The queueing logic in scsi_lib looks really odd right now.

- Defers calling elv_next_request() until we actually think we can queue
  something. Always want to do that, request will have been marked
  REQ_STARTED after this, so block layer cannot touch it or merge to it.

- Kill the queue empty check. If elv_next_request() returned a req,
  there's one to queue.