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

[Oded Gabbay]

On Mon, Nov 7, 2022 at 2:56 PM Jason Gunthorpe <jgg at nvidia.com> wrote:
>
> On Thu, Nov 03, 2022 at 10:39:36PM +0200, Oded Gabbay wrote:
> > On Thu, Nov 3, 2022 at 3:31 PM Oded Gabbay <ogabbay at kernel.org> wrote:
> > >
> > > On Thu, Nov 3, 2022 at 2:31 AM Greg Kroah-Hartman
> > > <gregkh at linuxfoundation.org> wrote:
> > > >
> > > > On Wed, Nov 02, 2022 at 10:34:03PM +0200, Oded Gabbay wrote:
> > > > > --- /dev/null
> > > > > +++ b/drivers/accel/Kconfig
> > > > > @@ -0,0 +1,24 @@
> > > > > +# SPDX-License-Identifier: GPL-2.0-only
> > > > > +#
> > > > > +# Compute Acceleration device configuration
> > > > > +#
> > > > > +# This framework provides support for compute acceleration devices, such
> > > > > +# as, but not limited to, Machine-Learning and Deep-Learning acceleration
> > > > > +# devices
> > > > > +#
> > > > > +menuconfig ACCEL
> > > > > +     tristate "Compute Acceleration Framework"
> > > > > +     depends on DRM
> > > > > +     help
> > > > > +       Framework for device drivers of compute acceleration devices, such
> > > > > +       as, but not limited to, Machine-Learning and Deep-Learning
> > > > > +       acceleration devices.
> > > > > +       If you say Y here, you need to select the module that's right for
> > > > > +       your acceleration device from the list below.
> > > > > +       This framework is integrated with the DRM subsystem as compute
> > > > > +       accelerators and GPUs share a lot in common and can use almost the
> > > > > +       same infrastructure code.
> > > > > +       Having said that, acceleration devices will have a different
> > > > > +       major number than GPUs, and will be exposed to user-space using
> > > > > +       different device files, called accel/accel* (in /dev, sysfs
> > > > > +       and debugfs)
> > > >
> > > > Module name if "M" is chosen?
> > > Will add
> > So, unfortunately, the path of doing accel as a kernel module won't
> > work cleanly (Thanks stanislaw for pointing this out to me).
> > The reason is the circular dependency between drm and accel. drm calls
> > accel exported symbols during init and when devices are registering
> > (all the minor handling), and accel calls drm exported symbols because
> > I don't want to duplicate the entire drm core code.
>
> I really don't think this is the right way to integrate with
> DRM. Accel should be a layer over top of DRM, not have these wakky
> co-dependencies.
>
> The fact you are running into stuff like this already smells really
> bad.
>
> Jason
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.

Please look at v3 of the patch-set. There I abandoned the idea of
having accel as a separate module and instead it is part of drm.ko, as
it should be because it is just a new drm minor.

The only alternative imo to that is to abandon the idea of reusing
drm, and just make an independant accel core code.

Oded