108 #endif // ndef DTRACE_ENABLED
109
110 // This exists only as a workaround of dtrace bug 6254741
111 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
112 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
113 return 0;
114 }
115
116 #define NINFLATIONLOCKS 256
117 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
118
119 // global list of blocks of monitors
120 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
121 // global monitor free list
122 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
123 // global monitor in-use list, for moribund threads,
124 // monitors they inflated need to be scanned for deflation
125 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
126 // count of entries in gOmInUseList
127 int ObjectSynchronizer::gOmInUseCount = 0;
128
129 static volatile intptr_t gListLock = 0; // protects global monitor lists
130 static volatile int gMonitorFreeCount = 0; // # on gFreeList
131 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
132
133 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
134
135
136 // =====================> Quick functions
137
138 // The quick_* forms are special fast-path variants used to improve
139 // performance. In the simplest case, a "quick_*" implementation could
140 // simply return false, in which case the caller will perform the necessary
141 // state transitions and call the slow-path form.
142 // The fast-path is designed to handle frequently arising cases in an efficient
143 // manner and is just a degenerate "optimistic" variant of the slow-path.
144 // returns true -- to indicate the call was satisfied.
145 // returns false -- to indicate the call needs the services of the slow-path.
146 // A no-loitering ordinance is in effect for code in the quick_* family
147 // operators: safepoints or indefinite blocking (blocking that might span a
194 }
195
196 // biased locking and any other IMS exception states take the slow-path
197 return false;
198 }
199
200
201 // The LockNode emitted directly at the synchronization site would have
202 // been too big if it were to have included support for the cases of inflated
203 // recursive enter and exit, so they go here instead.
204 // Note that we can't safely call AsyncPrintJavaStack() from within
205 // quick_enter() as our thread state remains _in_Java.
206
207 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
208 BasicLock * lock) {
209 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
210 assert(Self->is_Java_thread(), "invariant");
211 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
212 NoSafepointVerifier nsv;
213 if (obj == NULL) return false; // Need to throw NPE
214 const markOop mark = obj->mark();
215
216 if (mark->has_monitor()) {
217 ObjectMonitor * const m = mark->monitor();
218 assert(oopDesc::equals((oop) m->object(), obj), "invariant");
219 Thread * const owner = (Thread *) m->_owner;
220
221 // Lock contention and Transactional Lock Elision (TLE) diagnostics
222 // and observability
223 // Case: light contention possibly amenable to TLE
224 // Case: TLE inimical operations such as nested/recursive synchronization
225
226 if (owner == Self) {
227 m->_recursions++;
228 return true;
229 }
230
231 // This Java Monitor is inflated so obj's header will never be
232 // displaced to this thread's BasicLock. Make the displaced header
233 // non-NULL so this BasicLock is not seen as recursive nor as
234 // being locked. We do this unconditionally so that this thread's
235 // BasicLock cannot be mis-interpreted by any stack walkers. For
236 // performance reasons, stack walkers generally first check for
237 // Biased Locking in the object's header, the second check is for
238 // stack-locking in the object's header, the third check is for
239 // recursive stack-locking in the displaced header in the BasicLock,
240 // and last are the inflated Java Monitor (ObjectMonitor) checks.
241 lock->set_displaced_header(markOopDesc::unused_mark());
242
243 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
244 assert(m->_recursions == 0, "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,
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(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(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 in the BasicLock on a thread's stack
753 // is strictly immutable. It CANNOT be changed in ANY cases.
754 // So we have to inflate the stack lock into an ObjectMonitor
755 // even if the current thread owns the lock. The BasicLock on
756 // a thread's stack can be asynchronously read by other threads
757 // during an inflate() call so any change to that stack memory
758 // may not propagate to other threads correctly.
759 }
760
761 // Inflate the monitor to set hash code
762 monitor = inflate(Self, obj, inflate_cause_hash_code);
763 // Load displaced header and check it has hash code
764 mark = monitor->header();
765 assert(mark->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(mark));
766 hash = mark->hash();
767 if (hash == 0) {
768 hash = get_next_hash(Self, obj);
769 temp = mark->copy_set_hash(hash); // merge hash code into header
770 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(temp));
771 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark);
772 if (test != mark) {
773 // The only update to the ObjectMonitor's header/dmw field
774 // is to merge in the hash code. If someone adds a new usage
775 // of the header/dmw field, please update this code.
776 hash = test->hash();
777 assert(test->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(test));
778 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
779 }
780 }
781 // We finally get the hash
782 return hash;
783 }
784
785 // Deprecated -- use FastHashCode() instead.
786
787 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
788 return FastHashCode(Thread::current(), obj());
789 }
790
791
792 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
793 Handle h_obj) {
794 if (UseBiasedLocking) {
795 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
796 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
797 }
798
799 assert(thread == JavaThread::current(), "Can only be called on current thread");
800 oop obj = h_obj();
801
802 markOop mark = ReadStableMark(obj);
803
804 // Uncontended case, header points to stack
805 if (mark->has_locker()) {
806 return thread->is_lock_owned((address)mark->locker());
807 }
808 // Contended case, header points to ObjectMonitor (tagged pointer)
809 if (mark->has_monitor()) {
810 ObjectMonitor* monitor = mark->monitor();
811 return monitor->is_entered(thread) != 0;
812 }
813 // Unlocked case, header in place
814 assert(mark->is_neutral(), "sanity check");
815 return false;
816 }
817
818 // Be aware of this method could revoke bias of the lock object.
819 // This method queries the ownership of the lock handle specified by 'h_obj'.
820 // If the current thread owns the lock, it returns owner_self. If no
821 // thread owns the lock, it returns owner_none. Otherwise, it will return
822 // owner_other.
823 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
824 (JavaThread *self, Handle h_obj) {
825 // The caller must beware this method can revoke bias, and
826 // revocation can result in a safepoint.
827 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
828 assert(self->thread_state() != _thread_blocked, "invariant");
829
830 // Possible mark states: neutral, biased, stack-locked, inflated
831
832 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
833 // CASE: biased
834 BiasedLocking::revoke_and_rebias(h_obj, false, self);
835 assert(!h_obj->mark()->has_bias_pattern(),
836 "biases should be revoked by now");
837 }
838
839 assert(self == JavaThread::current(), "Can only be called on current thread");
840 oop obj = h_obj();
841 markOop mark = ReadStableMark(obj);
842
843 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
844 if (mark->has_locker()) {
845 return self->is_lock_owned((address)mark->locker()) ?
846 owner_self : owner_other;
847 }
848
849 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor.
850 // The Object:ObjectMonitor relationship is stable as long as we're
851 // not at a safepoint.
852 if (mark->has_monitor()) {
853 void * owner = mark->monitor()->_owner;
854 if (owner == NULL) return owner_none;
855 return (owner == self ||
856 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
857 }
858
859 // CASE: neutral
860 assert(mark->is_neutral(), "sanity check");
861 return owner_none; // it's unlocked
862 }
863
864 // FIXME: jvmti should call this
865 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
866 if (UseBiasedLocking) {
867 if (SafepointSynchronize::is_at_safepoint()) {
868 BiasedLocking::revoke_at_safepoint(h_obj);
869 } else {
870 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
871 }
872 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
873 }
874
875 oop obj = h_obj();
876 address owner = NULL;
877
878 markOop mark = ReadStableMark(obj);
879
880 // Uncontended case, header points to stack
881 if (mark->has_locker()) {
882 owner = (address) mark->locker();
883 }
884
885 // Contended case, header points to ObjectMonitor (tagged pointer)
886 else if (mark->has_monitor()) {
887 ObjectMonitor* monitor = mark->monitor();
888 assert(monitor != NULL, "monitor should be non-null");
889 owner = (address) monitor->owner();
890 }
891
892 if (owner != NULL) {
893 // owning_thread_from_monitor_owner() may also return NULL here
894 return Threads::owning_thread_from_monitor_owner(t_list, owner);
895 }
896
897 // Unlocked case, header in place
898 // Cannot have assertion since this object may have been
899 // locked by another thread when reaching here.
900 // assert(mark->is_neutral(), "sanity check");
901
902 return NULL;
903 }
904
905 // Visitors ...
906
907 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
908 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
909 while (block != NULL) {
910 assert(block->object() == CHAINMARKER, "must be a block header");
911 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
912 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
913 oop object = (oop)mid->object();
914 if (object != NULL) {
915 closure->do_monitor(mid);
916 }
917 }
918 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
919 }
920 }
921
922 // Get the next block in the block list.
923 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
924 assert(block->object() == CHAINMARKER, "must be a block header");
925 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
926 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
927 return block;
928 }
929
930 static bool monitors_used_above_threshold() {
931 if (gMonitorPopulation == 0) {
932 return false;
933 }
934 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
935 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation;
936 return monitor_usage > MonitorUsedDeflationThreshold;
937 }
938
939 bool ObjectSynchronizer::is_cleanup_needed() {
940 if (MonitorUsedDeflationThreshold > 0) {
941 return monitors_used_above_threshold();
942 }
943 return false;
944 }
945
946 void ObjectSynchronizer::oops_do(OopClosure* f) {
947 // We only scan the global used list here (for moribund threads), and
948 // the thread-local monitors in Thread::oops_do().
949 global_used_oops_do(f);
950 }
951
952 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
953 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
954 list_oops_do(gOmInUseList, f);
955 }
956
957 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
958 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
959 list_oops_do(thread->omInUseList, f);
960 }
961
962 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
963 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
964 ObjectMonitor* mid;
965 for (mid = list; mid != NULL; mid = mid->FreeNext) {
975 // -----------------------
976 // Inflation unlinks monitors from the global gFreeList and
977 // associates them with objects. Deflation -- which occurs at
978 // STW-time -- disassociates idle monitors from objects. Such
979 // scavenged monitors are returned to the gFreeList.
980 //
981 // The global list is protected by gListLock. All the critical sections
982 // are short and operate in constant-time.
983 //
984 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
985 //
986 // Lifecycle:
987 // -- unassigned and on the global free list
988 // -- unassigned and on a thread's private omFreeList
989 // -- assigned to an object. The object is inflated and the mark refers
990 // to the objectmonitor.
991
992
993 // Constraining monitor pool growth via MonitorBound ...
994 //
995 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
996 // the rate of scavenging is driven primarily by GC. As such, we can find
997 // an inordinate number of monitors in circulation.
998 // To avoid that scenario we can artificially induce a STW safepoint
999 // if the pool appears to be growing past some reasonable bound.
1000 // Generally we favor time in space-time tradeoffs, but as there's no
1001 // natural back-pressure on the # of extant monitors we need to impose some
1002 // type of limit. Beware that if MonitorBound is set to too low a value
1003 // we could just loop. In addition, if MonitorBound is set to a low value
1004 // we'll incur more safepoints, which are harmful to performance.
1005 // See also: GuaranteedSafepointInterval
1006 //
1007 // The current implementation uses asynchronous VM operations.
1008
1009 static void InduceScavenge(Thread * Self, const char * Whence) {
1010 // Induce STW safepoint to trim monitors
1011 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1012 // More precisely, trigger an asynchronous STW safepoint as the number
1013 // of active monitors passes the specified threshold.
1014 // TODO: assert thread state is reasonable
1015
1016 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1017 // Induce a 'null' safepoint to scavenge monitors
1018 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1019 // to the VMthread and have a lifespan longer than that of this activation record.
1020 // The VMThread will delete the op when completed.
1021 VMThread::execute(new VM_ScavengeMonitors());
1022 }
1023 }
1024
1025 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) {
1026 // A large MAXPRIVATE value reduces both list lock contention
1027 // and list coherency traffic, but also tends to increase the
1028 // number of objectMonitors in circulation as well as the STW
1029 // scavenge costs. As usual, we lean toward time in space-time
1030 // tradeoffs.
1031 const int MAXPRIVATE = 1024;
1032 stringStream ss;
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 take->Recycle();
1069 omRelease(Self, take, false);
1070 }
1071 Thread::muxRelease(&gListLock);
1072 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1073 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1074
1075 const int mx = MonitorBound;
1076 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1077 // We can't safely induce a STW safepoint from omAlloc() as our thread
1078 // state may not be appropriate for such activities and callers may hold
1079 // naked oops, so instead we defer the action.
1080 InduceScavenge(Self, "omAlloc");
1081 }
1082 continue;
1083 }
1084
1085 // 3: allocate a block of new ObjectMonitors
1086 // Both the local and global free lists are empty -- resort to malloc().
1087 // In the current implementation objectMonitors are TSM - immortal.
1088 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1089 // each ObjectMonitor to start at the beginning of a cache line,
1090 // so we use align_up().
1091 // A better solution would be to use C++ placement-new.
1092 // BEWARE: As it stands currently, we don't run the ctors!
1093 assert(_BLOCKSIZE > 1, "invariant");
1094 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE;
1095 PaddedEnd<ObjectMonitor> * temp;
1096 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1);
1101
1102 // NOTE: (almost) no way to recover if allocation failed.
1103 // We might be able to induce a STW safepoint and scavenge enough
1104 // objectMonitors to permit progress.
1105 if (temp == NULL) {
1106 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1107 "Allocate ObjectMonitors");
1108 }
1109 (void)memset((void *) temp, 0, neededsize);
1110
1111 // Format the block.
1112 // initialize the linked list, each monitor points to its next
1113 // forming the single linked free list, the very first monitor
1114 // will points to next block, which forms the block list.
1115 // The trick of using the 1st element in the block as gBlockList
1116 // linkage should be reconsidered. A better implementation would
1117 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1118
1119 for (int i = 1; i < _BLOCKSIZE; i++) {
1120 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1121 }
1122
1123 // terminate the last monitor as the end of list
1124 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1125
1126 // Element [0] is reserved for global list linkage
1127 temp[0].set_object(CHAINMARKER);
1128
1129 // Consider carving out this thread's current request from the
1130 // block in hand. This avoids some lock traffic and redundant
1131 // list activity.
1132
1133 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1134 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1135 Thread::muxAcquire(&gListLock, "omAlloc(2)");
1136 gMonitorPopulation += _BLOCKSIZE-1;
1137 gMonitorFreeCount += _BLOCKSIZE-1;
1138
1139 // Add the new block to the list of extant blocks (gBlockList).
1140 // The very first objectMonitor in a block is reserved and dedicated.
1143 // There are lock-free uses of gBlockList so make sure that
1144 // the previous stores happen before we update gBlockList.
1145 OrderAccess::release_store(&gBlockList, temp);
1146
1147 // Add the new string of objectMonitors to the global free list
1148 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1149 gFreeList = temp + 1;
1150 Thread::muxRelease(&gListLock);
1151 }
1152 }
1153
1154 // Place "m" on the caller's private per-thread omFreeList.
