107 #endif // ndef DTRACE_ENABLED
108
109 // This exists only as a workaround of dtrace bug 6254741
110 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
111 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
112 return 0;
113 }
114
115 #define NINFLATIONLOCKS 256
116 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
117
118 // global list of blocks of monitors
119 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
120 // global monitor free list
121 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
122 // global monitor in-use list, for moribund threads,
123 // monitors they inflated need to be scanned for deflation
124 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
125 // count of entries in gOmInUseList
126 int ObjectSynchronizer::gOmInUseCount = 0;
127
128 static volatile intptr_t gListLock = 0; // protects global monitor lists
129 static volatile int gMonitorFreeCount = 0; // # on gFreeList
130 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
131
132 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
133
134
135 // =====================> Quick functions
136
137 // The quick_* forms are special fast-path variants used to improve
138 // performance. In the simplest case, a "quick_*" implementation could
139 // simply return false, in which case the caller will perform the necessary
140 // state transitions and call the slow-path form.
141 // The fast-path is designed to handle frequently arising cases in an efficient
142 // manner and is just a degenerate "optimistic" variant of the slow-path.
143 // returns true -- to indicate the call was satisfied.
144 // returns false -- to indicate the call needs the services of the slow-path.
145 // A no-loitering ordinance is in effect for code in the quick_* family
146 // operators: safepoints or indefinite blocking (blocking that might span a
193 }
194
195 // biased locking and any other IMS exception states take the slow-path
196 return false;
197 }
198
199
200 // The LockNode emitted directly at the synchronization site would have
201 // been too big if it were to have included support for the cases of inflated
202 // recursive enter and exit, so they go here instead.
203 // Note that we can't safely call AsyncPrintJavaStack() from within
204 // quick_enter() as our thread state remains _in_Java.
205
206 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
207 BasicLock * lock) {
208 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
209 assert(Self->is_Java_thread(), "invariant");
210 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
211 NoSafepointVerifier nsv;
212 if (obj == NULL) return false; // Need to throw NPE
213 const markOop mark = obj->mark();
214
215 if (mark->has_monitor()) {
216 ObjectMonitor * const m = mark->monitor();
217 assert(oopDesc::equals((oop) m->object(), obj), "invariant");
218 Thread * const owner = (Thread *) m->_owner;
219
220 // Lock contention and Transactional Lock Elision (TLE) diagnostics
221 // and observability
222 // Case: light contention possibly amenable to TLE
223 // Case: TLE inimical operations such as nested/recursive synchronization
224
225 if (owner == Self) {
226 m->_recursions++;
227 return true;
228 }
229
230 // This Java Monitor is inflated so obj's header will never be
231 // displaced to this thread's BasicLock. Make the displaced header
232 // non-NULL so this BasicLock is not seen as recursive nor as
233 // being locked. We do this unconditionally so that this thread's
234 // BasicLock cannot be mis-interpreted by any stack walkers. For
235 // performance reasons, stack walkers generally first check for
236 // Biased Locking in the object's header, the second check is for
237 // stack-locking in the object's header, the third check is for
238 // recursive stack-locking in the displaced header in the BasicLock,
239 // and last are the inflated Java Monitor (ObjectMonitor) checks.
240 lock->set_displaced_header(markOopDesc::unused_mark());
241
242 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
243 assert(m->_recursions == 0, "invariant");
244 assert(m->_owner == Self, "invariant");
245 return true;
246 }
247 }
248
249 // Note that we could inflate in quick_enter.
250 // This is likely a useful optimization
251 // Critically, in quick_enter() we must not:
252 // -- perform bias revocation, or
253 // -- block indefinitely, or
254 // -- reach a safepoint
255
256 return false; // revert to slow-path
257 }
258
259 // -----------------------------------------------------------------------------
260 // Fast Monitor Enter/Exit
261 // This the fast monitor enter. The interpreter and compiler use
262 // some assembly copies of this code. Make sure update those code
263 // if the following function is changed. The implementation is
264 // extremely sensitive to race condition. Be careful.
265
266 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
267 bool attempt_rebias, TRAPS) {
310 // does not own the Java Monitor.
311 ObjectMonitor * m = mark->monitor();
312 assert(((oop)(m->object()))->mark() == mark, "invariant");
313 assert(m->is_entered(THREAD), "invariant");
314 }
315 }
316 #endif
317 return;
318 }
319
320 if (mark == (markOop) lock) {
321 // If the object is stack-locked by the current thread, try to
322 // swing the displaced header from the BasicLock back to the mark.
323 assert(dhw->is_neutral(), "invariant");
324 if (object->cas_set_mark(dhw, mark) == mark) {
325 return;
326 }
327 }
328
329 // We have to take the slow-path of possible inflation and then exit.
330 inflate(THREAD, object, inflate_cause_vm_internal)->exit(true, THREAD);
331 }
332
333 // -----------------------------------------------------------------------------
334 // Interpreter/Compiler Slow Case
335 // This routine is used to handle interpreter/compiler slow case
336 // We don't need to use fast path here, because it must have been
337 // failed in the interpreter/compiler code.
338 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
339 markOop mark = obj->mark();
340 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
341
342 if (mark->is_neutral()) {
343 // Anticipate successful CAS -- the ST of the displaced mark must
344 // be visible <= the ST performed by the CAS.
345 lock->set_displaced_header(mark);
346 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
347 return;
348 }
349 // Fall through to inflate() ...
350 } else if (mark->has_locker() &&
351 THREAD->is_lock_owned((address)mark->locker())) {
352 assert(lock != mark->locker(), "must not re-lock the same lock");
353 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
354 lock->set_displaced_header(NULL);
355 return;
356 }
357
358 // The object header will never be displaced to this lock,
359 // so it does not matter what the value is, except that it
360 // must be non-zero to avoid looking like a re-entrant lock,
361 // and must not look locked either.
362 lock->set_displaced_header(markOopDesc::unused_mark());
363 inflate(THREAD, obj(), inflate_cause_monitor_enter)->enter(THREAD);
364 }
365
366 // This routine is used to handle interpreter/compiler slow case
367 // We don't need to use fast path here, because it must have
368 // failed in the interpreter/compiler code. Simply use the heavy
369 // weight monitor should be ok, unless someone find otherwise.
370 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
371 fast_exit(object, lock, THREAD);
372 }
373
374 // -----------------------------------------------------------------------------
375 // Class Loader support to workaround deadlocks on the class loader lock objects
376 // Also used by GC
377 // complete_exit()/reenter() are used to wait on a nested lock
378 // i.e. to give up an outer lock completely and then re-enter
379 // Used when holding nested locks - lock acquisition order: lock1 then lock2
380 // 1) complete_exit lock1 - saving recursion count
381 // 2) wait on lock2
382 // 3) when notified on lock2, unlock lock2
383 // 4) reenter lock1 with original recursion count
384 // 5) lock lock2
385 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
386 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
387 if (UseBiasedLocking) {
388 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
389 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
390 }
391
392 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_vm_internal);
393
394 return monitor->complete_exit(THREAD);
395 }
396
397 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
398 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
399 if (UseBiasedLocking) {
400 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
401 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
402 }
403
404 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_vm_internal);
405
406 monitor->reenter(recursion, THREAD);
407 }
408 // -----------------------------------------------------------------------------
409 // JNI locks on java objects
410 // NOTE: must use heavy weight monitor to handle jni monitor enter
411 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
412 // the current locking is from JNI instead of Java code
413 if (UseBiasedLocking) {
414 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
415 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
416 }
417 THREAD->set_current_pending_monitor_is_from_java(false);
418 inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD);
419 THREAD->set_current_pending_monitor_is_from_java(true);
420 }
421
422 // NOTE: must use heavy weight monitor to handle jni monitor exit
423 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
424 if (UseBiasedLocking) {
425 Handle h_obj(THREAD, obj);
426 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
427 obj = h_obj();
428 }
429 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
430
431 ObjectMonitor* monitor = inflate(THREAD, obj, inflate_cause_jni_exit);
432 // If this thread has locked the object, exit the monitor. Note: can't use
433 // monitor->check(CHECK); must exit even if an exception is pending.
434 if (monitor->check(THREAD)) {
435 monitor->exit(true, THREAD);
436 }
437 }
438
439 // -----------------------------------------------------------------------------
440 // Internal VM locks on java objects
441 // standard constructor, allows locking failures
442 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
443 _dolock = doLock;
444 _thread = thread;
445 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
446 _obj = obj;
447
448 if (_dolock) {
449 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
450 }
451 }
452
453 ObjectLocker::~ObjectLocker() {
454 if (_dolock) {
455 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
456 }
457 }
458
459
460 // -----------------------------------------------------------------------------
461 // Wait/Notify/NotifyAll
462 // NOTE: must use heavy weight monitor to handle wait()
463 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
464 if (UseBiasedLocking) {
465 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
466 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
467 }
468 if (millis < 0) {
469 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
470 }
471 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_wait);
472
473 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
474 monitor->wait(millis, true, THREAD);
475
476 // This dummy call is in place to get around dtrace bug 6254741. Once
477 // that's fixed we can uncomment the following line, remove the call
478 // and change this function back into a "void" func.
479 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
480 return dtrace_waited_probe(monitor, obj, THREAD);
481 }
482
483 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
484 if (UseBiasedLocking) {
485 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
486 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
487 }
488 if (millis < 0) {
489 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
490 }
491 inflate(THREAD, obj(), inflate_cause_wait)->wait(millis, false, THREAD);
492 }
493
494 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
495 if (UseBiasedLocking) {
496 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
497 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
498 }
499
500 markOop mark = obj->mark();
501 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
502 return;
503 }
504 inflate(THREAD, obj(), inflate_cause_notify)->notify(THREAD);
505 }
506
507 // NOTE: see comment of notify()
508 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
509 if (UseBiasedLocking) {
510 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
511 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
512 }
513
514 markOop mark = obj->mark();
515 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
516 return;
517 }
518 inflate(THREAD, obj(), inflate_cause_notify)->notifyAll(THREAD);
519 }
520
521 // -----------------------------------------------------------------------------
522 // Hash Code handling
523 //
524 // Performance concern:
525 // OrderAccess::storestore() calls release() which at one time stored 0
526 // into the global volatile OrderAccess::dummy variable. This store was
527 // unnecessary for correctness. Many threads storing into a common location
528 // causes considerable cache migration or "sloshing" on large SMP systems.
529 // As such, I avoided using OrderAccess::storestore(). In some cases
530 // OrderAccess::fence() -- which incurs local latency on the executing
531 // processor -- is a better choice as it scales on SMP systems.
532 //
533 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
534 // a discussion of coherency costs. Note that all our current reference
535 // platforms provide strong ST-ST order, so the issue is moot on IA32,
536 // x64, and SPARC.
537 //
538 // As a general policy we use "volatile" to control compiler-based reordering
692 Handle hobj(Self, obj);
693 // Relaxing assertion for bug 6320749.
694 assert(Universe::verify_in_progress() ||
695 !SafepointSynchronize::is_at_safepoint(),
696 "biases should not be seen by VM thread here");
697 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
698 obj = hobj();
699 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
700 }
701 }
702
703 // hashCode() is a heap mutator ...
704 // Relaxing assertion for bug 6320749.
705 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
706 !SafepointSynchronize::is_at_safepoint(), "invariant");
707 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
708 Self->is_Java_thread() , "invariant");
709 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
710 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
711
712 ObjectMonitor* monitor = NULL;
713 markOop temp, test;
714 intptr_t hash;
715 markOop mark = ReadStableMark(obj);
716
717 // object should remain ineligible for biased locking
718 assert(!mark->has_bias_pattern(), "invariant");
719
720 if (mark->is_neutral()) {
721 hash = mark->hash(); // this is a normal header
722 if (hash != 0) { // if it has hash, just return it
723 return hash;
724 }
725 hash = get_next_hash(Self, obj); // allocate a new hash code
726 temp = mark->copy_set_hash(hash); // merge the hash code into header
727 // use (machine word version) atomic operation to install the hash
728 test = obj->cas_set_mark(temp, mark);
729 if (test == mark) {
730 return hash;
731 }
732 // If atomic operation failed, we must inflate the header
733 // into heavy weight monitor. We could add more code here
734 // for fast path, but it does not worth the complexity.
735 } else if (mark->has_monitor()) {
736 monitor = mark->monitor();
737 temp = monitor->header();
738 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)temp));
739 hash = temp->hash();
740 if (hash != 0) {
741 return hash;
742 }
743 // Skip to the following code to reduce code size
744 } else if (Self->is_lock_owned((address)mark->locker())) {
745 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
746 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)temp));
747 hash = temp->hash(); // by current thread, check if the displaced
748 if (hash != 0) { // header contains hash code
749 return hash;
750 }
751 // WARNING:
752 // The displaced header is strictly immutable.
753 // It can NOT be changed in ANY cases. So we have
754 // to inflate the header into heavyweight monitor
755 // even the current thread owns the lock. The reason
756 // is the BasicLock (stack slot) will be asynchronously
757 // read by other threads during the inflate() function.
758 // Any change to stack may not propagate to other threads
759 // correctly.
760 }
761
762 // Inflate the monitor to set hash code
763 monitor = inflate(Self, obj, inflate_cause_hash_code);
764 // Load displaced header and check it has hash code
765 mark = monitor->header();
766 assert(mark->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)mark));
767 hash = mark->hash();
768 if (hash == 0) {
769 hash = get_next_hash(Self, obj);
770 temp = mark->copy_set_hash(hash); // merge hash code into header
771 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)temp));
772 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark);
773 if (test != mark) {
774 // The only update to the header in the monitor (outside GC)
775 // is install the hash code. If someone add new usage of
776 // displaced header, please update this code
777 hash = test->hash();
778 assert(test->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)test));
779 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
780 }
781 }
782 // We finally get the hash
783 return hash;
784 }
785
786 // Deprecated -- use FastHashCode() instead.
787
788 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
789 return FastHashCode(Thread::current(), obj());
790 }
791
792
793 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
794 Handle h_obj) {
795 if (UseBiasedLocking) {
796 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
797 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
798 }
799
800 assert(thread == JavaThread::current(), "Can only be called on current thread");
801 oop obj = h_obj();
802
803 markOop mark = ReadStableMark(obj);
804
805 // Uncontended case, header points to stack
806 if (mark->has_locker()) {
807 return thread->is_lock_owned((address)mark->locker());
808 }
809 // Contended case, header points to ObjectMonitor (tagged pointer)
810 if (mark->has_monitor()) {
811 ObjectMonitor* monitor = mark->monitor();
812 return monitor->is_entered(thread) != 0;
813 }
814 // Unlocked case, header in place
815 assert(mark->is_neutral(), "sanity check");
816 return false;
817 }
818
819 // Be aware of this method could revoke bias of the lock object.
820 // This method queries the ownership of the lock handle specified by 'h_obj'.
