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 PaddedObjectMonitor* volatile ObjectSynchronizer::g_block_list = NULL;
121 // Global ObjectMonitor free list. Newly allocated and deflated
122 // ObjectMonitors are prepended here.
123 ObjectMonitor* volatile ObjectSynchronizer::g_free_list = NULL;
124 // Global ObjectMonitor in-use list. When a JavaThread is exiting,
125 // ObjectMonitors on its per-thread in-use list are prepended here.
126 ObjectMonitor* volatile ObjectSynchronizer::g_om_in_use_list = NULL;
127 int ObjectSynchronizer::g_om_in_use_count = 0; // # on g_om_in_use_list
128
129 static volatile intptr_t gListLock = 0; // protects global monitor lists
130 static volatile int g_om_free_count = 0; // # on g_free_list
131 static volatile int g_om_population = 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 markWord 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(markWord::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 == markWord::from_pointer(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 markWord 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(markWord::from_pointer(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().value(), "don't relock with same BasicLock");
354 lock->set_displaced_header(markWord::from_pointer(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(markWord::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. We
433 // intentionally do not use CHECK here because we must exit the
434 // monitor even if an exception is pending.
435 if (monitor->check_owner(THREAD)) {
436 monitor->exit(true, THREAD);
437 }
438 }
439
440 // -----------------------------------------------------------------------------
441 // Internal VM locks on java objects
442 // standard constructor, allows locking failures
443 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool do_lock) {
444 _dolock = do_lock;
445 _thread = thread;
446 _thread->check_for_valid_safepoint_state(false);
447 _obj = obj;
448
449 if (_dolock) {
450 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
451 }
452 }
453
454 ObjectLocker::~ObjectLocker() {
455 if (_dolock) {
456 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
457 }
458 }
459
460
461 // -----------------------------------------------------------------------------
462 // Wait/Notify/NotifyAll
463 // NOTE: must use heavy weight monitor to handle wait()
464 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
465 if (UseBiasedLocking) {
466 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
467 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
468 }
469 if (millis < 0) {
470 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
471 }
472 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_wait);
473
474 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
475 monitor->wait(millis, true, THREAD);
476
477 // This dummy call is in place to get around dtrace bug 6254741. Once
478 // that's fixed we can uncomment the following line, remove the call
479 // and change this function back into a "void" func.
480 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
481 return dtrace_waited_probe(monitor, obj, THREAD);
482 }
483
484 void ObjectSynchronizer::wait_uninterruptibly(Handle obj, jlong millis, TRAPS) {
485 if (UseBiasedLocking) {
486 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
487 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
488 }
489 if (millis < 0) {
490 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
491 }
492 inflate(THREAD, obj(), inflate_cause_wait)->wait(millis, false, THREAD);
493 }
494
495 void ObjectSynchronizer::notify(Handle obj, 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
501 markWord mark = obj->mark();
502 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) {
503 return;
504 }
505 inflate(THREAD, obj(), inflate_cause_notify)->notify(THREAD);
506 }
507
508 // NOTE: see comment of notify()
509 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
510 if (UseBiasedLocking) {
511 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
512 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
513 }
514
515 markWord mark = obj->mark();
516 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) {
517 return;
518 }
519 inflate(THREAD, obj(), inflate_cause_notify)->notifyAll(THREAD);
520 }
521
522 // -----------------------------------------------------------------------------
523 // Hash Code handling
524 //
525 // Performance concern:
526 // OrderAccess::storestore() calls release() which at one time stored 0
527 // into the global volatile OrderAccess::dummy variable. This store was
528 // unnecessary for correctness. Many threads storing into a common location
529 // causes considerable cache migration or "sloshing" on large SMP systems.
530 // As such, I avoided using OrderAccess::storestore(). In some cases
531 // OrderAccess::fence() -- which incurs local latency on the executing
532 // processor -- is a better choice as it scales on SMP systems.
533 //
534 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
535 // a discussion of coherency costs. Note that all our current reference
536 // platforms provide strong ST-ST order, so the issue is moot on IA32,
537 // x64, and SPARC.
538 //
539 // As a general policy we use "volatile" to control compiler-based reordering
693 Handle hobj(self, obj);
694 // Relaxing assertion for bug 6320749.
695 assert(Universe::verify_in_progress() ||
696 !SafepointSynchronize::is_at_safepoint(),
697 "biases should not be seen by VM thread here");
698 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
699 obj = hobj();
700 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
701 }
702 }
703
704 // hashCode() is a heap mutator ...
705 // Relaxing assertion for bug 6320749.
706 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
707 !SafepointSynchronize::is_at_safepoint(), "invariant");
708 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
709 self->is_Java_thread() , "invariant");
710 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
711 ((JavaThread *)self)->thread_state() != _thread_blocked, "invariant");
712
713 ObjectMonitor* monitor = NULL;
714 markWord temp, test;
715 intptr_t hash;
716 markWord mark = read_stable_mark(obj);
717
718 // object should remain ineligible for biased locking
719 assert(!mark.has_bias_pattern(), "invariant");
720
721 if (mark.is_neutral()) {
722 hash = mark.hash(); // this is a normal header
723 if (hash != 0) { // if it has hash, just return it
724 return hash;
725 }
726 hash = get_next_hash(self, obj); // allocate a new hash code
727 temp = mark.copy_set_hash(hash); // merge the hash code into header
728 // use (machine word version) atomic operation to install the hash
729 test = obj->cas_set_mark(temp, mark);
730 if (test == mark) {
731 return hash;
732 }
733 // If atomic operation failed, we must inflate the header
734 // into heavy weight monitor. We could add more code here
735 // for fast path, but it does not worth the complexity.
736 } else if (mark.has_monitor()) {
737 monitor = mark.monitor();
738 temp = monitor->header();
739 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
740 hash = temp.hash();
741 if (hash != 0) {
742 return hash;
743 }
744 // Skip to the following code to reduce code size
745 } else if (self->is_lock_owned((address)mark.locker())) {
746 temp = mark.displaced_mark_helper(); // this is a lightweight monitor owned
747 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
748 hash = temp.hash(); // by current thread, check if the displaced
749 if (hash != 0) { // header contains hash code
750 return hash;
751 }
752 // WARNING:
753 // The displaced header in the BasicLock on a thread's stack
754 // is strictly immutable. It CANNOT be changed in ANY cases.
755 // So we have to inflate the stack lock into an ObjectMonitor
756 // even if the current thread owns the lock. The BasicLock on
757 // a thread's stack can be asynchronously read by other threads
758 // during an inflate() call so any change to that stack memory
759 // may not propagate to other threads correctly.
760 }
761
762 // Inflate the monitor to set hash code
763 monitor = inflate(self, obj, inflate_cause_hash_code);
764 // Load displaced header and check it has hash code
765 mark = monitor->header();
766 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
767 hash = mark.hash();
768 if (hash == 0) {
769 hash = get_next_hash(self, obj);
770 temp = mark.copy_set_hash(hash); // merge hash code into header
771 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
772 uintptr_t v = Atomic::cmpxchg(temp.value(), (volatile uintptr_t*)monitor->header_addr(), mark.value());
773 test = markWord(v);
774 if (test != mark) {
775 // The only non-deflation update to the ObjectMonitor's
776 // header/dmw field is to merge in the hash code. If someone
777 // adds a new usage of the header/dmw field, please update
778 // this code.
779 hash = test.hash();
780 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
781 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
782 }
783 }
784 // We finally get the hash
785 return hash;
786 }
787
788 // Deprecated -- use FastHashCode() instead.
789
790 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
791 return FastHashCode(Thread::current(), obj());
792 }
793
794
795 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
796 Handle h_obj) {
797 if (UseBiasedLocking) {
798 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
799 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now");
800 }
801
802 assert(thread == JavaThread::current(), "Can only be called on current thread");
803 oop obj = h_obj();
804
805 markWord mark = read_stable_mark(obj);
806
807 // Uncontended case, header points to stack
808 if (mark.has_locker()) {
809 return thread->is_lock_owned((address)mark.locker());
810 }
811 // Contended case, header points to ObjectMonitor (tagged pointer)
812 if (mark.has_monitor()) {
813 ObjectMonitor* monitor = mark.monitor();
814 return monitor->is_entered(thread) != 0;
815 }
816 // Unlocked case, header in place
817 assert(mark.is_neutral(), "sanity check");
818 return false;
819 }
820
821 // Be aware of this method could revoke bias of the lock object.
822 // This method queries the ownership of the lock handle specified by 'h_obj'.
823 // If the current thread owns the lock, it returns owner_self. If no
824 // thread owns the lock, it returns owner_none. Otherwise, it will return
825 // owner_other.
826 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
827 (JavaThread *self, Handle h_obj) {
828 // The caller must beware this method can revoke bias, and
829 // revocation can result in a safepoint.
830 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
831 assert(self->thread_state() != _thread_blocked, "invariant");
832
833 // Possible mark states: neutral, biased, stack-locked, inflated
834
835 if (UseBiasedLocking && h_obj()->mark().has_bias_pattern()) {
836 // CASE: biased
837 BiasedLocking::revoke_and_rebias(h_obj, false, self);
838 assert(!h_obj->mark().has_bias_pattern(),
839 "biases should be revoked by now");
840 }
841
842 assert(self == JavaThread::current(), "Can only be called on current thread");
843 oop obj = h_obj();
844 markWord mark = read_stable_mark(obj);
845
846 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
847 if (mark.has_locker()) {
848 return self->is_lock_owned((address)mark.locker()) ?
849 owner_self : owner_other;
850 }
851
852 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor.
853 // The Object:ObjectMonitor relationship is stable as long as we're
854 // not at a safepoint.
855 if (mark.has_monitor()) {
856 void* owner = mark.monitor()->_owner;
857 if (owner == NULL) return owner_none;
858 return (owner == self ||
859 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
860 }
861
862 // CASE: neutral
863 assert(mark.is_neutral(), "sanity check");
864 return owner_none; // it's unlocked
865 }
866
867 // FIXME: jvmti should call this
868 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
869 if (UseBiasedLocking) {
870 if (SafepointSynchronize::is_at_safepoint()) {
871 BiasedLocking::revoke_at_safepoint(h_obj);
872 } else {
873 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
874 }
875 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now");
876 }
877
878 oop obj = h_obj();
879 address owner = NULL;
880
881 markWord mark = read_stable_mark(obj);
882
883 // Uncontended case, header points to stack
884 if (mark.has_locker()) {
885 owner = (address) mark.locker();
886 }
887
888 // Contended case, header points to ObjectMonitor (tagged pointer)
889 else if (mark.has_monitor()) {
890 ObjectMonitor* monitor = mark.monitor();
891 assert(monitor != NULL, "monitor should be non-null");
892 owner = (address) monitor->owner();
893 }
894
895 if (owner != NULL) {
896 // owning_thread_from_monitor_owner() may also return NULL here
897 return Threads::owning_thread_from_monitor_owner(t_list, owner);
898 }
899
900 // Unlocked case, header in place
901 // Cannot have assertion since this object may have been
902 // locked by another thread when reaching here.
903 // assert(mark.is_neutral(), "sanity check");
904
905 return NULL;
906 }
907
908 // Visitors ...
909
910 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
911 PaddedObjectMonitor* block = OrderAccess::load_acquire(&g_block_list);
912 while (block != NULL) {
913 assert(block->object() == CHAINMARKER, "must be a block header");
914 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
915 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
916 oop object = (oop)mid->object();
917 if (object != NULL) {
918 // Only process with closure if the object is set.
919 closure->do_monitor(mid);
920 }
921 }
922 block = (PaddedObjectMonitor*)block->_next_om;
923 }
924 }
925
926 static bool monitors_used_above_threshold() {
927 if (g_om_population == 0) {
928 return false;
929 }
930 int monitors_used = g_om_population - g_om_free_count;
931 int monitor_usage = (monitors_used * 100LL) / g_om_population;
932 return monitor_usage > MonitorUsedDeflationThreshold;
933 }
934
935 bool ObjectSynchronizer::is_cleanup_needed() {
936 if (MonitorUsedDeflationThreshold > 0) {
937 return monitors_used_above_threshold();
938 }
939 return false;
940 }
941
942 void ObjectSynchronizer::oops_do(OopClosure* f) {
943 // We only scan the global used list here (for moribund threads), and
944 // the thread-local monitors in Thread::oops_do().
945 global_used_oops_do(f);
946 }
947
948 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
949 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
950 list_oops_do(g_om_in_use_list, f);
951 }
952
953 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
954 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
955 list_oops_do(thread->om_in_use_list, f);
956 }
957
958 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
959 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
960 ObjectMonitor* mid;
961 for (mid = list; mid != NULL; mid = mid->_next_om) {
973 // associates them with objects. Deflation -- which occurs at
974 // STW-time -- disassociates idle monitors from objects. Such
975 // scavenged monitors are returned to the g_free_list.
976 //
977 // The global list is protected by gListLock. All the critical sections
978 // are short and operate in constant-time.
979 //
980 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
981 //
982 // Lifecycle:
983 // -- unassigned and on the global free list
984 // -- unassigned and on a thread's private om_free_list
985 // -- assigned to an object. The object is inflated and the mark refers
986 // to the objectmonitor.
987
988
989 // Constraining monitor pool growth via MonitorBound ...
990 //
991 // If MonitorBound is not set (<= 0), MonitorBound checks are disabled.
992 //
993 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
994 // the rate of scavenging is driven primarily by GC. As such, we can find
995 // an inordinate number of monitors in circulation.
996 // To avoid that scenario we can artificially induce a STW safepoint
997 // if the pool appears to be growing past some reasonable bound.
998 // Generally we favor time in space-time tradeoffs, but as there's no
999 // natural back-pressure on the # of extant monitors we need to impose some
1000 // type of limit. Beware that if MonitorBound is set to too low a value
1001 // we could just loop. In addition, if MonitorBound is set to a low value
1002 // we'll incur more safepoints, which are harmful to performance.
1003 // See also: GuaranteedSafepointInterval
1004 //
1005 // The current implementation uses asynchronous VM operations.
1006 //
1007 // If MonitorBound is set, the boundry applies to
1008 // (g_om_population - g_om_free_count)
1009 // i.e., if there are not enough ObjectMonitors on the global free list,
1010 // then a safepoint deflation is induced. Picking a good MonitorBound value
1011 // is non-trivial.
1012
1013 static void InduceScavenge(Thread* self, const char * Whence) {
1014 // Induce STW safepoint to trim monitors
1015 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1016 // More precisely, trigger an asynchronous STW safepoint as the number
1017 // of active monitors passes the specified threshold.
1018 // TODO: assert thread state is reasonable
1019
1020 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1021 // Induce a 'null' safepoint to scavenge monitors
1022 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1023 // to the VMthread and have a lifespan longer than that of this activation record.
1024 // The VMThread will delete the op when completed.
1025 VMThread::execute(new VM_ScavengeMonitors());
1026 }
1027 }
1028
1029 ObjectMonitor* ObjectSynchronizer::om_alloc(Thread* self) {
1030 // A large MAXPRIVATE value reduces both list lock contention
1031 // and list coherency traffic, but also tends to increase the
1032 // number of ObjectMonitors in circulation as well as the STW
1033 // scavenge costs. As usual, we lean toward time in space-time
1034 // tradeoffs.
1035 const int MAXPRIVATE = 1024;
1036 stringStream ss;
1037 for (;;) {
1038 ObjectMonitor* m;
1039
1040 // 1: try to allocate from the thread's local om_free_list.