1155 // In practice there's no need to clamp or limit the number of
1156 // monitors on a thread's omFreeList as the only time we'll call
1157 // omRelease is to return a monitor to the free list after a CAS
1158 // attempt failed. This doesn't allow unbounded #s of monitors to
1159 // accumulate on a thread's free list.
1160 //
1161 // Key constraint: all ObjectMonitors on a thread's free list and the global
1162 // free list must have their object field set to null. This prevents the
1163 // scavenger -- deflate_monitor_list() -- from reclaiming them.
1164
1165 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1166 bool fromPerThreadAlloc) {
1167 guarantee(m->header() == NULL, "invariant");
1168 guarantee(m->object() == NULL, "invariant");
1169 stringStream ss;
1170 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: "
1171 "%s, recursions=" INTPTR_FORMAT, m->is_busy_to_string(&ss),
1172 m->_recursions);
1173 // Remove from omInUseList
1174 if (fromPerThreadAlloc) {
1175 ObjectMonitor* cur_mid_in_use = NULL;
1176 bool extracted = false;
1177 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1178 if (m == mid) {
1179 // extract from per-thread in-use list
1180 if (mid == Self->omInUseList) {
1181 Self->omInUseList = mid->FreeNext;
1182 } else if (cur_mid_in_use != NULL) {
1183 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1184 }
1185 extracted = true;
1186 Self->omInUseCount--;
1187 break;
1188 }
1189 }
1190 assert(extracted, "Should have extracted from in-use list");
1191 }
1192
1193 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1194 m->FreeNext = Self->omFreeList;
1195 Self->omFreeList = m;
1196 Self->omFreeCount++;
1197 }
1198
1199 // Return the monitors of a moribund thread's local free list to
1200 // the global free list. Typically a thread calls omFlush() when
1201 // it's dying. We could also consider having the VM thread steal
1202 // monitors from threads that have not run java code over a few
1203 // consecutive STW safepoints. Relatedly, we might decay
1204 // omFreeProvision at STW safepoints.
1205 //
1206 // Also return the monitors of a moribund thread's omInUseList to
1207 // a global gOmInUseList under the global list lock so these
1208 // will continue to be scanned.
1209 //
1210 // We currently call omFlush() from Threads::remove() _before the thread
1211 // has been excised from the thread list and is no longer a mutator.
1212 // This means that omFlush() cannot run concurrently with a safepoint and
1213 // interleave with the deflate_idle_monitors scavenge operator. In particular,
1214 // this ensures that the thread's monitors are scanned by a GC safepoint,
1215 // either via Thread::oops_do() (if safepoint happens before omFlush()) or via
1216 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1217 // monitors have been transferred to the global in-use list).
1218
1219 void ObjectSynchronizer::omFlush(Thread * Self) {
1220 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1221 ObjectMonitor * tail = NULL;
1222 int tally = 0;
1223 if (list != NULL) {
1224 ObjectMonitor * s;
1225 // The thread is going away. Set 'tail' to the last per-thread free
1226 // monitor which will be linked to gFreeList below under the gListLock.
1227 stringStream ss;
1228 for (s = list; s != NULL; s = s->FreeNext) {
1229 tally++;
1230 tail = s;
1231 guarantee(s->object() == NULL, "invariant");
1232 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss));
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(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->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1383
1384 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1385 if (cmp != mark) {
1386 omRelease(Self, m, true);
1387 continue; // Interference -- just retry
1388 }
1389
1390 // We've successfully installed INFLATING (0) into the mark-word.
1391 // This is the only case where 0 will appear in a mark-word.
1392 // Only the singular thread that successfully swings the mark-word
1393 // to 0 can perform (or more precisely, complete) inflation.
1394 //
1395 // Why do we CAS a 0 into the mark-word instead of just CASing the
1396 // mark-word from the stack-locked value directly to the new inflated state?
1403 // value from the basiclock on the owner's stack to the objectMonitor, all
1404 // the while preserving the hashCode stability invariants. If the owner
1405 // decides to release the lock while the value is 0, the unlock will fail
1406 // and control will eventually pass from slow_exit() to inflate. The owner
1407 // will then spin, waiting for the 0 value to disappear. Put another way,
1408 // the 0 causes the owner to stall if the owner happens to try to
1409 // drop the lock (restoring the header from the basiclock to the object)
1410 // while inflation is in-progress. This protocol avoids races that might
1411 // would otherwise permit hashCode values to change or "flicker" for an object.
1412 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1413 // 0 serves as a "BUSY" inflate-in-progress indicator.
1414
1415
1416 // fetch the displaced mark from the owner's stack.
1417 // The owner can't die or unwind past the lock while our INFLATING
1418 // object is in the mark. Furthermore the owner can't complete
1419 // an unlock on the object, either.
1420 markOop dmw = mark->displaced_mark_helper();
1421 // Catch if the object's header is not neutral (not locked and
1422 // not marked is what we care about here).
1423 assert(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(dmw));
1424
1425 // Setup monitor fields to proper values -- prepare the monitor
1426 m->set_header(dmw);
1427
1428 // Optimization: if the mark->locker stack address is associated
1429 // with this thread we could simply set m->_owner = Self.
1430 // Note that a thread can inflate an object
1431 // that it has stack-locked -- as might happen in wait() -- directly
1432 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1433 m->set_owner(mark->locker());
1434 m->set_object(object);
1435 // TODO-FIXME: assert BasicLock->dhw != 0.
1436
1437 // Must preserve store ordering. The monitor state must
1438 // be stable at the time of publishing the monitor address.
1439 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1440 object->release_set_mark(markOopDesc::encode(m));
1441
1442 // Hopefully the performance counters are allocated on distinct cache lines
1443 // to avoid false sharing on MP systems ...
1444 OM_PERFDATA_OP(Inflations, inc());
1445 if (log_is_enabled(Trace, monitorinflation)) {
1446 ResourceMark rm(Self);
1447 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1448 INTPTR_FORMAT ", type='%s'", p2i(object),
1449 p2i(object->mark()), object->klass()->external_name());
1450 }
1451 if (event.should_commit()) {
1452 post_monitor_inflate_event(&event, object, cause);
1453 }
1454 return m;
1455 }
1456
1457 // CASE: neutral
1458 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1459 // If we know we're inflating for entry it's better to inflate by swinging a
1460 // pre-locked objectMonitor pointer into the object header. A successful
1461 // CAS inflates the object *and* confers ownership to the inflating thread.
1462 // In the current implementation we use a 2-step mechanism where we CAS()
1463 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1464 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1465 // would be useful.
1466
1467 // Catch if the object's header is not neutral (not locked and
1468 // not marked is what we care about here).
1469 assert(mark->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(mark));
1470 ObjectMonitor * m = omAlloc(Self);
1471 // prepare m for installation - set monitor to initial state
1472 m->Recycle();
1473 m->set_header(mark);
1474 m->set_object(object);
1475 m->_Responsible = NULL;
1476 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1477
1478 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1479 m->set_header(NULL);
1480 m->set_object(NULL);
1481 m->Recycle();
1482 omRelease(Self, m, true);
1483 m = NULL;
1484 continue;
1485 // interference - the markword changed - just retry.
1486 // The state-transitions are one-way, so there's no chance of
1487 // live-lock -- "Inflated" is an absorbing state.
1488 }
1489
1490 // Hopefully the performance counters are allocated on distinct
1491 // cache lines to avoid false sharing on MP systems ...
1492 OM_PERFDATA_OP(Inflations, inc());
1493 if (log_is_enabled(Trace, monitorinflation)) {
1494 ResourceMark rm(Self);
1495 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1496 INTPTR_FORMAT ", type='%s'", p2i(object),
1497 p2i(object->mark()), object->klass()->external_name());
1498 }
1499 if (event.should_commit()) {
1500 post_monitor_inflate_event(&event, object, cause);
1501 }
1502 return m;
1503 }
1504 }
1505
1506
1507 // We maintain a list of in-use monitors for each thread.
1508 //
1509 // deflate_thread_local_monitors() scans a single thread's in-use list, while
1510 // deflate_idle_monitors() scans only a global list of in-use monitors which
1511 // is populated only as a thread dies (see omFlush()).
1512 //
1513 // These operations are called at all safepoints, immediately after mutators
1514 // are stopped, but before any objects have moved. Collectively they traverse
1515 // the population of in-use monitors, deflating where possible. The scavenged
1516 // monitors are returned to the global monitor free list.
1517 //
1518 // Beware that we scavenge at *every* stop-the-world point. Having a large
1519 // number of monitors in-use could negatively impact performance. We also want
1520 // to minimize the total # of monitors in circulation, as they incur a small
1521 // footprint penalty.
1522 //
1523 // Perversely, the heap size -- and thus the STW safepoint rate --
1524 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1525 // which in turn can mean large(r) numbers of ObjectMonitors in circulation.
1526 // This is an unfortunate aspect of this design.
1527
1528 // Deflate a single monitor if not in-use
1529 // Return true if deflated, false if in-use
1530 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1531 ObjectMonitor** freeHeadp,
1532 ObjectMonitor** freeTailp) {
1533 bool deflated;
1534 // Normal case ... The monitor is associated with obj.
1535 const markOop mark = obj->mark();
1536 guarantee(mark == markOopDesc::encode(mid), "should match: mark="
1537 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, p2i(mark),
1538 p2i(markOopDesc::encode(mid)));
1539 // Make sure that mark->monitor() and markOopDesc::encode() agree:
1540 guarantee(mark->monitor() == mid, "should match: monitor()=" INTPTR_FORMAT
1541 ", mid=" INTPTR_FORMAT, p2i(mark->monitor()), p2i(mid));
1542 const markOop dmw = mid->header();
1543 guarantee(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(dmw));
1544
1545 if (mid->is_busy()) {
1546 deflated = false;
1547 } else {
1548 // Deflate the monitor if it is no longer being used
1549 // It's idle - scavenge and return to the global free list
1550 // plain old deflation ...
1551 if (log_is_enabled(Trace, monitorinflation)) {
1552 ResourceMark rm;
1553 log_trace(monitorinflation)("deflate_monitor: "
1554 "object=" INTPTR_FORMAT ", mark="
1555 INTPTR_FORMAT ", type='%s'", p2i(obj),
1556 p2i(mark), obj->klass()->external_name());
1557 }
1558
1559 // Restore the header back to obj
1560 obj->release_set_mark(dmw);
1561 mid->clear();
1562
1563 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT,
1564 p2i(mid->object()));
1565
1566 // Move the object to the working free list defined by freeHeadp, freeTailp
1567 if (*freeHeadp == NULL) *freeHeadp = mid;
1568 if (*freeTailp != NULL) {
1569 ObjectMonitor * prevtail = *freeTailp;
1570 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1571 prevtail->FreeNext = mid;
1572 }
1573 *freeTailp = mid;
1574 deflated = true;
1575 }
1576 return deflated;
1577 }
1578
1579 // Walk a given monitor list, and deflate idle monitors
1580 // The given list could be a per-thread list or a global list
1581 // Caller acquires gListLock as needed.
1582 //
1583 // In the case of parallel processing of thread local monitor lists,
1584 // work is done by Threads::parallel_threads_do() which ensures that
1585 // each Java thread is processed by exactly one worker thread, and
1586 // thus avoid conflicts that would arise when worker threads would
1587 // process the same monitor lists concurrently.
1588 //
1589 // See also ParallelSPCleanupTask and
1590 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1591 // Threads::parallel_java_threads_do() in thread.cpp.
1592 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1593 ObjectMonitor** freeHeadp,
1594 ObjectMonitor** freeTailp) {
1595 ObjectMonitor* mid;
1596 ObjectMonitor* next;
1597 ObjectMonitor* cur_mid_in_use = NULL;
1598 int deflated_count = 0;
1602 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1603 // if deflate_monitor succeeded,
1604 // extract from per-thread in-use list
1605 if (mid == *listHeadp) {
1606 *listHeadp = mid->FreeNext;
1607 } else if (cur_mid_in_use != NULL) {
1608 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1609 }
1610 next = mid->FreeNext;
1611 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1612 mid = next;
1613 deflated_count++;
1614 } else {
1615 cur_mid_in_use = mid;
1616 mid = mid->FreeNext;
1617 }
1618 }
1619 return deflated_count;
1620 }
1621
1622 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1623 counters->nInuse = 0; // currently associated with objects
1624 counters->nInCirculation = 0; // extant
1625 counters->nScavenged = 0; // reclaimed (global and per-thread)
1626 counters->perThreadScavenged = 0; // per-thread scavenge total
1627 counters->perThreadTimes = 0.0; // per-thread scavenge times
1628 }
1629
1630 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
1631 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1632 bool deflated = false;
1633
1634 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1635 ObjectMonitor * freeTailp = NULL;
1636 elapsedTimer timer;
1637
1638 if (log_is_enabled(Info, monitorinflation)) {
1639 timer.start();
1640 }
1641
1642 // Prevent omFlush from changing mids in Thread dtor's during deflation
1643 // And in case the vm thread is acquiring a lock during a safepoint
1644 // See e.g. 6320749
1645 Thread::muxAcquire(&gListLock, "deflate_idle_monitors");
1646
1647 // Note: the thread-local monitors lists get deflated in
1648 // a separate pass. See deflate_thread_local_monitors().
1649
1650 // For moribund threads, scan gOmInUseList
1651 int deflated_count = 0;
1664 // constant-time list splice - prepend scavenged segment to gFreeList
1665 freeTailp->FreeNext = gFreeList;
1666 gFreeList = freeHeadp;
1667 }
1668 Thread::muxRelease(&gListLock);
1669 timer.stop();
1670
1671 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1672 LogStreamHandle(Info, monitorinflation) lsh_info;
1673 LogStream * ls = NULL;
1674 if (log_is_enabled(Debug, monitorinflation)) {
1675 ls = &lsh_debug;
1676 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
1677 ls = &lsh_info;
1678 }
1679 if (ls != NULL) {
1680 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
1681 }
1682 }
1683
1684 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1685 // Report the cumulative time for deflating each thread's idle
1686 // monitors. Note: if the work is split among more than one
1687 // worker thread, then the reported time will likely be more
1688 // than a beginning to end measurement of the phase.
1689 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->perThreadTimes, counters->perThreadScavenged);
1690
1691 gMonitorFreeCount += counters->nScavenged;
1692
1693 if (log_is_enabled(Debug, monitorinflation)) {
1694 // exit_globals()'s call to audit_and_print_stats() is done
1695 // at the Info level.