821 // If the current thread owns the lock, it returns owner_self. If no
822 // thread owns the lock, it returns owner_none. Otherwise, it will return
823 // owner_other.
824 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
825 (JavaThread *self, Handle h_obj) {
826 // The caller must beware this method can revoke bias, and
827 // revocation can result in a safepoint.
828 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
829 assert(self->thread_state() != _thread_blocked, "invariant");
830
831 // Possible mark states: neutral, biased, stack-locked, inflated
832
833 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
834 // CASE: biased
835 BiasedLocking::revoke_and_rebias(h_obj, false, self);
836 assert(!h_obj->mark()->has_bias_pattern(),
837 "biases should be revoked by now");
838 }
839
840 assert(self == JavaThread::current(), "Can only be called on current thread");
841 oop obj = h_obj();
842 markOop mark = ReadStableMark(obj);
843
844 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
845 if (mark->has_locker()) {
846 return self->is_lock_owned((address)mark->locker()) ?
847 owner_self : owner_other;
848 }
849
850 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor.
851 // The Object:ObjectMonitor relationship is stable as long as we're
852 // not at a safepoint.
853 if (mark->has_monitor()) {
854 void * owner = mark->monitor()->_owner;
855 if (owner == NULL) return owner_none;
856 return (owner == self ||
857 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
858 }
859
860 // CASE: neutral
861 assert(mark->is_neutral(), "sanity check");
862 return owner_none; // it's unlocked
863 }
864
865 // FIXME: jvmti should call this
866 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
867 if (UseBiasedLocking) {
868 if (SafepointSynchronize::is_at_safepoint()) {
869 BiasedLocking::revoke_at_safepoint(h_obj);
870 } else {
871 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
872 }
873 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
874 }
875
876 oop obj = h_obj();
877 address owner = NULL;
878
879 markOop mark = ReadStableMark(obj);
880
881 // Uncontended case, header points to stack
882 if (mark->has_locker()) {
883 owner = (address) mark->locker();
884 }
885
886 // Contended case, header points to ObjectMonitor (tagged pointer)
887 else if (mark->has_monitor()) {
888 ObjectMonitor* monitor = mark->monitor();
889 assert(monitor != NULL, "monitor should be non-null");
890 owner = (address) monitor->owner();
891 }
892
893 if (owner != NULL) {
894 // owning_thread_from_monitor_owner() may also return NULL here
895 return Threads::owning_thread_from_monitor_owner(t_list, owner);
896 }
897
898 // Unlocked case, header in place
899 // Cannot have assertion since this object may have been
900 // locked by another thread when reaching here.
901 // assert(mark->is_neutral(), "sanity check");
902
903 return NULL;
904 }
905
906 // Visitors ...
907
908 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
909 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
910 while (block != NULL) {
911 assert(block->object() == CHAINMARKER, "must be a block header");
912 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
913 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
914 oop object = (oop)mid->object();
915 if (object != NULL) {
916 closure->do_monitor(mid);
917 }
918 }
919 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
920 }
921 }
922
923 // Get the next block in the block list.
924 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
925 assert(block->object() == CHAINMARKER, "must be a block header");
926 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
927 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
928 return block;
929 }
930
931 static bool monitors_used_above_threshold() {
932 if (gMonitorPopulation == 0) {
933 return false;
934 }
935 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
1006 // See also: GuaranteedSafepointInterval
1007 //
1008 // The current implementation uses asynchronous VM operations.
1009
1010 static void InduceScavenge(Thread * Self, const char * Whence) {
1011 // Induce STW safepoint to trim monitors
1012 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1013 // More precisely, trigger an asynchronous STW safepoint as the number
1014 // of active monitors passes the specified threshold.
1015 // TODO: assert thread state is reasonable
1016
1017 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1018 // Induce a 'null' safepoint to scavenge monitors
1019 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1020 // to the VMthread and have a lifespan longer than that of this activation record.
1021 // The VMThread will delete the op when completed.
1022 VMThread::execute(new VM_ScavengeMonitors());
1023 }
1024 }
1025
1026 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) {
1027 // A large MAXPRIVATE value reduces both list lock contention
1028 // and list coherency traffic, but also tends to increase the
1029 // number of objectMonitors in circulation as well as the STW
1030 // scavenge costs. As usual, we lean toward time in space-time
1031 // tradeoffs.
1032 const int MAXPRIVATE = 1024;
1033 for (;;) {
1034 ObjectMonitor * m;
1035
1036 // 1: try to allocate from the thread's local omFreeList.
1037 // Threads will attempt to allocate first from their local list, then
1038 // from the global list, and only after those attempts fail will the thread
1039 // attempt to instantiate new monitors. Thread-local free lists take
1040 // heat off the gListLock and improve allocation latency, as well as reducing
1041 // coherency traffic on the shared global list.
1042 m = Self->omFreeList;
1043 if (m != NULL) {
1044 Self->omFreeList = m->FreeNext;
1045 Self->omFreeCount--;
1046 guarantee(m->object() == NULL, "invariant");
1047 m->FreeNext = Self->omInUseList;
1048 Self->omInUseList = m;
1049 Self->omInUseCount++;
1050 return m;
1051 }
1052
1053 // 2: try to allocate from the global gFreeList
1054 // CONSIDER: use muxTry() instead of muxAcquire().
1055 // If the muxTry() fails then drop immediately into case 3.
1056 // If we're using thread-local free lists then try
1057 // to reprovision the caller's free list.
1058 if (gFreeList != NULL) {
1059 // Reprovision the thread's omFreeList.
1060 // Use bulk transfers to reduce the allocation rate and heat
1061 // on various locks.
1062 Thread::muxAcquire(&gListLock, "omAlloc(1)");
1063 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1064 gMonitorFreeCount--;
1065 ObjectMonitor * take = gFreeList;
1066 gFreeList = take->FreeNext;
1067 guarantee(take->object() == NULL, "invariant");
1068 guarantee(!take->is_busy(), "invariant");
1069 take->Recycle();
1070 omRelease(Self, take, false);
1071 }
1072 Thread::muxRelease(&gListLock);
1073 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1074 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1075
1076 const int mx = MonitorBound;
1077 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1078 // We can't safely induce a STW safepoint from omAlloc() as our thread
1079 // state may not be appropriate for such activities and callers may hold
1080 // naked oops, so instead we defer the action.
1081 InduceScavenge(Self, "omAlloc");
1082 }
1083 continue;
1084 }
1085
1086 // 3: allocate a block of new ObjectMonitors
1087 // Both the local and global free lists are empty -- resort to malloc().
1088 // In the current implementation objectMonitors are TSM - immortal.
1089 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1102
1103 // NOTE: (almost) no way to recover if allocation failed.
1104 // We might be able to induce a STW safepoint and scavenge enough
1105 // objectMonitors to permit progress.
1106 if (temp == NULL) {
1107 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1108 "Allocate ObjectMonitors");
1109 }
1110 (void)memset((void *) temp, 0, neededsize);
1111
1112 // Format the block.
1113 // initialize the linked list, each monitor points to its next
1114 // forming the single linked free list, the very first monitor
1115 // will points to next block, which forms the block list.
1116 // The trick of using the 1st element in the block as gBlockList
1117 // linkage should be reconsidered. A better implementation would
1118 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1119
1120 for (int i = 1; i < _BLOCKSIZE; i++) {
1121 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1122 }
1123
1124 // terminate the last monitor as the end of list
1125 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1126
1127 // Element [0] is reserved for global list linkage
1128 temp[0].set_object(CHAINMARKER);
1129
1130 // Consider carving out this thread's current request from the
1131 // block in hand. This avoids some lock traffic and redundant
1132 // list activity.
1133
1134 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1135 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1136 Thread::muxAcquire(&gListLock, "omAlloc(2)");
1137 gMonitorPopulation += _BLOCKSIZE-1;
1138 gMonitorFreeCount += _BLOCKSIZE-1;
1139
1140 // Add the new block to the list of extant blocks (gBlockList).
1141 // The very first objectMonitor in a block is reserved and dedicated.
1144 // There are lock-free uses of gBlockList so make sure that
1145 // the previous stores happen before we update gBlockList.
1146 OrderAccess::release_store(&gBlockList, temp);
1147
1148 // Add the new string of objectMonitors to the global free list
1149 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1150 gFreeList = temp + 1;
1151 Thread::muxRelease(&gListLock);
1152 }
1153 }
1154
1155 // Place "m" on the caller's private per-thread omFreeList.
1156 // In practice there's no need to clamp or limit the number of
1157 // monitors on a thread's omFreeList as the only time we'll call
1158 // omRelease is to return a monitor to the free list after a CAS
1159 // attempt failed. This doesn't allow unbounded #s of monitors to
1160 // accumulate on a thread's free list.
1161 //
1162 // Key constraint: all ObjectMonitors on a thread's free list and the global
1163 // free list must have their object field set to null. This prevents the
1164 // scavenger -- deflate_monitor_list() -- from reclaiming them.
1165
1166 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1167 bool fromPerThreadAlloc) {
1168 guarantee(m->header() == NULL, "invariant");
1169 guarantee(m->object() == NULL, "invariant");
1170 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor");
1171 // Remove from omInUseList
1172 if (fromPerThreadAlloc) {
1173 ObjectMonitor* cur_mid_in_use = NULL;
1174 bool extracted = false;
1175 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1176 if (m == mid) {
1177 // extract from per-thread in-use list
1178 if (mid == Self->omInUseList) {
1179 Self->omInUseList = mid->FreeNext;
1180 } else if (cur_mid_in_use != NULL) {
1181 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1182 }
1183 extracted = true;
1184 Self->omInUseCount--;
1185 break;
1186 }
1187 }
1188 assert(extracted, "Should have extracted from in-use list");
1189 }
1190
1191 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1192 m->FreeNext = Self->omFreeList;
1193 Self->omFreeList = m;
1194 Self->omFreeCount++;
1195 }
1196
1197 // Return the monitors of a moribund thread's local free list to
1198 // the global free list. Typically a thread calls omFlush() when
1199 // it's dying. We could also consider having the VM thread steal
1200 // monitors from threads that have not run java code over a few
1201 // consecutive STW safepoints. Relatedly, we might decay
1202 // omFreeProvision at STW safepoints.
1203 //
1204 // Also return the monitors of a moribund thread's omInUseList to
1205 // a global gOmInUseList under the global list lock so these
1206 // will continue to be scanned.
1207 //
1208 // We currently call omFlush() from Threads::remove() _before the thread
1209 // has been excised from the thread list and is no longer a mutator.
1210 // This means that omFlush() cannot run concurrently with a safepoint and
1211 // interleave with the deflate_idle_monitors scavenge operator. In particular,
1212 // this ensures that the thread's monitors are scanned by a GC safepoint,
1213 // either via Thread::oops_do() (if safepoint happens before omFlush()) or via
1214 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1215 // monitors have been transferred to the global in-use list).
1216
1217 void ObjectSynchronizer::omFlush(Thread * Self) {
1218 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1219 ObjectMonitor * tail = NULL;
1220 int tally = 0;
1221 if (list != NULL) {
1222 ObjectMonitor * s;
1223 // The thread is going away, the per-thread free monitors
1224 // are freed via set_owner(NULL)
1225 // Link them to tail, which will be linked into the global free list
1226 // gFreeList below, under the gListLock
1227 for (s = list; s != NULL; s = s->FreeNext) {
1228 tally++;
1229 tail = s;
1230 guarantee(s->object() == NULL, "invariant");
1231 guarantee(!s->is_busy(), "invariant");
1232 s->set_owner(NULL); // redundant but good hygiene
1233 }
1234 guarantee(tail != NULL, "invariant");
1235 assert(Self->omFreeCount == tally, "free-count off");
1236 Self->omFreeList = NULL;
1237 Self->omFreeCount = 0;
1238 }
1239
1240 ObjectMonitor * inUseList = Self->omInUseList;
1241 ObjectMonitor * inUseTail = NULL;
1242 int inUseTally = 0;
1243 if (inUseList != NULL) {
1244 ObjectMonitor *cur_om;
1245 // The thread is going away, however the omInUseList inflated
1246 // monitors may still be in-use by other threads.
1247 // Link them to inUseTail, which will be linked into the global in-use list
1248 // gOmInUseList below, under the gListLock
1249 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1250 inUseTail = cur_om;
1251 inUseTally++;
1252 }
1253 guarantee(inUseTail != NULL, "invariant");
1254 assert(Self->omInUseCount == inUseTally, "in-use count off");
1255 Self->omInUseList = NULL;
1256 Self->omInUseCount = 0;
1257 }
1258
1259 Thread::muxAcquire(&gListLock, "omFlush");
1260 if (tail != NULL) {
1261 tail->FreeNext = gFreeList;
1262 gFreeList = list;
1263 gMonitorFreeCount += tally;
1264 }
1265
1266 if (inUseTail != NULL) {
1267 inUseTail->FreeNext = gOmInUseList;
1268 gOmInUseList = inUseList;
1269 gOmInUseCount += inUseTally;
1270 }
1271
1272 Thread::muxRelease(&gListLock);
1273
1274 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1282 }
1283 if (ls != NULL) {
1284 ls->print_cr("omFlush: jt=" INTPTR_FORMAT ", free_monitor_tally=%d"
1285 ", in_use_monitor_tally=%d" ", omFreeProvision=%d",
1286 p2i(Self), tally, inUseTally, Self->omFreeProvision);
1287 }
1288 }
1289
1290 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1291 const oop obj,
1292 ObjectSynchronizer::InflateCause cause) {
1293 assert(event != NULL, "invariant");
1294 assert(event->should_commit(), "invariant");
1295 event->set_monitorClass(obj->klass());
1296 event->set_address((uintptr_t)(void*)obj);
1297 event->set_cause((u1)cause);
1298 event->commit();
1299 }
1300
1301 // Fast path code shared by multiple functions
1302 void ObjectSynchronizer::inflate_helper(oop obj) {
1303 markOop mark = obj->mark();
1304 if (mark->has_monitor()) {
1305 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1306 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1307 return;
1308 }
1309 inflate(Thread::current(), obj, inflate_cause_vm_internal);
1310 }
1311
1312 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self,
1313 oop object,
1314 const InflateCause cause) {
1315 // Inflate mutates the heap ...
1316 // Relaxing assertion for bug 6320749.
1317 assert(Universe::verify_in_progress() ||
1318 !SafepointSynchronize::is_at_safepoint(), "invariant");
1319
1320 EventJavaMonitorInflate event;
1321
1322 for (;;) {
1323 const markOop mark = object->mark();
1324 assert(!mark->has_bias_pattern(), "invariant");
1325
1326 // The mark can be in one of the following states:
1327 // * Inflated - just return
1328 // * Stack-locked - coerce it to inflated
1329 // * INFLATING - busy wait for conversion to complete
1330 // * Neutral - aggressively inflate the object.