1041 // Threads will attempt to allocate first from their local list, then
1042 // from the global list, and only after those attempts fail will the thread
1043 // attempt to instantiate new monitors. Thread-local free lists take
1044 // heat off the gListLock and improve allocation latency, as well as reducing
1045 // coherency traffic on the shared global list.
1046 m = self->om_free_list;
1047 if (m != NULL) {
1048 self->om_free_list = m->_next_om;
1049 self->om_free_count--;
1050 guarantee(m->object() == NULL, "invariant");
1051 m->_next_om = self->om_in_use_list;
1052 self->om_in_use_list = m;
1053 self->om_in_use_count++;
1054 return m;
1055 }
1056
1057 // 2: try to allocate from the global g_free_list
1058 // CONSIDER: use muxTry() instead of muxAcquire().
1059 // If the muxTry() fails then drop immediately into case 3.
1060 // If we're using thread-local free lists then try
1061 // to reprovision the caller's free list.
1062 if (g_free_list != NULL) {
1063 // Reprovision the thread's om_free_list.
1064 // Use bulk transfers to reduce the allocation rate and heat
1065 // on various locks.
1066 Thread::muxAcquire(&gListLock, "om_alloc(1)");
1067 for (int i = self->om_free_provision; --i >= 0 && g_free_list != NULL;) {
1068 g_om_free_count--;
1069 ObjectMonitor* take = g_free_list;
1070 g_free_list = take->_next_om;
1071 guarantee(take->object() == NULL, "invariant");
1072 take->Recycle();
1073 om_release(self, take, false);
1074 }
1075 Thread::muxRelease(&gListLock);
1076 self->om_free_provision += 1 + (self->om_free_provision/2);
1077 if (self->om_free_provision > MAXPRIVATE) self->om_free_provision = MAXPRIVATE;
1078
1079 const int mx = MonitorBound;
1080 if (mx > 0 && (g_om_population-g_om_free_count) > mx) {
1081 // Not enough ObjectMonitors on the global free list.
1082 // We can't safely induce a STW safepoint from om_alloc() as our thread
1083 // state may not be appropriate for such activities and callers may hold
1084 // naked oops, so instead we defer the action.
1085 InduceScavenge(self, "om_alloc");
1086 }
1087 continue;
1088 }
1089
1090 // 3: allocate a block of new ObjectMonitors
1091 // Both the local and global free lists are empty -- resort to malloc().
1092 // In the current implementation ObjectMonitors are TSM - immortal.
1093 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1094 // each ObjectMonitor to start at the beginning of a cache line,
1095 // so we use align_up().
1096 // A better solution would be to use C++ placement-new.
1097 // BEWARE: As it stands currently, we don't run the ctors!
1098 assert(_BLOCKSIZE > 1, "invariant");
1099 size_t neededsize = sizeof(PaddedObjectMonitor) * _BLOCKSIZE;
1100 PaddedObjectMonitor* temp;
1105
1106 // NOTE: (almost) no way to recover if allocation failed.
1107 // We might be able to induce a STW safepoint and scavenge enough
1108 // ObjectMonitors to permit progress.
1109 if (temp == NULL) {
1110 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1111 "Allocate ObjectMonitors");
1112 }
1113 (void)memset((void *) temp, 0, neededsize);
1114
1115 // Format the block.
1116 // initialize the linked list, each monitor points to its next
1117 // forming the single linked free list, the very first monitor
1118 // will points to next block, which forms the block list.
1119 // The trick of using the 1st element in the block as g_block_list
1120 // linkage should be reconsidered. A better implementation would
1121 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1122
1123 for (int i = 1; i < _BLOCKSIZE; i++) {
1124 temp[i]._next_om = (ObjectMonitor *)&temp[i+1];
1125 }
1126
1127 // terminate the last monitor as the end of list
1128 temp[_BLOCKSIZE - 1]._next_om = NULL;
1129
1130 // Element [0] is reserved for global list linkage
1131 temp[0].set_object(CHAINMARKER);
1132
1133 // Consider carving out this thread's current request from the
1134 // block in hand. This avoids some lock traffic and redundant
1135 // list activity.
1136
1137 // Acquire the gListLock to manipulate g_block_list and g_free_list.
1138 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1139 Thread::muxAcquire(&gListLock, "om_alloc(2)");
1140 g_om_population += _BLOCKSIZE-1;
1141 g_om_free_count += _BLOCKSIZE-1;
1142
1143 // Add the new block to the list of extant blocks (g_block_list).
1144 // The very first ObjectMonitor in a block is reserved and dedicated.
1147 // There are lock-free uses of g_block_list so make sure that
1148 // the previous stores happen before we update g_block_list.
1149 OrderAccess::release_store(&g_block_list, temp);
1150
1151 // Add the new string of ObjectMonitors to the global free list
1152 temp[_BLOCKSIZE - 1]._next_om = g_free_list;
1153 g_free_list = temp + 1;
1154 Thread::muxRelease(&gListLock);
1155 }
1156 }
1157
1158 // Place "m" on the caller's private per-thread om_free_list.
1159 // In practice there's no need to clamp or limit the number of
1160 // monitors on a thread's om_free_list as the only non-allocation time
1161 // we'll call om_release() is to return a monitor to the free list after
1162 // a CAS attempt failed. This doesn't allow unbounded #s of monitors to
1163 // accumulate on a thread's free list.
1164 //
1165 // Key constraint: all ObjectMonitors on a thread's free list and the global
1166 // free list must have their object field set to null. This prevents the
1167 // scavenger -- deflate_monitor_list() -- from reclaiming them while we
1168 // are trying to release them.
1169
1170 void ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m,
1171 bool from_per_thread_alloc) {
1172 guarantee(m->header().value() == 0, "invariant");
1173 guarantee(m->object() == NULL, "invariant");
1174 stringStream ss;
1175 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: "
1176 "%s, recursions=" INTPTR_FORMAT, m->is_busy_to_string(&ss),
1177 m->_recursions);
1178 // _next_om is used for both per-thread in-use and free lists so
1179 // we have to remove 'm' from the in-use list first (as needed).
1180 if (from_per_thread_alloc) {
1181 // Need to remove 'm' from om_in_use_list.
1182 ObjectMonitor* cur_mid_in_use = NULL;
1183 bool extracted = false;
1184 for (ObjectMonitor* mid = self->om_in_use_list; mid != NULL; cur_mid_in_use = mid, mid = mid->_next_om) {
1185 if (m == mid) {
1186 // extract from per-thread in-use list
1187 if (mid == self->om_in_use_list) {
1188 self->om_in_use_list = mid->_next_om;
1189 } else if (cur_mid_in_use != NULL) {
1190 cur_mid_in_use->_next_om = mid->_next_om; // maintain the current thread in-use list
1191 }
1192 extracted = true;
1193 self->om_in_use_count--;
1194 break;
1195 }
1196 }
1197 assert(extracted, "Should have extracted from in-use list");
1198 }
1199
1200 m->_next_om = self->om_free_list;
1201 self->om_free_list = m;
1202 self->om_free_count++;
1203 }
1204
1205 // Return ObjectMonitors on a moribund thread's free and in-use
1206 // lists to the appropriate global lists. The ObjectMonitors on the
1207 // per-thread in-use list may still be in use by other threads.
1208 //
1209 // We currently call om_flush() from Threads::remove() before the
1210 // thread has been excised from the thread list and is no longer a
1211 // mutator. This means that om_flush() cannot run concurrently with
1212 // a safepoint and interleave with deflate_idle_monitors(). In
1213 // particular, this ensures that the thread's in-use monitors are
1214 // scanned by a GC safepoint, either via Thread::oops_do() (before
1215 // om_flush() is called) or via ObjectSynchronizer::oops_do() (after
1216 // om_flush() is called).
1217
1218 void ObjectSynchronizer::om_flush(Thread* self) {
1219 ObjectMonitor* free_list = self->om_free_list;
1220 ObjectMonitor* free_tail = NULL;
1221 int free_count = 0;
1222 if (free_list != NULL) {
1223 ObjectMonitor* s;
1224 // The thread is going away. Set 'free_tail' to the last per-thread free
1225 // monitor which will be linked to g_free_list below under the gListLock.
1226 stringStream ss;
1227 for (s = free_list; s != NULL; s = s->_next_om) {
1228 free_count++;
1229 free_tail = s;
1230 guarantee(s->object() == NULL, "invariant");
1231 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss));
1232 }
1233 guarantee(free_tail != NULL, "invariant");
1234 assert(self->om_free_count == free_count, "free-count off");
1235 self->om_free_list = NULL;
1236 self->om_free_count = 0;
1237 }
1238
1239 ObjectMonitor* in_use_list = self->om_in_use_list;
1240 ObjectMonitor* in_use_tail = NULL;
1241 int in_use_count = 0;
1242 if (in_use_list != NULL) {
1243 // The thread is going away, however the ObjectMonitors on the
1244 // om_in_use_list may still be in-use by other threads. Link
1245 // them to in_use_tail, which will be linked into the global
1246 // in-use list g_om_in_use_list below, under the gListLock.
1247 ObjectMonitor *cur_om;
1248 for (cur_om = in_use_list; cur_om != NULL; cur_om = cur_om->_next_om) {
1249 in_use_tail = cur_om;
1250 in_use_count++;
1251 }
1252 guarantee(in_use_tail != NULL, "invariant");
1253 assert(self->om_in_use_count == in_use_count, "in-use count off");
1254 self->om_in_use_list = NULL;
1255 self->om_in_use_count = 0;
1256 }
1257
1258 Thread::muxAcquire(&gListLock, "om_flush");
1259 if (free_tail != NULL) {
1260 free_tail->_next_om = g_free_list;
1261 g_free_list = free_list;
1262 g_om_free_count += free_count;
1263 }
1264
1265 if (in_use_tail != NULL) {
1266 in_use_tail->_next_om = g_om_in_use_list;
1267 g_om_in_use_list = in_use_list;
1268 g_om_in_use_count += in_use_count;
1269 }
1270
1271 Thread::muxRelease(&gListLock);
1272
1273 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1281 }
1282 if (ls != NULL) {
1283 ls->print_cr("om_flush: jt=" INTPTR_FORMAT ", free_count=%d"
1284 ", in_use_count=%d" ", om_free_provision=%d",
1285 p2i(self), free_count, in_use_count, self->om_free_provision);
1286 }
1287 }
1288
1289 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1290 const oop obj,
1291 ObjectSynchronizer::InflateCause cause) {
1292 assert(event != NULL, "invariant");
1293 assert(event->should_commit(), "invariant");
1294 event->set_monitorClass(obj->klass());
1295 event->set_address((uintptr_t)(void*)obj);
1296 event->set_cause((u1)cause);
1297 event->commit();
1298 }
1299
1300 // Fast path code shared by multiple functions
1301 void ObjectSynchronizer::inflate_helper(oop obj) {
1302 markWord mark = obj->mark();
1303 if (mark.has_monitor()) {
1304 assert(ObjectSynchronizer::verify_objmon_isinpool(mark.monitor()), "monitor is invalid");
1305 assert(mark.monitor()->header().is_neutral(), "monitor must record a good object header");
1306 return;
1307 }
1308 inflate(Thread::current(), obj, inflate_cause_vm_internal);
1309 }
1310
1311 ObjectMonitor* ObjectSynchronizer::inflate(Thread* self,
1312 oop object,
1313 const InflateCause cause) {
1314 // Inflate mutates the heap ...
1315 // Relaxing assertion for bug 6320749.
1316 assert(Universe::verify_in_progress() ||
1317 !SafepointSynchronize::is_at_safepoint(), "invariant");
1318
1319 EventJavaMonitorInflate event;
1320
1321 for (;;) {
1322 const markWord mark = object->mark();
1323 assert(!mark.has_bias_pattern(), "invariant");
1324
1325 // The mark can be in one of the following states:
1326 // * Inflated - just return
1327 // * Stack-locked - coerce it to inflated
1328 // * INFLATING - busy wait for conversion to complete
1329 // * Neutral - aggressively inflate the object.
1330 // * BIASED - Illegal. We should never see this
1331
1332 // CASE: inflated
1333 if (mark.has_monitor()) {
1334 ObjectMonitor* inf = mark.monitor();
1335 markWord dmw = inf->header();
1336 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1337 assert(oopDesc::equals((oop) inf->object(), object), "invariant");
1338 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1339 return inf;
1340 }
1341
1342 // CASE: inflation in progress - inflating over a stack-lock.
1343 // Some other thread is converting from stack-locked to inflated.
1344 // Only that thread can complete inflation -- other threads must wait.
1345 // The INFLATING value is transient.
1346 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1347 // We could always eliminate polling by parking the thread on some auxiliary list.
1348 if (mark == markWord::INFLATING()) {
1349 read_stable_mark(object);
1350 continue;
1351 }
1352
1353 // CASE: stack-locked
1354 // Could be stack-locked either by this thread or by some other thread.
1355 //
1356 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1357 // to install INFLATING into the mark word. We originally installed INFLATING,
1358 // allocated the objectmonitor, and then finally STed the address of the
1359 // objectmonitor into the mark. This was correct, but artificially lengthened
1360 // the interval in which INFLATED appeared in the mark, thus increasing
1361 // the odds of inflation contention.
1362 //
1363 // We now use per-thread private objectmonitor free lists.
1364 // These list are reprovisioned from the global free list outside the
1365 // critical INFLATING...ST interval. A thread can transfer
1366 // multiple objectmonitors en-mass from the global free list to its local free list.
1367 // This reduces coherency traffic and lock contention on the global free list.
1368 // Using such local free lists, it doesn't matter if the om_alloc() call appears
1369 // before or after the CAS(INFLATING) operation.
1370 // See the comments in om_alloc().
1371
1372 LogStreamHandle(Trace, monitorinflation) lsh;
1373
1374 if (mark.has_locker()) {
1375 ObjectMonitor* m = om_alloc(self);
1376 // Optimistically prepare the objectmonitor - anticipate successful CAS
1377 // We do this before the CAS in order to minimize the length of time
1378 // in which INFLATING appears in the mark.
1379 m->Recycle();
1380 m->_Responsible = NULL;
1381 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1382
1383 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1384 if (cmp != mark) {
1385 om_release(self, m, true);
1386 continue; // Interference -- just retry
1387 }
1388
1389 // We've successfully installed INFLATING (0) into the mark-word.
1390 // This is the only case where 0 will appear in a mark-word.
1391 // Only the singular thread that successfully swings the mark-word
1392 // to 0 can perform (or more precisely, complete) inflation.
1393 //
1394 // Why do we CAS a 0 into the mark-word instead of just CASing the
1395 // mark-word from the stack-locked value directly to the new inflated state?
1402 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1403 // the while preserving the hashCode stability invariants. If the owner
1404 // decides to release the lock while the value is 0, the unlock will fail
1405 // and control will eventually pass from slow_exit() to inflate. The owner
1406 // will then spin, waiting for the 0 value to disappear. Put another way,
1407 // the 0 causes the owner to stall if the owner happens to try to
1408 // drop the lock (restoring the header from the BasicLock to the object)
1409 // while inflation is in-progress. This protocol avoids races that might
1410 // would otherwise permit hashCode values to change or "flicker" for an object.
1411 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1412 // 0 serves as a "BUSY" inflate-in-progress indicator.
1413
1414
1415 // fetch the displaced mark from the owner's stack.