1696 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */);
1697 } else if (log_is_enabled(Info, monitorinflation)) {
1698 Thread::muxAcquire(&gListLock, "finish_deflate_idle_monitors");
1699 log_info(monitorinflation)("gMonitorPopulation=%d, gOmInUseCount=%d, "
1700 "gMonitorFreeCount=%d", gMonitorPopulation,
1701 gOmInUseCount, gMonitorFreeCount);
1702 Thread::muxRelease(&gListLock);
1703 }
1704
1705 ForceMonitorScavenge = 0; // Reset
1706
1707 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
1708 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
1709
1710 GVars.stwRandom = os::random();
1711 GVars.stwCycle++;
1712 }
1713
1714 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
1715 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1716
1717 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1718 ObjectMonitor * freeTailp = NULL;
1719 elapsedTimer timer;
1720
1721 if (log_is_enabled(Info, safepoint, cleanup) ||
1722 log_is_enabled(Info, monitorinflation)) {
1723 timer.start();
1724 }
1725
1726 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
1727
1728 Thread::muxAcquire(&gListLock, "deflate_thread_local_monitors");
1729
1730 // Adjust counters
1731 counters->nInCirculation += thread->omInUseCount;
1732 thread->omInUseCount -= deflated_count;
1733 counters->nScavenged += deflated_count;
1734 counters->nInuse += thread->omInUseCount;
1735 counters->perThreadScavenged += deflated_count;
1736
1923 void ObjectSynchronizer::chk_free_entry(JavaThread * jt, ObjectMonitor * n,
1924 outputStream * out, int *error_cnt_p) {
1925 stringStream ss;
1926 if (n->is_busy()) {
1927 if (jt != NULL) {
1928 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1929 ": free per-thread monitor must not be busy: %s", p2i(jt),
1930 p2i(n), n->is_busy_to_string(&ss));
1931 } else {
1932 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1933 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss));
1934 }
1935 *error_cnt_p = *error_cnt_p + 1;
1936 }
1937 if (n->header() != NULL) {
1938 if (jt != NULL) {
1939 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1940 ": free per-thread monitor must have NULL _header "
1941 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n),
1942 p2i(n->header()));
1943 } else {
1944 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1945 "must have NULL _header field: _header=" INTPTR_FORMAT,
1946 p2i(n), p2i(n->header()));
1947 }
1948 *error_cnt_p = *error_cnt_p + 1;
1949 }
1950 if (n->object() != NULL) {
1951 if (jt != NULL) {
1952 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1953 ": free per-thread monitor must have NULL _object "
1954 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n),
1955 p2i(n->object()));
1956 } else {
1957 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1958 "must have NULL _object field: _object=" INTPTR_FORMAT,
1959 p2i(n), p2i(n->object()));
1960 }
1961 *error_cnt_p = *error_cnt_p + 1;
1962 }
1963 }
1964
1965 // Check the global free list and count; log the results of the checks.
1966 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out,
1967 int *error_cnt_p) {
1968 int chkMonitorFreeCount = 0;
1969 for (ObjectMonitor * n = gFreeList; n != NULL; n = n->FreeNext) {
2096 "equal to chkOmInUseCount=%d", p2i(jt), jt->omInUseCount,
2097 chkOmInUseCount);
2098 *error_cnt_p = *error_cnt_p + 1;
2099 }
2100 }
2101
2102 // Log details about ObjectMonitors on the in-use lists. The 'BHL'
2103 // flags indicate why the entry is in-use, 'object' and 'object type'
2104 // indicate the associated object and its type.
2105 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out,
2106 bool on_exit) {
2107 if (!on_exit) {
2108 // Not at VM exit so grab the global list lock.
2109 Thread::muxAcquire(&gListLock, "log_in_use_monitor_details");
2110 }
2111
2112 stringStream ss;
2113 if (gOmInUseCount > 0) {
2114 out->print_cr("In-use global monitor info:");
2115 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2116 out->print_cr("%18s %s %18s %18s",
2117 "monitor", "BHL", "object", "object type");
2118 out->print_cr("================== === ================== ==================");
2119 for (ObjectMonitor * n = gOmInUseList; n != NULL; n = n->FreeNext) {
2120 const oop obj = (oop) n->object();
2121 const markOop mark = n->header();
2122 ResourceMark rm;
2123 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(n),
2124 n->is_busy() != 0, mark->hash() != 0, n->owner() != NULL,
2125 p2i(obj), obj->klass()->external_name());
2126 if (n->is_busy() != 0) {
2127 out->print(" (%s)", n->is_busy_to_string(&ss));
2128 ss.reset();
2129 }
2130 out->cr();
2131 }
2132 }
2133
2134 if (!on_exit) {
2135 Thread::muxRelease(&gListLock);
2136 }
2137
2138 out->print_cr("In-use per-thread monitor info:");
2139 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2140 out->print_cr("%18s %18s %s %18s %18s",
2141 "jt", "monitor", "BHL", "object", "object type");
2142 out->print_cr("================== ================== === ================== ==================");
2143 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2144 for (ObjectMonitor * n = jt->omInUseList; n != NULL; n = n->FreeNext) {
2145 const oop obj = (oop) n->object();
2146 const markOop mark = n->header();
2147 ResourceMark rm;
2148 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT
2149 " %s", p2i(jt), p2i(n), n->is_busy() != 0,
2150 mark->hash() != 0, n->owner() != NULL, p2i(obj),
2151 obj->klass()->external_name());
2152 if (n->is_busy() != 0) {
2153 out->print(" (%s)", n->is_busy_to_string(&ss));
2154 ss.reset();
2155 }
2156 out->cr();
2157 }
2158 }
2159
2160 out->flush();
2161 }
2162
2163 // Log counts for the global and per-thread monitor lists and return
2164 // the population count.
2165 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) {
2166 int popCount = 0;
2167 out->print_cr("%18s %10s %10s %10s",
2168 "Global Lists:", "InUse", "Free", "Total");
2169 out->print_cr("================== ========== ========== ==========");
2170 out->print_cr("%18s %10d %10d %10d", "",
2171 gOmInUseCount, gMonitorFreeCount, gMonitorPopulation);
|
108 #endif // ndef DTRACE_ENABLED
109
110 // This exists only as a workaround of dtrace bug 6254741
111 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
112 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
113 return 0;
114 }
115
116 #define NINFLATIONLOCKS 256
117 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
118
119 // global list of blocks of monitors
120 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
121 // global monitor free list
122 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
123 // global monitor in-use list, for moribund threads,
124 // monitors they inflated need to be scanned for deflation
125 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
126 // count of entries in gOmInUseList
127 int ObjectSynchronizer::gOmInUseCount = 0;
128 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
129 bool volatile ObjectSynchronizer::_is_special_deflation_requested = false;
130 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
131
132 static volatile intptr_t gListLock = 0; // protects global monitor lists
133 static volatile int gMonitorFreeCount = 0; // # on gFreeList
134 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
135
136 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
137
138
139 // =====================> Quick functions
140
141 // The quick_* forms are special fast-path variants used to improve
142 // performance. In the simplest case, a "quick_*" implementation could
143 // simply return false, in which case the caller will perform the necessary
144 // state transitions and call the slow-path form.
145 // The fast-path is designed to handle frequently arising cases in an efficient
146 // manner and is just a degenerate "optimistic" variant of the slow-path.
147 // returns true -- to indicate the call was satisfied.
148 // returns false -- to indicate the call needs the services of the slow-path.
149 // A no-loitering ordinance is in effect for code in the quick_* family
150 // operators: safepoints or indefinite blocking (blocking that might span a
197 }
198
199 // biased locking and any other IMS exception states take the slow-path
200 return false;
201 }
202
203
204 // The LockNode emitted directly at the synchronization site would have
205 // been too big if it were to have included support for the cases of inflated
206 // recursive enter and exit, so they go here instead.
207 // Note that we can't safely call AsyncPrintJavaStack() from within
208 // quick_enter() as our thread state remains _in_Java.
209
210 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
211 BasicLock * lock) {
212 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
213 assert(Self->is_Java_thread(), "invariant");
214 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
215 NoSafepointVerifier nsv;
216 if (obj == NULL) return false; // Need to throw NPE
217
218 while (true) {
219 const markOop mark = obj->mark();
220
221 if (mark->has_monitor()) {
222 ObjectMonitorHandle omh;
223 if (!omh.save_om_ptr(obj, mark)) {
224 // Lost a race with async deflation so try again.
225 assert(AsyncDeflateIdleMonitors, "sanity check");
226 continue;
227 }
228 ObjectMonitor * const m = omh.om_ptr();
229 assert(oopDesc::equals((oop) m->object(), obj), "invariant");
230 Thread * const owner = (Thread *) m->_owner;
231
232 // Lock contention and Transactional Lock Elision (TLE) diagnostics
233 // and observability
234 // Case: light contention possibly amenable to TLE
235 // Case: TLE inimical operations such as nested/recursive synchronization
236
237 if (owner == Self) {
238 m->_recursions++;
239 return true;
240 }
241
242 // This Java Monitor is inflated so obj's header will never be
243 // displaced to this thread's BasicLock. Make the displaced header
244 // non-NULL so this BasicLock is not seen as recursive nor as
245 // being locked. We do this unconditionally so that this thread's
246 // BasicLock cannot be mis-interpreted by any stack walkers. For
247 // performance reasons, stack walkers generally first check for
248 // Biased Locking in the object's header, the second check is for
249 // stack-locking in the object's header, the third check is for
250 // recursive stack-locking in the displaced header in the BasicLock,
251 // and last are the inflated Java Monitor (ObjectMonitor) checks.
252 lock->set_displaced_header(markOopDesc::unused_mark());
253
254 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
255 assert(m->_recursions == 0, "invariant");
256 return true;
257 }
258
259 if (AsyncDeflateIdleMonitors &&
260 Atomic::cmpxchg(Self, &m->_owner, DEFLATER_MARKER) == DEFLATER_MARKER) {
261 // The deflation protocol finished the first part (setting owner),
262 // but it failed the second part (making ref_count negative) and
263 // bailed. Or the ObjectMonitor was async deflated and reused.
264 // Acquired the monitor.
265 assert(m->_recursions == 0, "invariant");
266 return true;
267 }
268 }
269 break;
270 }
271
272 // Note that we could inflate in quick_enter.
273 // This is likely a useful optimization
274 // Critically, in quick_enter() we must not:
275 // -- perform bias revocation, or
276 // -- block indefinitely, or
277 // -- reach a safepoint
278
279 return false; // revert to slow-path
280 }
281
282 // -----------------------------------------------------------------------------
283 // Fast Monitor Enter/Exit
284 // This the fast monitor enter. The interpreter and compiler use
285 // some assembly copies of this code. Make sure update those code
286 // if the following function is changed. The implementation is
287 // extremely sensitive to race condition. Be careful.
288
289 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
333 // does not own the Java Monitor.
334 ObjectMonitor * m = mark->monitor();
335 assert(((oop)(m->object()))->mark() == mark, "invariant");
336 assert(m->is_entered(THREAD), "invariant");
337 }
338 }
339 #endif
340 return;
341 }
342
343 if (mark == (markOop) lock) {
344 // If the object is stack-locked by the current thread, try to
345 // swing the displaced header from the BasicLock back to the mark.
346 assert(dhw->is_neutral(), "invariant");
347 if (object->cas_set_mark(dhw, mark) == mark) {
348 return;
349 }
350 }
351
352 // We have to take the slow-path of possible inflation and then exit.
353 ObjectMonitorHandle omh;
354 inflate(&omh, THREAD, object, inflate_cause_vm_internal);
355 omh.om_ptr()->exit(true, THREAD);
356 }
357
358 // -----------------------------------------------------------------------------
359 // Interpreter/Compiler Slow Case
360 // This routine is used to handle interpreter/compiler slow case
361 // We don't need to use fast path here, because it must have been
362 // failed in the interpreter/compiler code.
363 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
364 markOop mark = obj->mark();
365 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
366
367 if (mark->is_neutral()) {
368 // Anticipate successful CAS -- the ST of the displaced mark must
369 // be visible <= the ST performed by the CAS.
370 lock->set_displaced_header(mark);
371 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
372 return;
373 }
374 // Fall through to inflate() ...
375 } else if (mark->has_locker() &&
376 THREAD->is_lock_owned((address)mark->locker())) {
377 assert(lock != mark->locker(), "must not re-lock the same lock");
378 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
379 lock->set_displaced_header(NULL);
380 return;
381 }
382
383 // The object header will never be displaced to this lock,
384 // so it does not matter what the value is, except that it
385 // must be non-zero to avoid looking like a re-entrant lock,
386 // and must not look locked either.
387 lock->set_displaced_header(markOopDesc::unused_mark());
388 ObjectMonitorHandle omh;
389 inflate(&omh, THREAD, obj(), inflate_cause_monitor_enter);
390 omh.om_ptr()->enter(THREAD);
391 }
392
393 // This routine is used to handle interpreter/compiler slow case
394 // We don't need to use fast path here, because it must have
395 // failed in the interpreter/compiler code. Simply use the heavy
396 // weight monitor should be ok, unless someone find otherwise.
397 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
398 fast_exit(object, lock, THREAD);
399 }
400
401 // -----------------------------------------------------------------------------
402 // Class Loader support to workaround deadlocks on the class loader lock objects
403 // Also used by GC
404 // complete_exit()/reenter() are used to wait on a nested lock
405 // i.e. to give up an outer lock completely and then re-enter
406 // Used when holding nested locks - lock acquisition order: lock1 then lock2
407 // 1) complete_exit lock1 - saving recursion count
408 // 2) wait on lock2
409 // 3) when notified on lock2, unlock lock2
410 // 4) reenter lock1 with original recursion count
411 // 5) lock lock2
412 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
413 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
414 if (UseBiasedLocking) {
415 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
416 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
417 }
418
419 ObjectMonitorHandle omh;
420 inflate(&omh, THREAD, obj(), inflate_cause_vm_internal);
421 intptr_t ret_code = omh.om_ptr()->complete_exit(THREAD);
422 return ret_code;
423 }
424
425 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
426 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
427 if (UseBiasedLocking) {
428 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
429 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
430 }
431
432 ObjectMonitorHandle omh;
433 inflate(&omh, THREAD, obj(), inflate_cause_vm_internal);
434 omh.om_ptr()->reenter(recursion, THREAD);
435 }
436 // -----------------------------------------------------------------------------
437 // JNI locks on java objects
438 // NOTE: must use heavy weight monitor to handle jni monitor enter
439 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
440 // the current locking is from JNI instead of Java code
441 if (UseBiasedLocking) {
442 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
443 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
444 }
445 THREAD->set_current_pending_monitor_is_from_java(false);
446 ObjectMonitorHandle omh;
447 inflate(&omh, THREAD, obj(), inflate_cause_jni_enter);
448 omh.om_ptr()->enter(THREAD);
449 THREAD->set_current_pending_monitor_is_from_java(true);
450 }
451
452 // NOTE: must use heavy weight monitor to handle jni monitor exit
453 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
454 if (UseBiasedLocking) {
455 Handle h_obj(THREAD, obj);
456 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
457 obj = h_obj();
458 }
459 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
460
461 ObjectMonitorHandle omh;
462 inflate(&omh, THREAD, obj, inflate_cause_jni_exit);
463 ObjectMonitor * monitor = omh.om_ptr();
464 // If this thread has locked the object, exit the monitor. Note: can't use
465 // monitor->check(CHECK); must exit even if an exception is pending.