1331 // * BIASED - Illegal. We should never see this
1332
1333 // CASE: inflated
1334 if (mark->has_monitor()) {
1335 ObjectMonitor * inf = mark->monitor();
1336 markOop dmw = inf->header();
1337 assert(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)dmw));
1338 assert(oopDesc::equals((oop) inf->object(), object), "invariant");
1339 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1340 return inf;
1341 }
1342
1343 // CASE: inflation in progress - inflating over a stack-lock.
1344 // Some other thread is converting from stack-locked to inflated.
1345 // Only that thread can complete inflation -- other threads must wait.
1346 // The INFLATING value is transient.
1347 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1348 // We could always eliminate polling by parking the thread on some auxiliary list.
1349 if (mark == markOopDesc::INFLATING()) {
1350 ReadStableMark(object);
1351 continue;
1352 }
1353
1354 // CASE: stack-locked
1355 // Could be stack-locked either by this thread or by some other thread.
1356 //
1357 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1358 // to install INFLATING into the mark word. We originally installed INFLATING,
1359 // allocated the objectmonitor, and then finally STed the address of the
1360 // objectmonitor into the mark. This was correct, but artificially lengthened
1361 // the interval in which INFLATED appeared in the mark, thus increasing
1362 // the odds of inflation contention.
1363 //
1364 // We now use per-thread private objectmonitor free lists.
1365 // These list are reprovisioned from the global free list outside the
1366 // critical INFLATING...ST interval. A thread can transfer
1367 // multiple objectmonitors en-mass from the global free list to its local free list.
1368 // This reduces coherency traffic and lock contention on the global free list.
1369 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1370 // before or after the CAS(INFLATING) operation.
1371 // See the comments in omAlloc().
1372
1373 LogStreamHandle(Trace, monitorinflation) lsh;
1374
1375 if (mark->has_locker()) {
1376 ObjectMonitor * m = omAlloc(Self);
1377 // Optimistically prepare the objectmonitor - anticipate successful CAS
1378 // We do this before the CAS in order to minimize the length of time
1379 // in which INFLATING appears in the mark.
1380 m->Recycle();
1381 m->_Responsible = NULL;
1382 m->_recursions = 0;
1383 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1384
1385 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1386 if (cmp != mark) {
1387 omRelease(Self, m, true);
1388 continue; // Interference -- just retry
1389 }
1390
1391 // We've successfully installed INFLATING (0) into the mark-word.
1392 // This is the only case where 0 will appear in a mark-word.
1393 // Only the singular thread that successfully swings the mark-word
1394 // to 0 can perform (or more precisely, complete) inflation.
1395 //
1396 // Why do we CAS a 0 into the mark-word instead of just CASing the
1433 m->set_object(object);
1434 // TODO-FIXME: assert BasicLock->dhw != 0.
1435
1436 // Must preserve store ordering. The monitor state must
1437 // be stable at the time of publishing the monitor address.
1438 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1439 object->release_set_mark(markOopDesc::encode(m));
1440
1441 // Hopefully the performance counters are allocated on distinct cache lines
1442 // to avoid false sharing on MP systems ...
1443 OM_PERFDATA_OP(Inflations, inc());
1444 if (log_is_enabled(Trace, monitorinflation)) {
1445 ResourceMark rm(Self);
1446 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1447 INTPTR_FORMAT ", type='%s'", p2i(object),
1448 p2i(object->mark()), object->klass()->external_name());
1449 }
1450 if (event.should_commit()) {
1451 post_monitor_inflate_event(&event, object, cause);
1452 }
1453 return m;
1454 }
1455
1456 // CASE: neutral
1457 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1458 // If we know we're inflating for entry it's better to inflate by swinging a
1459 // pre-locked objectMonitor pointer into the object header. A successful
1460 // CAS inflates the object *and* confers ownership to the inflating thread.
1461 // In the current implementation we use a 2-step mechanism where we CAS()
1462 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1463 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1464 // would be useful.
1465
1466 assert(mark->is_neutral(), "invariant");
1467 ObjectMonitor * m = omAlloc(Self);
1468 // prepare m for installation - set monitor to initial state
1469 m->Recycle();
1470 m->set_header(mark);
1471 m->set_owner(NULL);
1472 m->set_object(object);
1473 m->_recursions = 0;
1474 m->_Responsible = NULL;
1475 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1476
1477 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1478 m->set_header(NULL);
1479 m->set_object(NULL);
1480 m->Recycle();
1481 omRelease(Self, m, true);
1482 m = NULL;
1483 continue;
1484 // interference - the markword changed - just retry.
1485 // The state-transitions are one-way, so there's no chance of
1486 // live-lock -- "Inflated" is an absorbing state.
1487 }
1488
1489 // Hopefully the performance counters are allocated on distinct
1490 // cache lines to avoid false sharing on MP systems ...
1491 OM_PERFDATA_OP(Inflations, inc());
1492 if (log_is_enabled(Trace, monitorinflation)) {
1493 ResourceMark rm(Self);
1494 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1495 INTPTR_FORMAT ", type='%s'", p2i(object),
1496 p2i(object->mark()), object->klass()->external_name());
1497 }
1498 if (event.should_commit()) {
1499 post_monitor_inflate_event(&event, object, cause);
1500 }
1501 return m;
1502 }
1503 }
1504
1505
1506 // We create a list of in-use monitors for each thread.
1507 //
1508 // deflate_thread_local_monitors() scans a single thread's in-use list, while
1509 // deflate_idle_monitors() scans only a global list of in-use monitors which
1510 // is populated only as a thread dies (see omFlush()).
1511 //
1512 // These operations are called at all safepoints, immediately after mutators
1513 // are stopped, but before any objects have moved. Collectively they traverse
1514 // the population of in-use monitors, deflating where possible. The scavenged
1515 // monitors are returned to the monitor free list.
1516 //
1517 // Beware that we scavenge at *every* stop-the-world point. Having a large
1518 // number of monitors in-use could negatively impact performance. We also want
1519 // to minimize the total # of monitors in circulation, as they incur a small
1520 // footprint penalty.
1521 //
1522 // Perversely, the heap size -- and thus the STW safepoint rate --
1523 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1524 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1525 // This is an unfortunate aspect of this design.
1526
1527 // Deflate a single monitor if not in-use
1528 // Return true if deflated, false if in-use
1529 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1530 ObjectMonitor** freeHeadp,
1531 ObjectMonitor** freeTailp) {
1532 bool deflated;
1533 // Normal case ... The monitor is associated with obj.
1534 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1535 guarantee(mid == obj->mark()->monitor(), "invariant");
1536 guarantee(mid->header()->is_neutral(), "invariant");
1537
1538 if (mid->is_busy()) {
1539 deflated = false;
1540 } else {
1541 // Deflate the monitor if it is no longer being used
1542 // It's idle - scavenge and return to the global free list
1543 // plain old deflation ...
1544 if (log_is_enabled(Trace, monitorinflation)) {
1545 ResourceMark rm;
1546 log_trace(monitorinflation)("deflate_monitor: "
1547 "object=" INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", type='%s'",
1548 p2i(obj), p2i(obj->mark()),
1549 obj->klass()->external_name());
1550 }
1551
1552 // Restore the header back to obj
1553 obj->release_set_mark(mid->header());
1554 mid->clear();
1555
1556 assert(mid->object() == NULL, "invariant");
1557
1558 // Move the object to the working free list defined by freeHeadp, freeTailp
1559 if (*freeHeadp == NULL) *freeHeadp = mid;
1560 if (*freeTailp != NULL) {
1561 ObjectMonitor * prevtail = *freeTailp;
1562 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1563 prevtail->FreeNext = mid;
1564 }
1565 *freeTailp = mid;
1566 deflated = true;
1567 }
1568 return deflated;
1569 }
1570
1571 // Walk a given monitor list, and deflate idle monitors
1572 // The given list could be a per-thread list or a global list
1573 // Caller acquires gListLock as needed.
1574 //
1575 // In the case of parallel processing of thread local monitor lists,
1576 // work is done by Threads::parallel_threads_do() which ensures that
1577 // each Java thread is processed by exactly one worker thread, and
1578 // thus avoid conflicts that would arise when worker threads would
1579 // process the same monitor lists concurrently.
1580 //
1581 // See also ParallelSPCleanupTask and
1582 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1583 // Threads::parallel_java_threads_do() in thread.cpp.
1584 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1585 ObjectMonitor** freeHeadp,
1586 ObjectMonitor** freeTailp) {
1587 ObjectMonitor* mid;
1588 ObjectMonitor* next;
1589 ObjectMonitor* cur_mid_in_use = NULL;
1590 int deflated_count = 0;
1594 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1595 // if deflate_monitor succeeded,
1596 // extract from per-thread in-use list
1597 if (mid == *listHeadp) {
1598 *listHeadp = mid->FreeNext;
1599 } else if (cur_mid_in_use != NULL) {
1600 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1601 }
1602 next = mid->FreeNext;
1603 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1604 mid = next;
1605 deflated_count++;
1606 } else {
1607 cur_mid_in_use = mid;
1608 mid = mid->FreeNext;
1609 }
1610 }
1611 return deflated_count;
1612 }
1613
1614 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1615 counters->nInuse = 0; // currently associated with objects
1616 counters->nInCirculation = 0; // extant
1617 counters->nScavenged = 0; // reclaimed (global and per-thread)
1618 counters->perThreadScavenged = 0; // per-thread scavenge total
1619 counters->perThreadTimes = 0.0; // per-thread scavenge times
1620 }
1621
1622 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
1623 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1624 bool deflated = false;
1625
1626 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1627 ObjectMonitor * freeTailp = NULL;
1628 elapsedTimer timer;
1629
1630 if (log_is_enabled(Info, monitorinflation)) {
1631 timer.start();
1632 }
1633
1634 // Prevent omFlush from changing mids in Thread dtor's during deflation
1635 // And in case the vm thread is acquiring a lock during a safepoint
1636 // See e.g. 6320749
1637 Thread::muxAcquire(&gListLock, "deflate_idle_monitors");
1638
1639 // Note: the thread-local monitors lists get deflated in
1640 // a separate pass. See deflate_thread_local_monitors().
1641
1642 // For moribund threads, scan gOmInUseList
1656 // constant-time list splice - prepend scavenged segment to gFreeList
1657 freeTailp->FreeNext = gFreeList;
1658 gFreeList = freeHeadp;
1659 }
1660 Thread::muxRelease(&gListLock);
1661 timer.stop();
1662
1663 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1664 LogStreamHandle(Info, monitorinflation) lsh_info;
1665 LogStream * ls = NULL;
1666 if (log_is_enabled(Debug, monitorinflation)) {
1667 ls = &lsh_debug;
1668 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
1669 ls = &lsh_info;
1670 }
1671 if (ls != NULL) {
1672 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
1673 }
1674 }
1675
1676 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1677 // Report the cumulative time for deflating each thread's idle
1678 // monitors. Note: if the work is split among more than one
1679 // worker thread, then the reported time will likely be more
1680 // than a beginning to end measurement of the phase.
1681 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->perThreadTimes, counters->perThreadScavenged);
1682
1683 gMonitorFreeCount += counters->nScavenged;
1684
1685 if (log_is_enabled(Debug, monitorinflation)) {
1686 // exit_globals()'s call to audit_and_print_stats() is done
1687 // at the Info level.
1688 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */);
1689 } else if (log_is_enabled(Info, monitorinflation)) {
1690 Thread::muxAcquire(&gListLock, "finish_deflate_idle_monitors");
1691 log_info(monitorinflation)("gMonitorPopulation=%d, gOmInUseCount=%d, "
1692 "gMonitorFreeCount=%d", gMonitorPopulation,
1693 gOmInUseCount, gMonitorFreeCount);
1694 Thread::muxRelease(&gListLock);
1695 }
1696
1697 ForceMonitorScavenge = 0; // Reset
1698
1699 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
1700 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
1701
1702 GVars.stwRandom = os::random();
1703 GVars.stwCycle++;
1704 }
1705
1706 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
1707 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1708
1709 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1710 ObjectMonitor * freeTailp = NULL;
1711 elapsedTimer timer;
1712
1713 if (log_is_enabled(Info, safepoint, cleanup) ||
1714 log_is_enabled(Info, monitorinflation)) {
1715 timer.start();
1716 }
1717
1718 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
1719
1720 Thread::muxAcquire(&gListLock, "deflate_thread_local_monitors(1)");
1721
1722 // Adjust counters
1723 counters->nInCirculation += thread->omInUseCount;
1724 thread->omInUseCount -= deflated_count;
1725 counters->nScavenged += deflated_count;
1726 counters->nInuse += thread->omInUseCount;
1727 counters->perThreadScavenged += deflated_count;
1728
1900 } else {
1901 log_error(monitorinflation)("found monitor list errors: error_cnt=%d", error_cnt);
1902 }
1903
1904 if ((on_exit && log_is_enabled(Info, monitorinflation)) ||
1905 (!on_exit && log_is_enabled(Trace, monitorinflation))) {
1906 // When exiting this log output is at the Info level. When called
1907 // at a safepoint, this log output is at the Trace level since
1908 // there can be a lot of it.
1909 log_in_use_monitor_details(ls, on_exit);
1910 }
1911
1912 ls->flush();
1913
1914 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
1915 }
1916
1917 // Check a free monitor entry; log any errors.
1918 void ObjectSynchronizer::chk_free_entry(JavaThread * jt, ObjectMonitor * n,
1919 outputStream * out, int *error_cnt_p) {
1920 if (n->is_busy()) {
1921 if (jt != NULL) {
1922 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1923 ": free per-thread monitor must not be busy.", p2i(jt),
1924 p2i(n));
1925 } else {
1926 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1927 "must not be busy.", p2i(n));
1928 }
1929 *error_cnt_p = *error_cnt_p + 1;
1930 }
1931 if (n->header() != NULL) {
1932 if (jt != NULL) {
1933 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1934 ": free per-thread monitor must have NULL _header "
1935 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n),
1936 p2i(n->header()));
1937 } else {
1938 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1939 "must have NULL _header field: _header=" INTPTR_FORMAT,
1940 p2i(n), p2i(n->header()));
2092 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": omInUseCount=%d is not "
2093 "equal to chkOmInUseCount=%d", p2i(jt), jt->omInUseCount,
2094 chkOmInUseCount);
2095 *error_cnt_p = *error_cnt_p + 1;
2096 }
2097 }
2098
2099 // Log details about ObjectMonitors on the in-use lists. The 'BHL'
2100 // flags indicate why the entry is in-use, 'object' and 'object type'
2101 // indicate the associated object and its type.
2102 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out,
2103 bool on_exit) {
2104 if (!on_exit) {
2105 // Not at VM exit so grab the global list lock.