1416 // The owner can't die or unwind past the lock while our INFLATING
1417 // object is in the mark. Furthermore the owner can't complete
1418 // an unlock on the object, either.
1419 markWord dmw = mark.displaced_mark_helper();
1420 // Catch if the object's header is not neutral (not locked and
1421 // not marked is what we care about here).
1422 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1423
1424 // Setup monitor fields to proper values -- prepare the monitor
1425 m->set_header(dmw);
1426
1427 // Optimization: if the mark.locker stack address is associated
1428 // with this thread we could simply set m->_owner = self.
1429 // Note that a thread can inflate an object
1430 // that it has stack-locked -- as might happen in wait() -- directly
1431 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1432 m->set_owner(mark.locker());
1433 m->set_object(object);
1434 // TODO-FIXME: assert BasicLock->dhw != 0.
1435
1436 // Must preserve store ordering. The monitor state must
1437 // be stable at the time of publishing the monitor address.
1438 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1439 object->release_set_mark(markWord::encode(m));
1440
1441 // Hopefully the performance counters are allocated on distinct cache lines
1442 // to avoid false sharing on MP systems ...
1443 OM_PERFDATA_OP(Inflations, inc());
1444 if (log_is_enabled(Trace, monitorinflation)) {
1445 ResourceMark rm(self);
1446 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1447 INTPTR_FORMAT ", type='%s'", p2i(object),
1448 object->mark().value(), object->klass()->external_name());
1449 }
1450 if (event.should_commit()) {
1451 post_monitor_inflate_event(&event, object, cause);
1452 }
1453 return m;
1454 }
1455
1456 // CASE: neutral
1457 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1458 // If we know we're inflating for entry it's better to inflate by swinging a
1459 // pre-locked ObjectMonitor pointer into the object header. A successful
1460 // CAS inflates the object *and* confers ownership to the inflating thread.
1461 // In the current implementation we use a 2-step mechanism where we CAS()
1462 // to inflate and then CAS() again to try to swing _owner from NULL to self.
1463 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1464 // would be useful.
1465
1466 // Catch if the object's header is not neutral (not locked and
1467 // not marked is what we care about here).
1468 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1469 ObjectMonitor* m = om_alloc(self);
1470 // prepare m for installation - set monitor to initial state
1471 m->Recycle();
1472 m->set_header(mark);
1473 m->set_object(object);
1474 m->_Responsible = NULL;
1475 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1476
1477 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1478 m->set_header(markWord::zero());
1479 m->set_object(NULL);
1480 m->Recycle();
1481 om_release(self, m, true);
1482 m = NULL;
1483 continue;
1484 // interference - the markword changed - just retry.
1485 // The state-transitions are one-way, so there's no chance of
1486 // live-lock -- "Inflated" is an absorbing state.
1487 }
1488
1489 // Hopefully the performance counters are allocated on distinct
1490 // cache lines to avoid false sharing on MP systems ...
1491 OM_PERFDATA_OP(Inflations, inc());
1492 if (log_is_enabled(Trace, monitorinflation)) {
1493 ResourceMark rm(self);
1494 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1495 INTPTR_FORMAT ", type='%s'", p2i(object),
1496 object->mark().value(), object->klass()->external_name());
1497 }
1498 if (event.should_commit()) {
1499 post_monitor_inflate_event(&event, object, cause);
1500 }
1501 return m;
1502 }
1503 }
1504
1505
1506 // We maintain a list of in-use monitors for each thread.
1507 //
1508 // deflate_thread_local_monitors() scans a single thread's in-use list, while
1509 // deflate_idle_monitors() scans only a global list of in-use monitors which
1510 // is populated only as a thread dies (see om_flush()).
1511 //
1512 // These operations are called at all safepoints, immediately after mutators
1513 // are stopped, but before any objects have moved. Collectively they traverse
1514 // the population of in-use monitors, deflating where possible. The scavenged
1515 // monitors are returned to the global monitor free list.
1516 //
1517 // Beware that we scavenge at *every* stop-the-world point. Having a large
1518 // number of monitors in-use could negatively impact performance. We also want
1519 // to minimize the total # of monitors in circulation, as they incur a small
1520 // footprint penalty.
1521 //
1522 // Perversely, the heap size -- and thus the STW safepoint rate --
1523 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1524 // which in turn can mean large(r) numbers of ObjectMonitors in circulation.
1525 // This is an unfortunate aspect of this design.
1526
1527 // Deflate a single monitor if not in-use
1528 // Return true if deflated, false if in-use
1529 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1530 ObjectMonitor** free_head_p,
1531 ObjectMonitor** free_tail_p) {
1532 bool deflated;
1533 // Normal case ... The monitor is associated with obj.
1534 const markWord mark = obj->mark();
1535 guarantee(mark == markWord::encode(mid), "should match: mark="
1536 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, mark.value(),
1537 markWord::encode(mid).value());
1538 // Make sure that mark.monitor() and markWord::encode() agree:
1539 guarantee(mark.monitor() == mid, "should match: monitor()=" INTPTR_FORMAT
1540 ", mid=" INTPTR_FORMAT, p2i(mark.monitor()), p2i(mid));
1541 const markWord dmw = mid->header();
1542 guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1543
1544 if (mid->is_busy()) {
1545 deflated = false;
1546 } else {
1547 // Deflate the monitor if it is no longer being used
1548 // It's idle - scavenge and return to the global free list
1549 // plain old deflation ...
1550 if (log_is_enabled(Trace, monitorinflation)) {
1551 ResourceMark rm;
1552 log_trace(monitorinflation)("deflate_monitor: "
1553 "object=" INTPTR_FORMAT ", mark="
1554 INTPTR_FORMAT ", type='%s'", p2i(obj),
1555 mark.value(), obj->klass()->external_name());
1556 }
1557
1558 // Restore the header back to obj
1559 obj->release_set_mark(dmw);
1560 mid->clear();
1561
1562 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT,
1563 p2i(mid->object()));
1564
1565 // Move the deflated ObjectMonitor to the working free list
1566 // defined by free_head_p and free_tail_p.
1567 if (*free_head_p == NULL) *free_head_p = mid;
1568 if (*free_tail_p != NULL) {
1569 // We append to the list so the caller can use mid->_next_om
1570 // to fix the linkages in its context.
1571 ObjectMonitor* prevtail = *free_tail_p;
1572 // Should have been cleaned up by the caller:
1573 assert(prevtail->_next_om == NULL, "cleaned up deflated?");
1574 prevtail->_next_om = mid;
1575 }
1576 *free_tail_p = mid;
1577 // At this point, mid->_next_om still refers to its current
1578 // value and another ObjectMonitor's _next_om field still
1579 // refers to this ObjectMonitor. Those linkages have to be
1580 // cleaned up by the caller who has the complete context.
1581 deflated = true;
1582 }
1583 return deflated;
1584 }
1585
1586 // Walk a given monitor list, and deflate idle monitors
1587 // The given list could be a per-thread list or a global list
1588 // Caller acquires gListLock as needed.
1589 //
1590 // In the case of parallel processing of thread local monitor lists,
1591 // work is done by Threads::parallel_threads_do() which ensures that
1592 // each Java thread is processed by exactly one worker thread, and
1593 // thus avoid conflicts that would arise when worker threads would
1594 // process the same monitor lists concurrently.
1595 //
1596 // See also ParallelSPCleanupTask and
1597 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1598 // Threads::parallel_java_threads_do() in thread.cpp.
1599 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** list_p,
1600 ObjectMonitor** free_head_p,
1601 ObjectMonitor** free_tail_p) {
1602 ObjectMonitor* mid;
1603 ObjectMonitor* next;
1604 ObjectMonitor* cur_mid_in_use = NULL;
1605 int deflated_count = 0;
1610 // Deflation succeeded and already updated free_head_p and
1611 // free_tail_p as needed. Finish the move to the local free list
1612 // by unlinking mid from the global or per-thread in-use list.
1613 if (mid == *list_p) {
1614 *list_p = mid->_next_om;
1615 } else if (cur_mid_in_use != NULL) {
1616 cur_mid_in_use->_next_om = mid->_next_om; // maintain the current thread in-use list
1617 }
1618 next = mid->_next_om;
1619 mid->_next_om = NULL; // This mid is current tail in the free_head_p list
1620 mid = next;
1621 deflated_count++;
1622 } else {
1623 cur_mid_in_use = mid;
1624 mid = mid->_next_om;
1625 }
1626 }
1627 return deflated_count;
1628 }
1629
1630 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1631 counters->n_in_use = 0; // currently associated with objects
1632 counters->n_in_circulation = 0; // extant
1633 counters->n_scavenged = 0; // reclaimed (global and per-thread)
1634 counters->per_thread_scavenged = 0; // per-thread scavenge total
1635 counters->per_thread_times = 0.0; // per-thread scavenge times
1636 }
1637
1638 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
1639 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1640 bool deflated = false;
1641
1642 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors
1643 ObjectMonitor* free_tail_p = NULL;
1644 elapsedTimer timer;
1645
1646 if (log_is_enabled(Info, monitorinflation)) {
1647 timer.start();
1648 }
1649
1650 // Prevent om_flush from changing mids in Thread dtor's during deflation
1651 // And in case the vm thread is acquiring a lock during a safepoint
1652 // See e.g. 6320749
1653 Thread::muxAcquire(&gListLock, "deflate_idle_monitors");
1654
1655 // Note: the thread-local monitors lists get deflated in
1656 // a separate pass. See deflate_thread_local_monitors().
1657
1658 // For moribund threads, scan g_om_in_use_list
1659 int deflated_count = 0;
1672 // constant-time list splice - prepend scavenged segment to g_free_list
1673 free_tail_p->_next_om = g_free_list;
1674 g_free_list = free_head_p;
1675 }
1676 Thread::muxRelease(&gListLock);
1677 timer.stop();
1678
1679 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1680 LogStreamHandle(Info, monitorinflation) lsh_info;
1681 LogStream* ls = NULL;
1682 if (log_is_enabled(Debug, monitorinflation)) {
1683 ls = &lsh_debug;
1684 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
1685 ls = &lsh_info;
1686 }
1687 if (ls != NULL) {
1688 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
1689 }
1690 }
1691
1692 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1693 // Report the cumulative time for deflating each thread's idle
1694 // monitors. Note: if the work is split among more than one
1695 // worker thread, then the reported time will likely be more
1696 // than a beginning to end measurement of the phase.
1697 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->per_thread_times, counters->per_thread_scavenged);
1698
1699 g_om_free_count += counters->n_scavenged;
1700
1701 if (log_is_enabled(Debug, monitorinflation)) {
1702 // exit_globals()'s call to audit_and_print_stats() is done
1703 // at the Info level.
1704 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */);
1705 } else if (log_is_enabled(Info, monitorinflation)) {
1706 Thread::muxAcquire(&gListLock, "finish_deflate_idle_monitors");
1707 log_info(monitorinflation)("g_om_population=%d, g_om_in_use_count=%d, "
1708 "g_om_free_count=%d", g_om_population,
1709 g_om_in_use_count, g_om_free_count);
1710 Thread::muxRelease(&gListLock);
1711 }
1712
1713 ForceMonitorScavenge = 0; // Reset
1714
1715 OM_PERFDATA_OP(Deflations, inc(counters->n_scavenged));
1716 OM_PERFDATA_OP(MonExtant, set_value(counters->n_in_circulation));
1717
1718 GVars.stw_random = os::random();
1719 GVars.stw_cycle++;
1720 }
1721
1722 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
1723 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1724
1725 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors
1726 ObjectMonitor* free_tail_p = NULL;
1727 elapsedTimer timer;
1728
1729 if (log_is_enabled(Info, safepoint, cleanup) ||
1730 log_is_enabled(Info, monitorinflation)) {
1731 timer.start();
1732 }
1733
1734 int deflated_count = deflate_monitor_list(thread->om_in_use_list_addr(), &free_head_p, &free_tail_p);
1735
1736 Thread::muxAcquire(&gListLock, "deflate_thread_local_monitors");
1737
1738 // Adjust counters
1739 counters->n_in_circulation += thread->om_in_use_count;
1740 thread->om_in_use_count -= deflated_count;
1741 counters->n_scavenged += deflated_count;
1742 counters->n_in_use += thread->om_in_use_count;
1743 counters->per_thread_scavenged += deflated_count;
1744
1931 void ObjectSynchronizer::chk_free_entry(JavaThread* jt, ObjectMonitor* n,
1932 outputStream * out, int *error_cnt_p) {
1933 stringStream ss;
1934 if (n->is_busy()) {
1935 if (jt != NULL) {
1936 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1937 ": free per-thread monitor must not be busy: %s", p2i(jt),
1938 p2i(n), n->is_busy_to_string(&ss));
1939 } else {
1940 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1941 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss));
1942 }
1943 *error_cnt_p = *error_cnt_p + 1;
1944 }
1945 if (n->header().value() != 0) {
1946 if (jt != NULL) {
1947 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1948 ": free per-thread monitor must have NULL _header "
1949 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n),
1950 n->header().value());
1951 } else {
1952 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1953 "must have NULL _header field: _header=" INTPTR_FORMAT,
1954 p2i(n), n->header().value());
1955 }
1956 *error_cnt_p = *error_cnt_p + 1;
1957 }
1958 if (n->object() != NULL) {
1959 if (jt != NULL) {
1960 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
1961 ": free per-thread monitor must have NULL _object "
1962 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n),
1963 p2i(n->object()));
1964 } else {
1965 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
1966 "must have NULL _object field: _object=" INTPTR_FORMAT,
1967 p2i(n), p2i(n->object()));
1968 }
1969 *error_cnt_p = *error_cnt_p + 1;
1970 }
1971 }
1972
1973 // Check the global free list and count; log the results of the checks.
1974 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out,
1975 int *error_cnt_p) {
1976 int chk_om_free_count = 0;
1977 for (ObjectMonitor* n = g_free_list; n != NULL; n = n->_next_om) {
2104 "equal to chk_om_in_use_count=%d", p2i(jt), jt->om_in_use_count,
2105 chk_om_in_use_count);
2106 *error_cnt_p = *error_cnt_p + 1;
2107 }
2108 }
2109
2110 // Log details about ObjectMonitors on the in-use lists. The 'BHL'
2111 // flags indicate why the entry is in-use, 'object' and 'object type'
2112 // indicate the associated object and its type.
2113 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out,
2114 bool on_exit) {
2115 if (!on_exit) {
2116 // Not at VM exit so grab the global list lock.
