseTail->FreeNext = gOmInUseList; 1131 gOmInUseList = InUseList; 1132 gOmInUseCount += InUseTally; 1133 } 1134 1135 Thread::muxRelease(&ListLock); 1136 TEVENT(omFlush); 1137} 1138 1139// Fast path code shared by multiple functions 1140ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { 1141 markOop mark = obj->mark(); 1142 if (mark->has_monitor()) { 1143 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); 1144 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); 1145 return mark->monitor(); 1146 } 1147 return ObjectSynchronizer::inflate(Thread::current(), obj); 1148} 1149 1150 1151// Note that we could encounter some performance loss through false-sharing as 1152// multiple locks occupy the same $ line. Padding might be appropriate. 1153 1154 1155ObjectMonitor * NOINLINE ObjectSynchronizer::inflate (Thread * Self, oop object) { 1156 // Inflate mutates the heap ... 1157 // Relaxing assertion for bug 6320749. 1158 assert(Universe::verify_in_progress() || 1159 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1160 1161 for (;;) { 1162 const markOop mark = object->mark(); 1163 assert(!mark->has_bias_pattern(), "invariant"); 1164 1165 // The mark can be in one of the following states: 1166 // * Inflated - just return 1167 // * Stack-locked - coerce it to inflated 1168 // * INFLATING - busy wait for conversion to complete 1169 // * Neutral - aggressively inflate the object. 1170 // * BIASED - Illegal. We should never see this 1171 1172 // CASE: inflated 1173 if (mark->has_monitor()) { 1174 ObjectMonitor * inf = mark->monitor(); 1175 assert(inf->header()->is_neutral(), "invariant"); 1176 assert(inf->object() == object, "invariant"); 1177 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1178 return inf; 1179 } 1180 1181 // CASE: inflation in progress - inflating over a stack-lock. 1182 // Some other thread is converting from stack-locked to inflated. 1183 // Only that thread can complete inflation -- other threads must wait. 1184 // The INFLATING value is transient. 1185 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1186 // We could always eliminate polling by parking the thread on some auxiliary list. 1187 if (mark == markOopDesc::INFLATING()) { 1188 TEVENT(Inflate: spin while INFLATING); 1189 ReadStableMark(object); 1190 continue; 1191 } 1192 1193 // CASE: stack-locked 1194 // Could be stack-locked either by this thread or by some other thread. 1195 // 1196 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1197 // to install INFLATING into the mark word. We originally installed INFLATING, 1198 // allocated the objectmonitor, and then finally STed the address of the 1199 // objectmonitor into the mark. This was correct, but artificially lengthened 1200 // the interval in which INFLATED appeared in the mark, thus increasing 1201 // the odds of inflation contention. 1202 // 1203 // We now use per-thread private objectmonitor free lists. 1204 // These list are reprovisioned from the global free list outside the 1205 // critical INFLATING...ST interval. A thread can transfer 1206 // multiple objectmonitors en-mass from the global free list to its local free list. 1207 // This reduces coherency traffic and lock contention on the global free list. 1208 // Using such local free lists, it doesn't matter if the omAlloc() call appears 1209 // before or after the CAS(INFLATING) operation. 1210 // See the comments in omAlloc(). 1211 1212 if (mark->has_locker()) { 1213 ObjectMonitor * m = omAlloc(Self); 1214 // Optimistically prepare the objectmonitor - anticipate successful CAS 1215 // We do this before the CAS in order to minimize the length of time 1216 // in which INFLATING appears in the mark. 1217 m->Recycle(); 1218 m->_Responsible = NULL; 1219 m->OwnerIsThread = 0; 1220 m->_recursions = 0; 1221 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1222 1223 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark); 1224 if (cmp != mark) { 1225 omRelease(Self, m, true); 1226 continue; // Interference -- just retry 1227 } 1228 1229 // We've successfully installed INFLATING (0) into the mark-word. 1230 // This is the only case where 0 will appear in a mark-work. 1231 // Only the singular thread that successfully swings the mark-word 1232 // to 0 can perform (or more precisely, complete) inflation. 1233 // 1234 // Why do we CAS a 0 into the mark-word instead of just CASing the 1235 // mark-word from the stack-locked value directly to the new inflated state? 