466 if (monitor->check(THREAD)) {
467 monitor->exit(true, THREAD);
468 }
469 }
470
471 // -----------------------------------------------------------------------------
472 // Internal VM locks on java objects
473 // standard constructor, allows locking failures
474 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
475 _dolock = doLock;
476 _thread = thread;
477 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
478 _obj = obj;
479
480 if (_dolock) {
481 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
482 }
483 }
484
485 ObjectLocker::~ObjectLocker() {
486 if (_dolock) {
487 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
488 }
489 }
490
491
492 // -----------------------------------------------------------------------------
493 // Wait/Notify/NotifyAll
494 // NOTE: must use heavy weight monitor to handle wait()
495 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
496 if (UseBiasedLocking) {
497 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
498 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
499 }
500 if (millis < 0) {
501 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
502 }
503 ObjectMonitorHandle omh;
504 inflate(&omh, THREAD, obj(), inflate_cause_wait);
505 ObjectMonitor * monitor = omh.om_ptr();
506
507 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
508 monitor->wait(millis, true, THREAD);
509
510 // This dummy call is in place to get around dtrace bug 6254741. Once
511 // that's fixed we can uncomment the following line, remove the call
512 // and change this function back into a "void" func.
513 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
514 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
515 return ret_code;
516 }
517
518 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
519 if (UseBiasedLocking) {
520 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
521 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
522 }
523 if (millis < 0) {
524 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
525 }
526 ObjectMonitorHandle omh;
527 inflate(&omh, THREAD, obj(), inflate_cause_wait);
528 omh.om_ptr()->wait(millis, false, THREAD);
529 }
530
531 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
532 if (UseBiasedLocking) {
533 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
534 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
535 }
536
537 markOop mark = obj->mark();
538 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
539 return;
540 }
541 ObjectMonitorHandle omh;
542 inflate(&omh, THREAD, obj(), inflate_cause_notify);
543 omh.om_ptr()->notify(THREAD);
544 }
545
546 // NOTE: see comment of notify()
547 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
548 if (UseBiasedLocking) {
549 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
550 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
551 }
552
553 markOop mark = obj->mark();
554 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
555 return;
556 }
557 ObjectMonitorHandle omh;
558 inflate(&omh, THREAD, obj(), inflate_cause_notify);
559 omh.om_ptr()->notifyAll(THREAD);
560 }
561
562 // -----------------------------------------------------------------------------
563 // Hash Code handling
564 //
565 // Performance concern:
566 // OrderAccess::storestore() calls release() which at one time stored 0
567 // into the global volatile OrderAccess::dummy variable. This store was
568 // unnecessary for correctness. Many threads storing into a common location
569 // causes considerable cache migration or "sloshing" on large SMP systems.
570 // As such, I avoided using OrderAccess::storestore(). In some cases
571 // OrderAccess::fence() -- which incurs local latency on the executing
572 // processor -- is a better choice as it scales on SMP systems.
573 //
574 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
575 // a discussion of coherency costs. Note that all our current reference
576 // platforms provide strong ST-ST order, so the issue is moot on IA32,
577 // x64, and SPARC.
578 //
579 // As a general policy we use "volatile" to control compiler-based reordering
733 Handle hobj(Self, obj);
734 // Relaxing assertion for bug 6320749.
735 assert(Universe::verify_in_progress() ||
736 !SafepointSynchronize::is_at_safepoint(),
737 "biases should not be seen by VM thread here");
738 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
739 obj = hobj();
740 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
741 }
742 }
743
744 // hashCode() is a heap mutator ...
745 // Relaxing assertion for bug 6320749.
746 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
747 !SafepointSynchronize::is_at_safepoint(), "invariant");
748 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
749 Self->is_Java_thread() , "invariant");
750 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
751 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
752
753 while (true) {
754 ObjectMonitor* monitor = NULL;
755 markOop temp, test;
756 intptr_t hash;
757 markOop mark = ReadStableMark(obj);
758
759 // object should remain ineligible for biased locking
760 assert(!mark->has_bias_pattern(), "invariant");
761
762 if (mark->is_neutral()) {
763 hash = mark->hash(); // this is a normal header
764 if (hash != 0) { // if it has hash, just return it
765 return hash;
766 }
767 hash = get_next_hash(Self, obj); // allocate a new hash code
768 temp = mark->copy_set_hash(hash); // merge the hash code into header
769 // use (machine word version) atomic operation to install the hash
770 test = obj->cas_set_mark(temp, mark);
771 if (test == mark) {
772 return hash;
773 }
774 // If atomic operation failed, we must inflate the header
775 // into heavy weight monitor. We could add more code here
776 // for fast path, but it does not worth the complexity.
777 } else if (mark->has_monitor()) {
778 ObjectMonitorHandle omh;
779 if (!omh.save_om_ptr(obj, mark)) {
780 // Lost a race with async deflation so try again.
781 assert(AsyncDeflateIdleMonitors, "sanity check");
782 continue;
783 }
784 monitor = omh.om_ptr();
785 temp = monitor->header();
786 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(temp));
787 hash = temp->hash();
788 if (hash != 0) {
789 return hash;
790 }
791 // Skip to the following code to reduce code size
792 } else if (Self->is_lock_owned((address)mark->locker())) {
793 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
794 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(temp));
795 hash = temp->hash(); // by current thread, check if the displaced
796 if (hash != 0) { // header contains hash code
797 return hash;
798 }
799 // WARNING:
800 // The displaced header in the BasicLock on a thread's stack
801 // is strictly immutable. It CANNOT be changed in ANY cases.
802 // So we have to inflate the stack lock into an ObjectMonitor
803 // even if the current thread owns the lock. The BasicLock on
804 // a thread's stack can be asynchronously read by other threads
805 // during an inflate() call so any change to that stack memory
806 // may not propagate to other threads correctly.
807 }
808
809 // Inflate the monitor to set hash code
810 ObjectMonitorHandle omh;
811 inflate(&omh, Self, obj, inflate_cause_hash_code);
812 monitor = omh.om_ptr();
813 // Load displaced header and check it has hash code
814 mark = monitor->header();
815 assert(mark->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(mark));
816 hash = mark->hash();
817 if (hash == 0) {
818 hash = get_next_hash(Self, obj);
819 temp = mark->copy_set_hash(hash); // merge hash code into header
820 assert(temp->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(temp));
821 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark);
822 if (test != mark) {
823 // The only non-deflation update to the ObjectMonitor's
824 // header/dmw field is to merge in the hash code. If someone
825 // adds a new usage of the header/dmw field, please update
826 // this code.
827 // ObjectMonitor::install_displaced_markword_in_object()
828 // does mark the header/dmw field as part of async deflation,
829 // but that protocol cannot happen now due to the
830 // ObjectMonitorHandle above.
831 hash = test->hash();
832 assert(test->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(test));
833 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
834 }
835 }
836 // We finally get the hash
837 return hash;
838 }
839 }
840
841 // Deprecated -- use FastHashCode() instead.
842
843 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
844 return FastHashCode(Thread::current(), obj());
845 }
846
847
848 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
849 Handle h_obj) {
850 if (UseBiasedLocking) {
851 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
852 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
853 }
854
855 assert(thread == JavaThread::current(), "Can only be called on current thread");
856 oop obj = h_obj();
857
858 while (true) {
859 markOop mark = ReadStableMark(obj);
860
861 // Uncontended case, header points to stack
862 if (mark->has_locker()) {
863 return thread->is_lock_owned((address)mark->locker());
864 }
865 // Contended case, header points to ObjectMonitor (tagged pointer)
866 if (mark->has_monitor()) {
867 ObjectMonitorHandle omh;
868 if (!omh.save_om_ptr(obj, mark)) {
869 // Lost a race with async deflation so try again.
870 assert(AsyncDeflateIdleMonitors, "sanity check");
871 continue;
872 }
873 bool ret_code = omh.om_ptr()->is_entered(thread) != 0;
874 return ret_code;
875 }
876 // Unlocked case, header in place
877 assert(mark->is_neutral(), "sanity check");
878 return false;
879 }
880 }
881
882 // Be aware of this method could revoke bias of the lock object.
883 // This method queries the ownership of the lock handle specified by 'h_obj'.
884 // If the current thread owns the lock, it returns owner_self. If no
885 // thread owns the lock, it returns owner_none. Otherwise, it will return
886 // owner_other.
887 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
888 (JavaThread *self, Handle h_obj) {
889 // The caller must beware this method can revoke bias, and
890 // revocation can result in a safepoint.
891 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
892 assert(self->thread_state() != _thread_blocked, "invariant");
893
894 // Possible mark states: neutral, biased, stack-locked, inflated
895
896 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
897 // CASE: biased
898 BiasedLocking::revoke_and_rebias(h_obj, false, self);
899 assert(!h_obj->mark()->has_bias_pattern(),
900 "biases should be revoked by now");
901 }
902
903 assert(self == JavaThread::current(), "Can only be called on current thread");
904 oop obj = h_obj();
905
906 while (true) {
907 markOop mark = ReadStableMark(obj);
908
909 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
910 if (mark->has_locker()) {
911 return self->is_lock_owned((address)mark->locker()) ?
912 owner_self : owner_other;
913 }
914
915 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor.
916 // The Object:ObjectMonitor relationship is stable as long as we're
917 // not at a safepoint and AsyncDeflateIdleMonitors is false.
918 if (mark->has_monitor()) {
919 ObjectMonitorHandle omh;
920 if (!omh.save_om_ptr(obj, mark)) {
921 // Lost a race with async deflation so try again.
922 assert(AsyncDeflateIdleMonitors, "sanity check");
923 continue;
924 }
925 ObjectMonitor * monitor = omh.om_ptr();
926 void * owner = monitor->_owner;
927 if (owner == NULL) return owner_none;
928 return (owner == self ||
929 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
930 }
931
932 // CASE: neutral
933 assert(mark->is_neutral(), "sanity check");
934 return owner_none; // it's unlocked
935 }
936 }
937
938 // FIXME: jvmti should call this
939 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
940 if (UseBiasedLocking) {
941 if (SafepointSynchronize::is_at_safepoint()) {
942 BiasedLocking::revoke_at_safepoint(h_obj);
943 } else {
944 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
945 }
946 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
947 }
948
949 oop obj = h_obj();
950
951 while (true) {
952 address owner = NULL;
953 markOop mark = ReadStableMark(obj);
954
955 // Uncontended case, header points to stack
956 if (mark->has_locker()) {
957 owner = (address) mark->locker();
958 }
959
960 // Contended case, header points to ObjectMonitor (tagged pointer)
961 else if (mark->has_monitor()) {
962 ObjectMonitorHandle omh;
963 if (!omh.save_om_ptr(obj, mark)) {
964 // Lost a race with async deflation so try again.
965 assert(AsyncDeflateIdleMonitors, "sanity check");
966 continue;
967 }
968 ObjectMonitor* monitor = omh.om_ptr();
969 assert(monitor != NULL, "monitor should be non-null");
970 owner = (address) monitor->owner();
971 }
972
973 if (owner != NULL) {
974 // owning_thread_from_monitor_owner() may also return NULL here
975 return Threads::owning_thread_from_monitor_owner(t_list, owner);
976 }
977
978 // Unlocked case, header in place
979 // Cannot have assertion since this object may have been
980 // locked by another thread when reaching here.
981 // assert(mark->is_neutral(), "sanity check");
982
983 return NULL;
984 }
985 }
986
987 // Visitors ...
988
989 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
990 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
991 while (block != NULL) {
992 assert(block->object() == CHAINMARKER, "must be a block header");
993 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
994 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
995 if (mid->is_active()) {
996 ObjectMonitorHandle omh(mid);
997
998 if (mid->object() == NULL ||
999 (AsyncDeflateIdleMonitors && mid->_owner == DEFLATER_MARKER)) {
1000 // Only process with closure if the object is set.
1001 // For async deflation, race here if monitor is not owned!
1002 // The above ref_count bump (in ObjectMonitorHandle ctr)
1003 // will cause subsequent async deflation to skip it.
1004 // However, previous or concurrent async deflation is a race.
1005 continue;
1006 }
1007 closure->do_monitor(mid);
1008 }
1009 }
1010 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1011 }
1012 }
1013
1014 // Get the next block in the block list.
1015 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
1016 assert(block->object() == CHAINMARKER, "must be a block header");
1017 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
1018 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
1019 return block;
1020 }
1021
1022 static bool monitors_used_above_threshold() {
1023 if (gMonitorPopulation == 0) {
1024 return false;
1025 }
1026 if (MonitorUsedDeflationThreshold > 0) {
1027 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
1028 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation;
1029 return monitor_usage > MonitorUsedDeflationThreshold;
1030 }
1031 return false;
1032 }
1033
1034 // Returns true if MonitorBound is set (> 0) and if the specified
1035 // cnt is > MonitorBound. Otherwise returns false.
1036 static bool is_MonitorBound_exceeded(const int cnt) {
1037 const int mx = MonitorBound;
1038 return mx > 0 && cnt > mx;
1039 }
1040
1041 bool ObjectSynchronizer::is_async_deflation_needed() {
1042 if (!AsyncDeflateIdleMonitors) {
1043 return false;
1044 }
1045 if (is_async_deflation_requested()) {
1046 // Async deflation request.
1047 return true;
1048 }
1049 if (AsyncDeflationInterval > 0 &&
1050 time_since_last_async_deflation_ms() > AsyncDeflationInterval &&
1051 monitors_used_above_threshold()) {
1052 // It's been longer than our specified deflate interval and there
1053 // are too many monitors in use. We don't deflate more frequently
1054 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1055 // in order to not swamp the ServiceThread.
1056 _last_async_deflation_time_ns = os::javaTimeNanos();
1057 return true;
1058 }
1059 if (is_MonitorBound_exceeded(gMonitorPopulation - gMonitorFreeCount)) {
1060 // Not enough ObjectMonitors on the global free list.