2117 Thread::muxAcquire(&gListLock, "log_in_use_monitor_details");
2118 }
2119
2120 stringStream ss;
2121 if (g_om_in_use_count > 0) {
2122 out->print_cr("In-use global monitor info:");
2123 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2124 out->print_cr("%18s %s %18s %18s",
2125 "monitor", "BHL", "object", "object type");
2126 out->print_cr("================== === ================== ==================");
2127 for (ObjectMonitor* n = g_om_in_use_list; n != NULL; n = n->_next_om) {
2128 const oop obj = (oop) n->object();
2129 const markWord mark = n->header();
2130 ResourceMark rm;
2131 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(n),
2132 n->is_busy() != 0, mark.hash() != 0, n->owner() != NULL,
2133 p2i(obj), obj->klass()->external_name());
2134 if (n->is_busy() != 0) {
2135 out->print(" (%s)", n->is_busy_to_string(&ss));
2136 ss.reset();
2137 }
2138 out->cr();
2139 }
2140 }
2141
2142 if (!on_exit) {
2143 Thread::muxRelease(&gListLock);
2144 }
2145
2146 out->print_cr("In-use per-thread monitor info:");
2147 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2148 out->print_cr("%18s %18s %s %18s %18s",
2149 "jt", "monitor", "BHL", "object", "object type");
2150 out->print_cr("================== ================== === ================== ==================");
2151 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2152 for (ObjectMonitor* n = jt->om_in_use_list; n != NULL; n = n->_next_om) {
2153 const oop obj = (oop) n->object();
2154 const markWord mark = n->header();
2155 ResourceMark rm;
2156 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT
2157 " %s", p2i(jt), p2i(n), n->is_busy() != 0,
2158 mark.hash() != 0, n->owner() != NULL, p2i(obj),
2159 obj->klass()->external_name());
2160 if (n->is_busy() != 0) {
2161 out->print(" (%s)", n->is_busy_to_string(&ss));
2162 ss.reset();
2163 }
2164 out->cr();
2165 }
2166 }
2167
2168 out->flush();
2169 }
2170
2171 // Log counts for the global and per-thread monitor lists and return
2172 // the population count.
2173 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) {
2174 int pop_count = 0;
2175 out->print_cr("%18s %10s %10s %10s",
2176 "Global Lists:", "InUse", "Free", "Total");
2177 out->print_cr("================== ========== ========== ==========");
2178 out->print_cr("%18s %10d %10d %10d", "",
2179 g_om_in_use_count, g_om_free_count, g_om_population);
|
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 PaddedObjectMonitor* volatile ObjectSynchronizer::g_block_list = NULL;
121 // Global ObjectMonitor free list. Newly allocated and deflated
122 // ObjectMonitors are prepended here.
123 ObjectMonitor* volatile ObjectSynchronizer::g_free_list = NULL;
124 // Global ObjectMonitor in-use list. When a JavaThread is exiting,
125 // ObjectMonitors on its per-thread in-use list are prepended here.
126 ObjectMonitor* volatile ObjectSynchronizer::g_om_in_use_list = NULL;
127 int ObjectSynchronizer::g_om_in_use_count = 0; // # on g_om_in_use_list
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 g_om_free_count = 0; // # on g_free_list
134 static volatile int g_om_population = 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 markWord 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(markWord::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 == markWord::from_pointer(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 markWord 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(markWord::from_pointer(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().value(), "don't relock with same BasicLock");
379 lock->set_displaced_header(markWord::from_pointer(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(markWord::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. We
465 // intentionally do not use CHECK here because we must exit the
466 // monitor even if an exception is pending.
467 if (monitor->check_owner(THREAD)) {
468 monitor->exit(true, THREAD);
469 }
470 }
471
472 // -----------------------------------------------------------------------------
473 // Internal VM locks on java objects
474 // standard constructor, allows locking failures
475 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool do_lock) {
476 _dolock = do_lock;
477 _thread = thread;
478 _thread->check_for_valid_safepoint_state(false);
479 _obj = obj;
480
481 if (_dolock) {
482 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
483 }
484 }
485
486 ObjectLocker::~ObjectLocker() {
487 if (_dolock) {
488 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
489 }
490 }
491
492
493 // -----------------------------------------------------------------------------
494 // Wait/Notify/NotifyAll
495 // NOTE: must use heavy weight monitor to handle wait()
496 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
497 if (UseBiasedLocking) {
498 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
499 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
500 }
501 if (millis < 0) {
502 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
503 }
504 ObjectMonitorHandle omh;
505 inflate(&omh, THREAD, obj(), inflate_cause_wait);
506 ObjectMonitor* monitor = omh.om_ptr();
507
508 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
509 monitor->wait(millis, true, THREAD);
510
511 // This dummy call is in place to get around dtrace bug 6254741. Once
512 // that's fixed we can uncomment the following line, remove the call
513 // and change this function back into a "void" func.
514 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
515 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
516 return ret_code;
517 }
518
519 void ObjectSynchronizer::wait_uninterruptibly(Handle obj, jlong millis, TRAPS) {
520 if (UseBiasedLocking) {
521 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
522 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
523 }
524 if (millis < 0) {
525 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
526 }
527 ObjectMonitorHandle omh;
528 inflate(&omh, THREAD, obj(), inflate_cause_wait);
529 omh.om_ptr()->wait(millis, false, THREAD);
530 }
531
532 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
533 if (UseBiasedLocking) {
534 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
535 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
536 }
537
538 markWord mark = obj->mark();
539 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) {
540 return;
541 }
542 ObjectMonitorHandle omh;
543 inflate(&omh, THREAD, obj(), inflate_cause_notify);
544 omh.om_ptr()->notify(THREAD);
545 }
546
547 // NOTE: see comment of notify()
548 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
549 if (UseBiasedLocking) {
550 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
551 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
552 }
553
554 markWord mark = obj->mark();
555 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) {
556 return;
557 }
558 ObjectMonitorHandle omh;
559 inflate(&omh, THREAD, obj(), inflate_cause_notify);
560 omh.om_ptr()->notifyAll(THREAD);
561 }
562
563 // -----------------------------------------------------------------------------
564 // Hash Code handling
565 //
566 // Performance concern:
567 // OrderAccess::storestore() calls release() which at one time stored 0
568 // into the global volatile OrderAccess::dummy variable. This store was
569 // unnecessary for correctness. Many threads storing into a common location
570 // causes considerable cache migration or "sloshing" on large SMP systems.
571 // As such, I avoided using OrderAccess::storestore(). In some cases
572 // OrderAccess::fence() -- which incurs local latency on the executing
573 // processor -- is a better choice as it scales on SMP systems.
574 //
575 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
576 // a discussion of coherency costs. Note that all our current reference
577 // platforms provide strong ST-ST order, so the issue is moot on IA32,
578 // x64, and SPARC.
579 //
580 // As a general policy we use "volatile" to control compiler-based reordering
734 Handle hobj(self, obj);
735 // Relaxing assertion for bug 6320749.
736 assert(Universe::verify_in_progress() ||
737 !SafepointSynchronize::is_at_safepoint(),
738 "biases should not be seen by VM thread here");
739 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
740 obj = hobj();
741 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
742 }
743 }
744
745 // hashCode() is a heap mutator ...
746 // Relaxing assertion for bug 6320749.
747 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
748 !SafepointSynchronize::is_at_safepoint(), "invariant");
749 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
750 self->is_Java_thread() , "invariant");
751 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
752 ((JavaThread *)self)->thread_state() != _thread_blocked, "invariant");
753
754 while (true) {
755 ObjectMonitor* monitor = NULL;
756 markWord temp, test;
757 intptr_t hash;
758 markWord mark = read_stable_mark(obj);
759
760 // object should remain ineligible for biased locking
761 assert(!mark.has_bias_pattern(), "invariant");
762
763 if (mark.is_neutral()) {
764 hash = mark.hash(); // this is a normal header
765 if (hash != 0) { // if it has hash, just return it
766 return hash;
767 }
768 hash = get_next_hash(self, obj); // allocate a new hash code
769 temp = mark.copy_set_hash(hash); // merge the hash code into header
770 // use (machine word version) atomic operation to install the hash
771 test = obj->cas_set_mark(temp, mark);
772 if (test == mark) {
773 return hash;
774 }
775 // If atomic operation failed, we must inflate the header
776 // into heavy weight monitor. We could add more code here
777 // for fast path, but it does not worth the complexity.
778 } else if (mark.has_monitor()) {
779 ObjectMonitorHandle omh;
780 if (!omh.save_om_ptr(obj, mark)) {
781 // Lost a race with async deflation so try again.
782 assert(AsyncDeflateIdleMonitors, "sanity check");
783 continue;
784 }
785 monitor = omh.om_ptr();
786 temp = monitor->header();
787 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
788 hash = temp.hash();
789 if (hash != 0) {
790 return hash;
791 }
792 // Skip to the following code to reduce code size
793 } else if (self->is_lock_owned((address)mark.locker())) {
794 temp = mark.displaced_mark_helper(); // this is a lightweight monitor owned
795 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
796 hash = temp.hash(); // by current thread, check if the displaced
797 if (hash != 0) { // header contains hash code
798 return hash;
799 }
800 // WARNING:
801 // The displaced header in the BasicLock on a thread's stack
802 // is strictly immutable. It CANNOT be changed in ANY cases.
803 // So we have to inflate the stack lock into an ObjectMonitor
804 // even if the current thread owns the lock. The BasicLock on
805 // a thread's stack can be asynchronously read by other threads
806 // during an inflate() call so any change to that stack memory
807 // may not propagate to other threads correctly.
808 }
809
810 // Inflate the monitor to set hash code
811 ObjectMonitorHandle omh;
812 inflate(&omh, self, obj, inflate_cause_hash_code);
813 monitor = omh.om_ptr();
814 // Load displaced header and check it has hash code
815 mark = monitor->header();
816 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
817 hash = mark.hash();
818 if (hash == 0) {
819 hash = get_next_hash(self, obj);
820 temp = mark.copy_set_hash(hash); // merge hash code into header
821 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
822 uintptr_t v = Atomic::cmpxchg(temp.value(), (volatile uintptr_t*)monitor->header_addr(), mark.value());
823 test = markWord(v);
824 if (test != mark) {
825 // The only non-deflation update to the ObjectMonitor's
826 // header/dmw field is to merge in the hash code. If someone
827 // adds a new usage of the header/dmw field, please update
828 // this code.
829 // ObjectMonitor::install_displaced_markword_in_object()
830 // does mark the header/dmw field as part of async deflation,
831 // but that protocol cannot happen now due to the
832 // ObjectMonitorHandle above.
833 hash = test.hash();
834 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
835 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
836 }
837 }
838 // We finally get the hash
839 return hash;
840 }
841 }
842
843 // Deprecated -- use FastHashCode() instead.
844
845 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
846 return FastHashCode(Thread::current(), obj());
847 }
848
849
850 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
851 Handle h_obj) {
852 if (UseBiasedLocking) {
853 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
854 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now");
855 }
856
857 assert(thread == JavaThread::current(), "Can only be called on current thread");
858 oop obj = h_obj();
859
860 while (true) {
861 markWord mark = read_stable_mark(obj);
862
863 // Uncontended case, header points to stack
864 if (mark.has_locker()) {
865 return thread->is_lock_owned((address)mark.locker());
866 }
867 // Contended case, header points to ObjectMonitor (tagged pointer)
868 if (mark.has_monitor()) {
869 ObjectMonitorHandle omh;
870 if (!omh.save_om_ptr(obj, mark)) {
871 // Lost a race with async deflation so try again.
872 assert(AsyncDeflateIdleMonitors, "sanity check");
873 continue;
874 }
875 bool ret_code = omh.om_ptr()->is_entered(thread) != 0;
876 return ret_code;
877 }
878 // Unlocked case, header in place
879 assert(mark.is_neutral(), "sanity check");
880 return false;
881 }
882 }
883
884 // Be aware of this method could revoke bias of the lock object.
885 // This method queries the ownership of the lock handle specified by 'h_obj'.
886 // If the current thread owns the lock, it returns owner_self. If no
887 // thread owns the lock, it returns owner_none. Otherwise, it will return
888 // owner_other.
889 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
890 (JavaThread *self, Handle h_obj) {
891 // The caller must beware this method can revoke bias, and
892 // revocation can result in a safepoint.
893 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
894 assert(self->thread_state() != _thread_blocked, "invariant");
895
896 // Possible mark states: neutral, biased, stack-locked, inflated
897
898 if (UseBiasedLocking && h_obj()->mark().has_bias_pattern()) {
899 // CASE: biased
900 BiasedLocking::revoke_and_rebias(h_obj, false, self);
901 assert(!h_obj->mark().has_bias_pattern(),
902 "biases should be revoked by now");
903 }
904
905 assert(self == JavaThread::current(), "Can only be called on current thread");
906 oop obj = h_obj();
907
908 while (true) {
909 markWord mark = read_stable_mark(obj);
910
911 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
912 if (mark.has_locker()) {
913 return self->is_lock_owned((address)mark.locker()) ?
914 owner_self : owner_other;
915 }
916
917 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor.
918 // The Object:ObjectMonitor relationship is stable as long as we're
919 // not at a safepoint and AsyncDeflateIdleMonitors is false.
920 if (mark.has_monitor()) {
921 ObjectMonitorHandle omh;
922 if (!omh.save_om_ptr(obj, mark)) {
923 // Lost a race with async deflation so try again.
924 assert(AsyncDeflateIdleMonitors, "sanity check");
925 continue;
926 }
927 ObjectMonitor* monitor = omh.om_ptr();
928 void* owner = monitor->_owner;
929 if (owner == NULL) return owner_none;
930 return (owner == self ||
931 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
932 }
933
934 // CASE: neutral
935 assert(mark.is_neutral(), "sanity check");
936 return owner_none; // it's unlocked
937 }
938 }
939
940 // FIXME: jvmti should call this
941 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
942 if (UseBiasedLocking) {
943 if (SafepointSynchronize::is_at_safepoint()) {
944 BiasedLocking::revoke_at_safepoint(h_obj);
945 } else {
946 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
947 }
948 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now");
949 }
950
951 oop obj = h_obj();
952
953 while (true) {
954 address owner = NULL;
955 markWord mark = read_stable_mark(obj);
956
957 // Uncontended case, header points to stack
958 if (mark.has_locker()) {
959 owner = (address) mark.locker();
960 }
961
962 // Contended case, header points to ObjectMonitor (tagged pointer)
963 else if (mark.has_monitor()) {
964 ObjectMonitorHandle omh;
965 if (!omh.save_om_ptr(obj, mark)) {
966 // Lost a race with async deflation so try again.
967 assert(AsyncDeflateIdleMonitors, "sanity check");
968 continue;
969 }
970 ObjectMonitor* monitor = omh.om_ptr();
971 assert(monitor != NULL, "monitor should be non-null");
972 owner = (address) monitor->owner();
973 }
974
975 if (owner != NULL) {
976 // owning_thread_from_monitor_owner() may also return NULL here
977 return Threads::owning_thread_from_monitor_owner(t_list, owner);
978 }
979
980 // Unlocked case, header in place
981 // Cannot have assertion since this object may have been
982 // locked by another thread when reaching here.
983 // assert(mark.is_neutral(), "sanity check");
984
985 return NULL;
986 }
987 }
988
989 // Visitors ...
990
991 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
992 PaddedObjectMonitor* block = OrderAccess::load_acquire(&g_block_list);
993 while (block != NULL) {
994 assert(block->object() == CHAINMARKER, "must be a block header");
995 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
996 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
997 if (mid->is_active()) {
998 ObjectMonitorHandle omh(mid);
999
1000 if (mid->object() == NULL ||
1001 (AsyncDeflateIdleMonitors && mid->_owner == DEFLATER_MARKER)) {
1002 // Only process with closure if the object is set.
1003 // For async deflation, race here if monitor is not owned!
1004 // The above ref_count bump (in ObjectMonitorHandle ctr)
1005 // will cause subsequent async deflation to skip it.
1006 // However, previous or concurrent async deflation is a race.
