Re: [RFC PATCH v2 1/3] drivers/accel: define kconfig and register a new major

[Jason Gunthorpe <jgg@nvidia.com>]

On Mon, Nov 07, 2022 at 04:02:01PM +0200, Oded Gabbay wrote:
> On Mon, Nov 7, 2022 at 3:10 PM Jason Gunthorpe <jgg@nvidia.com> wrote:
> >
> > On Mon, Nov 07, 2022 at 03:01:08PM +0200, Oded Gabbay wrote:
> > > I don't agree with your statement that it should be "a layer over top of DRM".
> > > Anything on top of DRM is a device driver.
> > > Accel is not a device driver, it is a new type of drm minor / drm driver.
> >
> > Yeah, I still think this is not the right way, you are getting almost
> > nothing from DRM and making everything more complicated in the
> > process.
> >
> > > The only alternative imo to that is to abandon the idea of reusing
> > > drm, and just make an independant accel core code.
> >
> > Not quite really, layer it properly and librarize parts of DRM into
> > things accel can re-use so they are not intimately tied to the DRM
> > struct device notion.
> >
> > IMHO this is much better, because accel has very little need of DRM to
> > manage a struct device/cdev in the first place.
> >
> > Jason
> I'm not following. How can an accel device be a new type of drm_minor,
> if it doesn't have access to all its functions and members ?

"drm_minor" is not necessary anymore. Strictly managing minor numbers
lost its value years ago when /dev/ was reorganized. Just use
dynamic minors fully.

> How will accel device leverage, for example, the GEM code without
> being a drm_minor ?

Split GEM into a library so it doesn't require that.

> Librarizing parts of DRM sounds nice in theory but the reality is that
> everything there is interconnected, all the structures are
> interdependent.

Yes, the kernel is full of stuff that needs improving. Let's not take
shortcuts.

> I would have to re-write the entire DRM library to make such a thing
> work. I don't think this was the intention.

Not necessarily you, whoever someday needs GEM would have to do some
work.

> The current design makes the accel device an integral part of drm,
> with very minimal code duplication and without re-writing DRM.

And it smells bad, you can't even make it into a proper module. Who
knows what other problems will come.

Jason