1236 // Consider what happens when a thread unlocks a stack-locked object. 1237 // It attempts to use CAS to swing the displaced header value from the 1238 // on-stack basiclock back into the object header. Recall also that the 1239 // header value (hashcode, etc) can reside in (a) the object header, or 1240 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1241 // header in an objectMonitor. The inflate() routine must copy the header 1242 // value from the basiclock on the owner's stack to the objectMonitor, all 1243 // the while preserving the hashCode stability invariants. If the owner 1244 // decides to release the lock while the value is 0, the unlock will fail 1245 // and control will eventually pass from slow_exit() to inflate. The owner 1246 // will then spin, waiting for the 0 value to disappear. Put another way, 1247 // the 0 causes the owner to stall if the owner happens to try to 1248 // drop the lock (restoring the header from the basiclock to the object) 1249 // while inflation is in-progress. This protocol avoids races that might 1250 // would otherwise permit hashCode values to change or "flicker" for an object. 1251 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. 1252 // 0 serves as a "BUSY" inflate-in-progress indicator. 1253 1254 1255 // fetch the displaced mark from the owner's stack. 1256 // The owner can't die or unwind past the lock while our INFLATING 1257 // object is in the mark. Furthermore the owner can't complete 1258 // an unlock on the object, either. 1259 markOop dmw = mark->displaced_mark_helper(); 1260 assert(dmw->is_neutral(), "invariant"); 1261 1262 // Setup monitor fields to proper values -- prepare the monitor 1263 m->set_header(dmw); 1264 1265 // Optimization: if the mark->locker stack address is associated 1266 // with this thread we could simply set m->_owner = Self and 1267 // m->OwnerIsThread = 1. Note that a thread can inflate an object 1268 // that it has stack-locked -- as might happen in wait() -- directly 1269 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1270 m->set_owner(mark->locker()); 1271 m->set_object(object); 1272 // TODO-FIXME: assert BasicLock->dhw != 0. 1273 1274 // Must preserve store ordering. The monitor state must 1275 // be stable at the time of publishing the monitor address. 1276 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant"); 1277 object->release_set_mark(markOopDesc::encode(m)); 1278 1279 // Hopefully the performance counters are allocated on distinct cache lines 1280 // to avoid false sharing on MP systems ... 1281 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc(); 1282 TEVENT(Inflate: overwrite stacklock); 1283 if (TraceMonitorInflation) { 1284 if (object->is_instance()) { 1285 ResourceMark rm; 1286 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1287 (void *) object, (intptr_t) object->mark(), 1288 object->klass()->external_name()); 1289 } 1290 } 1291 return m; 1292 } 1293 1294 // CASE: neutral 1295 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1296 // If we know we're inflating for entry it's better to inflate by swinging a 1297 // pre-locked objectMonitor pointer into the object header. A successful 1298 // CAS inflates the object *and* confers ownership to the inflating thread. 1299 // In the current implementation we use a 2-step mechanism where we CAS() 1300 // to inflate and then CAS() again to try to swing _owner from NULL to Self. 1301 // An inflateTry() method that we could call from fast_enter() and slow_enter() 1302 // would be useful. 1303 1304 assert(mark->is_neutral(), "invariant"); 1305 ObjectMonitor * m = omAlloc(Self); 1306 // prepare m for installation - set monitor to initial state 1307 m->Recycle(); 1308 m->set_header(mark); 1309 m->set_owner(NULL); 1310 m->set_object(object); 1311 m->OwnerIsThread = 1; 1312 m->_recursions = 0; 1313 m->_Responsible = NULL; 1314 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1315 1316 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { 1317 m->set_object(NULL); 1318 m->set_owner(NULL); 1319 m->OwnerIsThread = 0; 1320 m->Recycle(); 1321 omRelease(Self, m, true); 1322 m = NULL; 1323 continue; 1324 // interference - the markword changed - just retry. 1325 // The state-transitions are one-way, so there's no chance of 1326 // live-lock -- "Inflated" is an absorbing state. 