1007 continue;
1008 }
1009 closure->do_monitor(mid);
1010 }
1011 }
1012 block = (PaddedObjectMonitor*)block->_next_om;
1013 }
1014 }
1015
1016 static bool monitors_used_above_threshold() {
1017 if (g_om_population == 0) {
1018 return false;
1019 }
1020 if (MonitorUsedDeflationThreshold > 0) {
1021 int monitors_used = g_om_population - g_om_free_count;
1022 int monitor_usage = (monitors_used * 100LL) / g_om_population;
1023 return monitor_usage > MonitorUsedDeflationThreshold;
1024 }
1025 return false;
1026 }
1027
1028 // Returns true if MonitorBound is set (> 0) and if the specified
1029 // cnt is > MonitorBound. Otherwise returns false.
1030 static bool is_MonitorBound_exceeded(const int cnt) {
1031 const int mx = MonitorBound;
1032 return mx > 0 && cnt > mx;
1033 }
1034
1035 bool ObjectSynchronizer::is_async_deflation_needed() {
1036 if (!AsyncDeflateIdleMonitors) {
1037 return false;
1038 }
1039 if (is_async_deflation_requested()) {
1040 // Async deflation request.
1041 return true;
1042 }
1043 if (AsyncDeflationInterval > 0 &&
1044 time_since_last_async_deflation_ms() > AsyncDeflationInterval &&
1045 monitors_used_above_threshold()) {
1046 // It's been longer than our specified deflate interval and there
1047 // are too many monitors in use. We don't deflate more frequently
1048 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1049 // in order to not swamp the ServiceThread.
1050 _last_async_deflation_time_ns = os::javaTimeNanos();
1051 return true;
1052 }
1053 if (is_MonitorBound_exceeded(g_om_population - g_om_free_count)) {
1054 // Not enough ObjectMonitors on the global free list.
1055 return true;
1056 }
1057 return false;
1058 }
1059
1060 bool ObjectSynchronizer::is_safepoint_deflation_needed() {
1061 if (!AsyncDeflateIdleMonitors) {
1062 if (monitors_used_above_threshold()) {
1063 // Too many monitors in use.
1064 return true;
1065 }
1066 return false;
1067 }
1068 if (is_special_deflation_requested()) {
1069 // For AsyncDeflateIdleMonitors only do a safepoint deflation
1070 // if there is a special deflation request.
1071 return true;
1072 }
1073 return false;
1074 }
1075
1076 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1077 return (os::javaTimeNanos() - _last_async_deflation_time_ns) / (NANOUNITS / MILLIUNITS);
1078 }
1079
1080 void ObjectSynchronizer::oops_do(OopClosure* f) {
1081 // We only scan the global used list here (for moribund threads), and
1082 // the thread-local monitors in Thread::oops_do().
1083 global_used_oops_do(f);
1084 }
1085
1086 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
1087 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1088 list_oops_do(g_om_in_use_list, f);
1089 }
1090
1091 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
1092 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1093 list_oops_do(thread->om_in_use_list, f);
1094 }
1095
1096 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
1097 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1098 ObjectMonitor* mid;
1099 for (mid = list; mid != NULL; mid = mid->_next_om) {
1111 // associates them with objects. Deflation -- which occurs at
1112 // STW-time -- disassociates idle monitors from objects. Such
1113 // scavenged monitors are returned to the g_free_list.
1114 //
1115 // The global list is protected by gListLock. All the critical sections
1116 // are short and operate in constant-time.
1117 //
1118 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
1119 //
1120 // Lifecycle:
1121 // -- unassigned and on the global free list
1122 // -- unassigned and on a thread's private om_free_list
1123 // -- assigned to an object. The object is inflated and the mark refers
1124 // to the objectmonitor.
1125
1126
1127 // Constraining monitor pool growth via MonitorBound ...
1128 //
1129 // If MonitorBound is not set (<= 0), MonitorBound checks are disabled.
1130 //
1131 // When safepoint deflation is being used (!AsyncDeflateIdleMonitors):
1132 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
1133 // the rate of scavenging is driven primarily by GC. As such, we can find
1134 // an inordinate number of monitors in circulation.
1135 // To avoid that scenario we can artificially induce a STW safepoint
1136 // if the pool appears to be growing past some reasonable bound.
1137 // Generally we favor time in space-time tradeoffs, but as there's no
1138 // natural back-pressure on the # of extant monitors we need to impose some
1139 // type of limit. Beware that if MonitorBound is set to too low a value
1140 // we could just loop. In addition, if MonitorBound is set to a low value
1141 // we'll incur more safepoints, which are harmful to performance.
1142 // See also: GuaranteedSafepointInterval
1143 //
1144 // The current implementation uses asynchronous VM operations.
1145 //
1146 // When safepoint deflation is being used and MonitorBound is set, the
1147 // boundry applies to
1148 // (g_om_population - g_om_free_count)
1149 // i.e., if there are not enough ObjectMonitors on the global free list,
1150 // then a safepoint deflation is induced. Picking a good MonitorBound value
1151 // is non-trivial.
1152 //
1153 // When async deflation is being used:
1154 // The monitor pool is still grow-only. Async deflation is requested
1155 // by a safepoint's cleanup phase or by the ServiceThread at periodic
1156 // intervals when is_async_deflation_needed() returns true. In
1157 // addition to other policies that are checked, if there are not
1158 // enough ObjectMonitors on the global free list, then
1159 // is_async_deflation_needed() will return true. The ServiceThread
1160 // calls deflate_global_idle_monitors_using_JT() and also sets the
1161 // per-thread om_request_deflation flag as needed.
1162
1163 static void InduceScavenge(Thread* self, const char * Whence) {
1164 assert(!AsyncDeflateIdleMonitors, "is not used by async deflation");
1165
1166 // Induce STW safepoint to trim monitors
1167 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1168 // More precisely, trigger an asynchronous STW safepoint as the number
1169 // of active monitors passes the specified threshold.
1170 // TODO: assert thread state is reasonable
1171
1172 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1173 // Induce a 'null' safepoint to scavenge monitors
1174 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1175 // to the VMthread and have a lifespan longer than that of this activation record.
1176 // The VMThread will delete the op when completed.
1177 VMThread::execute(new VM_ScavengeMonitors());
1178 }
1179 }
1180
1181 ObjectMonitor* ObjectSynchronizer::om_alloc(Thread* self,
1182 const InflateCause cause) {
1183 // A large MAXPRIVATE value reduces both list lock contention
1184 // and list coherency traffic, but also tends to increase the
1185 // number of ObjectMonitors in circulation as well as the STW
1186 // scavenge costs. As usual, we lean toward time in space-time
1187 // tradeoffs.
1188 const int MAXPRIVATE = 1024;
1189
1190 if (AsyncDeflateIdleMonitors) {
1191 JavaThread* jt = (JavaThread *)self;
1192 if (jt->om_request_deflation && jt->om_in_use_count > 0 &&
1193 cause != inflate_cause_vm_internal) {
1194 // Deflate any per-thread idle monitors for this JavaThread if
1195 // this is not an internal inflation; internal inflations can
1196 // occur in places where it is not safe to pause for a safepoint.
1197 // Clean up your own mess (Gibbs Rule 45). Otherwise, skip this
1198 // deflation. deflate_global_idle_monitors_using_JT() is called
1199 // by the ServiceThread. Per-thread async deflation is triggered
1200 // by the ServiceThread via om_request_deflation.
1201 debug_only(jt->check_for_valid_safepoint_state(false);)
1202 ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT();
1203 }
1204 }
1205
1206 stringStream ss;
1207 for (;;) {
1208 ObjectMonitor* m;
1209
1210 // 1: try to allocate from the thread's local om_free_list.
1211 // Threads will attempt to allocate first from their local list, then
1212 // from the global list, and only after those attempts fail will the thread
1213 // attempt to instantiate new monitors. Thread-local free lists take
1214 // heat off the gListLock and improve allocation latency, as well as reducing
1215 // coherency traffic on the shared global list.
1216 m = self->om_free_list;
1217 if (m != NULL) {
1218 self->om_free_list = m->_next_om;
1219 self->om_free_count--;
1220 guarantee(m->object() == NULL, "invariant");
1221 m->set_allocation_state(ObjectMonitor::New);
1222 m->_next_om = self->om_in_use_list;
1223 self->om_in_use_list = m;
1224 self->om_in_use_count++;
1225 return m;
1226 }
1227
1228 // 2: try to allocate from the global g_free_list
1229 // CONSIDER: use muxTry() instead of muxAcquire().
1230 // If the muxTry() fails then drop immediately into case 3.
1231 // If we're using thread-local free lists then try
1232 // to reprovision the caller's free list.
1233 if (g_free_list != NULL) {
1234 // Reprovision the thread's om_free_list.
1235 // Use bulk transfers to reduce the allocation rate and heat
1236 // on various locks.
1237 Thread::muxAcquire(&gListLock, "om_alloc(1)");
1238 for (int i = self->om_free_provision; --i >= 0 && g_free_list != NULL;) {
1239 g_om_free_count--;
1240 ObjectMonitor* take = g_free_list;
1241 g_free_list = take->_next_om;
1242 guarantee(take->object() == NULL, "invariant");
1243 if (AsyncDeflateIdleMonitors) {
1244 // We allowed 3 field values to linger during async deflation.
1245 // We clear header and restore ref_count here, but we leave
1246 // owner == DEFLATER_MARKER so the simple C2 ObjectMonitor
1247 // enter optimization can no longer race with async deflation
1248 // and reuse.
1249 take->set_header(markWord::zero());
1250 if (take->ref_count() < 0) {
1251 // Add back max_jint to restore the ref_count field to its
1252 // proper value.
1253 Atomic::add(max_jint, &take->_ref_count);
1254
1255 assert(take->ref_count() >= 0, "must not be negative: ref_count=%d",
1256 take->ref_count());
1257 }
1258 }
1259 take->Recycle();
1260 assert(take->is_free(), "invariant");
1261 om_release(self, take, false);
1262 }
1263 Thread::muxRelease(&gListLock);
1264 self->om_free_provision += 1 + (self->om_free_provision/2);
1265 if (self->om_free_provision > MAXPRIVATE) self->om_free_provision = MAXPRIVATE;
1266
1267 if (!AsyncDeflateIdleMonitors &&
1268 is_MonitorBound_exceeded(g_om_population - g_om_free_count)) {
1269 // Not enough ObjectMonitors on the global free list.
1270 // We can't safely induce a STW safepoint from om_alloc() as our thread
1271 // state may not be appropriate for such activities and callers may hold
1272 // naked oops, so instead we defer the action.
1273 InduceScavenge(self, "om_alloc");
1274 }
1275 continue;
1276 }
1277
1278 // 3: allocate a block of new ObjectMonitors
1279 // Both the local and global free lists are empty -- resort to malloc().
1280 // In the current implementation ObjectMonitors are TSM - immortal.
1281 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1282 // each ObjectMonitor to start at the beginning of a cache line,
1283 // so we use align_up().
1284 // A better solution would be to use C++ placement-new.
1285 // BEWARE: As it stands currently, we don't run the ctors!
1286 assert(_BLOCKSIZE > 1, "invariant");
1287 size_t neededsize = sizeof(PaddedObjectMonitor) * _BLOCKSIZE;
1288 PaddedObjectMonitor* temp;
1293
1294 // NOTE: (almost) no way to recover if allocation failed.
1295 // We might be able to induce a STW safepoint and scavenge enough
1296 // ObjectMonitors to permit progress.
1297 if (temp == NULL) {
1298 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1299 "Allocate ObjectMonitors");
1300 }
1301 (void)memset((void *) temp, 0, neededsize);
1302
1303 // Format the block.
1304 // initialize the linked list, each monitor points to its next
1305 // forming the single linked free list, the very first monitor
1306 // will points to next block, which forms the block list.
1307 // The trick of using the 1st element in the block as g_block_list
1308 // linkage should be reconsidered. A better implementation would
1309 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1310
1311 for (int i = 1; i < _BLOCKSIZE; i++) {
1312 temp[i]._next_om = (ObjectMonitor *)&temp[i+1];
1313 assert(temp[i].is_free(), "invariant");
1314 }
1315
1316 // terminate the last monitor as the end of list
1317 temp[_BLOCKSIZE - 1]._next_om = NULL;
1318
1319 // Element [0] is reserved for global list linkage
1320 temp[0].set_object(CHAINMARKER);
1321
1322 // Consider carving out this thread's current request from the
1323 // block in hand. This avoids some lock traffic and redundant
1324 // list activity.
1325
1326 // Acquire the gListLock to manipulate g_block_list and g_free_list.
1327 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1328 Thread::muxAcquire(&gListLock, "om_alloc(2)");
1329 g_om_population += _BLOCKSIZE-1;
1330 g_om_free_count += _BLOCKSIZE-1;
1331
1332 // Add the new block to the list of extant blocks (g_block_list).
1333 // The very first ObjectMonitor in a block is reserved and dedicated.
1336 // There are lock-free uses of g_block_list so make sure that
1337 // the previous stores happen before we update g_block_list.
1338 OrderAccess::release_store(&g_block_list, temp);
1339
1340 // Add the new string of ObjectMonitors to the global free list
1341 temp[_BLOCKSIZE - 1]._next_om = g_free_list;
1342 g_free_list = temp + 1;
1343 Thread::muxRelease(&gListLock);
1344 }
1345 }
1346
1347 // Place "m" on the caller's private per-thread om_free_list.
1348 // In practice there's no need to clamp or limit the number of
1349 // monitors on a thread's om_free_list as the only non-allocation time
1350 // we'll call om_release() is to return a monitor to the free list after
1351 // a CAS attempt failed. This doesn't allow unbounded #s of monitors to
1352 // accumulate on a thread's free list.
1353 //
1354 // Key constraint: all ObjectMonitors on a thread's free list and the global
1355 // free list must have their object field set to null. This prevents the
1356 // scavenger -- deflate_monitor_list() or deflate_monitor_list_using_JT()
1357 // -- from reclaiming them while we are trying to release them.
1358
1359 void ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m,
1360 bool from_per_thread_alloc) {
1361 guarantee(m->header().value() == 0, "invariant");
1362 guarantee(m->object() == NULL, "invariant");
1363 stringStream ss;
1364 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: "
1365 "%s, recursions=" INTPTR_FORMAT, m->is_busy_to_string(&ss),
1366 m->_recursions);
1367 m->set_allocation_state(ObjectMonitor::Free);
1368 // _next_om is used for both per-thread in-use and free lists so
1369 // we have to remove 'm' from the in-use list first (as needed).
1370 if (from_per_thread_alloc) {
1371 // Need to remove 'm' from om_in_use_list.
1372 ObjectMonitor* cur_mid_in_use = NULL;
1373 bool extracted = false;
1374 for (ObjectMonitor* mid = self->om_in_use_list; mid != NULL; cur_mid_in_use = mid, mid = mid->_next_om) {
1375 if (m == mid) {
1376 // extract from per-thread in-use list
1377 if (mid == self->om_in_use_list) {
1378 self->om_in_use_list = mid->_next_om;
1379 } else if (cur_mid_in_use != NULL) {
1380 cur_mid_in_use->_next_om = mid->_next_om; // maintain the current thread in-use list
1381 }
1382 extracted = true;
1383 self->om_in_use_count--;
1384 break;
1385 }
1386 }
1387 assert(extracted, "Should have extracted from in-use list");
1388 }
1389
1390 m->_next_om = self->om_free_list;
1391 guarantee(m->is_free(), "invariant");
1392 self->om_free_list = m;
1393 self->om_free_count++;
1394 }
1395
1396 // Return ObjectMonitors on a moribund thread's free and in-use
1397 // lists to the appropriate global lists. The ObjectMonitors on the
1398 // per-thread in-use list may still be in use by other threads.
