When looking at it with John Cuthbertson, we noticed that ConcurrentG1Refine::clean_up_cache() might miss processing some cards. The code looks like this:

  int start_ind = _hot_cache_idx-1;
  for (int i = 0; i < _n_hot; i++) {
    int ind = start_ind - i;
    if (ind < 0) ind = ind + _hot_cache_size;
    jbyte* entry = _hot_cache[ind];
    if (entry != NULL) {
      g1rs->concurrentRefineOneCard(entry, worker_i);
    }
  }

_hot_cache_size is the max size of the cache, _n_hot is how many entries we've added to it and all the entries we need to process are at the bottom of the array [0, _n_hot). But, the above will miss processing [_hot_cache_idx, _n_hot) and instead process (incorrectly) the very top of the array.