Btw, I really start to wonder if the request level is the right place
to do this on-stack plugging.  Wouldn't it be better to just plug
bios in the on-stack queue?  That way we could also stop doing the
special case merging when adding to the plug list, and leave all the
merging / I/O schedule logic in the __make_request path.  Probably
not .39 material, but worth a prototype?

Also what this dicussion brought up is that the block layer data
structures are highly confusing.  Using a small subset of the
request_queue also for make_request based driver just doesn't make
sense.  It seems like we should try to migrate the required state
to struct gendisk, and submit I/O through a block_device_ops.submit
method, leaving the request_queue as an internal abstraction for
the request based drivers.