1399 //
1400 // We currently call om_flush() from Threads::remove() before the
1401 // thread has been excised from the thread list and is no longer a
1402 // mutator. This means that om_flush() cannot run concurrently with
1403 // a safepoint and interleave with deflate_idle_monitors(). In
1404 // particular, this ensures that the thread's in-use monitors are
1405 // scanned by a GC safepoint, either via Thread::oops_do() (before
1406 // om_flush() is called) or via ObjectSynchronizer::oops_do() (after
1407 // om_flush() is called).
1408 //
1409 // With AsyncDeflateIdleMonitors, deflate_global_idle_monitors_using_JT()
1410 // and deflate_per_thread_idle_monitors_using_JT() (in another thread) can
1411 // run at the same time as om_flush() so we have to be careful.
1412
1413 void ObjectSynchronizer::om_flush(Thread* self) {
1414 ObjectMonitor* free_list = self->om_free_list;
1415 ObjectMonitor* free_tail = NULL;
1416 int free_count = 0;
1417 if (free_list != NULL) {
1418 ObjectMonitor* s;
1419 // The thread is going away. Set 'free_tail' to the last per-thread free
1420 // monitor which will be linked to g_free_list below under the gListLock.
1421 stringStream ss;
1422 for (s = free_list; s != NULL; s = s->_next_om) {
1423 free_count++;
1424 free_tail = s;
1425 guarantee(s->object() == NULL, "invariant");
1426 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss));
1427 }
1428 guarantee(free_tail != NULL, "invariant");
1429 ADIM_guarantee(self->om_free_count == free_count, "free-count off");
1430 self->om_free_list = NULL;
1431 self->om_free_count = 0;
1432 }
1433
1434 ObjectMonitor* in_use_list = self->om_in_use_list;
1435 ObjectMonitor* in_use_tail = NULL;
1436 int in_use_count = 0;
1437 if (in_use_list != NULL) {
1438 // The thread is going away, however the ObjectMonitors on the
1439 // om_in_use_list may still be in-use by other threads. Link
1440 // them to in_use_tail, which will be linked into the global
1441 // in-use list g_om_in_use_list below, under the gListLock.
1442 ObjectMonitor *cur_om;
1443 for (cur_om = in_use_list; cur_om != NULL; cur_om = cur_om->_next_om) {
1444 in_use_tail = cur_om;
1445 in_use_count++;
1446 ADIM_guarantee(cur_om->is_active(), "invariant");
1447 }
1448 guarantee(in_use_tail != NULL, "invariant");
1449 ADIM_guarantee(self->om_in_use_count == in_use_count, "in-use count off");
1450 self->om_in_use_list = NULL;
1451 self->om_in_use_count = 0;
1452 }
1453
1454 Thread::muxAcquire(&gListLock, "om_flush");
1455 if (free_tail != NULL) {
1456 free_tail->_next_om = g_free_list;
1457 g_free_list = free_list;
1458 g_om_free_count += free_count;
1459 }
1460
1461 if (in_use_tail != NULL) {
1462 in_use_tail->_next_om = g_om_in_use_list;
1463 g_om_in_use_list = in_use_list;
1464 g_om_in_use_count += in_use_count;
1465 }
1466
1467 Thread::muxRelease(&gListLock);
1468
1469 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1477 }
1478 if (ls != NULL) {
1479 ls->print_cr("om_flush: jt=" INTPTR_FORMAT ", free_count=%d"
1480 ", in_use_count=%d" ", om_free_provision=%d",
1481 p2i(self), free_count, in_use_count, self->om_free_provision);
1482 }
1483 }
1484
1485 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1486 const oop obj,
1487 ObjectSynchronizer::InflateCause cause) {
1488 assert(event != NULL, "invariant");
1489 assert(event->should_commit(), "invariant");
1490 event->set_monitorClass(obj->klass());
1491 event->set_address((uintptr_t)(void*)obj);
1492 event->set_cause((u1)cause);
1493 event->commit();
1494 }
1495
1496 // Fast path code shared by multiple functions
1497 void ObjectSynchronizer::inflate_helper(ObjectMonitorHandle* omh_p, oop obj) {
1498 while (true) {
1499 markWord mark = obj->mark();
1500 if (mark.has_monitor()) {
1501 if (!omh_p->save_om_ptr(obj, mark)) {
1502 // Lost a race with async deflation so try again.
1503 assert(AsyncDeflateIdleMonitors, "sanity check");
1504 continue;
1505 }
1506 ObjectMonitor* monitor = omh_p->om_ptr();
1507 assert(ObjectSynchronizer::verify_objmon_isinpool(monitor), "monitor is invalid");
1508 markWord dmw = monitor->header();
1509 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1510 return;
1511 }
1512 inflate(omh_p, Thread::current(), obj, inflate_cause_vm_internal);
1513 return;
1514 }
1515 }
1516
1517 void ObjectSynchronizer::inflate(ObjectMonitorHandle* omh_p, Thread* self,
1518 oop object, const InflateCause cause) {
1519 // Inflate mutates the heap ...
1520 // Relaxing assertion for bug 6320749.
1521 assert(Universe::verify_in_progress() ||
1522 !SafepointSynchronize::is_at_safepoint(), "invariant");
1523
1524 EventJavaMonitorInflate event;
1525
1526 for (;;) {
1527 const markWord mark = object->mark();
1528 assert(!mark.has_bias_pattern(), "invariant");
1529
1530 // The mark can be in one of the following states:
1531 // * Inflated - just return
1532 // * Stack-locked - coerce it to inflated
1533 // * INFLATING - busy wait for conversion to complete
1534 // * Neutral - aggressively inflate the object.
1535 // * BIASED - Illegal. We should never see this
1536
1537 // CASE: inflated
1538 if (mark.has_monitor()) {
1539 if (!omh_p->save_om_ptr(object, mark)) {
1540 // Lost a race with async deflation so try again.
1541 assert(AsyncDeflateIdleMonitors, "sanity check");
1542 continue;
1543 }
1544 ObjectMonitor* inf = omh_p->om_ptr();
1545 markWord dmw = inf->header();
1546 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1547 assert(oopDesc::equals((oop) inf->object(), object), "invariant");
1548 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1549 return;
1550 }
1551
1552 // CASE: inflation in progress - inflating over a stack-lock.
1553 // Some other thread is converting from stack-locked to inflated.
1554 // Only that thread can complete inflation -- other threads must wait.
1555 // The INFLATING value is transient.
1556 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1557 // We could always eliminate polling by parking the thread on some auxiliary list.
1558 if (mark == markWord::INFLATING()) {
1559 read_stable_mark(object);
1560 continue;
1561 }
1562
1563 // CASE: stack-locked
1564 // Could be stack-locked either by this thread or by some other thread.
1565 //
1566 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1567 // to install INFLATING into the mark word. We originally installed INFLATING,
1568 // allocated the objectmonitor, and then finally STed the address of the
1569 // objectmonitor into the mark. This was correct, but artificially lengthened
1570 // the interval in which INFLATED appeared in the mark, thus increasing
1571 // the odds of inflation contention.
1572 //
1573 // We now use per-thread private objectmonitor free lists.
1574 // These list are reprovisioned from the global free list outside the
1575 // critical INFLATING...ST interval. A thread can transfer
1576 // multiple objectmonitors en-mass from the global free list to its local free list.
1577 // This reduces coherency traffic and lock contention on the global free list.
1578 // Using such local free lists, it doesn't matter if the om_alloc() call appears
1579 // before or after the CAS(INFLATING) operation.
1580 // See the comments in om_alloc().
1581
1582 LogStreamHandle(Trace, monitorinflation) lsh;
1583
1584 if (mark.has_locker()) {
1585 ObjectMonitor* m;
1586 if (!AsyncDeflateIdleMonitors || cause == inflate_cause_vm_internal) {
1587 // If !AsyncDeflateIdleMonitors or if an internal inflation, then
1588 // we won't stop for a potential safepoint in om_alloc.
1589 m = om_alloc(self, cause);
1590 } else {
1591 // If AsyncDeflateIdleMonitors and not an internal inflation, then
1592 // we may stop for a safepoint in om_alloc() so protect object.
1593 Handle h_obj(self, object);
1594 m = om_alloc(self, cause);
1595 object = h_obj(); // Refresh object.
1596 }
1597 // Optimistically prepare the objectmonitor - anticipate successful CAS
1598 // We do this before the CAS in order to minimize the length of time
1599 // in which INFLATING appears in the mark.
1600 m->Recycle();
1601 m->_Responsible = NULL;
1602 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1603
1604 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1605 if (cmp != mark) {
1606 om_release(self, m, true);
1607 continue; // Interference -- just retry
1608 }
1609
1610 // We've successfully installed INFLATING (0) into the mark-word.
1611 // This is the only case where 0 will appear in a mark-word.
1612 // Only the singular thread that successfully swings the mark-word
1613 // to 0 can perform (or more precisely, complete) inflation.
1614 //
1615 // Why do we CAS a 0 into the mark-word instead of just CASing the
1616 // mark-word from the stack-locked value directly to the new inflated state?
1623 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1624 // the while preserving the hashCode stability invariants. If the owner
1625 // decides to release the lock while the value is 0, the unlock will fail
1626 // and control will eventually pass from slow_exit() to inflate. The owner
1627 // will then spin, waiting for the 0 value to disappear. Put another way,
1628 // the 0 causes the owner to stall if the owner happens to try to
1629 // drop the lock (restoring the header from the BasicLock to the object)
1630 // while inflation is in-progress. This protocol avoids races that might
1631 // would otherwise permit hashCode values to change or "flicker" for an object.
1632 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1633 // 0 serves as a "BUSY" inflate-in-progress indicator.
1634
1635
1636 // fetch the displaced mark from the owner's stack.
1637 // The owner can't die or unwind past the lock while our INFLATING
1638 // object is in the mark. Furthermore the owner can't complete
1639 // an unlock on the object, either.
1640 markWord dmw = mark.displaced_mark_helper();
1641 // Catch if the object's header is not neutral (not locked and
1642 // not marked is what we care about here).
1643 ADIM_guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1644
1645 // Setup monitor fields to proper values -- prepare the monitor
1646 m->set_header(dmw);
1647
1648 // Optimization: if the mark.locker stack address is associated
1649 // with this thread we could simply set m->_owner = self.
1650 // Note that a thread can inflate an object
1651 // that it has stack-locked -- as might happen in wait() -- directly
1652 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1653 m->set_owner(mark.locker());
1654 m->set_object(object);
1655 // TODO-FIXME: assert BasicLock->dhw != 0.
1656
1657 omh_p->set_om_ptr(m);
1658 assert(m->is_new(), "freshly allocated monitor must be new");
1659 m->set_allocation_state(ObjectMonitor::Old);
1660
1661 // Must preserve store ordering. The monitor state must
1662 // be stable at the time of publishing the monitor address.
1663 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1664 object->release_set_mark(markWord::encode(m));
1665
1666 // Hopefully the performance counters are allocated on distinct cache lines
1667 // to avoid false sharing on MP systems ...
1668 OM_PERFDATA_OP(Inflations, inc());
1669 if (log_is_enabled(Trace, monitorinflation)) {
1670 ResourceMark rm(self);
1671 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1672 INTPTR_FORMAT ", type='%s'", p2i(object),
1673 object->mark().value(), object->klass()->external_name());
1674 }
1675 if (event.should_commit()) {
1676 post_monitor_inflate_event(&event, object, cause);
1677 }
1678 ADIM_guarantee(!m->is_free(), "inflated monitor to be returned cannot be free");
1679 return;
1680 }
1681
1682 // CASE: neutral
1683 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1684 // If we know we're inflating for entry it's better to inflate by swinging a
1685 // pre-locked ObjectMonitor pointer into the object header. A successful
1686 // CAS inflates the object *and* confers ownership to the inflating thread.
1687 // In the current implementation we use a 2-step mechanism where we CAS()
1688 // to inflate and then CAS() again to try to swing _owner from NULL to self.
1689 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1690 // would be useful.
1691
1692 // Catch if the object's header is not neutral (not locked and
1693 // not marked is what we care about here).
1694 ADIM_guarantee(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1695 ObjectMonitor* m;
1696 if (!AsyncDeflateIdleMonitors || cause == inflate_cause_vm_internal) {
1697 // If !AsyncDeflateIdleMonitors or if an internal inflation, then
1698 // we won't stop for a potential safepoint in om_alloc.
1699 m = om_alloc(self, cause);
1700 } else {
1701 // If AsyncDeflateIdleMonitors and not an internal inflation, then
1702 // we may stop for a safepoint in om_alloc() so protect object.
1703 Handle h_obj(self, object);
1704 m = om_alloc(self, cause);
1705 object = h_obj(); // Refresh object.
1706 }
1707 // prepare m for installation - set monitor to initial state
1708 m->Recycle();
1709 m->set_header(mark);
1710 // If we leave _owner == DEFLATER_MARKER here, then the simple C2
1711 // ObjectMonitor enter optimization can no longer race with async
1712 // deflation and reuse.
1713 m->set_object(object);
1714 m->_Responsible = NULL;
1715 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1716
1717 omh_p->set_om_ptr(m);
1718 assert(m->is_new(), "freshly allocated monitor must be new");
1719 m->set_allocation_state(ObjectMonitor::Old);
1720
1721 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1722 m->set_header(markWord::zero());
1723 m->set_object(NULL);
1724 m->Recycle();
1725 omh_p->set_om_ptr(NULL);
1726 // om_release() will reset the allocation state
1727 om_release(self, m, true);
1728 m = NULL;
1729 continue;
1730 // interference - the markword changed - just retry.
1731 // The state-transitions are one-way, so there's no chance of
1732 // live-lock -- "Inflated" is an absorbing state.
1733 }
1734
1735 // Hopefully the performance counters are allocated on distinct
1736 // cache lines to avoid false sharing on MP systems ...
1737 OM_PERFDATA_OP(Inflations, inc());
1738 if (log_is_enabled(Trace, monitorinflation)) {
1739 ResourceMark rm(self);
1740 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1741 INTPTR_FORMAT ", type='%s'", p2i(object),
1742 object->mark().value(), object->klass()->external_name());
1743 }
1744 if (event.should_commit()) {
1745 post_monitor_inflate_event(&event, object, cause);
1746 }
1747 ADIM_guarantee(!m->is_free(), "inflated monitor to be returned cannot be free");
1748 return;
1749 }
1750 }
1751
1752
1753 // We maintain a list of in-use monitors for each thread.
1754 //
1755 // For safepoint based deflation:
1756 // deflate_thread_local_monitors() scans a single thread's in-use list, while
1757 // deflate_idle_monitors() scans only a global list of in-use monitors which
1758 // is populated only as a thread dies (see om_flush()).
1759 //
1760 // These operations are called at all safepoints, immediately after mutators
1761 // are stopped, but before any objects have moved. Collectively they traverse
1762 // the population of in-use monitors, deflating where possible. The scavenged
1763 // monitors are returned to the global monitor free list.
1764 //
1765 // Beware that we scavenge at *every* stop-the-world point. Having a large
1766 // number of monitors in-use could negatively impact performance. We also want
1767 // to minimize the total # of monitors in circulation, as they incur a small
1768 // footprint penalty.
1769 //
1770 // Perversely, the heap size -- and thus the STW safepoint rate --
1771 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1772 // which in turn can mean large(r) numbers of ObjectMonitors in circulation.
