Re: nvme: allow ANA support to be independent of native multipathing

[Mike Snitzer]

On Fri, Nov 16 2018 at  5:17am -0500,
Christoph Hellwig <hch@xxxxxx> wrote:

> On Fri, Nov 16, 2018 at 11:06:32AM +0100, Hannes Reinecke wrote:
> > Ok, so would you be happy with making ANA support configurable?
> 
> I've looked a bit over the whole situation, and what I think we need
> to do is:
> 
>  a) warn if we see a ANA capable device without multipath support
>     so people know it is not going to work properly.

I disagree with your cynicism but v2 of this patch now emits a warning
accordingly.

>  b) deprecate the multipath module option.  It was only intended as
>     a migration for any pre-existing PCIe multipath user if there
>     were any, not to support any new functionality.  So for 4.20
>     put in a patch that prints a clear warning when it is used,
>     including a link to the nvme list, and then for 4.25 or so
>     remove it entirely unless something unexpected come up.

You rejected the idea of allowing fine-grained control over whether
native NVMe multipathing is enabled or not on a per-namespace basis.
All we have is the coarse-grained nvme_core.multipath=N knob.  Now
you're forecasting removing even that.  Please don't do that.

> This whole drama of optional multipath use has wasted way too much
> of everyones time already.

It has wasted _way_ too much time.

But the drama is born out of you rejecting that we need to preserve
multipath-tools and dm-multipath's ability to work across any
transport.  You don't need to do that work: Hannes, myself and others
have always been willing and able -- if you'd let us.

IIRC it was at 2016's LSF in Boston where Ewan Milne and I had a
face-to-face conversation with you in the hallway track where you agreed
that ANA support would be activated if the capability was advertised by
the target.  The model we discussed is that it would be comparable to
how ALUA gets enabled during SCSI LUN discovery.

I hope you can see your way forward to be more accommodating now.
Especially given the proposed changes are backed by NVMe standards.

Please, PLEASE take v2 of this patch.. please? ;)

Thanks,
Mike