2106 Thread::muxAcquire(&gListLock, "log_in_use_monitor_details");
2107 }
2108
2109 if (gOmInUseCount > 0) {
2110 out->print_cr("In-use global monitor info:");
2111 out->print_cr("(B -> is_busy, H -> has hashcode, L -> lock status)");
2112 out->print_cr("%18s %s %18s %18s",
2113 "monitor", "BHL", "object", "object type");
2114 out->print_cr("================== === ================== ==================");
2115 for (ObjectMonitor * n = gOmInUseList; n != NULL; n = n->FreeNext) {
2116 const oop obj = (oop) n->object();
2117 const markOop mark = n->header();
2118 ResourceMark rm;
2119 out->print_cr(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(n),
2120 n->is_busy() != 0, mark->hash() != 0, n->owner() != NULL,
2121 p2i(obj), obj->klass()->external_name());
2122 }
2123 }
2124
2125 if (!on_exit) {
2126 Thread::muxRelease(&gListLock);
2127 }
2128
2129 out->print_cr("In-use per-thread monitor info:");
2130 out->print_cr("(B -> is_busy, H -> has hashcode, L -> lock status)");
2131 out->print_cr("%18s %18s %s %18s %18s",
2132 "jt", "monitor", "BHL", "object", "object type");
2133 out->print_cr("================== ================== === ================== ==================");
2134 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2135 for (ObjectMonitor * n = jt->omInUseList; n != NULL; n = n->FreeNext) {
2136 const oop obj = (oop) n->object();
2137 const markOop mark = n->header();
2138 ResourceMark rm;
2139 out->print_cr(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT
2140 " %s", p2i(jt), p2i(n), n->is_busy() != 0,
2141 mark->hash() != 0, n->owner() != NULL, p2i(obj),
2142 obj->klass()->external_name());
2143 }
2144 }
2145
2146 out->flush();
2147 }
2148
2149 // Log counts for the global and per-thread monitor lists and return
2150 // the population count.
2151 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) {
2152 int popCount = 0;
2153 out->print_cr("%18s %10s %10s %10s",
2154 "Global Lists:", "InUse", "Free", "Total");
2155 out->print_cr("================== ========== ========== ==========");
2156 out->print_cr("%18s %10d %10d %10d", "",
2157 gOmInUseCount, gMonitorFreeCount, gMonitorPopulation);
2158 popCount += gOmInUseCount + gMonitorFreeCount;
2159
2160 out->print_cr("%18s %10s %10s %10s",
2161 "Per-Thread Lists:", "InUse", "Free", "Provision");
2162 out->print_cr("================== ========== ========== ==========");
|
107 #endif // ndef DTRACE_ENABLED
108
109 // This exists only as a workaround of dtrace bug 6254741
110 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
111 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
112 return 0;
113 }
114
115 #define NINFLATIONLOCKS 256
116 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
117
118 // global list of blocks of monitors
119 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
120 // global monitor free list
121 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
122 // global monitor in-use list, for moribund threads,
123 // monitors they inflated need to be scanned for deflation
124 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
125 // count of entries in gOmInUseList
126 int ObjectSynchronizer::gOmInUseCount = 0;
127 bool ObjectSynchronizer::_gOmShouldDeflateIdleMonitors = false;
128 bool volatile ObjectSynchronizer::_is_cleanup_requested = false;
129
130 static volatile intptr_t gListLock = 0; // protects global monitor lists
131 static volatile int gMonitorFreeCount = 0; // # on gFreeList
132 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
133
134 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
135
136
137 // =====================> Quick functions
138
139 // The quick_* forms are special fast-path variants used to improve
140 // performance. In the simplest case, a "quick_*" implementation could
141 // simply return false, in which case the caller will perform the necessary
142 // state transitions and call the slow-path form.
143 // The fast-path is designed to handle frequently arising cases in an efficient
144 // manner and is just a degenerate "optimistic" variant of the slow-path.
145 // returns true -- to indicate the call was satisfied.
146 // returns false -- to indicate the call needs the services of the slow-path.
147 // A no-loitering ordinance is in effect for code in the quick_* family
148 // operators: safepoints or indefinite blocking (blocking that might span a
195 }
196
197 // biased locking and any other IMS exception states take the slow-path
198 return false;
199 }
200
201
202 // The LockNode emitted directly at the synchronization site would have
203 // been too big if it were to have included support for the cases of inflated
204 // recursive enter and exit, so they go here instead.
205 // Note that we can't safely call AsyncPrintJavaStack() from within
206 // quick_enter() as our thread state remains _in_Java.
207
208 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
209 BasicLock * lock) {
210 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
211 assert(Self->is_Java_thread(), "invariant");
212 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
213 NoSafepointVerifier nsv;
214 if (obj == NULL) return false; // Need to throw NPE
215
216 while (true) {
217 const markOop mark = obj->mark();
218
219 if (mark->has_monitor()) {
220 ObjectMonitorHandle omh;
221 if (!omh.save_om_ptr(obj, mark)) {
222 // Lost a race with async deflation so try again.
223 assert(AsyncDeflateIdleMonitors, "sanity check");
224 continue;
225 }
226 ObjectMonitor * const m = omh.om_ptr();
227 assert(oopDesc::equals((oop) m->object(), obj), "invariant");
228 Thread * const owner = (Thread *) m->_owner;
229
230 // Lock contention and Transactional Lock Elision (TLE) diagnostics
231 // and observability
232 // Case: light contention possibly amenable to TLE
233 // Case: TLE inimical operations such as nested/recursive synchronization
234
235 if (owner == Self) {
236 m->_recursions++;
237 return true;
238 }
239
240 // This Java Monitor is inflated so obj's header will never be
241 // displaced to this thread's BasicLock. Make the displaced header
242 // non-NULL so this BasicLock is not seen as recursive nor as
243 // being locked. We do this unconditionally so that this thread's
244 // BasicLock cannot be mis-interpreted by any stack walkers. For
245 // performance reasons, stack walkers generally first check for
246 // Biased Locking in the object's header, the second check is for
247 // stack-locking in the object's header, the third check is for
248 // recursive stack-locking in the displaced header in the BasicLock,
249 // and last are the inflated Java Monitor (ObjectMonitor) checks.
250 lock->set_displaced_header(markOopDesc::unused_mark());
251
252 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
253 assert(m->_recursions == 0, "invariant");
254 assert(m->_owner == Self, "invariant");
255 return true;
256 }
257 }
258 break;
259 }
260
261 // Note that we could inflate in quick_enter.
262 // This is likely a useful optimization
263 // Critically, in quick_enter() we must not:
264 // -- perform bias revocation, or
265 // -- block indefinitely, or
266 // -- reach a safepoint
267
268 return false; // revert to slow-path
269 }
270
271 // -----------------------------------------------------------------------------
272 // Fast Monitor Enter/Exit
273 // This the fast monitor enter. The interpreter and compiler use
274 // some assembly copies of this code. Make sure update those code
275 // if the following function is changed. The implementation is
276 // extremely sensitive to race condition. Be careful.
277
278 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
279 bool attempt_rebias, TRAPS) {
322 // does not own the Java Monitor.
323 ObjectMonitor * m = mark->monitor();
324 assert(((oop)(m->object()))->mark() == mark, "invariant");
325 assert(m->is_entered(THREAD), "invariant");
326 }
327 }
328 #endif
329 return;
330 }
331
332 if (mark == (markOop) lock) {
333 // If the object is stack-locked by the current thread, try to
334 // swing the displaced header from the BasicLock back to the mark.
335 assert(dhw->is_neutral(), "invariant");
336 if (object->cas_set_mark(dhw, mark) == mark) {
337 return;
338 }
339 }
340
341 // We have to take the slow-path of possible inflation and then exit.
342 ObjectMonitorHandle omh;
343 inflate(&omh, THREAD, object, inflate_cause_vm_internal);
344 omh.om_ptr()->exit(true, THREAD);
345 }
346
347 // -----------------------------------------------------------------------------
348 // Interpreter/Compiler Slow Case
349 // This routine is used to handle interpreter/compiler slow case
350 // We don't need to use fast path here, because it must have been
351 // failed in the interpreter/compiler code.
352 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
353 bool do_loop = true;
354 while (do_loop) {
355 markOop mark = obj->mark();
356 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
357
358 if (mark->is_neutral()) {
359 // Anticipate successful CAS -- the ST of the displaced mark must
360 // be visible <= the ST performed by the CAS.
361 lock->set_displaced_header(mark);
362 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
363 return;
364 }
365 // Fall through to inflate() ...
366 } else if (mark->has_locker() &&
367 THREAD->is_lock_owned((address)mark->locker())) {
368 assert(lock != mark->locker(), "must not re-lock the same lock");
369 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
370 lock->set_displaced_header(NULL);
371 return;
372 }
373
374 // The object header will never be displaced to this lock,
375 // so it does not matter what the value is, except that it
376 // must be non-zero to avoid looking like a re-entrant lock,
377 // and must not look locked either.
378 lock->set_displaced_header(markOopDesc::unused_mark());
379 ObjectMonitorHandle omh;
380 inflate(&omh, THREAD, obj(), inflate_cause_monitor_enter);
381 do_loop = !omh.om_ptr()->enter(THREAD);
382 }
383 }
384
385 // This routine is used to handle interpreter/compiler slow case
386 // We don't need to use fast path here, because it must have
387 // failed in the interpreter/compiler code. Simply use the heavy
388 // weight monitor should be ok, unless someone find otherwise.
389 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
390 fast_exit(object, lock, THREAD);
391 }
392
393 // -----------------------------------------------------------------------------
394 // Class Loader support to workaround deadlocks on the class loader lock objects
395 // Also used by GC
396 // complete_exit()/reenter() are used to wait on a nested lock
397 // i.e. to give up an outer lock completely and then re-enter
398 // Used when holding nested locks - lock acquisition order: lock1 then lock2
399 // 1) complete_exit lock1 - saving recursion count
400 // 2) wait on lock2
401 // 3) when notified on lock2, unlock lock2
402 // 4) reenter lock1 with original recursion count
403 // 5) lock lock2
404 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
405 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
406 if (UseBiasedLocking) {
407 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
408 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
409 }
410
411 ObjectMonitorHandle omh;
412 inflate(&omh, THREAD, obj(), inflate_cause_vm_internal);
413 intptr_t ret_code = omh.om_ptr()->complete_exit(THREAD);
414 return ret_code;
415 }
416
417 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
418 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
419 if (UseBiasedLocking) {
420 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
421 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
422 }
423
424 bool do_loop = true;
425 while (do_loop) {
426 ObjectMonitorHandle omh;
427 inflate(&omh, THREAD, obj(), inflate_cause_vm_internal);
428 do_loop = !omh.om_ptr()->reenter(recursion, THREAD);
429 }
430 }
431 // -----------------------------------------------------------------------------
432 // JNI locks on java objects
433 // NOTE: must use heavy weight monitor to handle jni monitor enter
434 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
435 // the current locking is from JNI instead of Java code
436 if (UseBiasedLocking) {
437 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
438 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
439 }
440 THREAD->set_current_pending_monitor_is_from_java(false);
441 bool do_loop = true;
442 while (do_loop) {
443 ObjectMonitorHandle omh;
444 inflate(&omh, THREAD, obj(), inflate_cause_jni_enter);
445 do_loop = !omh.om_ptr()->enter(THREAD);
446 }
447 THREAD->set_current_pending_monitor_is_from_java(true);
448 }
449
450 // NOTE: must use heavy weight monitor to handle jni monitor exit
451 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
452 if (UseBiasedLocking) {
453 Handle h_obj(THREAD, obj);
454 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
455 obj = h_obj();
456 }
457 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
458
459 ObjectMonitorHandle omh;
460 inflate(&omh, THREAD, obj, inflate_cause_jni_exit);
461 ObjectMonitor * monitor = omh.om_ptr();
462 // If this thread has locked the object, exit the monitor. Note: can't use
463 // monitor->check(CHECK); must exit even if an exception is pending.
464 if (monitor->check(THREAD)) {
465 monitor->exit(true, THREAD);
466 }
467 }
468
469 // -----------------------------------------------------------------------------
470 // Internal VM locks on java objects
471 // standard constructor, allows locking failures
472 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
473 _dolock = doLock;
474 _thread = thread;
475 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
476 _obj = obj;
477
478 if (_dolock) {
479 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
480 }
481 }
482
483 ObjectLocker::~ObjectLocker() {
484 if (_dolock) {
485 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
486 }
487 }
488
489
490 // -----------------------------------------------------------------------------
491 // Wait/Notify/NotifyAll
492 // NOTE: must use heavy weight monitor to handle wait()
493 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
494 if (UseBiasedLocking) {
495 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
496 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
497 }
498 if (millis < 0) {
499 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
500 }
501 ObjectMonitorHandle omh;
502 inflate(&omh, THREAD, obj(), inflate_cause_wait);
503 ObjectMonitor * monitor = omh.om_ptr();
504
505 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
506 monitor->wait(millis, true, THREAD);
507
508 // This dummy call is in place to get around dtrace bug 6254741. Once
509 // that's fixed we can uncomment the following line, remove the call
510 // and change this function back into a "void" func.
511 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
512 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
513 return ret_code;
514 }
515
516 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
517 if (UseBiasedLocking) {
518 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
519 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
520 }
521 if (millis < 0) {
522 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
523 }
524 ObjectMonitorHandle omh;
525 inflate(&omh, THREAD, obj(), inflate_cause_wait);
526 omh.om_ptr()->wait(millis, false, THREAD);
527 }
528
529 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
530 if (UseBiasedLocking) {
531 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
532 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
533 }
534
535 markOop mark = obj->mark();
536 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
537 return;
538 }
539 ObjectMonitorHandle omh;
540 inflate(&omh, THREAD, obj(), inflate_cause_notify);
541 omh.om_ptr()->notify(THREAD);
542 }
543
544 // NOTE: see comment of notify()
545 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
546 if (UseBiasedLocking) {
547 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
548 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
549 }
550
551 markOop mark = obj->mark();
552 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
553 return;
554 }
555 ObjectMonitorHandle omh;
556 inflate(&omh, THREAD, obj(), inflate_cause_notify);
557 omh.om_ptr()->notifyAll(THREAD);
558 }
559
560 // -----------------------------------------------------------------------------
561 // Hash Code handling
562 //
563 // Performance concern:
564 // OrderAccess::storestore() calls release() which at one time stored 0
565 // into the global volatile OrderAccess::dummy variable. This store was
566 // unnecessary for correctness. Many threads storing into a common location
567 // causes considerable cache migration or "sloshing" on large SMP systems.
568 // As such, I avoided using OrderAccess::storestore(). In some cases
569 // OrderAccess::fence() -- which incurs local latency on the executing
570 // processor -- is a better choice as it scales on SMP systems.
571 //
572 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
573 // a discussion of coherency costs. Note that all our current reference
574 // platforms provide strong ST-ST order, so the issue is moot on IA32,
575 // x64, and SPARC.
576 //
577 // As a general policy we use "volatile" to control compiler-based reordering
731 Handle hobj(Self, obj);
732 // Relaxing assertion for bug 6320749.