1061 return true;
1062 }
1063 return false;
1064 }
1065
1066 bool ObjectSynchronizer::is_safepoint_deflation_needed() {
1067 if (!AsyncDeflateIdleMonitors) {
1068 if (monitors_used_above_threshold()) {
1069 // Too many monitors in use.
1070 return true;
1071 }
1072 return false;
1073 }
1074 if (is_special_deflation_requested()) {
1075 // For AsyncDeflateIdleMonitors only do a safepoint deflation
1076 // if there is a special deflation request.
1077 return true;
1078 }
1079 return false;
1080 }
1081
1082 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1083 return (os::javaTimeNanos() - _last_async_deflation_time_ns) / (NANOUNITS / MILLIUNITS);
1084 }
1085
1086 void ObjectSynchronizer::oops_do(OopClosure* f) {
1087 // We only scan the global used list here (for moribund threads), and
1088 // the thread-local monitors in Thread::oops_do().
1089 global_used_oops_do(f);
1090 }
1091
1092 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
1093 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1094 list_oops_do(gOmInUseList, f);
1095 }
1096
1097 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
1098 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1099 list_oops_do(thread->omInUseList, f);
1100 }
1101
1102 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
1103 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1104 ObjectMonitor* mid;
1105 for (mid = list; mid != NULL; mid = mid->FreeNext) {
1115 // -----------------------
1116 // Inflation unlinks monitors from the global gFreeList and
1117 // associates them with objects. Deflation -- which occurs at
1118 // STW-time -- disassociates idle monitors from objects. Such
1119 // scavenged monitors are returned to the gFreeList.
1120 //
1121 // The global list is protected by gListLock. All the critical sections
1122 // are short and operate in constant-time.
1123 //
1124 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
1125 //
1126 // Lifecycle:
1127 // -- unassigned and on the global free list
1128 // -- unassigned and on a thread's private omFreeList
1129 // -- assigned to an object. The object is inflated and the mark refers
1130 // to the objectmonitor.
1131
1132
1133 // Constraining monitor pool growth via MonitorBound ...
1134 //
1135 // If MonitorBound is not set (<= 0), MonitorBound checks are disabled.
1136 //
1137 // When safepoint deflation is being used (!AsyncDeflateIdleMonitors):
1138 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
1139 // the rate of scavenging is driven primarily by GC. As such, we can find
1140 // an inordinate number of monitors in circulation.
1141 // To avoid that scenario we can artificially induce a STW safepoint
1142 // if the pool appears to be growing past some reasonable bound.
1143 // Generally we favor time in space-time tradeoffs, but as there's no
1144 // natural back-pressure on the # of extant monitors we need to impose some
1145 // type of limit. Beware that if MonitorBound is set to too low a value
1146 // we could just loop. In addition, if MonitorBound is set to a low value
1147 // we'll incur more safepoints, which are harmful to performance.
1148 // See also: GuaranteedSafepointInterval
1149 //
1150 // The current implementation uses asynchronous VM operations.
1151 //
1152 // When safepoint deflation is being used and MonitorBound is set, the
1153 // boundry applies to (gMonitorPopulation - gMonitorFreeCount), i.e.,
1154 // if there are not enough ObjectMonitors on the global free list, then
1155 // a safepoint deflation is induced. Picking a good MonitorBound value
1156 // is non-trivial.
1157 //
1158 // When async deflation is being used:
1159 // The monitor pool is still grow-only. Async deflation is requested
1160 // by a safepoint's cleanup phase or by the ServiceThread at periodic
1161 // intervals when is_async_deflation_needed() returns true. In
1162 // addition to other policies that are checked, if there are not
1163 // enough ObjectMonitors on the global free list, then
1164 // is_async_deflation_needed() will return true. The ServiceThread
1165 // calls deflate_global_idle_monitors_using_JT() and also sets the
1166 // per-thread omShouldDeflateIdleMonitors flag as needed.
1167
1168 static void InduceScavenge(Thread * Self, const char * Whence) {
1169 assert(!AsyncDeflateIdleMonitors, "is not used by async deflation");
1170
1171 // Induce STW safepoint to trim monitors
1172 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1173 // More precisely, trigger an asynchronous STW safepoint as the number
1174 // of active monitors passes the specified threshold.
1175 // TODO: assert thread state is reasonable
1176
1177 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1178 // Induce a 'null' safepoint to scavenge monitors
1179 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1180 // to the VMthread and have a lifespan longer than that of this activation record.
1181 // The VMThread will delete the op when completed.
1182 VMThread::execute(new VM_ScavengeMonitors());
1183 }
1184 }
1185
1186 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self,
1187 const InflateCause cause) {
1188 // A large MAXPRIVATE value reduces both list lock contention
1189 // and list coherency traffic, but also tends to increase the
1190 // number of objectMonitors in circulation as well as the STW
1191 // scavenge costs. As usual, we lean toward time in space-time
1192 // tradeoffs.
1193 const int MAXPRIVATE = 1024;
1194
1195 if (AsyncDeflateIdleMonitors) {
1196 JavaThread * jt = (JavaThread *)Self;
1197 if (jt->omShouldDeflateIdleMonitors && jt->omInUseCount > 0 &&
1198 cause != inflate_cause_vm_internal) {
1199 // Deflate any per-thread idle monitors for this JavaThread if
1200 // this is not an internal inflation; internal inflations can
1201 // occur in places where it is not safe to pause for a safepoint.
1202 // Clean up your own mess (Gibbs Rule 45). Otherwise, skip this
1203 // deflation. deflate_global_idle_monitors_using_JT() is called
1204 // by the ServiceThread. Per-thread async deflation is triggered
1205 // by the ServiceThread via omShouldDeflateIdleMonitors.
1206 debug_only(jt->check_for_valid_safepoint_state(false);)
1207 ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT();
1208 }
1209 }
1210
1211 stringStream ss;
1212 for (;;) {
1213 ObjectMonitor * m;
1214
1215 // 1: try to allocate from the thread's local omFreeList.
1216 // Threads will attempt to allocate first from their local list, then
1217 // from the global list, and only after those attempts fail will the thread
1218 // attempt to instantiate new monitors. Thread-local free lists take
1219 // heat off the gListLock and improve allocation latency, as well as reducing
1220 // coherency traffic on the shared global list.
1221 m = Self->omFreeList;
1222 if (m != NULL) {
1223 Self->omFreeList = m->FreeNext;
1224 Self->omFreeCount--;
1225 guarantee(m->object() == NULL, "invariant");
1226 m->set_allocation_state(ObjectMonitor::New);
1227 m->FreeNext = Self->omInUseList;
1228 Self->omInUseList = m;
1229 Self->omInUseCount++;
1230 return m;
1231 }
1232
1233 // 2: try to allocate from the global gFreeList
1234 // CONSIDER: use muxTry() instead of muxAcquire().
1235 // If the muxTry() fails then drop immediately into case 3.
1236 // If we're using thread-local free lists then try
1237 // to reprovision the caller's free list.
1238 if (gFreeList != NULL) {
1239 // Reprovision the thread's omFreeList.
1240 // Use bulk transfers to reduce the allocation rate and heat
1241 // on various locks.
1242 Thread::muxAcquire(&gListLock, "omAlloc(1)");
1243 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1244 gMonitorFreeCount--;
1245 ObjectMonitor * take = gFreeList;
1246 gFreeList = take->FreeNext;
1247 guarantee(take->object() == NULL, "invariant");
1248 if (AsyncDeflateIdleMonitors) {
1249 // We allowed 3 field values to linger during async deflation.
1250 // We clear header and restore ref_count here, but we leave
1251 // owner == DEFLATER_MARKER so the simple C2 ObjectMonitor
1252 // enter optimization can no longer race with async deflation
1253 // and reuse.
1254 take->_header = NULL;
1255 if (take->ref_count() < 0) {
1256 // Add back max_jint to restore the ref_count field to its
1257 // proper value.
1258 Atomic::add(max_jint, &take->_ref_count);
1259
1260 assert(take->ref_count() >= 0, "must not be negative: ref_count=%d",
1261 take->ref_count());
1262 }
1263 }
1264 take->Recycle();
1265 assert(take->is_free(), "invariant");
1266 omRelease(Self, take, false);
1267 }
1268 Thread::muxRelease(&gListLock);
1269 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1270 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1271
1272 if (!AsyncDeflateIdleMonitors &&
1273 is_MonitorBound_exceeded(gMonitorPopulation - gMonitorFreeCount)) {
1274 // Not enough ObjectMonitors on the global free list.
1275 // We can't safely induce a STW safepoint from omAlloc() as our thread
1276 // state may not be appropriate for such activities and callers may hold
1277 // naked oops, so instead we defer the action.
1278 InduceScavenge(Self, "omAlloc");
1279 }
1280 continue;
1281 }
1282
1283 // 3: allocate a block of new ObjectMonitors
1284 // Both the local and global free lists are empty -- resort to malloc().
1285 // In the current implementation objectMonitors are TSM - immortal.
1286 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1287 // each ObjectMonitor to start at the beginning of a cache line,
1288 // so we use align_up().
1289 // A better solution would be to use C++ placement-new.
1290 // BEWARE: As it stands currently, we don't run the ctors!
1291 assert(_BLOCKSIZE > 1, "invariant");
1292 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE;
1293 PaddedEnd<ObjectMonitor> * temp;
1294 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1);
1299
1300 // NOTE: (almost) no way to recover if allocation failed.
1301 // We might be able to induce a STW safepoint and scavenge enough
1302 // objectMonitors to permit progress.
1303 if (temp == NULL) {
1304 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1305 "Allocate ObjectMonitors");
1306 }
1307 (void)memset((void *) temp, 0, neededsize);
1308
1309 // Format the block.
1310 // initialize the linked list, each monitor points to its next
1311 // forming the single linked free list, the very first monitor
1312 // will points to next block, which forms the block list.
1313 // The trick of using the 1st element in the block as gBlockList
1314 // linkage should be reconsidered. A better implementation would
1315 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1316
1317 for (int i = 1; i < _BLOCKSIZE; i++) {
1318 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1319 assert(temp[i].is_free(), "invariant");
1320 }
1321
1322 // terminate the last monitor as the end of list
1323 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1324
1325 // Element [0] is reserved for global list linkage
1326 temp[0].set_object(CHAINMARKER);
1327
1328 // Consider carving out this thread's current request from the
1329 // block in hand. This avoids some lock traffic and redundant
1330 // list activity.
1331
1332 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1333 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1334 Thread::muxAcquire(&gListLock, "omAlloc(2)");
1335 gMonitorPopulation += _BLOCKSIZE-1;
1336 gMonitorFreeCount += _BLOCKSIZE-1;
1337
1338 // Add the new block to the list of extant blocks (gBlockList).
1339 // The very first objectMonitor in a block is reserved and dedicated.
1342 // There are lock-free uses of gBlockList so make sure that
1343 // the previous stores happen before we update gBlockList.
1344 OrderAccess::release_store(&gBlockList, temp);
1345
1346 // Add the new string of objectMonitors to the global free list
1347 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1348 gFreeList = temp + 1;
1349 Thread::muxRelease(&gListLock);
1350 }
1351 }
1352
1353 // Place "m" on the caller's private per-thread omFreeList.
1354 // In practice there's no need to clamp or limit the number of
1355 // monitors on a thread's omFreeList as the only time we'll call
1356 // omRelease is to return a monitor to the free list after a CAS
1357 // attempt failed. This doesn't allow unbounded #s of monitors to
1358 // accumulate on a thread's free list.
1359 //
1360 // Key constraint: all ObjectMonitors on a thread's free list and the global
1361 // free list must have their object field set to null. This prevents the
1362 // scavenger -- deflate_monitor_list() or deflate_monitor_list_using_JT()
1363 // -- from reclaiming them while we are trying to release them.
1364
1365 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1366 bool fromPerThreadAlloc) {
1367 guarantee(m->header() == NULL, "invariant");
1368 guarantee(m->object() == NULL, "invariant");
1369 stringStream ss;
1370 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: "
1371 "%s, recursions=" INTPTR_FORMAT, m->is_busy_to_string(&ss),
1372 m->_recursions);
1373 m->set_allocation_state(ObjectMonitor::Free);
1374 // Remove from omInUseList
1375 if (fromPerThreadAlloc) {
1376 ObjectMonitor* cur_mid_in_use = NULL;
1377 bool extracted = false;
1378 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1379 if (m == mid) {
1380 // extract from per-thread in-use list
1381 if (mid == Self->omInUseList) {
1382 Self->omInUseList = mid->FreeNext;
1383 } else if (cur_mid_in_use != NULL) {
1384 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1385 }
1386 extracted = true;
1387 Self->omInUseCount--;
1388 break;
1389 }
1390 }
1391 assert(extracted, "Should have extracted from in-use list");
1392 }
1393
1394 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1395 m->FreeNext = Self->omFreeList;
1396 guarantee(m->is_free(), "invariant");
1397 Self->omFreeList = m;
1398 Self->omFreeCount++;
1399 }
1400
1401 // Return the monitors of a moribund thread's local free list to
1402 // the global free list. Typically a thread calls omFlush() when
1403 // it's dying. We could also consider having the VM thread steal
1404 // monitors from threads that have not run java code over a few
1405 // consecutive STW safepoints. Relatedly, we might decay
1406 // omFreeProvision at STW safepoints.
1407 //
1408 // Also return the monitors of a moribund thread's omInUseList to
1409 // a global gOmInUseList under the global list lock so these
1410 // will continue to be scanned.
1411 //
1412 // We currently call omFlush() from Threads::remove() _before the thread
1413 // has been excised from the thread list and is no longer a mutator.
1414 // This means that omFlush() cannot run concurrently with a safepoint and
1415 // interleave with the deflate_idle_monitors scavenge operator. In particular,
1416 // this ensures that the thread's monitors are scanned by a GC safepoint,
1417 // either via Thread::oops_do() (if safepoint happens before omFlush()) or via
1418 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1419 // monitors have been transferred to the global in-use list).
1420 //
1421 // With AsyncDeflateIdleMonitors, deflate_global_idle_monitors_using_JT()
1422 // and deflate_per_thread_idle_monitors_using_JT() (in another thread) can
1423 // run at the same time as omFlush() so we have to be careful.
1424
1425 void ObjectSynchronizer::omFlush(Thread * Self) {
1426 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1427 ObjectMonitor * tail = NULL;
1428 int tally = 0;
1429 if (list != NULL) {
1430 ObjectMonitor * s;
1431 // The thread is going away. Set 'tail' to the last per-thread free
1432 // monitor which will be linked to gFreeList below under the gListLock.