1327 } 1328 1329 // Hopefully the performance counters are allocated on distinct 1330 // cache lines to avoid false sharing on MP systems ... 1331 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc(); 1332 TEVENT(Inflate: overwrite neutral); 1333 if (TraceMonitorInflation) { 1334 if (object->is_instance()) { 1335 ResourceMark rm; 1336 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1337 (void *) object, (intptr_t) object->mark(), 1338 object->klass()->external_name()); 1339 } 1340 } 1341 return m; 1342 } 1343} 1344 1345// Note that we could encounter some performance loss through false-sharing as 1346// multiple locks occupy the same $ line. Padding might be appropriate. 1347 1348 1349// Deflate_idle_monitors() is called at all safepoints, immediately 1350// after all mutators are stopped, but before any objects have moved. 1351// It traverses the list of known monitors, deflating where possible. 1352// The scavenged monitor are returned to the monitor free list. 1353// 1354// Beware that we scavenge at *every* stop-the-world point. 1355// Having a large number of monitors in-circulation negatively 1356// impacts the performance of some applications (e.g., PointBase). 1357// Broadly, we want to minimize the # of monitors in circulation. 1358// 1359// We have added a flag, MonitorInUseLists, which creates a list 1360// of active monitors for each thread. deflate_idle_monitors() 1361// only scans the per-thread inuse lists. omAlloc() puts all 1362// assigned monitors on the per-thread list. deflate_idle_monitors() 1363// returns the non-busy monitors to the global free list. 1364// When a thread dies, omFlush() adds the list of active monitors for 1365// that thread to a global gOmInUseList acquiring the 1366// global list lock. deflate_idle_monitors() acquires the global 1367// list lock to scan for non-busy monitors to the global free list. 1368// An alternative could have used a single global inuse list. The 1369// downside would have been the additional cost of acquiring the global list lock 1370// for every omAlloc(). 1371// 1372// Perversely, the heap size -- and thus the STW safepoint rate -- 1373// typically drives the scavenge rate. Large heaps can mean infrequent GC, 1374// which in turn can mean large(r) numbers of objectmonitors in circulation. 1375// This is an unfortunate aspect of this design. 1376// 1377 1378enum ManifestConstants { 1379 ClearResponsibleAtSTW = 0, 1380 MaximumRecheckInterval = 1000 1381}; 1382 1383// Deflate a single monitor if not in use 1384// Return true if deflated, false if in use 1385bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1386 ObjectMonitor** freeHeadp, ObjectMonitor** freeTailp) { 1387 bool deflated; 1388 // Normal case ... The monitor is associated with obj. 1389 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant"); 1390 guarantee(mid == obj->mark()->monitor(), "invariant"); 1391 guarantee(mid->header()->is_neutral(), "invariant"); 1392 1393 if (mid->is_busy()) { 1394 if (ClearResponsibleAtSTW) mid->_Responsible = NULL; 1395 deflated = false; 1396 } else { 1397 // Deflate the monitor if it is no longer being used 1398 // It's idle - scavenge and return to the global free list 1399 // plain old deflation ... 1400 TEVENT(deflate_idle_monitors - scavenge1); 1401 if (TraceMonitorInflation) { 1402 if (obj->is_instance()) { 1403 ResourceMark rm; 1404 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", 1405 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name()); 1406 } 1407 } 1408 1409 // Restore the header back to obj 1410 obj->release_set_mark(mid->header()); 1411 mid->clear(); 1412 1413 assert(mid->object() == NULL, "invariant"); 1414 1415 // Move the object to the working free list defined by FreeHead,FreeTail. 