733 assert(Universe::verify_in_progress() ||
734 !SafepointSynchronize::is_at_safepoint(),
735 "biases should not be seen by VM thread here");
736 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
737 obj = hobj();
738 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
739 }
740 }
741
742 // hashCode() is a heap mutator ...
743 // Relaxing assertion for bug 6320749.
744 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
745 !SafepointSynchronize::is_at_safepoint(), "invariant");
746 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
747 Self->is_Java_thread() , "invariant");
748 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
749 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
750
751 Retry:
752 ObjectMonitor* monitor = NULL;
753 markOop temp, test;
754 intptr_t hash;
755 markOop mark = ReadStableMark(obj);
756
757 // object should remain ineligible for biased locking
758 assert(!mark->has_bias_pattern(), "invariant");
759
760 if (mark->is_neutral()) {
761 hash = mark->hash(); // this is a normal header
762 if (hash != 0) { // if it has hash, just return it
763 return hash;
764 }
765 hash = get_next_hash(Self, obj); // allocate a new hash code
766 temp = mark->copy_set_hash(hash); // merge the hash code into header
767 // use (machine word version) atomic operation to install the hash
768 test = obj->cas_set_mark(temp, mark);
769 if (test == mark) {
770 return hash;
771 }
772 // If atomic operation failed, we must inflate the header
773 // into heavy weight monitor. We could add more code here
774 // for fast path, but it does not worth the complexity.
775 } else if (mark->has_monitor()) {
776 ObjectMonitorHandle omh;
777 if (!omh.save_om_ptr(obj, mark)) {
778 // Lost a race with async deflation so try again.
779 assert(AsyncDeflateIdleMonitors, "sanity check");
780 goto Retry;
781 }
782 monitor = omh.om_ptr();
783 temp = monitor->header();
784 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)temp));
785 hash = temp->hash();
786 if (hash != 0) {
787 return hash;
788 }
789 // Skip to the following code to reduce code size
790 } else if (Self->is_lock_owned((address)mark->locker())) {
791 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
792 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)temp));
793 hash = temp->hash(); // by current thread, check if the displaced
794 if (hash != 0) { // header contains hash code
795 return hash;
796 }
797 // WARNING:
798 // The displaced header is strictly immutable.
799 // It can NOT be changed in ANY cases. So we have
800 // to inflate the header into heavyweight monitor
801 // even the current thread owns the lock. The reason
802 // is the BasicLock (stack slot) will be asynchronously
803 // read by other threads during the inflate() function.
804 // Any change to stack may not propagate to other threads
805 // correctly.
806 }
807
808 // Inflate the monitor to set hash code
809 ObjectMonitorHandle omh;
810 inflate(&omh, Self, obj, inflate_cause_hash_code);
811 monitor = omh.om_ptr();
812 // Load displaced header and check it has hash code
813 mark = monitor->header();
814 assert(mark->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)mark));
815 hash = mark->hash();
816 if (hash == 0) {
817 hash = get_next_hash(Self, obj);
818 temp = mark->copy_set_hash(hash); // merge hash code into header
819 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)temp));
820 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark);
821 if (test != mark) {
822 // The only update to the header in the monitor (outside GC)
823 // is install the hash code. If someone add new usage of
824 // displaced header, please update this code
825 hash = test->hash();
826 assert(test->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)test));
827 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
828 }
829 }
830 // We finally get the hash
831 return hash;
832 }
833
834 // Deprecated -- use FastHashCode() instead.
835
836 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
837 return FastHashCode(Thread::current(), obj());
838 }
839
840
841 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
842 Handle h_obj) {
843 if (UseBiasedLocking) {
844 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
845 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
846 }
847
848 assert(thread == JavaThread::current(), "Can only be called on current thread");
849 oop obj = h_obj();
850
851 while (true) {
852 markOop mark = ReadStableMark(obj);
853
854 // Uncontended case, header points to stack
855 if (mark->has_locker()) {
856 return thread->is_lock_owned((address)mark->locker());
857 }
858 // Contended case, header points to ObjectMonitor (tagged pointer)
859 if (mark->has_monitor()) {
860 ObjectMonitorHandle omh;
861 if (!omh.save_om_ptr(obj, mark)) {
862 // Lost a race with async deflation so try again.
863 assert(AsyncDeflateIdleMonitors, "sanity check");
864 continue;
865 }
866 bool ret_code = omh.om_ptr()->is_entered(thread) != 0;
867 return ret_code;
868 }
869 // Unlocked case, header in place
870 assert(mark->is_neutral(), "sanity check");
871 return false;
872 }
873 }
874
875 // Be aware of this method could revoke bias of the lock object.
876 // This method queries the ownership of the lock handle specified by 'h_obj'.
877 // If the current thread owns the lock, it returns owner_self. If no
878 // thread owns the lock, it returns owner_none. Otherwise, it will return
879 // owner_other.
880 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
881 (JavaThread *self, Handle h_obj) {
882 // The caller must beware this method can revoke bias, and
883 // revocation can result in a safepoint.
884 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
885 assert(self->thread_state() != _thread_blocked, "invariant");
886
887 // Possible mark states: neutral, biased, stack-locked, inflated
888
889 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
890 // CASE: biased
891 BiasedLocking::revoke_and_rebias(h_obj, false, self);
892 assert(!h_obj->mark()->has_bias_pattern(),
893 "biases should be revoked by now");
894 }
895
896 assert(self == JavaThread::current(), "Can only be called on current thread");
897 oop obj = h_obj();
898
899 while (true) {
900 markOop mark = ReadStableMark(obj);
901
902 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
903 if (mark->has_locker()) {
904 return self->is_lock_owned((address)mark->locker()) ?
905 owner_self : owner_other;
906 }
907
908 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor.
909 // The Object:ObjectMonitor relationship is stable as long as we're
910 // not at a safepoint and AsyncDeflateIdleMonitors is false.
911 if (mark->has_monitor()) {
912 ObjectMonitorHandle omh;
913 if (!omh.save_om_ptr(obj, mark)) {
914 // Lost a race with async deflation so try again.
915 assert(AsyncDeflateIdleMonitors, "sanity check");
916 continue;
917 }
918 ObjectMonitor * monitor = omh.om_ptr();
919 void * owner = monitor->_owner;
920 if (owner == NULL) return owner_none;
921 return (owner == self ||
922 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
923 }
924
925 // CASE: neutral
926 assert(mark->is_neutral(), "sanity check");
927 return owner_none; // it's unlocked
928 }
929 }
930
931 // FIXME: jvmti should call this
932 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
933 if (UseBiasedLocking) {
934 if (SafepointSynchronize::is_at_safepoint()) {
935 BiasedLocking::revoke_at_safepoint(h_obj);
936 } else {
937 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
938 }
939 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
940 }
941
942 oop obj = h_obj();
943
944 while (true) {
945 address owner = NULL;
946 markOop mark = ReadStableMark(obj);
947
948 // Uncontended case, header points to stack
949 if (mark->has_locker()) {
950 owner = (address) mark->locker();
951 }
952
953 // Contended case, header points to ObjectMonitor (tagged pointer)
954 else if (mark->has_monitor()) {
955 ObjectMonitorHandle omh;
956 if (!omh.save_om_ptr(obj, mark)) {
957 // Lost a race with async deflation so try again.
958 assert(AsyncDeflateIdleMonitors, "sanity check");
959 continue;
960 }
961 ObjectMonitor* monitor = omh.om_ptr();
962 assert(monitor != NULL, "monitor should be non-null");
963 owner = (address) monitor->owner();
964 }
965
966 if (owner != NULL) {
967 // owning_thread_from_monitor_owner() may also return NULL here
968 return Threads::owning_thread_from_monitor_owner(t_list, owner);
969 }
970
971 // Unlocked case, header in place
972 // Cannot have assertion since this object may have been
973 // locked by another thread when reaching here.
974 // assert(mark->is_neutral(), "sanity check");
975
976 return NULL;
977 }
978 }
979
980 // Visitors ...
981
982 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
983 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
984 while (block != NULL) {
985 assert(block->object() == CHAINMARKER, "must be a block header");
986 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
987 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
988 if (mid->is_active()) {
989 ObjectMonitorHandle omh(mid);
990
991 if (mid->object() == NULL ||
992 (AsyncDeflateIdleMonitors && mid->_owner == DEFLATER_MARKER)) {
993 // Only process with closure if the object is set.
994 // For async deflation, race here if monitor is not owned!
995 // The above ref_count bump (in ObjectMonitorHandle ctr)
996 // will cause subsequent async deflation to skip it.
997 // However, previous or concurrent async deflation is a race.
998 continue;
999 }
1000 closure->do_monitor(mid);
1001 }
1002 }
1003 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1004 }
1005 }
1006
1007 // Get the next block in the block list.
1008 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
1009 assert(block->object() == CHAINMARKER, "must be a block header");
1010 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
1011 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
1012 return block;
1013 }
1014
1015 static bool monitors_used_above_threshold() {
1016 if (gMonitorPopulation == 0) {
1017 return false;
1018 }
1019 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
1090 // See also: GuaranteedSafepointInterval
1091 //
1092 // The current implementation uses asynchronous VM operations.
1093
1094 static void InduceScavenge(Thread * Self, const char * Whence) {
1095 // Induce STW safepoint to trim monitors
1096 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1097 // More precisely, trigger an asynchronous STW safepoint as the number
1098 // of active monitors passes the specified threshold.
1099 // TODO: assert thread state is reasonable
1100
1101 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1102 // Induce a 'null' safepoint to scavenge monitors
1103 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1104 // to the VMthread and have a lifespan longer than that of this activation record.
1105 // The VMThread will delete the op when completed.
1106 VMThread::execute(new VM_ScavengeMonitors());
1107 }
1108 }
1109
1110 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self,
1111 const InflateCause cause) {
1112 // A large MAXPRIVATE value reduces both list lock contention
1113 // and list coherency traffic, but also tends to increase the
1114 // number of objectMonitors in circulation as well as the STW
1115 // scavenge costs. As usual, we lean toward time in space-time
1116 // tradeoffs.
1117 const int MAXPRIVATE = 1024;
1118
1119 if (AsyncDeflateIdleMonitors) {
1120 JavaThread * jt = (JavaThread *)Self;
1121 if (jt->omShouldDeflateIdleMonitors && jt->omInUseCount > 0 &&
1122 cause != inflate_cause_vm_internal) {
1123 // Deflate any per-thread idle monitors for this JavaThread if
1124 // this is not an internal inflation. Clean up your own mess.
1125 // (Gibbs Rule 45) Otherwise, skip this cleanup.
1126 // deflate_global_idle_monitors_using_JT() is called by the ServiceThread.
1127 debug_only(jt->check_for_valid_safepoint_state(false);)
1128 ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT();
1129 }
1130 }
1131
1132 for (;;) {
1133 ObjectMonitor * m;
1134
1135 // 1: try to allocate from the thread's local omFreeList.
1136 // Threads will attempt to allocate first from their local list, then
1137 // from the global list, and only after those attempts fail will the thread
1138 // attempt to instantiate new monitors. Thread-local free lists take
1139 // heat off the gListLock and improve allocation latency, as well as reducing
1140 // coherency traffic on the shared global list.
1141 m = Self->omFreeList;
1142 if (m != NULL) {
1143 Self->omFreeList = m->FreeNext;
1144 Self->omFreeCount--;
1145 guarantee(m->object() == NULL, "invariant");
1146 m->set_allocation_state(ObjectMonitor::New);
1147 m->FreeNext = Self->omInUseList;
1148 Self->omInUseList = m;
1149 Self->omInUseCount++;
1150 return m;
1151 }
1152
1153 // 2: try to allocate from the global gFreeList
1154 // CONSIDER: use muxTry() instead of muxAcquire().
1155 // If the muxTry() fails then drop immediately into case 3.
1156 // If we're using thread-local free lists then try
1157 // to reprovision the caller's free list.
1158 if (gFreeList != NULL) {
1159 // Reprovision the thread's omFreeList.
1160 // Use bulk transfers to reduce the allocation rate and heat
1161 // on various locks.
1162 Thread::muxAcquire(&gListLock, "omAlloc(1)");
1163 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1164 gMonitorFreeCount--;
1165 ObjectMonitor * take = gFreeList;
1166 gFreeList = take->FreeNext;
1167 guarantee(take->object() == NULL, "invariant");
1168 if (AsyncDeflateIdleMonitors) {
1169 take->set_owner(NULL);
1170 take->_count = 0;
1171 }
1172 guarantee(!take->is_busy(), "invariant");
1173 take->Recycle();
1174 assert(take->is_free(), "invariant");
1175 omRelease(Self, take, false);
1176 }
1177 Thread::muxRelease(&gListLock);
1178 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1179 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1180
1181 const int mx = MonitorBound;
1182 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1183 // We can't safely induce a STW safepoint from omAlloc() as our thread
1184 // state may not be appropriate for such activities and callers may hold
1185 // naked oops, so instead we defer the action.
1186 InduceScavenge(Self, "omAlloc");
1187 }
1188 continue;
1189 }
1190
1191 // 3: allocate a block of new ObjectMonitors
1192 // Both the local and global free lists are empty -- resort to malloc().
1193 // In the current implementation objectMonitors are TSM - immortal.
1194 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1207
1208 // NOTE: (almost) no way to recover if allocation failed.
1209 // We might be able to induce a STW safepoint and scavenge enough
1210 // objectMonitors to permit progress.
1211 if (temp == NULL) {
1212 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1213 "Allocate ObjectMonitors");
1214 }
1215 (void)memset((void *) temp, 0, neededsize);
1216
1217 // Format the block.
1218 // initialize the linked list, each monitor points to its next
1219 // forming the single linked free list, the very first monitor
1220 // will points to next block, which forms the block list.
1221 // The trick of using the 1st element in the block as gBlockList
1222 // linkage should be reconsidered. A better implementation would
1223 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1224
1225 for (int i = 1; i < _BLOCKSIZE; i++) {
1226 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1227 assert(temp[i].is_free(), "invariant");
1228 }
1229
1230 // terminate the last monitor as the end of list
1231 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1232
1233 // Element [0] is reserved for global list linkage
1234 temp[0].set_object(CHAINMARKER);
1235
1236 // Consider carving out this thread's current request from the
1237 // block in hand. This avoids some lock traffic and redundant
1238 // list activity.
1239
1240 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1241 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1242 Thread::muxAcquire(&gListLock, "omAlloc(2)");
1243 gMonitorPopulation += _BLOCKSIZE-1;
1244 gMonitorFreeCount += _BLOCKSIZE-1;
1245
1246 // Add the new block to the list of extant blocks (gBlockList).
1247 // The very first objectMonitor in a block is reserved and dedicated.
1250 // There are lock-free uses of gBlockList so make sure that
1251 // the previous stores happen before we update gBlockList.
1252 OrderAccess::release_store(&gBlockList, temp);
1253
1254 // Add the new string of objectMonitors to the global free list
1255 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1256 gFreeList = temp + 1;
1257 Thread::muxRelease(&gListLock);
1258 }
1259 }
1260
1261 // Place "m" on the caller's private per-thread omFreeList.
