Re: [PATCH RFC 0/3] blk-mq/nvme: use blk_mq_alloc_request() for NVMe's connect request

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

On Mon, Nov 18, 2019 at 04:05:56PM -0800, Sagi Grimberg wrote:
> 
> > Hi Sagi,
> > 
> > On Fri, Nov 15, 2019 at 02:38:44PM -0800, Sagi Grimberg wrote:
> > > 
> > > > Hi,
> > > 
> > > Hey Ming,
> > > 
> > > > Use blk_mq_alloc_request() for allocating NVMe loop, fc, rdma and tcp's
> > > > connect request, and selecting transport queue runtime for connect
> > > > request.
> > > > 
> > > > Then kill blk_mq_alloc_request_hctx().
> > > 
> > > Is it really so wrong to have an API to allocate a tag that belongs to
> > > a specific queue? Why must the tags allocation always correlate to the
> > > running cpu? Its true that NVMe is the only consumer of this at the
> > > moment, but does this mean that the interface should be removed because
> > > it has one (or rather four) consumer(s)?
> > 
> > Now blk-mq takes a static queue mapping between CPU and hw queues, given
> > CPU hotplug may happen any time, so the specified hw queue may become
> > inactive any time.
> > 
> > Queue mapping from CPU to hw queue is one core model of blk-mq which
> > relies a lot on the fact that hw queue active or not depends on
> > request's submission CPU. And we always can retrieve one active hw
> > queue if there is at least one online CPU.
> > 
> > IO request is always mapped to the proper hw queue via the submission
> > CPU and queue type.
> > 
> > So blk_mq_alloc_request_hctx() is really weird from the above blk-mq's
> > model.
> > 
> > Actually the 4 consumer is just one single type of usage for submitting
> > connect command, seems no one explain this special usage before. And the
> > driver can handle well enough without this interface just like this
> > patch, can't it?
> 
> Does removing the cpumask_and with cpu_online_mask fix your test?

It can be workaround this way, or return NULL if the hctx becomes
inactive.

But there is big problem to dispatch such request to inactive hctx, as
I explained.

> 
> this check is wrong to begin with because it can not be true right after
> the check.
> 
> This is a much simpler fix that does not create this churn local to
> every driver. Also, I don't like the assumptions about tag reservations
> that the drivers is taking locally (that the connect will have tag 0
> for example). All this makes this look like a hack.

The patch I posted can be applied to non-reserved tag too, and the connect
request can figured by rq->private_rq_data simply.

Also, we can provide blk_mq_rq_is_reserved() helper if you think 'rq->tag == 0'
is like a hack.

> 
> There is also the question of what happens when we want to make connects
> parallel, which is not the case at the moment...

There are several solutions for this:

1) schedule the connection on selected CPUs, so that all hw queues can be
covered

2) use single tag_set for connect command only.


Thanks, 
Ming