1433 stringStream ss;
1434 for (s = list; s != NULL; s = s->FreeNext) {
1435 tally++;
1436 tail = s;
1437 guarantee(s->object() == NULL, "invariant");
1438 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss));
1439 }
1440 guarantee(tail != NULL, "invariant");
1441 ADIM_guarantee(Self->omFreeCount == tally, "free-count off");
1442 Self->omFreeList = NULL;
1443 Self->omFreeCount = 0;
1444 }
1445
1446 ObjectMonitor * inUseList = Self->omInUseList;
1447 ObjectMonitor * inUseTail = NULL;
1448 int inUseTally = 0;
1449 if (inUseList != NULL) {
1450 ObjectMonitor *cur_om;
1451 // The thread is going away, however the omInUseList inflated
1452 // monitors may still be in-use by other threads.
1453 // Link them to inUseTail, which will be linked into the global in-use list
1454 // gOmInUseList below, under the gListLock
1455 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1456 inUseTail = cur_om;
1457 inUseTally++;
1458 ADIM_guarantee(cur_om->is_active(), "invariant");
1459 }
1460 guarantee(inUseTail != NULL, "invariant");
1461 ADIM_guarantee(Self->omInUseCount == inUseTally, "in-use count off");
1462 Self->omInUseList = NULL;
1463 Self->omInUseCount = 0;
1464 }
1465
1466 Thread::muxAcquire(&gListLock, "omFlush");
1467 if (tail != NULL) {
1468 tail->FreeNext = gFreeList;
1469 gFreeList = list;
1470 gMonitorFreeCount += tally;
1471 }
1472
1473 if (inUseTail != NULL) {
1474 inUseTail->FreeNext = gOmInUseList;
1475 gOmInUseList = inUseList;
1476 gOmInUseCount += inUseTally;
1477 }
1478
1479 Thread::muxRelease(&gListLock);
1480
1481 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1489 }
1490 if (ls != NULL) {
1491 ls->print_cr("omFlush: jt=" INTPTR_FORMAT ", free_monitor_tally=%d"
1492 ", in_use_monitor_tally=%d" ", omFreeProvision=%d",
1493 p2i(Self), tally, inUseTally, Self->omFreeProvision);
1494 }
1495 }
1496
1497 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1498 const oop obj,
1499 ObjectSynchronizer::InflateCause cause) {
1500 assert(event != NULL, "invariant");
1501 assert(event->should_commit(), "invariant");
1502 event->set_monitorClass(obj->klass());
1503 event->set_address((uintptr_t)(void*)obj);
1504 event->set_cause((u1)cause);
1505 event->commit();
1506 }
1507
1508 // Fast path code shared by multiple functions
1509 void ObjectSynchronizer::inflate_helper(ObjectMonitorHandle * omh_p, oop obj) {
1510 while (true) {
1511 markOop mark = obj->mark();
1512 if (mark->has_monitor()) {
1513 if (!omh_p->save_om_ptr(obj, mark)) {
1514 // Lost a race with async deflation so try again.
1515 assert(AsyncDeflateIdleMonitors, "sanity check");
1516 continue;
1517 }
1518 ObjectMonitor * monitor = omh_p->om_ptr();
1519 assert(ObjectSynchronizer::verify_objmon_isinpool(monitor), "monitor is invalid");
1520 markOop dmw = monitor->header();
1521 assert(dmw->is_neutral(), "sanity check: header=" INTPTR_FORMAT, p2i(dmw));
1522 return;
1523 }
1524 inflate(omh_p, Thread::current(), obj, inflate_cause_vm_internal);
1525 return;
1526 }
1527 }
1528
1529 void ObjectSynchronizer::inflate(ObjectMonitorHandle * omh_p, Thread * Self,
1530 oop object, const InflateCause cause) {
1531 // Inflate mutates the heap ...
1532 // Relaxing assertion for bug 6320749.
1533 assert(Universe::verify_in_progress() ||
1534 !SafepointSynchronize::is_at_safepoint(), "invariant");
1535
1536 EventJavaMonitorInflate event;
1537
1538 for (;;) {
1539 const markOop mark = object->mark();
1540 assert(!mark->has_bias_pattern(), "invariant");
1541
1542 // The mark can be in one of the following states:
1543 // * Inflated - just return
1544 // * Stack-locked - coerce it to inflated
1545 // * INFLATING - busy wait for conversion to complete
1546 // * Neutral - aggressively inflate the object.
1547 // * BIASED - Illegal. We should never see this
1548
1549 // CASE: inflated
1550 if (mark->has_monitor()) {
1551 if (!omh_p->save_om_ptr(object, mark)) {
1552 // Lost a race with async deflation so try again.
1553 assert(AsyncDeflateIdleMonitors, "sanity check");
1554 continue;
1555 }
1556 ObjectMonitor * inf = omh_p->om_ptr();
1557 markOop dmw = inf->header();
1558 assert(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(dmw));
1559 assert(oopDesc::equals((oop) inf->object(), object), "invariant");
1560 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1561 return;
1562 }
1563
1564 // CASE: inflation in progress - inflating over a stack-lock.
1565 // Some other thread is converting from stack-locked to inflated.
1566 // Only that thread can complete inflation -- other threads must wait.
1567 // The INFLATING value is transient.
1568 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1569 // We could always eliminate polling by parking the thread on some auxiliary list.
1570 if (mark == markOopDesc::INFLATING()) {
1571 ReadStableMark(object);
1572 continue;
1573 }
1574
1575 // CASE: stack-locked
1576 // Could be stack-locked either by this thread or by some other thread.
1577 //
1578 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1579 // to install INFLATING into the mark word. We originally installed INFLATING,
1580 // allocated the objectmonitor, and then finally STed the address of the
1581 // objectmonitor into the mark. This was correct, but artificially lengthened
1582 // the interval in which INFLATED appeared in the mark, thus increasing
1583 // the odds of inflation contention.
1584 //
1585 // We now use per-thread private objectmonitor free lists.
1586 // These list are reprovisioned from the global free list outside the
1587 // critical INFLATING...ST interval. A thread can transfer
1588 // multiple objectmonitors en-mass from the global free list to its local free list.
1589 // This reduces coherency traffic and lock contention on the global free list.
1590 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1591 // before or after the CAS(INFLATING) operation.
1592 // See the comments in omAlloc().
1593
1594 LogStreamHandle(Trace, monitorinflation) lsh;
1595
1596 if (mark->has_locker()) {
1597 ObjectMonitor * m;
1598 if (!AsyncDeflateIdleMonitors || cause == inflate_cause_vm_internal) {
1599 // If !AsyncDeflateIdleMonitors or if an internal inflation, then
1600 // we won't stop for a potential safepoint in omAlloc.
1601 m = omAlloc(Self, cause);
1602 } else {
1603 // If AsyncDeflateIdleMonitors and not an internal inflation, then
1604 // we may stop for a safepoint in omAlloc() so protect object.
1605 Handle h_obj(Self, object);
1606 m = omAlloc(Self, cause);
1607 object = h_obj(); // Refresh object.
1608 }
1609 // Optimistically prepare the objectmonitor - anticipate successful CAS
1610 // We do this before the CAS in order to minimize the length of time
1611 // in which INFLATING appears in the mark.
1612 m->Recycle();
1613 m->_Responsible = NULL;
1614 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1615
1616 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1617 if (cmp != mark) {
1618 omRelease(Self, m, true);
1619 continue; // Interference -- just retry
1620 }
1621
1622 // We've successfully installed INFLATING (0) into the mark-word.
1623 // This is the only case where 0 will appear in a mark-word.
1624 // Only the singular thread that successfully swings the mark-word
1625 // to 0 can perform (or more precisely, complete) inflation.
1626 //
1627 // Why do we CAS a 0 into the mark-word instead of just CASing the
1628 // mark-word from the stack-locked value directly to the new inflated state?
1635 // value from the basiclock on the owner's stack to the objectMonitor, all
1636 // the while preserving the hashCode stability invariants. If the owner
1637 // decides to release the lock while the value is 0, the unlock will fail
1638 // and control will eventually pass from slow_exit() to inflate. The owner
1639 // will then spin, waiting for the 0 value to disappear. Put another way,
1640 // the 0 causes the owner to stall if the owner happens to try to
1641 // drop the lock (restoring the header from the basiclock to the object)
1642 // while inflation is in-progress. This protocol avoids races that might
1643 // would otherwise permit hashCode values to change or "flicker" for an object.
1644 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1645 // 0 serves as a "BUSY" inflate-in-progress indicator.
1646
1647
1648 // fetch the displaced mark from the owner's stack.
1649 // The owner can't die or unwind past the lock while our INFLATING
1650 // object is in the mark. Furthermore the owner can't complete
1651 // an unlock on the object, either.
1652 markOop dmw = mark->displaced_mark_helper();
1653 // Catch if the object's header is not neutral (not locked and
1654 // not marked is what we care about here).
1655 ADIM_guarantee(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(dmw));
1656
1657 // Setup monitor fields to proper values -- prepare the monitor
1658 m->set_header(dmw);
1659
1660 // Optimization: if the mark->locker stack address is associated
1661 // with this thread we could simply set m->_owner = Self.
1662 // Note that a thread can inflate an object
1663 // that it has stack-locked -- as might happen in wait() -- directly
1664 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1665 m->set_owner(mark->locker());
1666 m->set_object(object);
1667 // TODO-FIXME: assert BasicLock->dhw != 0.
1668
1669 omh_p->set_om_ptr(m);
1670 assert(m->is_new(), "freshly allocated monitor must be new");
1671 m->set_allocation_state(ObjectMonitor::Old);
1672
1673 // Must preserve store ordering. The monitor state must
1674 // be stable at the time of publishing the monitor address.
1675 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1676 object->release_set_mark(markOopDesc::encode(m));
1677
1678 // Hopefully the performance counters are allocated on distinct cache lines
1679 // to avoid false sharing on MP systems ...
1680 OM_PERFDATA_OP(Inflations, inc());
1681 if (log_is_enabled(Trace, monitorinflation)) {
1682 ResourceMark rm(Self);
1683 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1684 INTPTR_FORMAT ", type='%s'", p2i(object),
1685 p2i(object->mark()), object->klass()->external_name());
1686 }
1687 if (event.should_commit()) {
1688 post_monitor_inflate_event(&event, object, cause);
1689 }
1690 ADIM_guarantee(!m->is_free(), "inflated monitor to be returned cannot be free");
1691 return;
1692 }
1693
1694 // CASE: neutral
1695 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1696 // If we know we're inflating for entry it's better to inflate by swinging a
1697 // pre-locked objectMonitor pointer into the object header. A successful
1698 // CAS inflates the object *and* confers ownership to the inflating thread.
1699 // In the current implementation we use a 2-step mechanism where we CAS()
1700 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1701 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1702 // would be useful.
1703
1704 // Catch if the object's header is not neutral (not locked and
1705 // not marked is what we care about here).
1706 ADIM_guarantee(mark->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(mark));
1707 ObjectMonitor * m;
1708 if (!AsyncDeflateIdleMonitors || cause == inflate_cause_vm_internal) {
1709 // If !AsyncDeflateIdleMonitors or if an internal inflation, then
1710 // we won't stop for a potential safepoint in omAlloc.
1711 m = omAlloc(Self, cause);
1712 } else {
1713 // If AsyncDeflateIdleMonitors and not an internal inflation, then
1714 // we may stop for a safepoint in omAlloc() so protect object.
1715 Handle h_obj(Self, object);
1716 m = omAlloc(Self, cause);
1717 object = h_obj(); // Refresh object.
1718 }
1719 // prepare m for installation - set monitor to initial state
1720 m->Recycle();
1721 m->set_header(mark);
1722 // If we leave _owner == DEFLATER_MARKER here, then the simple C2
1723 // ObjectMonitor enter optimization can no longer race with async
1724 // deflation and reuse.
1725 m->set_object(object);
1726 m->_Responsible = NULL;
1727 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1728
1729 omh_p->set_om_ptr(m);
1730 assert(m->is_new(), "freshly allocated monitor must be new");
1731 m->set_allocation_state(ObjectMonitor::Old);
1732
1733 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1734 m->set_header(NULL);
1735 m->set_object(NULL);
1736 m->Recycle();
1737 omh_p->set_om_ptr(NULL);
1738 // omRelease() will reset the allocation state
1739 omRelease(Self, m, true);
1740 m = NULL;
1741 continue;
1742 // interference - the markword changed - just retry.
1743 // The state-transitions are one-way, so there's no chance of
1744 // live-lock -- "Inflated" is an absorbing state.
1745 }
1746
1747 // Hopefully the performance counters are allocated on distinct
1748 // cache lines to avoid false sharing on MP systems ...
1749 OM_PERFDATA_OP(Inflations, inc());
1750 if (log_is_enabled(Trace, monitorinflation)) {
1751 ResourceMark rm(Self);
1752 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1753 INTPTR_FORMAT ", type='%s'", p2i(object),
1754 p2i(object->mark()), object->klass()->external_name());
1755 }
1756 if (event.should_commit()) {
1757 post_monitor_inflate_event(&event, object, cause);
1758 }
1759 ADIM_guarantee(!m->is_free(), "inflated monitor to be returned cannot be free");
1760 return;
1761 }
1762 }
1763
1764
1765 // We maintain a list of in-use monitors for each thread.
1766 //
1767 // For safepoint based deflation:
1768 // deflate_thread_local_monitors() scans a single thread's in-use list, while
1769 // deflate_idle_monitors() scans only a global list of in-use monitors which
1770 // is populated only as a thread dies (see omFlush()).
1771 //
1772 // These operations are called at all safepoints, immediately after mutators
1773 // are stopped, but before any objects have moved. Collectively they traverse
1774 // the population of in-use monitors, deflating where possible. The scavenged
1775 // monitors are returned to the global monitor free list.
1776 //
1777 // Beware that we scavenge at *every* stop-the-world point. Having a large
1778 // number of monitors in-use could negatively impact performance. We also want
1779 // to minimize the total # of monitors in circulation, as they incur a small
1780 // footprint penalty.
1781 //
1782 // Perversely, the heap size -- and thus the STW safepoint rate --
1783 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1784 // which in turn can mean large(r) numbers of ObjectMonitors in circulation.
1785 // This is an unfortunate aspect of this design.
1786 //
1787 // For async deflation:
1788 // If a special deflation request is made, then the safepoint based
1789 // deflation mechanism is used. Otherwise, an async deflation request
1790 // is registered with the ServiceThread and it is notified.
1791
1792 void ObjectSynchronizer::do_safepoint_work(DeflateMonitorCounters* _counters) {
1793 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1794
1795 // The per-thread in-use lists are handled in
1796 // ParallelSPCleanupThreadClosure::do_thread().