1262 // In practice there's no need to clamp or limit the number of
1263 // monitors on a thread's omFreeList as the only time we'll call
1264 // omRelease is to return a monitor to the free list after a CAS
1265 // attempt failed. This doesn't allow unbounded #s of monitors to
1266 // accumulate on a thread's free list.
1267 //
1268 // Key constraint: all ObjectMonitors on a thread's free list and the global
1269 // free list must have their object field set to null. This prevents the
1270 // scavenger -- deflate_monitor_list() or deflate_monitor_list_using_JT()
1271 // -- from reclaiming them while we are trying to release them.
1272
1273 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1274 bool fromPerThreadAlloc) {
1275 guarantee(m->header() == NULL, "invariant");
1276 guarantee(m->object() == NULL, "invariant");
1277 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor");
1278 m->set_allocation_state(ObjectMonitor::Free);
1279 // Remove from omInUseList
1280 if (fromPerThreadAlloc) {
1281 ObjectMonitor* cur_mid_in_use = NULL;
1282 bool extracted = false;
1283 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1284 if (m == mid) {
1285 // extract from per-thread in-use list
1286 if (mid == Self->omInUseList) {
1287 Self->omInUseList = mid->FreeNext;
1288 } else if (cur_mid_in_use != NULL) {
1289 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1290 }
1291 extracted = true;
1292 Self->omInUseCount--;
1293 break;
1294 }
1295 }
1296 assert(extracted, "Should have extracted from in-use list");
1297 }
1298
1299 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1300 m->FreeNext = Self->omFreeList;
1301 guarantee(m->is_free(), "invariant");
1302 Self->omFreeList = m;
1303 Self->omFreeCount++;
1304 }
1305
1306 // Return the monitors of a moribund thread's local free list to
1307 // the global free list. Typically a thread calls omFlush() when
1308 // it's dying. We could also consider having the VM thread steal
1309 // monitors from threads that have not run java code over a few
1310 // consecutive STW safepoints. Relatedly, we might decay
1311 // omFreeProvision at STW safepoints.
1312 //
1313 // Also return the monitors of a moribund thread's omInUseList to
1314 // a global gOmInUseList under the global list lock so these
1315 // will continue to be scanned.
1316 //
1317 // We currently call omFlush() from Threads::remove() _before the thread
1318 // has been excised from the thread list and is no longer a mutator.
1319 // This means that omFlush() cannot run concurrently with a safepoint and
1320 // interleave with the deflate_idle_monitors scavenge operator. In particular,
1321 // this ensures that the thread's monitors are scanned by a GC safepoint,
1322 // either via Thread::oops_do() (if safepoint happens before omFlush()) or via
1323 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1324 // monitors have been transferred to the global in-use list).
1325 //
1326 // With AsyncDeflateIdleMonitors, deflate_global_idle_monitors_using_JT()
1327 // and deflate_per_thread_idle_monitors_using_JT() (in another thread) can
1328 // run at the same time as omFlush() so we have to be careful.
1329
1330 void ObjectSynchronizer::omFlush(Thread * Self) {
1331 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1332 ObjectMonitor * tail = NULL;
1333 int tally = 0;
1334 if (list != NULL) {
1335 ObjectMonitor * s;
1336 // The thread is going away, the per-thread free monitors
1337 // are freed via set_owner(NULL)
1338 // Link them to tail, which will be linked into the global free list
1339 // gFreeList below, under the gListLock
1340 for (s = list; s != NULL; s = s->FreeNext) {
1341 tally++;
1342 tail = s;
1343 guarantee(s->object() == NULL, "invariant");
1344 guarantee(!s->is_busy(), "invariant");
1345 s->set_owner(NULL); // redundant but good hygiene
1346 }
1347 guarantee(tail != NULL, "invariant");
1348 guarantee(Self->omFreeCount == tally, "free-count off");
1349 Self->omFreeList = NULL;
1350 Self->omFreeCount = 0;
1351 }
1352
1353 ObjectMonitor * inUseList = Self->omInUseList;
1354 ObjectMonitor * inUseTail = NULL;
1355 int inUseTally = 0;
1356 if (inUseList != NULL) {
1357 ObjectMonitor *cur_om;
1358 // The thread is going away, however the omInUseList inflated
1359 // monitors may still be in-use by other threads.
1360 // Link them to inUseTail, which will be linked into the global in-use list
1361 // gOmInUseList below, under the gListLock
1362 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1363 inUseTail = cur_om;
1364 inUseTally++;
1365 guarantee(cur_om->is_active(), "invariant");
1366 }
1367 guarantee(inUseTail != NULL, "invariant");
1368 guarantee(Self->omInUseCount == inUseTally, "in-use count off");
1369 Self->omInUseList = NULL;
1370 Self->omInUseCount = 0;
1371 }
1372
1373 Thread::muxAcquire(&gListLock, "omFlush");
1374 if (tail != NULL) {
1375 tail->FreeNext = gFreeList;
1376 gFreeList = list;
1377 gMonitorFreeCount += tally;
1378 }
1379
1380 if (inUseTail != NULL) {
1381 inUseTail->FreeNext = gOmInUseList;
1382 gOmInUseList = inUseList;
1383 gOmInUseCount += inUseTally;
1384 }
1385
1386 Thread::muxRelease(&gListLock);
1387
1388 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1396 }
1397 if (ls != NULL) {
1398 ls->print_cr("omFlush: jt=" INTPTR_FORMAT ", free_monitor_tally=%d"
1399 ", in_use_monitor_tally=%d" ", omFreeProvision=%d",
1400 p2i(Self), tally, inUseTally, Self->omFreeProvision);
1401 }
1402 }
1403
1404 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1405 const oop obj,
1406 ObjectSynchronizer::InflateCause cause) {
1407 assert(event != NULL, "invariant");
1408 assert(event->should_commit(), "invariant");
1409 event->set_monitorClass(obj->klass());
1410 event->set_address((uintptr_t)(void*)obj);
1411 event->set_cause((u1)cause);
1412 event->commit();
1413 }
1414
1415 // Fast path code shared by multiple functions
1416 void ObjectSynchronizer::inflate_helper(ObjectMonitorHandle * omh_p, oop obj) {
1417 while (true) {
1418 markOop mark = obj->mark();
1419 if (mark->has_monitor()) {
1420 if (!omh_p->save_om_ptr(obj, mark)) {
1421 // Lost a race with async deflation so try again.
1422 assert(AsyncDeflateIdleMonitors, "sanity check");
1423 continue;
1424 }
1425 ObjectMonitor * monitor = omh_p->om_ptr();
1426 assert(ObjectSynchronizer::verify_objmon_isinpool(monitor), "monitor is invalid");
1427 markOop dmw = monitor->header();
1428 assert(dmw->is_neutral(), "sanity check: header=" INTPTR_FORMAT, p2i((address)dmw));
1429 return;
1430 }
1431 inflate(omh_p, Thread::current(), obj, inflate_cause_vm_internal);
1432 return;
1433 }
1434 }
1435
1436 void ObjectSynchronizer::inflate(ObjectMonitorHandle * omh_p, Thread * Self,
1437 oop object, const InflateCause cause) {
1438 // Inflate mutates the heap ...
1439 // Relaxing assertion for bug 6320749.
1440 assert(Universe::verify_in_progress() ||
1441 !SafepointSynchronize::is_at_safepoint(), "invariant");
1442
1443 EventJavaMonitorInflate event;
1444
1445 for (;;) {
1446 const markOop mark = object->mark();
1447 assert(!mark->has_bias_pattern(), "invariant");
1448
1449 // The mark can be in one of the following states:
1450 // * Inflated - just return
1451 // * Stack-locked - coerce it to inflated
1452 // * INFLATING - busy wait for conversion to complete
1453 // * Neutral - aggressively inflate the object.
1454 // * BIASED - Illegal. We should never see this
1455
1456 // CASE: inflated
1457 if (mark->has_monitor()) {
1458 if (!omh_p->save_om_ptr(object, mark)) {
1459 // Lost a race with async deflation so try again.
1460 assert(AsyncDeflateIdleMonitors, "sanity check");
1461 continue;
1462 }
1463 ObjectMonitor * inf = omh_p->om_ptr();
1464 markOop dmw = inf->header();
1465 assert(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i((address)dmw));
1466 assert(oopDesc::equals((oop) inf->object(), object), "invariant");
1467 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1468 return;
1469 }
1470
1471 // CASE: inflation in progress - inflating over a stack-lock.
1472 // Some other thread is converting from stack-locked to inflated.
1473 // Only that thread can complete inflation -- other threads must wait.
1474 // The INFLATING value is transient.
1475 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1476 // We could always eliminate polling by parking the thread on some auxiliary list.
1477 if (mark == markOopDesc::INFLATING()) {
1478 ReadStableMark(object);
1479 continue;
1480 }
1481
1482 // CASE: stack-locked
1483 // Could be stack-locked either by this thread or by some other thread.
1484 //
1485 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1486 // to install INFLATING into the mark word. We originally installed INFLATING,
1487 // allocated the objectmonitor, and then finally STed the address of the
1488 // objectmonitor into the mark. This was correct, but artificially lengthened
1489 // the interval in which INFLATED appeared in the mark, thus increasing
1490 // the odds of inflation contention.
1491 //
1492 // We now use per-thread private objectmonitor free lists.
1493 // These list are reprovisioned from the global free list outside the
1494 // critical INFLATING...ST interval. A thread can transfer
1495 // multiple objectmonitors en-mass from the global free list to its local free list.
1496 // This reduces coherency traffic and lock contention on the global free list.
1497 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1498 // before or after the CAS(INFLATING) operation.
1499 // See the comments in omAlloc().
1500
1501 LogStreamHandle(Trace, monitorinflation) lsh;
1502
1503 if (mark->has_locker()) {
1504 ObjectMonitor * m;
1505 if (!AsyncDeflateIdleMonitors || cause == inflate_cause_vm_internal) {
1506 // If !AsyncDeflateIdleMonitors or if an internal inflation, then
1507 // we won't stop for a potential safepoint in omAlloc.
1508 m = omAlloc(Self, cause);
1509 } else {
1510 // If AsyncDeflateIdleMonitors and not an internal inflation, then
1511 // we may stop for a safepoint in omAlloc() so protect object.
1512 Handle h_obj(Self, object);
1513 m = omAlloc(Self, cause);
1514 object = h_obj(); // Refresh object.
1515 }
1516 // Optimistically prepare the objectmonitor - anticipate successful CAS
1517 // We do this before the CAS in order to minimize the length of time
1518 // in which INFLATING appears in the mark.
1519 m->Recycle();
1520 m->_Responsible = NULL;
1521 m->_recursions = 0;
1522 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1523
1524 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1525 if (cmp != mark) {
1526 omRelease(Self, m, true);
1527 continue; // Interference -- just retry
1528 }
1529
1530 // We've successfully installed INFLATING (0) into the mark-word.
1531 // This is the only case where 0 will appear in a mark-word.
1532 // Only the singular thread that successfully swings the mark-word
1533 // to 0 can perform (or more precisely, complete) inflation.
1534 //
1535 // Why do we CAS a 0 into the mark-word instead of just CASing the
1572 m->set_object(object);
1573 // TODO-FIXME: assert BasicLock->dhw != 0.
1574
1575 // Must preserve store ordering. The monitor state must
1576 // be stable at the time of publishing the monitor address.
1577 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1578 object->release_set_mark(markOopDesc::encode(m));
1579
1580 // Hopefully the performance counters are allocated on distinct cache lines
1581 // to avoid false sharing on MP systems ...
1582 OM_PERFDATA_OP(Inflations, inc());
1583 if (log_is_enabled(Trace, monitorinflation)) {
1584 ResourceMark rm(Self);
1585 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1586 INTPTR_FORMAT ", type='%s'", p2i(object),
1587 p2i(object->mark()), object->klass()->external_name());
1588 }
1589 if (event.should_commit()) {
1590 post_monitor_inflate_event(&event, object, cause);
1591 }
1592 assert(!m->is_free(), "post-condition");
1593 omh_p->set_om_ptr(m);
1594 return;
1595 }
1596
1597 // CASE: neutral
1598 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1599 // If we know we're inflating for entry it's better to inflate by swinging a
1600 // pre-locked objectMonitor pointer into the object header. A successful
1601 // CAS inflates the object *and* confers ownership to the inflating thread.
1602 // In the current implementation we use a 2-step mechanism where we CAS()
1603 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1604 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1605 // would be useful.
1606
1607 assert(mark->is_neutral(), "invariant");
1608 ObjectMonitor * m;
1609 if (!AsyncDeflateIdleMonitors || cause == inflate_cause_vm_internal) {
1610 // If !AsyncDeflateIdleMonitors or if an internal inflation, then
1611 // we won't stop for a potential safepoint in omAlloc.
1612 m = omAlloc(Self, cause);
1613 } else {
1614 // If AsyncDeflateIdleMonitors and not an internal inflation, then
1615 // we may stop for a safepoint in omAlloc() so protect object.
1616 Handle h_obj(Self, object);
1617 m = omAlloc(Self, cause);
1618 object = h_obj(); // Refresh object.
1619 }
1620 // prepare m for installation - set monitor to initial state
1621 m->Recycle();
1622 m->set_header(mark);
1623 m->set_owner(NULL);
1624 m->set_object(object);
1625 m->_recursions = 0;
1626 m->_Responsible = NULL;
1627 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1628
1629 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1630 m->set_header(NULL);
1631 m->set_object(NULL);
1632 m->Recycle();
1633 omRelease(Self, m, true);
1634 m = NULL;
1635 continue;
1636 // interference - the markword changed - just retry.
1637 // The state-transitions are one-way, so there's no chance of
1638 // live-lock -- "Inflated" is an absorbing state.
1639 }
1640
1641 // Hopefully the performance counters are allocated on distinct
1642 // cache lines to avoid false sharing on MP systems ...
1643 OM_PERFDATA_OP(Inflations, inc());
1644 if (log_is_enabled(Trace, monitorinflation)) {
1645 ResourceMark rm(Self);
1646 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1647 INTPTR_FORMAT ", type='%s'", p2i(object),
1648 p2i(object->mark()), object->klass()->external_name());
1649 }
1650 if (event.should_commit()) {
1651 post_monitor_inflate_event(&event, object, cause);
1652 }
1653 omh_p->set_om_ptr(m);
1654 return;
1655 }
1656 }
1657
1658
1659 // We create a list of in-use monitors for each thread.
1660 //
1661 // deflate_thread_local_monitors() scans a single thread's in-use list, while
1662 // deflate_idle_monitors() scans only a global list of in-use monitors which
1663 // is populated only as a thread dies (see omFlush()).
