The Mutex rank checking is excluded during a safepoint likely for historical reasons when the Threads_lock or Safepoint_lock were held.  It's better to not have this condition while rank checking, but removing it requires reassigning some Mutex ranks:

    owns -> takes (so reduce rank)

    SystemDictionary_lock(leaf) -> DumpTimeTable_lock(leaf-1) -> Zip_lock(leaf-2)
                                                                                                     -> Uncommit_lock(leaf-2)
                                                                                                     -> ConcurrentHashTableResize_lock(leaf-2)
                                                                                                     -> VtableStubs_lock(leaf-2)
                                                                                                     -> MetaspaceAllocation_lock(leaf -2)      
                                                                                                                         ->Metaspace_lock(leaf-3)

    CGC_lock(leaf) -> RootRegionScan_lock(leaf-1)
                             -> NonJavaThreadsListLock(leaf-1) -> G1DetachedRefinementStats_lock(leaf-2)

    FreeList_lock(leaf) -> Uncommit_lock(leaf-2)                       -> G1Mapper_lock(leaf-3)
                                       -> HeapRegionParAlloc_lock(leaf-1)


If the Service_lock is only taken if  not at_safepoint for oops_do/nmethods_do functions so we don't have to reduce the ranks of the G1 threads because of G1 locks taken at a safepoint.  The code adding to the ServiceThread's jvmti deferred event queue are either at a safepoint for G1 or not for concurrent GCs.

     Service_lock(service) -> HeapRegionRemSet#n_lock(service-1)
                                            -> FreeList_lock(should be service -1 now)

    StackWaterMark_lock(stackwatermark) -> Service_lock(service) for ServiceThread::oops_do_no_frames.