1797
1798 if (!AsyncDeflateIdleMonitors || is_special_deflation_requested()) {
1799 // Use the older mechanism for the global in-use list or if a
1800 // special deflation has been requested before the safepoint.
1801 ObjectSynchronizer::deflate_idle_monitors(_counters);
1802 return;
1803 }
1804
1805 log_debug(monitorinflation)("requesting async deflation of idle monitors.");
1806 // Request deflation of idle monitors by the ServiceThread:
1807 set_is_async_deflation_requested(true);
1808 MonitorLocker ml(Service_lock, Mutex::_no_safepoint_check_flag);
1809 ml.notify_all();
1810 }
1811
1812 // Deflate a single monitor if not in-use
1813 // Return true if deflated, false if in-use
1814 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1815 ObjectMonitor** freeHeadp,
1816 ObjectMonitor** freeTailp) {
1817 bool deflated;
1818 // Normal case ... The monitor is associated with obj.
1819 const markOop mark = obj->mark();
1820 guarantee(mark == markOopDesc::encode(mid), "should match: mark="
1821 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, p2i(mark),
1822 p2i(markOopDesc::encode(mid)));
1823 // Make sure that mark->monitor() and markOopDesc::encode() agree:
1824 guarantee(mark->monitor() == mid, "should match: monitor()=" INTPTR_FORMAT
1825 ", mid=" INTPTR_FORMAT, p2i(mark->monitor()), p2i(mid));
1826 const markOop dmw = mid->header();
1827 guarantee(dmw->is_neutral(), "invariant: header=" INTPTR_FORMAT, p2i(dmw));
1828
1829 if (mid->is_busy() || mid->ref_count() != 0) {
1830 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor*
1831 // is in use so no deflation.
1832 deflated = false;
1833 } else {
1834 // Deflate the monitor if it is no longer being used
1835 // It's idle - scavenge and return to the global free list
1836 // plain old deflation ...
1837 if (log_is_enabled(Trace, monitorinflation)) {
1838 ResourceMark rm;
1839 log_trace(monitorinflation)("deflate_monitor: "
1840 "object=" INTPTR_FORMAT ", mark="
1841 INTPTR_FORMAT ", type='%s'", p2i(obj),
1842 p2i(mark), obj->klass()->external_name());
1843 }
1844
1845 // Restore the header back to obj
1846 obj->release_set_mark(dmw);
1847 if (AsyncDeflateIdleMonitors) {
1848 // clear() expects the owner field to be NULL and we won't race
1849 // with the simple C2 ObjectMonitor enter optimization since
1850 // we're at a safepoint.
1851 mid->set_owner(NULL);
1852 }
1853 mid->clear();
1854
1855 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT,
1856 p2i(mid->object()));
1857 assert(mid->is_free(), "invariant");
1858
1859 // Move the object to the working free list defined by freeHeadp, freeTailp
1860 if (*freeHeadp == NULL) *freeHeadp = mid;
1861 if (*freeTailp != NULL) {
1862 ObjectMonitor * prevtail = *freeTailp;
1863 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1864 prevtail->FreeNext = mid;
1865 }
1866 *freeTailp = mid;
1867 deflated = true;
1868 }
1869 return deflated;
1870 }
1871
1872 // Deflate the specified ObjectMonitor if not in-use using a JavaThread.
1873 // Returns true if it was deflated and false otherwise.
1874 //
1875 // The async deflation protocol sets owner to DEFLATER_MARKER and
1876 // makes ref_count negative as signals to contending threads that
1877 // an async deflation is in progress. There are a number of checks
1878 // as part of the protocol to make sure that the calling thread has
1879 // not lost the race to a contending thread or to a thread that just
1880 // wants to use the ObjectMonitor*.
1881 //
1882 // The ObjectMonitor has been successfully async deflated when:
1883 // (owner == DEFLATER_MARKER && ref_count < 0)
1884 // Contending threads or ObjectMonitor* using threads that see those
1885 // values know to retry their operation.
1886 //
1887 bool ObjectSynchronizer::deflate_monitor_using_JT(ObjectMonitor* mid,
1888 ObjectMonitor** freeHeadp,
1889 ObjectMonitor** freeTailp) {
1890 assert(AsyncDeflateIdleMonitors, "sanity check");
1891 assert(Thread::current()->is_Java_thread(), "precondition");
1892 // A newly allocated ObjectMonitor should not be seen here so we
1893 // avoid an endless inflate/deflate cycle.
1894 assert(mid->is_old(), "must be old: allocation_state=%d",
1895 (int) mid->allocation_state());
1896
1897 if (mid->is_busy() || mid->ref_count() != 0) {
1898 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor*
1899 // is in use so no deflation.
1900 return false;
1901 }
1902
1903 if (Atomic::replace_if_null(DEFLATER_MARKER, &(mid->_owner))) {
1904 // ObjectMonitor is not owned by another thread. Our setting
1905 // owner to DEFLATER_MARKER forces any contending thread through
1906 // the slow path. This is just the first part of the async
1907 // deflation dance.
1908
1909 if (mid->_contentions != 0 || mid->_waiters != 0) {
1910 // Another thread has raced to enter the ObjectMonitor after
1911 // mid->is_busy() above or has already entered and waited on
1912 // it which makes it busy so no deflation. Restore owner to
1913 // NULL if it is still DEFLATER_MARKER.
1914 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER);
1915 return false;
1916 }
1917
1918 if (Atomic::cmpxchg(-max_jint, &mid->_ref_count, (jint)0) == 0) {
1919 // Make ref_count negative to force any contending threads or
1920 // ObjectMonitor* using threads to retry. This is the second
1921 // part of the async deflation dance.
1922
1923 if (mid->owner_is_DEFLATER_MARKER()) {
1924 // If owner is still DEFLATER_MARKER, then we have successfully
1925 // signaled any contending threads to retry. If it is not, then we
1926 // have lost the race to an entering thread and the ObjectMonitor
1927 // is now busy. This is the third and final part of the async
1928 // deflation dance.
1929 // Note: This owner check solves the ABA problem with ref_count
1930 // where another thread acquired the ObjectMonitor, finished
1931 // using it and restored the ref_count to zero.
1932
1933 // Sanity checks for the races:
1934 guarantee(mid->_contentions == 0, "must be 0: contentions=%d",
1935 mid->_contentions);
1936 guarantee(mid->_waiters == 0, "must be 0: waiters=%d", mid->_waiters);
1937 guarantee(mid->_cxq == NULL, "must be no contending threads: cxq="
1938 INTPTR_FORMAT, p2i(mid->_cxq));
1939 guarantee(mid->_EntryList == NULL,
1940 "must be no entering threads: EntryList=" INTPTR_FORMAT,
1941 p2i(mid->_EntryList));
1942
1943 const oop obj = (oop) mid->object();
1944 if (log_is_enabled(Trace, monitorinflation)) {
1945 ResourceMark rm;
1946 log_trace(monitorinflation)("deflate_monitor_using_JT: "
1947 "object=" INTPTR_FORMAT ", mark="
1948 INTPTR_FORMAT ", type='%s'",
1949 p2i(obj), p2i(obj->mark()),
1950 obj->klass()->external_name());
1951 }
1952
1953 // Install the old mark word if nobody else has already done it.
1954 mid->install_displaced_markword_in_object(obj);
1955 mid->clear_using_JT();
1956
1957 assert(mid->object() == NULL, "must be NULL: object=" INTPTR_FORMAT,
1958 p2i(mid->object()));
1959 assert(mid->is_free(), "must be free: allocation_state=%d",
1960 (int) mid->allocation_state());
1961
1962 // Move the deflated ObjectMonitor to the working free list
1963 // defined by freeHeadp and freeTailp.
1964 if (*freeHeadp == NULL) {
1965 // First one on the list.
1966 *freeHeadp = mid;
1967 }
1968 if (*freeTailp != NULL) {
1969 // We append to the list so the caller can use mid->FreeNext
1970 // to fix the linkages in its context.
1971 ObjectMonitor * prevtail = *freeTailp;
1972 // Should have been cleaned up by the caller:
1973 assert(prevtail->FreeNext == NULL, "must be NULL: FreeNext="
1974 INTPTR_FORMAT, p2i(prevtail->FreeNext));
1975 prevtail->FreeNext = mid;
1976 }
1977 *freeTailp = mid;
1978
1979 // At this point, mid->FreeNext still refers to its current
1980 // value and another ObjectMonitor's FreeNext field still
1981 // refers to this ObjectMonitor. Those linkages have to be
1982 // cleaned up by the caller who has the complete context.
1983
1984 // We leave owner == DEFLATER_MARKER and ref_count < 0
1985 // to force any racing threads to retry.
1986 return true; // Success, ObjectMonitor has been deflated.
1987 }
1988
1989 // The owner was changed from DEFLATER_MARKER so we lost the
1990 // race since the ObjectMonitor is now busy.
1991
1992 // Add back max_jint to restore the ref_count field to its
1993 // proper value (which may not be what we saw above):
1994 Atomic::add(max_jint, &mid->_ref_count);
1995
1996 assert(mid->ref_count() >= 0, "must not be negative: ref_count=%d",
1997 mid->ref_count());
1998 return false;
1999 }
2000
2001 // The ref_count was no longer 0 so we lost the race since the
2002 // ObjectMonitor is now busy or the ObjectMonitor* is now is use.
2003 // Restore owner to NULL if it is still DEFLATER_MARKER:
2004 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER);
2005 }
2006
2007 // The owner field is no longer NULL so we lost the race since the
2008 // ObjectMonitor is now busy.
2009 return false;
2010 }
2011
2012 // Walk a given monitor list, and deflate idle monitors
2013 // The given list could be a per-thread list or a global list
2014 // Caller acquires gListLock as needed.
2015 //
2016 // In the case of parallel processing of thread local monitor lists,
2017 // work is done by Threads::parallel_threads_do() which ensures that
2018 // each Java thread is processed by exactly one worker thread, and
2019 // thus avoid conflicts that would arise when worker threads would
2020 // process the same monitor lists concurrently.
2021 //
2022 // See also ParallelSPCleanupTask and
2023 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
2024 // Threads::parallel_java_threads_do() in thread.cpp.
2025 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
2026 ObjectMonitor** freeHeadp,
2027 ObjectMonitor** freeTailp) {
2028 ObjectMonitor* mid;
2029 ObjectMonitor* next;
2030 ObjectMonitor* cur_mid_in_use = NULL;
2031 int deflated_count = 0;
2035 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
2036 // if deflate_monitor succeeded,
2037 // extract from per-thread in-use list
2038 if (mid == *listHeadp) {
2039 *listHeadp = mid->FreeNext;
2040 } else if (cur_mid_in_use != NULL) {
2041 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
2042 }
2043 next = mid->FreeNext;
2044 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
2045 mid = next;
2046 deflated_count++;
2047 } else {
2048 cur_mid_in_use = mid;
2049 mid = mid->FreeNext;
2050 }
2051 }
2052 return deflated_count;
2053 }
2054
2055 // Walk a given ObjectMonitor list and deflate idle ObjectMonitors using
2056 // a JavaThread. Returns the number of deflated ObjectMonitors. The given
2057 // list could be a per-thread in-use list or the global in-use list.
2058 // Caller acquires gListLock as appropriate. If a safepoint has started,
2059 // then we save state via savedMidInUsep and return to the caller to
2060 // honor the safepoint.
2061 //
2062 int ObjectSynchronizer::deflate_monitor_list_using_JT(ObjectMonitor** listHeadp,
2063 ObjectMonitor** freeHeadp,
2064 ObjectMonitor** freeTailp,
2065 ObjectMonitor** savedMidInUsep) {
2066 assert(AsyncDeflateIdleMonitors, "sanity check");
2067 assert(Thread::current()->is_Java_thread(), "precondition");
2068
2069 ObjectMonitor* mid;
2070 ObjectMonitor* next;
2071 ObjectMonitor* cur_mid_in_use = NULL;
2072 int deflated_count = 0;
2073
2074 if (*savedMidInUsep == NULL) {
2075 // No saved state so start at the beginning.
2076 mid = *listHeadp;
2077 } else {
2078 // We're restarting after a safepoint so restore the necessary state
2079 // before we resume.
2080 cur_mid_in_use = *savedMidInUsep;
2081 mid = cur_mid_in_use->FreeNext;
2082 }
2083 while (mid != NULL) {
2084 // Only try to deflate if there is an associated Java object and if
2085 // mid is old (is not newly allocated and is not newly freed).
2086 if (mid->object() != NULL && mid->is_old() &&
2087 deflate_monitor_using_JT(mid, freeHeadp, freeTailp)) {
2088 // Deflation succeeded so update the in-use list.
2089 if (mid == *listHeadp) {
2090 *listHeadp = mid->FreeNext;
2091 } else if (cur_mid_in_use != NULL) {
2092 // Maintain the current in-use list.
2093 cur_mid_in_use->FreeNext = mid->FreeNext;
2094 }
2095 next = mid->FreeNext;
2096 mid->FreeNext = NULL;
2097 // At this point mid is disconnected from the in-use list
2098 // and is the current tail in the freeHeadp list.
2099 mid = next;
2100 deflated_count++;
2101 } else {
2102 // mid is considered in-use if it does not have an associated
2103 // Java object or mid is not old or deflation did not succeed.
2104 // A mid->is_new() node can be seen here when it is freshly
2105 // returned by omAlloc() (and skips the deflation code path).
2106 // A mid->is_old() node can be seen here when deflation failed.
2107 // A mid->is_free() node can be seen here when a fresh node from
2108 // omAlloc() is released by omRelease() due to losing the race
2109 // in inflate().
2110
2111 cur_mid_in_use = mid;
2112 mid = mid->FreeNext;
2113
2114 if (SafepointSynchronize::is_synchronizing() &&
2115 cur_mid_in_use != *listHeadp && cur_mid_in_use->is_old()) {
2116 // If a safepoint has started and cur_mid_in_use is not the list
2117 // head and is old, then it is safe to use as saved state. Return
2118 // to the caller so gListLock can be dropped as appropriate
2119 // before blocking.
2120 *savedMidInUsep = cur_mid_in_use;
2121 return deflated_count;
2122 }
2123 }
2124 }
2125 // We finished the list without a safepoint starting so there's
2126 // no need to save state.
2127 *savedMidInUsep = NULL;
2128 return deflated_count;
2129 }
2130
2131 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
2132 counters->nInuse = 0; // currently associated with objects
2133 counters->nInCirculation = 0; // extant
2134 counters->nScavenged = 0; // reclaimed (global and per-thread)
2135 counters->perThreadScavenged = 0; // per-thread scavenge total
2136 counters->perThreadTimes = 0.0; // per-thread scavenge times
2137 }
2138
2139 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
2140 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
2141
2142 if (AsyncDeflateIdleMonitors) {
2143 // Nothing to do when global idle ObjectMonitors are deflated using
2144 // a JavaThread unless a special deflation has been requested.