1773 // This is an unfortunate aspect of this design.
1774 //
1775 // For async deflation:
1776 // If a special deflation request is made, then the safepoint based
1777 // deflation mechanism is used. Otherwise, an async deflation request
1778 // is registered with the ServiceThread and it is notified.
1779
1780 void ObjectSynchronizer::do_safepoint_work(DeflateMonitorCounters* _counters) {
1781 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1782
1783 // The per-thread in-use lists are handled in
1784 // ParallelSPCleanupThreadClosure::do_thread().
1785
1786 if (!AsyncDeflateIdleMonitors || is_special_deflation_requested()) {
1787 // Use the older mechanism for the global in-use list or if a
1788 // special deflation has been requested before the safepoint.
1789 ObjectSynchronizer::deflate_idle_monitors(_counters);
1790 return;
1791 }
1792
1793 log_debug(monitorinflation)("requesting async deflation of idle monitors.");
1794 // Request deflation of idle monitors by the ServiceThread:
1795 set_is_async_deflation_requested(true);
1796 MonitorLocker ml(Service_lock, Mutex::_no_safepoint_check_flag);
1797 ml.notify_all();
1798 }
1799
1800 // Deflate a single monitor if not in-use
1801 // Return true if deflated, false if in-use
1802 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1803 ObjectMonitor** free_head_p,
1804 ObjectMonitor** free_tail_p) {
1805 bool deflated;
1806 // Normal case ... The monitor is associated with obj.
1807 const markWord mark = obj->mark();
1808 guarantee(mark == markWord::encode(mid), "should match: mark="
1809 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, mark.value(),
1810 markWord::encode(mid).value());
1811 // Make sure that mark.monitor() and markWord::encode() agree:
1812 guarantee(mark.monitor() == mid, "should match: monitor()=" INTPTR_FORMAT
1813 ", mid=" INTPTR_FORMAT, p2i(mark.monitor()), p2i(mid));
1814 const markWord dmw = mid->header();
1815 guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1816
1817 if (mid->is_busy() || mid->ref_count() != 0) {
1818 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor*
1819 // is in use so no deflation.
1820 deflated = false;
1821 } else {
1822 // Deflate the monitor if it is no longer being used
1823 // It's idle - scavenge and return to the global free list
1824 // plain old deflation ...
1825 if (log_is_enabled(Trace, monitorinflation)) {
1826 ResourceMark rm;
1827 log_trace(monitorinflation)("deflate_monitor: "
1828 "object=" INTPTR_FORMAT ", mark="
1829 INTPTR_FORMAT ", type='%s'", p2i(obj),
1830 mark.value(), obj->klass()->external_name());
1831 }
1832
1833 // Restore the header back to obj
1834 obj->release_set_mark(dmw);
1835 if (AsyncDeflateIdleMonitors) {
1836 // clear() expects the owner field to be NULL and we won't race
1837 // with the simple C2 ObjectMonitor enter optimization since
1838 // we're at a safepoint.
1839 mid->set_owner(NULL);
1840 }
1841 mid->clear();
1842
1843 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT,
1844 p2i(mid->object()));
1845 assert(mid->is_free(), "invariant");
1846
1847 // Move the deflated ObjectMonitor to the working free list
1848 // defined by free_head_p and free_tail_p.
1849 if (*free_head_p == NULL) *free_head_p = mid;
1850 if (*free_tail_p != NULL) {
1851 // We append to the list so the caller can use mid->_next_om
1852 // to fix the linkages in its context.
1853 ObjectMonitor* prevtail = *free_tail_p;
1854 // Should have been cleaned up by the caller:
1855 assert(prevtail->_next_om == NULL, "cleaned up deflated?");
1856 prevtail->_next_om = mid;
1857 }
1858 *free_tail_p = mid;
1859 // At this point, mid->_next_om still refers to its current
1860 // value and another ObjectMonitor's _next_om field still
1861 // refers to this ObjectMonitor. Those linkages have to be
1862 // cleaned up by the caller who has the complete context.
1863 deflated = true;
1864 }
1865 return deflated;
1866 }
1867
1868 // Deflate the specified ObjectMonitor if not in-use using a JavaThread.
1869 // Returns true if it was deflated and false otherwise.
1870 //
1871 // The async deflation protocol sets owner to DEFLATER_MARKER and
1872 // makes ref_count negative as signals to contending threads that
1873 // an async deflation is in progress. There are a number of checks
1874 // as part of the protocol to make sure that the calling thread has
1875 // not lost the race to a contending thread or to a thread that just
1876 // wants to use the ObjectMonitor*.
1877 //
1878 // The ObjectMonitor has been successfully async deflated when:
1879 // (owner == DEFLATER_MARKER && ref_count < 0)
1880 // Contending threads or ObjectMonitor* using threads that see those
1881 // values know to retry their operation.
1882 //
1883 bool ObjectSynchronizer::deflate_monitor_using_JT(ObjectMonitor* mid,
1884 ObjectMonitor** free_head_p,
1885 ObjectMonitor** free_tail_p) {
1886 assert(AsyncDeflateIdleMonitors, "sanity check");
1887 assert(Thread::current()->is_Java_thread(), "precondition");
1888 // A newly allocated ObjectMonitor should not be seen here so we
1889 // avoid an endless inflate/deflate cycle.
1890 assert(mid->is_old(), "must be old: allocation_state=%d",
1891 (int) mid->allocation_state());
1892
1893 if (mid->is_busy() || mid->ref_count() != 0) {
1894 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor*
1895 // is in use so no deflation.
1896 return false;
1897 }
1898
1899 if (Atomic::replace_if_null(DEFLATER_MARKER, &(mid->_owner))) {
1900 // ObjectMonitor is not owned by another thread. Our setting
1901 // owner to DEFLATER_MARKER forces any contending thread through
1902 // the slow path. This is just the first part of the async
1903 // deflation dance.
1904
1905 if (mid->_contentions != 0 || mid->_waiters != 0) {
1906 // Another thread has raced to enter the ObjectMonitor after
1907 // mid->is_busy() above or has already entered and waited on
1908 // it which makes it busy so no deflation. Restore owner to
1909 // NULL if it is still DEFLATER_MARKER.
1910 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER);
1911 return false;
1912 }
1913
1914 if (Atomic::cmpxchg(-max_jint, &mid->_ref_count, (jint)0) == 0) {
1915 // Make ref_count negative to force any contending threads or
1916 // ObjectMonitor* using threads to retry. This is the second
1917 // part of the async deflation dance.
1918
1919 if (mid->owner_is_DEFLATER_MARKER()) {
1920 // If owner is still DEFLATER_MARKER, then we have successfully
1921 // signaled any contending threads to retry. If it is not, then we
1922 // have lost the race to an entering thread and the ObjectMonitor
1923 // is now busy. This is the third and final part of the async
1924 // deflation dance.
1925 // Note: This owner check solves the ABA problem with ref_count
1926 // where another thread acquired the ObjectMonitor, finished
1927 // using it and restored the ref_count to zero.
1928
1929 // Sanity checks for the races:
1930 guarantee(mid->_contentions == 0, "must be 0: contentions=%d",
1931 mid->_contentions);
1932 guarantee(mid->_waiters == 0, "must be 0: waiters=%d", mid->_waiters);
1933 guarantee(mid->_cxq == NULL, "must be no contending threads: cxq="
1934 INTPTR_FORMAT, p2i(mid->_cxq));
1935 guarantee(mid->_EntryList == NULL,
1936 "must be no entering threads: EntryList=" INTPTR_FORMAT,
1937 p2i(mid->_EntryList));
1938
1939 const oop obj = (oop) mid->object();
1940 if (log_is_enabled(Trace, monitorinflation)) {
1941 ResourceMark rm;
1942 log_trace(monitorinflation)("deflate_monitor_using_JT: "
1943 "object=" INTPTR_FORMAT ", mark="
1944 INTPTR_FORMAT ", type='%s'",
1945 p2i(obj), obj->mark().value(),
1946 obj->klass()->external_name());
1947 }
1948
1949 // Install the old mark word if nobody else has already done it.
1950 mid->install_displaced_markword_in_object(obj);
1951 mid->clear_using_JT();
1952
1953 assert(mid->object() == NULL, "must be NULL: object=" INTPTR_FORMAT,
1954 p2i(mid->object()));
1955 assert(mid->is_free(), "must be free: allocation_state=%d",
1956 (int) mid->allocation_state());
1957
1958 // Move the deflated ObjectMonitor to the working free list
1959 // defined by free_head_p and free_tail_p.
1960 if (*free_head_p == NULL) {
1961 // First one on the list.
1962 *free_head_p = mid;
1963 }
1964 if (*free_tail_p != NULL) {
1965 // We append to the list so the caller can use mid->_next_om
1966 // to fix the linkages in its context.
1967 ObjectMonitor* prevtail = *free_tail_p;
1968 // Should have been cleaned up by the caller:
1969 assert(prevtail->_next_om == NULL, "must be NULL: _next_om="
1970 INTPTR_FORMAT, p2i(prevtail->_next_om));
1971 prevtail->_next_om = mid;
1972 }
1973 *free_tail_p = mid;
1974
1975 // At this point, mid->_next_om still refers to its current
1976 // value and another ObjectMonitor's _next_om field still
1977 // refers to this ObjectMonitor. Those linkages have to be
1978 // cleaned up by the caller who has the complete context.
1979
1980 // We leave owner == DEFLATER_MARKER and ref_count < 0
1981 // to force any racing threads to retry.
1982 return true; // Success, ObjectMonitor has been deflated.
1983 }
1984
1985 // The owner was changed from DEFLATER_MARKER so we lost the
1986 // race since the ObjectMonitor is now busy.
1987
1988 // Add back max_jint to restore the ref_count field to its
1989 // proper value (which may not be what we saw above):
1990 Atomic::add(max_jint, &mid->_ref_count);
1991
1992 assert(mid->ref_count() >= 0, "must not be negative: ref_count=%d",
1993 mid->ref_count());
1994 return false;
1995 }
1996
1997 // The ref_count was no longer 0 so we lost the race since the
1998 // ObjectMonitor is now busy or the ObjectMonitor* is now is use.
1999 // Restore owner to NULL if it is still DEFLATER_MARKER:
2000 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER);
2001 }
2002
2003 // The owner field is no longer NULL so we lost the race since the
2004 // ObjectMonitor is now busy.
2005 return false;
2006 }
2007
2008 // Walk a given monitor list, and deflate idle monitors
2009 // The given list could be a per-thread list or a global list
2010 // Caller acquires gListLock as needed.
2011 //
2012 // In the case of parallel processing of thread local monitor lists,
2013 // work is done by Threads::parallel_threads_do() which ensures that
2014 // each Java thread is processed by exactly one worker thread, and
2015 // thus avoid conflicts that would arise when worker threads would
2016 // process the same monitor lists concurrently.
2017 //
2018 // See also ParallelSPCleanupTask and
2019 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
2020 // Threads::parallel_java_threads_do() in thread.cpp.
2021 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** list_p,
2022 ObjectMonitor** free_head_p,
2023 ObjectMonitor** free_tail_p) {
2024 ObjectMonitor* mid;
2025 ObjectMonitor* next;
2026 ObjectMonitor* cur_mid_in_use = NULL;
2027 int deflated_count = 0;
2032 // Deflation succeeded and already updated free_head_p and
2033 // free_tail_p as needed. Finish the move to the local free list
2034 // by unlinking mid from the global or per-thread in-use list.
2035 if (mid == *list_p) {
2036 *list_p = mid->_next_om;
2037 } else if (cur_mid_in_use != NULL) {
2038 cur_mid_in_use->_next_om = mid->_next_om; // maintain the current thread in-use list
2039 }
2040 next = mid->_next_om;
2041 mid->_next_om = NULL; // This mid is current tail in the free_head_p list
2042 mid = next;
2043 deflated_count++;
2044 } else {
2045 cur_mid_in_use = mid;
2046 mid = mid->_next_om;
2047 }
2048 }
2049 return deflated_count;
2050 }
2051
2052 // Walk a given ObjectMonitor list and deflate idle ObjectMonitors using
2053 // a JavaThread. Returns the number of deflated ObjectMonitors. The given
2054 // list could be a per-thread in-use list or the global in-use list.
2055 // Caller acquires gListLock as appropriate. If a safepoint has started,
2056 // then we save state via saved_mid_in_use_p and return to the caller to
2057 // honor the safepoint.
2058 //
2059 int ObjectSynchronizer::deflate_monitor_list_using_JT(ObjectMonitor** list_p,
2060 ObjectMonitor** free_head_p,
2061 ObjectMonitor** free_tail_p,
2062 ObjectMonitor** saved_mid_in_use_p) {
2063 assert(AsyncDeflateIdleMonitors, "sanity check");
2064 assert(Thread::current()->is_Java_thread(), "precondition");
2065
2066 ObjectMonitor* mid;
2067 ObjectMonitor* next;
2068 ObjectMonitor* cur_mid_in_use = NULL;
2069 int deflated_count = 0;
2070
2071 if (*saved_mid_in_use_p == NULL) {
2072 // No saved state so start at the beginning.
2073 mid = *list_p;
2074 } else {
2075 // We're restarting after a safepoint so restore the necessary state
2076 // before we resume.
2077 cur_mid_in_use = *saved_mid_in_use_p;
2078 mid = cur_mid_in_use->_next_om;
2079 }
2080 while (mid != NULL) {
2081 // Only try to deflate if there is an associated Java object and if
2082 // mid is old (is not newly allocated and is not newly freed).
2083 if (mid->object() != NULL && mid->is_old() &&
2084 deflate_monitor_using_JT(mid, free_head_p, free_tail_p)) {
2085 // Deflation succeeded so update the in-use list.
2086 if (mid == *list_p) {
2087 *list_p = mid->_next_om;
2088 } else if (cur_mid_in_use != NULL) {
2089 // Maintain the current in-use list.
2090 cur_mid_in_use->_next_om = mid->_next_om;
2091 }
2092 next = mid->_next_om;
2093 mid->_next_om = NULL;
2094 // At this point mid is disconnected from the in-use list
2095 // and is the current tail in the free_head_p list.
2096 mid = next;
2097 deflated_count++;
2098 } else {
2099 // mid is considered in-use if it does not have an associated
2100 // Java object or mid is not old or deflation did not succeed.
2101 // A mid->is_new() node can be seen here when it is freshly
2102 // returned by om_alloc() (and skips the deflation code path).
2103 // A mid->is_old() node can be seen here when deflation failed.
2104 // A mid->is_free() node can be seen here when a fresh node from
2105 // om_alloc() is released by om_release() due to losing the race
2106 // in inflate().
2107
2108 cur_mid_in_use = mid;
2109 mid = mid->_next_om;
2110
2111 if (SafepointSynchronize::is_synchronizing() &&
2112 cur_mid_in_use != *list_p && cur_mid_in_use->is_old()) {
2113 // If a safepoint has started and cur_mid_in_use is not the list
2114 // head and is old, then it is safe to use as saved state. Return
2115 // to the caller so gListLock can be dropped as appropriate
2116 // before blocking.
2117 *saved_mid_in_use_p = cur_mid_in_use;
2118 return deflated_count;
2119 }
2120 }
2121 }
2122 // We finished the list without a safepoint starting so there's
2123 // no need to save state.