1664 //
1665 // These operations are called at all safepoints, immediately after mutators
1666 // are stopped, but before any objects have moved. Collectively they traverse
1667 // the population of in-use monitors, deflating where possible. The scavenged
1668 // monitors are returned to the monitor free list.
1669 //
1670 // Beware that we scavenge at *every* stop-the-world point. Having a large
1671 // number of monitors in-use could negatively impact performance. We also want
1672 // to minimize the total # of monitors in circulation, as they incur a small
1673 // footprint penalty.
1674 //
1675 // Perversely, the heap size -- and thus the STW safepoint rate --
1676 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1677 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1678 // This is an unfortunate aspect of this design.
1679
1680 void ObjectSynchronizer::do_safepoint_work(DeflateMonitorCounters* _counters) {
1681 if (!AsyncDeflateIdleMonitors) {
1682 // Use the older mechanism for the global in-use list.
1683 ObjectSynchronizer::deflate_idle_monitors(_counters);
1684 return;
1685 }
1686
1687 assert(_counters == NULL, "not used with AsyncDeflateIdleMonitors");
1688
1689 log_debug(monitorinflation)("requesting deflation of idle monitors.");
1690 // Request deflation of global idle monitors by the ServiceThread:
1691 _gOmShouldDeflateIdleMonitors = true;
1692 MutexLockerEx ml(Service_lock, Mutex::_no_safepoint_check_flag);
1693 Service_lock->notify_all();
1694
1695 // Request deflation of per-thread idle monitors by each JavaThread:
1696 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
1697 if (jt->omInUseCount > 0) {
1698 // This JavaThread is using monitors so check it.
1699 jt->omShouldDeflateIdleMonitors = true;
1700 }
1701 }
1702 }
1703
1704 // Deflate a single monitor if not in-use
1705 // Return true if deflated, false if in-use
1706 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1707 ObjectMonitor** freeHeadp,
1708 ObjectMonitor** freeTailp) {
1709 bool deflated;
1710 // Normal case ... The monitor is associated with obj.
1711 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1712 guarantee(mid == obj->mark()->monitor(), "invariant");
1713 guarantee(mid->header()->is_neutral(), "invariant");
1714
1715 if (mid->is_busy()) {
1716 deflated = false;
1717 } else {
1718 // Deflate the monitor if it is no longer being used
1719 // It's idle - scavenge and return to the global free list
1720 // plain old deflation ...
1721 if (log_is_enabled(Trace, monitorinflation)) {
1722 ResourceMark rm;
1723 log_trace(monitorinflation)("deflate_monitor: "
1724 "object=" INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", type='%s'",
1725 p2i(obj), p2i(obj->mark()),
1726 obj->klass()->external_name());
1727 }
1728
1729 // Restore the header back to obj
1730 obj->release_set_mark(mid->header());
1731 mid->clear();
1732
1733 assert(mid->object() == NULL, "invariant");
1734 assert(mid->is_free(), "invariant");
1735
1736 // Move the object to the working free list defined by freeHeadp, freeTailp
1737 if (*freeHeadp == NULL) *freeHeadp = mid;
1738 if (*freeTailp != NULL) {
1739 ObjectMonitor * prevtail = *freeTailp;
1740 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1741 prevtail->FreeNext = mid;
1742 }
1743 *freeTailp = mid;
1744 deflated = true;
1745 }
1746 return deflated;
1747 }
1748
1749 // Deflate the specified ObjectMonitor if not in-use using a JavaThread.
1750 // Returns true if it was deflated and false otherwise.
1751 //
1752 // The async deflation protocol sets _owner to DEFLATER_MARKER and
1753 // makes _count negative as signals to contending threads that an
1754 // async deflation is in progress. There are a number of checks as
1755 // part of the protocol to make sure that the calling thread has
1756 // not lost the race to a contending thread.
1757 //
1758 // The ObjectMonitor has been successfully async deflated when:
1759 // (_owner == DEFLATER_MARKER && _count < 0). Contending threads that
1760 // see those values know to retry their operation.
1761 //
1762 bool ObjectSynchronizer::deflate_monitor_using_JT(ObjectMonitor* mid,
1763 ObjectMonitor** freeHeadp,
1764 ObjectMonitor** freeTailp) {
1765 assert(AsyncDeflateIdleMonitors, "sanity check");
1766 assert(Thread::current()->is_Java_thread(), "precondition");
1767 // A newly allocated ObjectMonitor should not be seen here so we
1768 // avoid an endless inflate/deflate cycle.
1769 assert(mid->is_old(), "precondition");
1770
1771 if (mid->is_busy() || mid->ref_count() != 0) {
1772 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor*
1773 // is in use so no deflation.
1774 return false;
1775 }
1776
1777 if (Atomic::cmpxchg(DEFLATER_MARKER, &mid->_owner, (void*)NULL) == NULL) {
1778 // ObjectMonitor is not owned by another thread. Our setting
1779 // _owner to DEFLATER_MARKER forces any contending thread through
1780 // the slow path. This is just the first part of the async
1781 // deflation dance.
1782
1783 if (mid->_waiters != 0 || mid->ref_count() != 0) {
1784 // Another thread has raced to enter the ObjectMonitor after
1785 // mid->is_busy() above and has already waited on it which
1786 // makes it busy so no deflation. Or the ObjectMonitor* is
1787 // in use for some other operation like inflate(). Restore
1788 // _owner to NULL if it is still DEFLATER_MARKER.
1789 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER);
1790 return false;
1791 }
1792
1793 if (Atomic::cmpxchg(-max_jint, &mid->_count, (jint)0) == 0) {
1794 // Make _count negative to force racing threads to retry.
1795 // This is the second part of the async deflation dance.
1796
1797 if (mid->_owner == DEFLATER_MARKER) {
1798 // If _owner is still DEFLATER_MARKER, then we have successfully
1799 // signaled any racing threads to retry. If it is not, then we
1800 // have lost the race to another thread and the ObjectMonitor is
1801 // now busy. This is the third and final part of the async
1802 // deflation dance.
1803 // Note: This _owner check solves the ABA problem with _count
1804 // where another thread acquired the ObjectMonitor, finished
1805 // using it and restored the _count to zero.
1806
1807 // Sanity checks for the races:
1808 guarantee(mid->_waiters == 0, "should be no waiters");
1809 guarantee(mid->_cxq == NULL, "should be no contending threads");
1810 guarantee(mid->_EntryList == NULL, "should be no entering threads");
1811
1812 if (log_is_enabled(Trace, monitorinflation)) {
1813 oop obj = (oop) mid->object();
1814 assert(obj != NULL, "sanity check");
1815 if (obj->is_instance()) {
1816 ResourceMark rm;
1817 log_trace(monitorinflation)("deflate_monitor_using_JT: "
1818 "object=" INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", type='%s'",
1819 p2i(obj), p2i(obj->mark()),
1820 obj->klass()->external_name());
1821 }
1822 }
1823
1824 // Install the old mark word if nobody else has already done it.
1825 mid->install_displaced_markword_in_object();
1826 mid->clear_using_JT();
1827
1828 assert(mid->object() == NULL, "invariant");
1829 assert(mid->is_free(), "invariant");
1830
1831 // Move the deflated ObjectMonitor to the working free list
1832 // defined by freeHeadp and freeTailp.
1833 if (*freeHeadp == NULL) {
1834 // First one on the list.
1835 *freeHeadp = mid;
1836 }
1837 if (*freeTailp != NULL) {
1838 // We append to the list so the caller can use mid->FreeNext
1839 // to fix the linkages in its context.
1840 ObjectMonitor * prevtail = *freeTailp;
1841 assert(prevtail->FreeNext == NULL, "not cleaned up by the caller");
1842 prevtail->FreeNext = mid;
1843 }
1844 *freeTailp = mid;
1845
1846 // At this point, mid->FreeNext still refers to its current
1847 // value and another ObjectMonitor's FreeNext field still
1848 // refers to this ObjectMonitor. Those linkages have to be
1849 // cleaned up by the caller who has the complete context.
1850
1851 // We leave _owner == DEFLATER_MARKER and _count < 0 to
1852 // force any racing threads to retry.
1853 return true; // Success, ObjectMonitor has been deflated.
1854 }
1855
1856 // The _owner was changed from DEFLATER_MARKER so we lost the
1857 // race since the ObjectMonitor is now busy. Add back max_jint
1858 // to restore the _count field to its proper value (which may
1859 // not be what we saw above).
1860 Atomic::add(max_jint, &mid->_count);
1861
1862 assert(mid->_count >= 0, "_count should not be negative");
1863 }
1864
1865 // The _count was no longer 0 so we lost the race since the
1866 // ObjectMonitor is now busy.
1867 assert(mid->_owner != DEFLATER_MARKER, "should no longer be set");
1868 }
1869
1870 // The _owner field is no longer NULL so we lost the race since the
1871 // ObjectMonitor is now busy.
1872 return false;
1873 }
1874
1875 // Walk a given monitor list, and deflate idle monitors
1876 // The given list could be a per-thread list or a global list
1877 // Caller acquires gListLock as needed.
1878 //
1879 // In the case of parallel processing of thread local monitor lists,
1880 // work is done by Threads::parallel_threads_do() which ensures that
1881 // each Java thread is processed by exactly one worker thread, and
1882 // thus avoid conflicts that would arise when worker threads would
1883 // process the same monitor lists concurrently.
1884 //
1885 // See also ParallelSPCleanupTask and
1886 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1887 // Threads::parallel_java_threads_do() in thread.cpp.
1888 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1889 ObjectMonitor** freeHeadp,
1890 ObjectMonitor** freeTailp) {
1891 ObjectMonitor* mid;
1892 ObjectMonitor* next;
1893 ObjectMonitor* cur_mid_in_use = NULL;
1894 int deflated_count = 0;
1898 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1899 // if deflate_monitor succeeded,
1900 // extract from per-thread in-use list
1901 if (mid == *listHeadp) {
1902 *listHeadp = mid->FreeNext;
1903 } else if (cur_mid_in_use != NULL) {
1904 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1905 }
1906 next = mid->FreeNext;
1907 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1908 mid = next;
1909 deflated_count++;
1910 } else {
1911 cur_mid_in_use = mid;
1912 mid = mid->FreeNext;
1913 }
1914 }
1915 return deflated_count;
1916 }
1917
1918 // Walk a given ObjectMonitor list and deflate idle ObjectMonitors using
1919 // a JavaThread. Returns the number of deflated ObjectMonitors. The given
1920 // list could be a per-thread in-use list or the global in-use list.
1921 // Caller acquires gListLock as appropriate. If a safepoint has started,
1922 // then we save state via savedMidInUsep and return to the caller to
1923 // honor the safepoint.
1924 //
1925 int ObjectSynchronizer::deflate_monitor_list_using_JT(ObjectMonitor** listHeadp,
1926 ObjectMonitor** freeHeadp,
1927 ObjectMonitor** freeTailp,
1928 ObjectMonitor** savedMidInUsep) {
1929 assert(AsyncDeflateIdleMonitors, "sanity check");
1930 assert(Thread::current()->is_Java_thread(), "precondition");
1931
1932 ObjectMonitor* mid;
1933 ObjectMonitor* next;
1934 ObjectMonitor* cur_mid_in_use = NULL;
1935 int deflated_count = 0;
1936
1937 if (*savedMidInUsep == NULL) {
1938 // No saved state so start at the beginning.
1939 mid = *listHeadp;
1940 } else {
1941 // We're restarting after a safepoint so restore the necessary state
1942 // before we resume.
1943 cur_mid_in_use = *savedMidInUsep;
1944 mid = cur_mid_in_use->FreeNext;
1945 }
1946 while (mid != NULL) {
1947 // Only try to deflate if there is an associated Java object and if
1948 // mid is old (is not newly allocated and is not newly freed).
1949 if (mid->object() != NULL && mid->is_old() &&
1950 deflate_monitor_using_JT(mid, freeHeadp, freeTailp)) {
1951 // Deflation succeeded so update the in-use list.
1952 if (mid == *listHeadp) {
1953 *listHeadp = mid->FreeNext;
1954 } else if (cur_mid_in_use != NULL) {
1955 // Maintain the current in-use list.
1956 cur_mid_in_use->FreeNext = mid->FreeNext;
1957 }
1958 next = mid->FreeNext;
1959 mid->FreeNext = NULL;
1960 // At this point mid is disconnected from the in-use list
1961 // and is the current tail in the freeHeadp list.
1962 mid = next;
1963 deflated_count++;
1964 } else {
1965 // mid is considered in-use if it does not have an associated
1966 // Java object or mid is not old or deflation did not succeed.
1967 // A mid->is_new() node can be seen here when it is freshly returned
1968 // by omAlloc() (and skips the deflation code path).
1969 // A mid->is_old() node can be seen here when deflation failed.
1970 // A mid->is_free() node can be seen here when a fresh node from
1971 // omAlloc() is released by omRelease() due to losing the race
1972 // in inflate().
1973
1974 if (mid->object() != NULL && mid->is_new()) {
1975 // mid has an associated Java object and has now been seen
1976 // as newly allocated so mark it as "old".
1977 mid->set_allocation_state(ObjectMonitor::Old);
1978 }
1979 cur_mid_in_use = mid;
1980 mid = mid->FreeNext;
1981
1982 if (SafepointSynchronize::is_synchronizing() &&
1983 cur_mid_in_use != *listHeadp && cur_mid_in_use->is_old()) {
1984 // If a safepoint has started and cur_mid_in_use is not the list
1985 // head and is old, then it is safe to use as saved state. Return
1986 // to the caller so gListLock can be dropped as appropriate
1987 // before blocking.
1988 *savedMidInUsep = cur_mid_in_use;
1989 return deflated_count;
1990 }
1991 }
1992 }
1993 // We finished the list without a safepoint starting so there's
1994 // no need to save state.
1995 *savedMidInUsep = NULL;
1996 return deflated_count;
1997 }
1998
1999 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
2000 counters->nInuse = 0; // currently associated with objects
2001 counters->nInCirculation = 0; // extant
2002 counters->nScavenged = 0; // reclaimed (global and per-thread)
2003 counters->perThreadScavenged = 0; // per-thread scavenge total
2004 counters->perThreadTimes = 0.0; // per-thread scavenge times
2005 }
2006
2007 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
2008 assert(!AsyncDeflateIdleMonitors, "sanity check");
2009 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
2010 bool deflated = false;
2011
2012 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
2013 ObjectMonitor * freeTailp = NULL;
2014 elapsedTimer timer;
2015
2016 if (log_is_enabled(Info, monitorinflation)) {
2017 timer.start();
2018 }
2019
2020 // Prevent omFlush from changing mids in Thread dtor's during deflation
2021 // And in case the vm thread is acquiring a lock during a safepoint
2022 // See e.g. 6320749
2023 Thread::muxAcquire(&gListLock, "deflate_idle_monitors");
2024
2025 // Note: the thread-local monitors lists get deflated in
2026 // a separate pass. See deflate_thread_local_monitors().
