loop: are parallel requests serialized by the single workqueue?

[Eric Wheeler <linux-block@xxxxxxxxxxxxxxxxxx>]

Hi Ming,

I was studying the loop.c DIO & AIO changes you made back in 2015 that 
increased loop performance and reduced the memory footprint 
(bc07c10a3603a5ab3ef01ba42b3d41f9ac63d1b6).  

I have a few questions if you are able to comment, here is a quick 
summary:

The direct IO path starts by queuing the work:

  .queue_rq       = loop_queue_rq:

	-> loop_queue_work(lo, cmd);
	-> INIT_WORK(&worker->work, loop_workfn);
		... queue_work(lo->workqueue, work);

Then from within the workqueue:

	-> loop_workfn()
	-> loop_process_work(worker, &worker->cmd_list, worker->lo);
	-> loop_handle_cmd(cmd);
	-> do_req_filebacked(lo, blk_mq_rq_from_pdu(cmd) );
	-> lo_rw_aio(lo, cmd, pos, READ) // (or WRITE)