2124 *saved_mid_in_use_p = NULL;
2125 return deflated_count;
2126 }
2127
2128 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
2129 counters->n_in_use = 0; // currently associated with objects
2130 counters->n_in_circulation = 0; // extant
2131 counters->n_scavenged = 0; // reclaimed (global and per-thread)
2132 counters->per_thread_scavenged = 0; // per-thread scavenge total
2133 counters->per_thread_times = 0.0; // per-thread scavenge times
2134 }
2135
2136 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
2137 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
2138
2139 if (AsyncDeflateIdleMonitors) {
2140 // Nothing to do when global idle ObjectMonitors are deflated using
2141 // a JavaThread unless a special deflation has been requested.
2142 if (!is_special_deflation_requested()) {
2143 return;
2144 }
2145 }
2146
2147 bool deflated = false;
2148
2149 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors
2150 ObjectMonitor* free_tail_p = NULL;
2151 elapsedTimer timer;
2152
2153 if (log_is_enabled(Info, monitorinflation)) {
2154 timer.start();
2155 }
2156
2157 // Prevent om_flush from changing mids in Thread dtor's during deflation
2158 // And in case the vm thread is acquiring a lock during a safepoint
2159 // See e.g. 6320749
2160 Thread::muxAcquire(&gListLock, "deflate_idle_monitors");
2161
2162 // Note: the thread-local monitors lists get deflated in
2163 // a separate pass. See deflate_thread_local_monitors().
2164
2165 // For moribund threads, scan g_om_in_use_list
2166 int deflated_count = 0;
2179 // constant-time list splice - prepend scavenged segment to g_free_list
2180 free_tail_p->_next_om = g_free_list;
2181 g_free_list = free_head_p;
2182 }
2183 Thread::muxRelease(&gListLock);
2184 timer.stop();
2185
2186 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2187 LogStreamHandle(Info, monitorinflation) lsh_info;
2188 LogStream* ls = NULL;
2189 if (log_is_enabled(Debug, monitorinflation)) {
2190 ls = &lsh_debug;
2191 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2192 ls = &lsh_info;
2193 }
2194 if (ls != NULL) {
2195 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
2196 }
2197 }
2198
2199 // Deflate global idle ObjectMonitors using a JavaThread.
2200 //
2201 void ObjectSynchronizer::deflate_global_idle_monitors_using_JT() {
2202 assert(AsyncDeflateIdleMonitors, "sanity check");
2203 assert(Thread::current()->is_Java_thread(), "precondition");
2204 JavaThread* self = JavaThread::current();
2205
2206 deflate_common_idle_monitors_using_JT(true /* is_global */, self);
2207 }
2208
2209 // Deflate per-thread idle ObjectMonitors using a JavaThread.
2210 //
2211 void ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT() {
2212 assert(AsyncDeflateIdleMonitors, "sanity check");
2213 assert(Thread::current()->is_Java_thread(), "precondition");
2214 JavaThread* self = JavaThread::current();
2215
2216 self->om_request_deflation = false;
2217
2218 deflate_common_idle_monitors_using_JT(false /* !is_global */, self);
2219 }
2220
2221 // Deflate global or per-thread idle ObjectMonitors using a JavaThread.
2222 //
2223 void ObjectSynchronizer::deflate_common_idle_monitors_using_JT(bool is_global, JavaThread* self) {
2224 int deflated_count = 0;
2225 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged ObjectMonitors
2226 ObjectMonitor* free_tail_p = NULL;
2227 ObjectMonitor* saved_mid_in_use_p = NULL;
2228 elapsedTimer timer;
2229
2230 if (log_is_enabled(Info, monitorinflation)) {
2231 timer.start();
2232 }
2233
2234 if (is_global) {
2235 Thread::muxAcquire(&gListLock, "deflate_global_idle_monitors_using_JT(1)");
2236 OM_PERFDATA_OP(MonExtant, set_value(g_om_in_use_count));
2237 } else {
2238 OM_PERFDATA_OP(MonExtant, inc(self->om_in_use_count));
2239 }
2240
2241 do {
2242 int local_deflated_count;
2243 if (is_global) {
2244 local_deflated_count = deflate_monitor_list_using_JT((ObjectMonitor **)&g_om_in_use_list, &free_head_p, &free_tail_p, &saved_mid_in_use_p);
2245 g_om_in_use_count -= local_deflated_count;
2246 } else {
2247 local_deflated_count = deflate_monitor_list_using_JT(self->om_in_use_list_addr(), &free_head_p, &free_tail_p, &saved_mid_in_use_p);
2248 self->om_in_use_count -= local_deflated_count;
2249 }
2250 deflated_count += local_deflated_count;
2251
2252 if (free_head_p != NULL) {
2253 // Move the scavenged ObjectMonitors to the global free list.
2254 guarantee(free_tail_p != NULL && local_deflated_count > 0, "free_tail_p=" INTPTR_FORMAT ", local_deflated_count=%d", p2i(free_tail_p), local_deflated_count);
2255 assert(free_tail_p->_next_om == NULL, "invariant");
2256
2257 if (!is_global) {
2258 Thread::muxAcquire(&gListLock, "deflate_per_thread_idle_monitors_using_JT(2)");
2259 }
2260 // Constant-time list splice - prepend scavenged segment to g_free_list.
2261 free_tail_p->_next_om = g_free_list;
2262 g_free_list = free_head_p;
2263
2264 g_om_free_count += local_deflated_count;
2265 OM_PERFDATA_OP(Deflations, inc(local_deflated_count));
2266 if (!is_global) {
2267 Thread::muxRelease(&gListLock);
2268 }
2269 }
2270
2271 if (saved_mid_in_use_p != NULL) {
2272 // deflate_monitor_list_using_JT() detected a safepoint starting.
2273 if (is_global) {
2274 Thread::muxRelease(&gListLock);
2275 }
2276 timer.stop();
2277 {
2278 if (is_global) {
2279 log_debug(monitorinflation)("pausing deflation of global idle monitors for a safepoint.");
2280 } else {
2281 log_debug(monitorinflation)("jt=" INTPTR_FORMAT ": pausing deflation of per-thread idle monitors for a safepoint.", p2i(self));
2282 }
2283 assert(SafepointSynchronize::is_synchronizing(), "sanity check");
2284 ThreadBlockInVM blocker(self);
2285 }
2286 // Prepare for another loop after the safepoint.
2287 free_head_p = NULL;
2288 free_tail_p = NULL;
2289 if (log_is_enabled(Info, monitorinflation)) {
2290 timer.start();
2291 }
2292 if (is_global) {
2293 Thread::muxAcquire(&gListLock, "deflate_global_idle_monitors_using_JT(3)");
2294 }
2295 }
2296 } while (saved_mid_in_use_p != NULL);
2297 if (is_global) {
2298 Thread::muxRelease(&gListLock);
2299 }
2300 timer.stop();
2301
2302 LogStreamHandle(Debug, monitorinflation) lsh_debug;
2303 LogStreamHandle(Info, monitorinflation) lsh_info;
2304 LogStream* ls = NULL;
2305 if (log_is_enabled(Debug, monitorinflation)) {
2306 ls = &lsh_debug;
2307 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) {
2308 ls = &lsh_info;
2309 }
2310 if (ls != NULL) {
2311 if (is_global) {
2312 ls->print_cr("async-deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count);
2313 } else {
2314 ls->print_cr("jt=" INTPTR_FORMAT ": async-deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(self), timer.seconds(), deflated_count);
2315 }
2316 }
2317 }
2318
2319 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
2320 // Report the cumulative time for deflating each thread's idle
2321 // monitors. Note: if the work is split among more than one
2322 // worker thread, then the reported time will likely be more
2323 // than a beginning to end measurement of the phase.
2324 // Note: AsyncDeflateIdleMonitors only deflates per-thread idle
2325 // monitors at a safepoint when a special deflation has been requested.
2326 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->per_thread_times, counters->per_thread_scavenged);
2327
2328 bool needs_special_deflation = is_special_deflation_requested();
2329 if (!AsyncDeflateIdleMonitors || needs_special_deflation) {
2330 // AsyncDeflateIdleMonitors does not use these counters unless
2331 // there is a special deflation request.
2332
2333 g_om_free_count += counters->n_scavenged;
2334
2335 OM_PERFDATA_OP(Deflations, inc(counters->n_scavenged));
2336 OM_PERFDATA_OP(MonExtant, set_value(counters->n_in_circulation));
2337 }
2338
2339 if (log_is_enabled(Debug, monitorinflation)) {
2340 // exit_globals()'s call to audit_and_print_stats() is done
2341 // at the Info level.
2342 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */);
2343 } else if (log_is_enabled(Info, monitorinflation)) {
2344 Thread::muxAcquire(&gListLock, "finish_deflate_idle_monitors");
2345 log_info(monitorinflation)("g_om_population=%d, g_om_in_use_count=%d, "
2346 "g_om_free_count=%d", g_om_population,
2347 g_om_in_use_count, g_om_free_count);
2348 Thread::muxRelease(&gListLock);
2349 }
2350
2351 ForceMonitorScavenge = 0; // Reset
2352 GVars.stw_random = os::random();
2353 GVars.stw_cycle++;
2354 if (needs_special_deflation) {
2355 set_is_special_deflation_requested(false); // special deflation is done
2356 }
2357 }
2358
2359 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
2360 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
2361
2362 if (AsyncDeflateIdleMonitors) {
2363 if (!is_special_deflation_requested()) {
2364 // Mark the JavaThread for idle monitor deflation if a special
2365 // deflation has NOT been requested.
2366 if (thread->om_in_use_count > 0) {
2367 // This JavaThread is using monitors so mark it.
2368 thread->om_request_deflation = true;
2369 }
2370 return;
2371 }
2372 }
2373
2374 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors
2375 ObjectMonitor* free_tail_p = NULL;
2376 elapsedTimer timer;
2377
2378 if (log_is_enabled(Info, safepoint, cleanup) ||
2379 log_is_enabled(Info, monitorinflation)) {
2380 timer.start();
2381 }
2382
2383 int deflated_count = deflate_monitor_list(thread->om_in_use_list_addr(), &free_head_p, &free_tail_p);
2384
2385 Thread::muxAcquire(&gListLock, "deflate_thread_local_monitors");
2386
2387 // Adjust counters
2388 counters->n_in_circulation += thread->om_in_use_count;
2389 thread->om_in_use_count -= deflated_count;
2390 counters->n_scavenged += deflated_count;
2391 counters->n_in_use += thread->om_in_use_count;
2392 counters->per_thread_scavenged += deflated_count;
2393
2580 void ObjectSynchronizer::chk_free_entry(JavaThread* jt, ObjectMonitor* n,
2581 outputStream * out, int *error_cnt_p) {
2582 stringStream ss;
2583 if (n->is_busy()) {
2584 if (jt != NULL) {
2585 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2586 ": free per-thread monitor must not be busy: %s", p2i(jt),
2587 p2i(n), n->is_busy_to_string(&ss));
2588 } else {
2589 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2590 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss));
2591 }
2592 *error_cnt_p = *error_cnt_p + 1;
2593 }
2594 if (n->header().value() != 0) {
2595 if (jt != NULL) {
2596 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2597 ": free per-thread monitor must have NULL _header "
2598 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n),
2599 n->header().value());
2600 *error_cnt_p = *error_cnt_p + 1;
2601 } else if (!AsyncDeflateIdleMonitors) {
2602 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2603 "must have NULL _header field: _header=" INTPTR_FORMAT,
2604 p2i(n), n->header().value());
2605 *error_cnt_p = *error_cnt_p + 1;
2606 }
2607 }
2608 if (n->object() != NULL) {
2609 if (jt != NULL) {
2610 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT
2611 ": free per-thread monitor must have NULL _object "
2612 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n),
2613 p2i(n->object()));
2614 } else {
2615 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor "
2616 "must have NULL _object field: _object=" INTPTR_FORMAT,
2617 p2i(n), p2i(n->object()));
2618 }
2619 *error_cnt_p = *error_cnt_p + 1;
2620 }
2621 }
2622
2623 // Check the global free list and count; log the results of the checks.
2624 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out,
2625 int *error_cnt_p) {
2626 int chk_om_free_count = 0;
2627 for (ObjectMonitor* n = g_free_list; n != NULL; n = n->_next_om) {
2754 "equal to chk_om_in_use_count=%d", p2i(jt), jt->om_in_use_count,
2755 chk_om_in_use_count);
2756 *error_cnt_p = *error_cnt_p + 1;
2757 }
2758 }
2759
2760 // Log details about ObjectMonitors on the in-use lists. The 'BHL'
2761 // flags indicate why the entry is in-use, 'object' and 'object type'
2762 // indicate the associated object and its type.
2763 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out,
2764 bool on_exit) {
2765 if (!on_exit) {
2766 // Not at VM exit so grab the global list lock.
2767 Thread::muxAcquire(&gListLock, "log_in_use_monitor_details");
2768 }
2769
2770 stringStream ss;
2771 if (g_om_in_use_count > 0) {
2772 out->print_cr("In-use global monitor info:");
2773 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2774 out->print_cr("%18s %s %7s %18s %18s",
2775 "monitor", "BHL", "ref_cnt", "object", "object type");
2776 out->print_cr("================== === ======= ================== ==================");
2777 for (ObjectMonitor* n = g_om_in_use_list; n != NULL; n = n->_next_om) {
2778 const oop obj = (oop) n->object();
2779 const markWord mark = n->header();
2780 ResourceMark rm;
2781 out->print(INTPTR_FORMAT " %d%d%d %7d " INTPTR_FORMAT " %s",
2782 p2i(n), n->is_busy() != 0, mark.hash() != 0,
2783 n->owner() != NULL, (int)n->ref_count(), p2i(obj),
2784 obj->klass()->external_name());
2785 if (n->is_busy() != 0) {
2786 out->print(" (%s)", n->is_busy_to_string(&ss));
2787 ss.reset();
2788 }
2789 out->cr();
2790 }
2791 }
2792
2793 if (!on_exit) {
2794 Thread::muxRelease(&gListLock);
2795 }
2796
2797 out->print_cr("In-use per-thread monitor info:");
2798 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2799 out->print_cr("%18s %18s %s %7s %18s %18s",
2800 "jt", "monitor", "BHL", "ref_cnt", "object", "object type");
2801 out->print_cr("================== ================== === ======= ================== ==================");
2802 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2803 for (ObjectMonitor* n = jt->om_in_use_list; n != NULL; n = n->_next_om) {
2804 const oop obj = (oop) n->object();
2805 const markWord mark = n->header();
2806 ResourceMark rm;
2807 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d %7d "
2808 INTPTR_FORMAT " %s", p2i(jt), p2i(n), n->is_busy() != 0,
2809 mark.hash() != 0, n->owner() != NULL, (int)n->ref_count(),
2810 p2i(obj), obj->klass()->external_name());
2811 if (n->is_busy() != 0) {
2812 out->print(" (%s)", n->is_busy_to_string(&ss));
2813 ss.reset();
2814 }
2815 out->cr();
2816 }
2817 }
2818
2819 out->flush();
2820 }
2821
2822 // Log counts for the global and per-thread monitor lists and return
2823 // the population count.
2824 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) {
2825 int pop_count = 0;
2826 out->print_cr("%18s %10s %10s %10s",
2827 "Global Lists:", "InUse", "Free", "Total");
2828 out->print_cr("================== ========== ========== ==========");
2829 out->print_cr("%18s %10d %10d %10d", "",
2830 g_om_in_use_count, g_om_free_count, g_om_population);
|