2027
2028 // For moribund threads, scan gOmInUseList
2042 // constant-time list splice - prepend scavenged segment to gFreeList
2043 freeTailp->FreeNext = gFreeList;
2044 gFreeList = freeHeadp;
2045 }
2046 Thread::muxRelease(&gListLock);
2047 timer.stop();
2048
2049 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2050 LogStreamHandle(Info, monitorinflation) lsh_info;
2051 LogStream * ls = NULL;
2052 if (log_is_enabled(Debug, monitorinflation)) {
2053 ls = &lsh_debug;
2054 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2055 ls = &lsh_info;
2056 }
2057 if (ls != NULL) {
2058 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
2059 }
2060 }
2061
2062 // Deflate global idle ObjectMonitors using a JavaThread.
2063 //
2064 void ObjectSynchronizer::deflate_global_idle_monitors_using_JT() {
2065 assert(AsyncDeflateIdleMonitors, "sanity check");
2066 assert(Thread::current()->is_Java_thread(), "precondition");
2067 JavaThread * cur_jt = JavaThread::current();
2068
2069 _gOmShouldDeflateIdleMonitors = false;
2070
2071 int deflated_count = 0;
2072 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged ObjectMonitors
2073 ObjectMonitor * freeTailp = NULL;
2074 ObjectMonitor * savedMidInUsep = NULL;
2075 elapsedTimer timer;
2076
2077 if (log_is_enabled(Info, monitorinflation)) {
2078 timer.start();
2079 }
2080 Thread::muxAcquire(&gListLock, "deflate_global_idle_monitors_using_JT(1)");
2081 OM_PERFDATA_OP(MonExtant, set_value(gOmInUseCount));
2082
2083 do {
2084 int local_deflated_count = deflate_monitor_list_using_JT((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp, &savedMidInUsep);
2085 gOmInUseCount -= local_deflated_count;
2086 deflated_count += local_deflated_count;
2087
2088 if (freeHeadp != NULL) {
2089 // Move the scavenged ObjectMonitors to the global free list.
2090 guarantee(freeTailp != NULL && local_deflated_count > 0, "freeTailp=" INTPTR_FORMAT ", local_deflated_count=%d", p2i(freeTailp), local_deflated_count);
2091 assert(freeTailp->FreeNext == NULL, "invariant");
2092
2093 // Constant-time list splice - prepend scavenged segment to gFreeList.
2094 freeTailp->FreeNext = gFreeList;
2095 gFreeList = freeHeadp;
2096
2097 gMonitorFreeCount += local_deflated_count;
2098 OM_PERFDATA_OP(Deflations, inc(local_deflated_count));
2099 }
2100
2101 if (savedMidInUsep != NULL) {
2102 // deflate_monitor_list_using_JT() detected a safepoint starting.
2103 Thread::muxRelease(&gListLock);
2104 timer.stop();
2105 {
2106 log_debug(monitorinflation)("pausing deflation of global idle monitors for a safepoint.");
2107 assert(SafepointSynchronize::is_synchronizing(), "sanity check");
2108 ThreadBlockInVM blocker(cur_jt);
2109 }
2110 // Prepare for another loop after the safepoint.
2111 freeHeadp = NULL;
2112 freeTailp = NULL;
2113 if (log_is_enabled(Info, monitorinflation)) {
2114 timer.start();
2115 }
2116 Thread::muxAcquire(&gListLock, "deflate_global_idle_monitors_using_JT(2)");
2117 }
2118 } while (savedMidInUsep != NULL);
2119 Thread::muxRelease(&gListLock);
2120 timer.stop();
2121
2122 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2123 LogStreamHandle(Info, monitorinflation) lsh_info;
2124 LogStream * ls = NULL;
2125 if (log_is_enabled(Debug, monitorinflation)) {
2126 ls = &lsh_debug;
2127 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2128 ls = &lsh_info;
2129 }
2130 if (ls != NULL) {
2131 ls->print_cr("async-deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
2132 }
2133 }
2134
2135 // Deflate per-thread idle ObjectMonitors using a JavaThread.
2136 //
2137 void ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT() {
2138 assert(AsyncDeflateIdleMonitors, "sanity check");
2139 assert(Thread::current()->is_Java_thread(), "precondition");
2140 JavaThread * cur_jt = JavaThread::current();
2141
2142 cur_jt->omShouldDeflateIdleMonitors = false;
2143
2144 int deflated_count = 0;
2145 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged ObjectMonitors
2146 ObjectMonitor * freeTailp = NULL;
2147 ObjectMonitor * savedMidInUsep = NULL;
2148 elapsedTimer timer;
2149
2150 if (log_is_enabled(Info, monitorinflation)) {
2151 timer.start();
2152 }
2153
2154 OM_PERFDATA_OP(MonExtant, inc(cur_jt->omInUseCount));
2155 do {
2156 int local_deflated_count = deflate_monitor_list_using_JT(cur_jt->omInUseList_addr(), &freeHeadp, &freeTailp, &savedMidInUsep);
2157 cur_jt->omInUseCount -= local_deflated_count;
2158 deflated_count += local_deflated_count;
2159
2160 if (freeHeadp != NULL) {
2161 // Move the scavenged ObjectMonitors to the global free list.
2162 Thread::muxAcquire(&gListLock, "deflate_per_thread_idle_monitors_using_JT");
2163 guarantee(freeTailp != NULL && local_deflated_count > 0, "freeTailp=" INTPTR_FORMAT ", local_deflated_count=%d", p2i(freeTailp), local_deflated_count);
2164 assert(freeTailp->FreeNext == NULL, "invariant");
2165
2166 // Constant-time list splice - prepend scavenged segment to gFreeList.
2167 freeTailp->FreeNext = gFreeList;
2168 gFreeList = freeHeadp;
2169
2170 gMonitorFreeCount += local_deflated_count;
2171 OM_PERFDATA_OP(Deflations, inc(local_deflated_count));
2172 Thread::muxRelease(&gListLock);
2173 // Prepare for another loop on the current JavaThread.
2174 freeHeadp = NULL;
2175 freeTailp = NULL;
2176 }
2177 timer.stop();
2178
2179 if (savedMidInUsep != NULL) {
2180 // deflate_monitor_list_using_JT() detected a safepoint starting.
2181 {
2182 log_debug(monitorinflation)("jt=" INTPTR_FORMAT ": pausing deflation of per-thread idle monitors for a safepoint.", p2i(cur_jt));
2183 assert(SafepointSynchronize::is_synchronizing(), "sanity check");
2184 ThreadBlockInVM blocker(cur_jt);
2185 }
2186 // Prepare for another loop on the current JavaThread after
2187 // the safepoint.
2188 if (log_is_enabled(Info, monitorinflation)) {
2189 timer.start();
2190 }
2191 }
2192 } while (savedMidInUsep != NULL);
2193
2194 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2195 LogStreamHandle(Info, monitorinflation) lsh_info;
2196 LogStream * ls = NULL;
2197 if (log_is_enabled(Debug, monitorinflation)) {
2198 ls = &lsh_debug;
2199 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2200 ls = &lsh_info;
2201 }
2202 if (ls != NULL) {
2203 ls->print_cr("jt=" INTPTR_FORMAT ": async-deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(cur_jt), timer.seconds(), deflated_count);
2204 }
2205 }
2206
2207 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
2208 // Report the cumulative time for deflating each thread's idle
2209 // monitors. Note: if the work is split among more than one
2210 // worker thread, then the reported time will likely be more
2211 // than a beginning to end measurement of the phase.
2212 // Note: AsyncDeflateIdleMonitors only deflates per-thread idle
2213 // monitors at a safepoint when a special cleanup has been requested.
2214 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->perThreadTimes, counters->perThreadScavenged);
2215
2216 bool needs_special_cleanup = is_cleanup_requested();
2217 if (!AsyncDeflateIdleMonitors || needs_special_cleanup) {
2218 // AsyncDeflateIdleMonitors does not use these counters unless
2219 // there is a special cleanup request.
2220
2221 gMonitorFreeCount += counters->nScavenged;
2222
2223 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
2224 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
2225 }
2226
2227 if (log_is_enabled(Debug, monitorinflation)) {
2228 // exit_globals()'s call to audit_and_print_stats() is done
2229 // at the Info level.
2230 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */);
2231 } else if (log_is_enabled(Info, monitorinflation)) {
2232 Thread::muxAcquire(&gListLock, "finish_deflate_idle_monitors");
2233 log_info(monitorinflation)("gMonitorPopulation=%d, gOmInUseCount=%d, "
2234 "gMonitorFreeCount=%d", gMonitorPopulation,
2235 gOmInUseCount, gMonitorFreeCount);
2236 Thread::muxRelease(&gListLock);
2237 }
2238
2239 ForceMonitorScavenge = 0; // Reset
2240 GVars.stwRandom = os::random();
2241 GVars.stwCycle++;
2242 if (needs_special_cleanup) {
2243 set_is_cleanup_requested(false); // special clean up is done
2244 }
2245 }
2246
2247 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
2248 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
2249
2250 if (AsyncDeflateIdleMonitors) {
2251 // Nothing to do when idle ObjectMonitors are deflated using a
2252 // JavaThread unless a special cleanup has been requested.
2253 if (!is_cleanup_requested()) {
2254 return;
2255 }
2256 }
2257
2258 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
2259 ObjectMonitor * freeTailp = NULL;
2260 elapsedTimer timer;
2261
2262 if (log_is_enabled(Info, safepoint, cleanup) ||
2263 log_is_enabled(Info, monitorinflation)) {
2264 timer.start();
2265 }
2266
2267 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
2268
2269 Thread::muxAcquire(&gListLock, "deflate_thread_local_monitors(1)");
2270
2271 // Adjust counters
2272 counters->nInCirculation += thread->omInUseCount;
2273 thread->omInUseCount -= deflated_count;
2274 counters->nScavenged += deflated_count;
2275 counters->nInuse += thread->omInUseCount;
2276 counters->perThreadScavenged += deflated_count;
2277
2449 } else {
2450 log_error(monitorinflation)("found monitor list errors: error_cnt=%d", error_cnt);
2451 }
2452
2453 if ((on_exit && log_is_enabled(Info, monitorinflation)) ||
2454 (!on_exit && log_is_enabled(Trace, monitorinflation))) {
2455 // When exiting this log output is at the Info level. When called
2456 // at a safepoint, this log output is at the Trace level since
2457 // there can be a lot of it.
2458 log_in_use_monitor_details(ls, on_exit);
2459 }
2460
2461 ls->flush();
2462
2463 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
2464 }
2465
2466 // Check a free monitor entry; log any errors.
2467 void ObjectSynchronizer::chk_free_entry(JavaThread * jt, ObjectMonitor * n,
2468 outputStream * out, int *error_cnt_p) {
2469 if ((!AsyncDeflateIdleMonitors && n->is_busy()) ||
2470 (AsyncDeflateIdleMonitors && n->is_busy_async())) {
2471 if (jt != NULL) {
2472 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2473 ": free per-thread monitor must not be busy.", p2i(jt),
2474 p2i(n));
2475 } else {
2476 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2477 "must not be busy.", p2i(n));
2478 }
2479 *error_cnt_p = *error_cnt_p + 1;
2480 }
2481 if (n->header() != NULL) {
2482 if (jt != NULL) {
2483 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2484 ": free per-thread monitor must have NULL _header "
2485 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n),
2486 p2i(n->header()));
2487 } else {
2488 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2489 "must have NULL _header field: _header=" INTPTR_FORMAT,
2490 p2i(n), p2i(n->header()));
2642 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": omInUseCount=%d is not "
2643 "equal to chkOmInUseCount=%d", p2i(jt), jt->omInUseCount,
2644 chkOmInUseCount);
2645 *error_cnt_p = *error_cnt_p + 1;
2646 }
2647 }
2648
2649 // Log details about ObjectMonitors on the in-use lists. The 'BHL'
2650 // flags indicate why the entry is in-use, 'object' and 'object type'
2651 // indicate the associated object and its type.
2652 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out,
2653 bool on_exit) {
2654 if (!on_exit) {
2655 // Not at VM exit so grab the global list lock.
2656 Thread::muxAcquire(&gListLock, "log_in_use_monitor_details");
2657 }
2658
2659 if (gOmInUseCount > 0) {
2660 out->print_cr("In-use global monitor info:");
2661 out->print_cr("(B -> is_busy, H -> has hashcode, L -> lock status)");
2662 out->print_cr("%18s %s %7s %18s %18s",
2663 "monitor", "BHL", "ref_cnt", "object", "object type");
2664 out->print_cr("================== === ======= ================== ==================");
2665 for (ObjectMonitor * n = gOmInUseList; n != NULL; n = n->FreeNext) {
2666 const oop obj = (oop) n->object();
2667 const markOop mark = n->header();
2668 ResourceMark rm;
2669 out->print_cr(INTPTR_FORMAT " %d%d%d %7d " INTPTR_FORMAT " %s",
2670 p2i(n), n->is_busy() != 0, mark->hash() != 0,
2671 n->owner() != NULL, (int)n->ref_count(), p2i(obj),
2672 obj->klass()->external_name());
2673 }
2674 }
2675
2676 if (!on_exit) {
2677 Thread::muxRelease(&gListLock);
2678 }
2679
2680 out->print_cr("In-use per-thread monitor info:");
2681 out->print_cr("(B -> is_busy, H -> has hashcode, L -> lock status)");
2682 out->print_cr("%18s %18s %s %7s %18s %18s",
2683 "jt", "monitor", "BHL", "ref_cnt", "object", "object type");
2684 out->print_cr("================== ================== === ======= ================== ==================");
2685 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2686 for (ObjectMonitor * n = jt->omInUseList; n != NULL; n = n->FreeNext) {
2687 const oop obj = (oop) n->object();
2688 const markOop mark = n->header();
2689 ResourceMark rm;
2690 out->print_cr(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d %7d "
2691 INTPTR_FORMAT " %s", p2i(jt), p2i(n), n->is_busy() != 0,
2692 mark->hash() != 0, n->owner() != NULL, (int)n->ref_count(),
2693 p2i(obj), obj->klass()->external_name());
2694 }
2695 }
2696
2697 out->flush();
2698 }
2699
2700 // Log counts for the global and per-thread monitor lists and return
2701 // the population count.
2702 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) {
2703 int popCount = 0;
2704 out->print_cr("%18s %10s %10s %10s",
2705 "Global Lists:", "InUse", "Free", "Total");
2706 out->print_cr("================== ========== ========== ==========");
2707 out->print_cr("%18s %10d %10d %10d", "",
2708 gOmInUseCount, gMonitorFreeCount, gMonitorPopulation);
2709 popCount += gOmInUseCount + gMonitorFreeCount;
2710
2711 out->print_cr("%18s %10s %10s %10s",
2712 "Per-Thread Lists:", "InUse", "Free", "Provision");
2713 out->print_cr("================== ========== ========== ==========");
|