In osd_trans_stop()

     } else {
                osd_trans_stop_cb(oh, th->th_result);
                OBD_FREE_PTR(oh);
     }

It should call dcb commit callback as well, because the dcb callback is registered before transaction start(). Otherwise once error happens before transaction start, local commit callback will never be called, which will stop the llog cancellation, and cumulate a lot of update logs, cause long time recovery. This issue might be related with endless recovery found in soak-test.