1416 if (*freeHeadp == NULL) *freeHeadp = mid; 1417 if (*freeTailp != NULL) { 1418 ObjectMonitor * prevtail = *freeTailp; 1419 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK 1420 prevtail->FreeNext = mid; 1421 } 1422 *freeTailp = mid; 1423 deflated = true; 1424 } 1425 return deflated; 1426} 1427 1428// Caller acquires ListLock 1429int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp, 1430 ObjectMonitor** freeHeadp, ObjectMonitor** freeTailp) { 1431 ObjectMonitor* mid; 1432 ObjectMonitor* next; 1433 ObjectMonitor* curmidinuse = NULL; 1434 int deflatedcount = 0; 1435 1436 for (mid = *listheadp; mid != NULL;) { 1437 oop obj = (oop) mid->object(); 1438 bool deflated = false; 1439 if (obj != NULL) { 1440 deflated = deflate_monitor(mid, obj, freeHeadp, freeTailp); 1441 } 1442 if (deflated) { 1443 // extract from per-thread in-use-list 1444 if (mid == *listheadp) { 1445 *listheadp = mid->FreeNext; 1446 } else if (curmidinuse != NULL) { 1447 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist 1448 } 1449 next = mid->FreeNext; 1450 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list 1451 mid = next; 1452 deflatedcount++; 1453 } else { 1454 curmidinuse = mid; 1455 mid = mid->FreeNext; 1456 } 1457 } 1458 return deflatedcount; 1459} 1460 1461void ObjectSynchronizer::deflate_idle_monitors() { 1462 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1463 int nInuse = 0; // currently associated with objects 1464 int nInCirculation = 0; // extant 1465 int nScavenged = 0; // reclaimed 1466 bool deflated = false; 1467 1468 ObjectMonitor * FreeHead = NULL; // Local SLL of scavenged monitors 1469 ObjectMonitor * FreeTail = NULL; 1470 1471 TEVENT(deflate_idle_monitors); 1472 // Prevent omFlush from changing mids in Thread dtor's during deflation 1473 // And in case the vm thread is acquiring a lock during a safepoint 1474 // See e.g. 6320749 1475 Thread::muxAcquire(&ListLock, "scavenge - return"); 1476 1477 if (MonitorInUseLists) { 1478 int inUse = 0; 1479 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { 1480 nInCirculation+= cur->omInUseCount; 1481 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail); 1482 cur->omInUseCount-= deflatedcount; 1483 if (ObjectMonitor::Knob_VerifyInUse) { 1484 verifyInUse(cur); 1485 } 1486 nScavenged += deflatedcount; 1487 nInuse += cur->omInUseCount; 1488 } 1489 1490 // For moribund threads, scan gOmInUseList 1491 if (gOmInUseList) { 1492 nInCirculation += gOmInUseCount; 1493 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail); 1494 gOmInUseCount-= deflatedcount; 1495 nScavenged += deflatedcount; 1496 nInuse += gOmInUseCount; 1497 } 1498 1499 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { 1500 // Iterate over all extant monitors - Scavenge all idle monitors. 1501 assert(block->object() == CHAINMARKER, "must be a block header"); 1502 nInCirculation += _BLOCKSIZE; 1503 for (int i = 1; i < _BLOCKSIZE; i++) { 1504 ObjectMonitor* mid = &block[i]; 1505 oop obj = (oop) mid->object(); 1506 1507 if (obj == NULL) { 1508 // The monitor is not associated with an object. 1509 // The monitor should either be a thread-specific private 1510 // free list or the global free list. 1511 // obj == NULL IMPLIES mid->is_busy() == 0 1512 guarantee(!mid->is_busy(), "invariant"); 1513 continue; 1514 } 1515 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail); 1516 1517 if (deflated) { 1518 mid->FreeNext = NULL; 1519 nScavenged++; 1520 } else { 1521 nInuse++; 1522 } 1523 } 1524 } 1525 1526 MonitorFreeCount += nScavenged; 1527 1528 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree. 1529 1530 if (ObjectMonitor::Knob_Verbose) { 1531 ::printf("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n", 1532 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, 1533 MonitorPopulation, MonitorFreeCount); 1534 ::fflush(stdout); 1535 } 1536 1537 ForceMonitorScavenge = 0; // Reset 1538 1539 // Move the scavenged monitors back to the global free list. 1540 if (FreeHead != NULL) { 1541 guarantee(FreeTail != NULL && nScavenged > 0, "invariant"); 1542 assert(FreeTail->FreeNext == NULL, "invariant"); 1543 // constant-time list splice - prepend scavenged segment to gFreeList 1544 FreeTail->FreeNext = gFreeList; 1545 gFreeList = FreeHead; 1546 } 1547 Thread::muxRelease(&ListLock); 1548 1549 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged); 1550 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation); 1551 1552 // TODO: Add objectMonitor leak detection. 1553 // Audit/inventory the objectMonitors -- make sure they're all accounted for. 1554 GVars.stwRandom = os::random(); 1555 GVars.stwCycle++; 1556} 1557 1558// Monitor cleanup on JavaThread::exit 1559 1560// Iterate through monitor cache and attempt to release thread's monitors 1561// Gives up on a particular monitor if an exception occurs, but continues 1562// the overall iteration, swallowing the exception. 