Re: [PATCH] ublk: eliminate unnecessary io_cmds queue

[Ming Lei <ming.lei@redhat.com>]

On Tue, Oct 15, 2024 at 02:51:14PM -0600, Uday Shankar wrote:
> On Thu, Oct 10, 2024 at 11:45:03AM +0800, Ming Lei wrote:
> > >  static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
> > >  {
> > > -	struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
> > > -
> > > -	if (llist_add(&data->node, &ubq->io_cmds)) {
> > > -		struct ublk_io *io = &ubq->ios[rq->tag];
> > > +	struct ublk_io *io = &ubq->ios[rq->tag];
> > >  
> > > -		io_uring_cmd_complete_in_task(io->cmd, ublk_rq_task_work_cb);
> > > -	}
> > > +	ublk_get_uring_cmd_pdu(io->cmd)->req = rq;
> > > +	io_uring_cmd_complete_in_task(io->cmd, __ublk_rq_task_work);
> > >  }
> > 
> > I'd suggest to comment that io_uring_cmd_complete_in_task() needs to
> > maintain io command order.
> 
> Sorry, can you explain why this is important? Generally speaking
> out-of-order completion of I/Os is considered okay, so what's the issue
> if the dispatch to the ublk server here is not in order?

It is just okay, but proper implementation requires to keep IO order.

Please see:

1) 7d4a93176e01 ("ublk_drv: don't forward io commands in reserve order")

2) [Report] requests are submitted to hardware in reverse order from nvme/virtio-blk queue_rqs()

https://lore.kernel.org/linux-block/ZbD7ups50ryrlJ%2FG@fedora/


I am also working on ublk-bpf which needs this extra list for submitting
IO in batch, please hold on this patch now.

I plan to send out bpf patches in this cycle or next, and we can restart
the cleanup if the bpf thing turns out not doable.


Thanks, 
Ming