Re: [PATCH 1/6] cxl/region: Add region creation ABI

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Jun 18, 2021 at 8:08 AM Ben Widawsky <ben.widawsky@intel.com> wrote:
[..]
> > It's a bit horrible, but maybe a more intuitive interface would be to make the
> > targetX visibility dynamic and hence make interleave an attribute of the region?
> >
> > sysfs_update_group() is rarely used, but I think the intent is to support
> > this sort of case.
> >
> > I've not fully thought through the effects of that though so might
> > well be missing something.
>
> I really would have liked something dynamic like you described. I searched for
> other drivers that this without much luck. I'm relatively unfamiliar with a lot
> of the device model machinery in Linux. Dan, are you familiar with this, is it
> feasible? If it sounds reasonable I can spend some more time figuring out how to
> do it.

I would prefer not to go the sysfs_update_group() route. That tends to
make it harder for tooling to reason about the arrival of attributes
relative to KOBJ_ADD events.

That said "create" may indeed be too generic an attribute to reason
about what it does. We could make all the setup parameters that do not
involve targets their own attribute group and then have a trigger
attribute that adds the next region device. Something like:

decoderX.Y/setup_region/uuid
decoderX.Y/setup_region/size
decoderX.Y/setup_region/interleave
decoderX.Y/setup_region/commit

...and then @commit atomically adds the region and clears the
parameters for the next setup?

Alternatively, renaming create to something like add_interleave sounds
ok to me too... trying to approach the border with configfs without
crossing over.