1563class ReleaseJavaMonitorsClosure: public MonitorClosure { 1564 private: 1565 TRAPS; 1566 1567 public: 1568 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 1569 void do_monitor(ObjectMonitor* mid) { 1570 if (mid->owner() == THREAD) { 1571 (void)mid->complete_exit(CHECK); 1572 } 1573 } 1574}; 1575 1576// Release all inflated monitors owned by THREAD. Lightweight monitors are 1577// ignored. This is meant to be called during JNI thread detach which assumes 1578// all remaining monitors are heavyweight. All exceptions are swallowed. 1579// Scanning the extant monitor list can be time consuming. 1580// A simple optimization is to add a per-thread flag that indicates a thread 1581// called jni_monitorenter() during its lifetime. 1582// 1583// Instead of No_Savepoint_Verifier it might be cheaper to 1584// use an idiom of the form: 1585// auto int tmp = SafepointSynchronize::_safepoint_counter ; 1586// <code that must not run at safepoint> 1587// guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1588// Since the tests are extremely cheap we could leave them enabled 1589// for normal product builds. 1590 1591void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 1592 assert(THREAD == JavaThread::current(), "must be current Java thread"); 1593 No_Safepoint_Verifier nsv; 1594 ReleaseJavaMonitorsClosure rjmc(THREAD); 1595 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); 1596 ObjectSynchronizer::monitors_iterate(&rjmc); 1597 Thread::muxRelease(&ListLock); 1598 THREAD->clear_pending_exception(); 1599} 1600 1601//------------------------------------------------------------------------------ 1602// Debugging code 1603 1604void ObjectSynchronizer::sanity_checks(const bool verbose, 1605 const uint cache_line_size, 1606 int *error_cnt_ptr, 1607 int *warning_cnt_ptr) { 1608 u_char *addr_begin = (u_char*)&GVars; 1609 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom; 1610 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence; 1611 1612 if (verbose) { 1613 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT, 1614 sizeof(SharedGlobals)); 1615 } 1616 1617 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin); 1618 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom); 1619 1620 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin); 1621 if (verbose) { 1622 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence); 1623 } 1624 1625 if (cache_line_size != 0) { 1626 // We were able to determine the L1 data cache line size so 1627 // do some cache line specific sanity checks 1628 1629 if (offset_stwRandom < cache_line_size) { 1630 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer " 1631 "to the struct beginning than a cache line which permits " 1632 "false sharing."); 1633 (*warning_cnt_ptr)++; 1634 } 1635 1636 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) { 1637 tty->print_cr("WARNING: the SharedGlobals.stwRandom and " 1638 "SharedGlobals.hcSequence fields are closer than a cache " 1639 "line which permits false sharing."); 1640 (*warning_cnt_ptr)++; 1641 } 1642 1643 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) { 1644 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer " 1645 "to the struct end than a cache line which permits false " 1646 "sharing."); 1647 (*warning_cnt_ptr)++; 1648 } 1649 } 1650} 1651 1652#ifndef PRODUCT 1653 1654// Verify all monitors in the monitor cache, the verification is weak. 1655void ObjectSynchronizer::verify() { 1656 ObjectMonitor* block = gBlockList; 1657 ObjectMonitor* mid; 1658 while (block) { 1659 assert(block->object() == CHAINMARKER, "must be a block header"); 1660 for (int i = 1; i < _BLOCKSIZE; i++) { 1661 mid = block + i; 1662 oop object = (oop) mid->object(); 1663 if (object != NULL) { 1664 mid->verify(); 1665 } 1666 } 1667 block = (ObjectMonitor*) block->FreeNext; 1668 } 1669} 1670 1671// Check if monitor belongs to the monitor cache 1672// The list is grow-only so it's *relatively* safe to traverse 1673// the list of extant blocks without taking a lock. 1674 1675int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 1676 ObjectMonitor* block = gBlockList; 1677 1678 while (block) { 1679 assert(block->object() == CHAINMARKER, "must be a block header"); 1680 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 1681 address mon = (address) monitor; 1682 address blk = (address) block; 1683 size_t diff = mon - blk; 1684 assert((diff % sizeof(ObjectMonitor)) == 0, "check"); 1685 return 1; 1686 } 1687 block = (ObjectMonitor*) block->FreeNext; 1688 } 1689 return 0; 1690} 1691 1692#endif 1693