On Fri, 2019-02-15 at 19:48 +0300, Andrey Kuzmin wrote:
> Ben,
> 
> you had indicated that you'd like to have a discussion of the multipath
> driver. Let me know what you think.

I haven't had a chance to through the actual implementation of the patch, but
multipathing in general is certainly of interest and warrants inclusion in SPDK.
I'd love to move ahead on this.

The big challenge is figuring out the best place to implement it. There are at
least 3 different options that I see.

1) Implement it as a bdev module. When the module gets a claim callback, it
checks to see if the new bdev is simply a different path to an existing bdev
(using some mechanism that isn't defined yet?) and then claims them both and
exposes a single bdev to the upper layer.

2) Implement it in the generic bdev code so that each individual bdev module can
call some function to add additional paths to a given bdev.

3) Implement it inside the nvme driver. There are a number of really cool
features for multipathing in NVMe (ANA and such) that don't translate well to a
generic bdev. The only way to take advantage of some of these might be to add
support directly into the nvme driver.

It could be that we end up doing a combination of 1 or 2 with number 3 as well.
I'd love some other opinions here from interested parties.

> 
> Thanks,
> Andrey
> 
> On Fri, Feb 1, 2019, 21:12 Andrey Kuzmin <akuzmin(a)jetstreamsoft.com wrote:
> 
> > I will appreciate an input on if the bdev-level multipathing is of
> > general interest and is worth inclusion. On positive determination I
> > can find time to rebase and resubmit.
> > 
> > Notice also that, in this case, decoupling open in write mode from
> > claims patch (https://review.gerrithub.io/c/spdk/spdk/+/422177) should
> > be revisited and included, as otherwise implementing multiple paths to
> > the same base bdev as  path bdevs is rather problematic (as this
> > obviously involves multiple writers).
> > 
> > Thanks,
> > Andrey
> > 
> 
> _______________________________________________
> SPDK mailing list
> SPDK(a)lists.01.org
> https://lists.01.org/mailman/listinfo/spdk