2145 if (!is_special_deflation_requested()) {
2146 return;
2147 }
2148 }
2149
2150 bool deflated = false;
2151
2152 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
2153 ObjectMonitor * freeTailp = NULL;
2154 elapsedTimer timer;
2155
2156 if (log_is_enabled(Info, monitorinflation)) {
2157 timer.start();
2158 }
2159
2160 // Prevent omFlush from changing mids in Thread dtor's during deflation
2161 // And in case the vm thread is acquiring a lock during a safepoint
2162 // See e.g. 6320749
2163 Thread::muxAcquire(&gListLock, "deflate_idle_monitors");
2164
2165 // Note: the thread-local monitors lists get deflated in
2166 // a separate pass. See deflate_thread_local_monitors().
2167
2168 // For moribund threads, scan gOmInUseList
2169 int deflated_count = 0;
2182 // constant-time list splice - prepend scavenged segment to gFreeList
2183 freeTailp->FreeNext = gFreeList;
2184 gFreeList = freeHeadp;
2185 }
2186 Thread::muxRelease(&gListLock);
2187 timer.stop();
2188
2189 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2190 LogStreamHandle(Info, monitorinflation) lsh_info;
2191 LogStream * ls = NULL;
2192 if (log_is_enabled(Debug, monitorinflation)) {
2193 ls = &lsh_debug;
2194 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2195 ls = &lsh_info;
2196 }
2197 if (ls != NULL) {
2198 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
2199 }
2200 }
2201
2202 // Deflate global idle ObjectMonitors using a JavaThread.
2203 //
2204 void ObjectSynchronizer::deflate_global_idle_monitors_using_JT() {
2205 assert(AsyncDeflateIdleMonitors, "sanity check");
2206 assert(Thread::current()->is_Java_thread(), "precondition");
2207 JavaThread * self = JavaThread::current();
2208
2209 deflate_common_idle_monitors_using_JT(true /* is_global */, self);
2210 }
2211
2212 // Deflate per-thread idle ObjectMonitors using a JavaThread.
2213 //
2214 void ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT() {
2215 assert(AsyncDeflateIdleMonitors, "sanity check");
2216 assert(Thread::current()->is_Java_thread(), "precondition");
2217 JavaThread * self = JavaThread::current();
2218
2219 self->omShouldDeflateIdleMonitors = false;
2220
2221 deflate_common_idle_monitors_using_JT(false /* !is_global */, self);
2222 }
2223
2224 // Deflate global or per-thread idle ObjectMonitors using a JavaThread.
2225 //
2226 void ObjectSynchronizer::deflate_common_idle_monitors_using_JT(bool is_global, JavaThread * self) {
2227 int deflated_count = 0;
2228 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged ObjectMonitors
2229 ObjectMonitor * freeTailp = NULL;
2230 ObjectMonitor * savedMidInUsep = NULL;
2231 elapsedTimer timer;
2232
2233 if (log_is_enabled(Info, monitorinflation)) {
2234 timer.start();
2235 }
2236
2237 if (is_global) {
2238 Thread::muxAcquire(&gListLock, "deflate_global_idle_monitors_using_JT(1)");
2239 OM_PERFDATA_OP(MonExtant, set_value(gOmInUseCount));
2240 } else {
2241 OM_PERFDATA_OP(MonExtant, inc(self->omInUseCount));
2242 }
2243
2244 do {
2245 int local_deflated_count;
2246 if (is_global) {
2247 local_deflated_count = deflate_monitor_list_using_JT((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp, &savedMidInUsep);
2248 gOmInUseCount -= local_deflated_count;
2249 } else {
2250 local_deflated_count = deflate_monitor_list_using_JT(self->omInUseList_addr(), &freeHeadp, &freeTailp, &savedMidInUsep);
2251 self->omInUseCount -= local_deflated_count;
2252 }
2253 deflated_count += local_deflated_count;
2254
2255 if (freeHeadp != NULL) {
2256 // Move the scavenged ObjectMonitors to the global free list.
2257 guarantee(freeTailp != NULL && local_deflated_count > 0, "freeTailp=" INTPTR_FORMAT ", local_deflated_count=%d", p2i(freeTailp), local_deflated_count);
2258 assert(freeTailp->FreeNext == NULL, "invariant");
2259
2260 if (!is_global) {
2261 Thread::muxAcquire(&gListLock, "deflate_per_thread_idle_monitors_using_JT(2)");
2262 }
2263 // Constant-time list splice - prepend scavenged segment to gFreeList.
2264 freeTailp->FreeNext = gFreeList;
2265 gFreeList = freeHeadp;
2266
2267 gMonitorFreeCount += local_deflated_count;
2268 OM_PERFDATA_OP(Deflations, inc(local_deflated_count));
2269 if (!is_global) {
2270 Thread::muxRelease(&gListLock);
2271 }
2272 }
2273
2274 if (savedMidInUsep != NULL) {
2275 // deflate_monitor_list_using_JT() detected a safepoint starting.
2276 if (is_global) {
2277 Thread::muxRelease(&gListLock);
2278 }
2279 timer.stop();
2280 {
2281 if (is_global) {
2282 log_debug(monitorinflation)("pausing deflation of global idle monitors for a safepoint.");
2283 } else {
2284 log_debug(monitorinflation)("jt=" INTPTR_FORMAT ": pausing deflation of per-thread idle monitors for a safepoint.", p2i(self));
2285 }
2286 assert(SafepointSynchronize::is_synchronizing(), "sanity check");
2287 ThreadBlockInVM blocker(self);
2288 }
2289 // Prepare for another loop after the safepoint.
2290 freeHeadp = NULL;
2291 freeTailp = NULL;
2292 if (log_is_enabled(Info, monitorinflation)) {
2293 timer.start();
2294 }
2295 if (is_global) {
2296 Thread::muxAcquire(&gListLock, "deflate_global_idle_monitors_using_JT(3)");
2297 }
2298 }
2299 } while (savedMidInUsep != NULL);
2300 if (is_global) {
2301 Thread::muxRelease(&gListLock);
2302 }
2303 timer.stop();
2304
2305 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2306 LogStreamHandle(Info, monitorinflation) lsh_info;
2307 LogStream * ls = NULL;
2308 if (log_is_enabled(Debug, monitorinflation)) {
2309 ls = &lsh_debug;
2310 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2311 ls = &lsh_info;
2312 }
2313 if (ls != NULL) {
2314 if (is_global) {
2315 ls->print_cr("async-deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
2316 } else {
2317 ls->print_cr("jt=" INTPTR_FORMAT ": async-deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(self), timer.seconds(), deflated_count);
2318 }
2319 }
2320 }
2321
2322 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
2323 // Report the cumulative time for deflating each thread's idle
2324 // monitors. Note: if the work is split among more than one
2325 // worker thread, then the reported time will likely be more
2326 // than a beginning to end measurement of the phase.
2327 // Note: AsyncDeflateIdleMonitors only deflates per-thread idle
2328 // monitors at a safepoint when a special deflation has been requested.
2329 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->perThreadTimes, counters->perThreadScavenged);
2330
2331 bool needs_special_deflation = is_special_deflation_requested();
2332 if (!AsyncDeflateIdleMonitors || needs_special_deflation) {
2333 // AsyncDeflateIdleMonitors does not use these counters unless
2334 // there is a special deflation request.
2335
2336 gMonitorFreeCount += counters->nScavenged;
2337
2338 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
2339 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
2340 }
2341
2342 if (log_is_enabled(Debug, monitorinflation)) {
2343 // exit_globals()'s call to audit_and_print_stats() is done
2344 // at the Info level.
2345 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */);
2346 } else if (log_is_enabled(Info, monitorinflation)) {
2347 Thread::muxAcquire(&gListLock, "finish_deflate_idle_monitors");
2348 log_info(monitorinflation)("gMonitorPopulation=%d, gOmInUseCount=%d, "
2349 "gMonitorFreeCount=%d", gMonitorPopulation,
2350 gOmInUseCount, gMonitorFreeCount);
2351 Thread::muxRelease(&gListLock);
2352 }
2353
2354 ForceMonitorScavenge = 0; // Reset
2355 GVars.stwRandom = os::random();
2356 GVars.stwCycle++;
2357 if (needs_special_deflation) {
2358 set_is_special_deflation_requested(false); // special deflation is done
2359 }
2360 }
2361
2362 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
2363 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
2364
2365 if (AsyncDeflateIdleMonitors) {
2366 if (!is_special_deflation_requested()) {
2367 // Mark the JavaThread for idle monitor deflation if a special
2368 // deflation has NOT been requested.
2369 if (thread->omInUseCount > 0) {
2370 // This JavaThread is using monitors so mark it.
2371 thread->omShouldDeflateIdleMonitors = true;
2372 }
2373 return;
2374 }
2375 }
2376
2377 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
2378 ObjectMonitor * freeTailp = NULL;
2379 elapsedTimer timer;
2380
2381 if (log_is_enabled(Info, safepoint, cleanup) ||
2382 log_is_enabled(Info, monitorinflation)) {
2383 timer.start();
2384 }
2385
2386 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
2387
2388 Thread::muxAcquire(&gListLock, "deflate_thread_local_monitors");
2389
2390 // Adjust counters
2391 counters->nInCirculation += thread->omInUseCount;
2392 thread->omInUseCount -= deflated_count;
2393 counters->nScavenged += deflated_count;
2394 counters->nInuse += thread->omInUseCount;
2395 counters->perThreadScavenged += deflated_count;
2396
2583 void ObjectSynchronizer::chk_free_entry(JavaThread * jt, ObjectMonitor * n,
2584 outputStream * out, int *error_cnt_p) {
2585 stringStream ss;
2586 if (n->is_busy()) {
2587 if (jt != NULL) {
2588 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2589 ": free per-thread monitor must not be busy: %s", p2i(jt),
2590 p2i(n), n->is_busy_to_string(&ss));
2591 } else {
2592 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2593 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss));
2594 }
2595 *error_cnt_p = *error_cnt_p + 1;
2596 }
2597 if (n->header() != NULL) {
2598 if (jt != NULL) {
2599 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2600 ": free per-thread monitor must have NULL _header "
2601 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n),
2602 p2i(n->header()));
2603 *error_cnt_p = *error_cnt_p + 1;
2604 } else if (!AsyncDeflateIdleMonitors) {
2605 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2606 "must have NULL _header field: _header=" INTPTR_FORMAT,
2607 p2i(n), p2i(n->header()));
2608 *error_cnt_p = *error_cnt_p + 1;
2609 }
2610 }
2611 if (n->object() != NULL) {
2612 if (jt != NULL) {
2613 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2614 ": free per-thread monitor must have NULL _object "
2615 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n),
2616 p2i(n->object()));
2617 } else {
2618 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2619 "must have NULL _object field: _object=" INTPTR_FORMAT,
2620 p2i(n), p2i(n->object()));
2621 }
2622 *error_cnt_p = *error_cnt_p + 1;
2623 }
2624 }
2625
2626 // Check the global free list and count; log the results of the checks.
2627 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out,
2628 int *error_cnt_p) {
2629 int chkMonitorFreeCount = 0;
2630 for (ObjectMonitor * n = gFreeList; n != NULL; n = n->FreeNext) {
2757 "equal to chkOmInUseCount=%d", p2i(jt), jt->omInUseCount,
2758 chkOmInUseCount);
2759 *error_cnt_p = *error_cnt_p + 1;
2760 }
2761 }
2762
2763 // Log details about ObjectMonitors on the in-use lists. The 'BHL'
2764 // flags indicate why the entry is in-use, 'object' and 'object type'
2765 // indicate the associated object and its type.
2766 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out,
2767 bool on_exit) {
2768 if (!on_exit) {
2769 // Not at VM exit so grab the global list lock.
2770 Thread::muxAcquire(&gListLock, "log_in_use_monitor_details");
2771 }
2772
2773 stringStream ss;
2774 if (gOmInUseCount > 0) {
2775 out->print_cr("In-use global monitor info:");
2776 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2777 out->print_cr("%18s %s %7s %18s %18s",
2778 "monitor", "BHL", "ref_cnt", "object", "object type");
2779 out->print_cr("================== === ======= ================== ==================");
2780 for (ObjectMonitor * n = gOmInUseList; n != NULL; n = n->FreeNext) {
2781 const oop obj = (oop) n->object();
2782 const markOop mark = n->header();
2783 ResourceMark rm;
2784 out->print(INTPTR_FORMAT " %d%d%d %7d " INTPTR_FORMAT " %s",
2785 p2i(n), n->is_busy() != 0, mark->hash() != 0,
2786 n->owner() != NULL, (int)n->ref_count(), p2i(obj),
2787 obj->klass()->external_name());
2788 if (n->is_busy() != 0) {
2789 out->print(" (%s)", n->is_busy_to_string(&ss));
2790 ss.reset();
2791 }
2792 out->cr();
2793 }
2794 }
2795
2796 if (!on_exit) {
2797 Thread::muxRelease(&gListLock);
2798 }
2799
2800 out->print_cr("In-use per-thread monitor info:");
2801 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2802 out->print_cr("%18s %18s %s %7s %18s %18s",
2803 "jt", "monitor", "BHL", "ref_cnt", "object", "object type");
2804 out->print_cr("================== ================== === ======= ================== ==================");
2805 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2806 for (ObjectMonitor * n = jt->omInUseList; n != NULL; n = n->FreeNext) {
2807 const oop obj = (oop) n->object();
2808 const markOop mark = n->header();
2809 ResourceMark rm;
2810 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d %7d "
2811 INTPTR_FORMAT " %s", p2i(jt), p2i(n), n->is_busy() != 0,
2812 mark->hash() != 0, n->owner() != NULL, (int)n->ref_count(),
2813 p2i(obj), obj->klass()->external_name());
2814 if (n->is_busy() != 0) {
2815 out->print(" (%s)", n->is_busy_to_string(&ss));
2816 ss.reset();
2817 }
2818 out->cr();
2819 }
2820 }
2821
2822 out->flush();
2823 }
2824
2825 // Log counts for the global and per-thread monitor lists and return
2826 // the population count.
2827 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) {
2828 int popCount = 0;
2829 out->print_cr("%18s %10s %10s %10s",
2830 "Global Lists:", "InUse", "Free", "Total");
2831 out->print_cr("================== ========== ========== ==========");
2832 out->print_cr("%18s %10d %10d %10d", "",
2833 gOmInUseCount, gMonitorFreeCount, gMonitorPopulation);
|