Received-SPF: pass (google.com: domain of linux-kernel+bounces-174556-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Precedence: bulk
MIME-Version: 1.0
References: <20240424063753.3740664-1-tomeu@tomeuvizoso.net>
 <8c55dba5-6308-685e-13da-e728197d8101@quicinc.com> <CAAObsKD4-k7Ya4Mi=vEPaC9DucbnVGDO5SaEUt-_o2_Bg+_FgA@mail.gmail.com>
In-Reply-To: <CAAObsKD4-k7Ya4Mi=vEPaC9DucbnVGDO5SaEUt-_o2_Bg+_FgA@mail.gmail.com>
From: Tomeu Vizoso <tomeu@tomeuvizoso.net>
Date: Thu, 9 May 2024 15:53:01 +0200
Message-ID: <CAAObsKCm49y-nUph=m9c+-eG37SaGKG93-1etwOQab4f5MXxOg@mail.gmail.com>
Subject: Re: [PATCH] drm/etnaviv: Create an accel device node if compute-only
To: Jeffrey Hugo <quic_jhugo@quicinc.com>
Cc: linux-kernel@vger.kernel.org, Oded Gabbay <ogabbay@kernel.org>, 
	Lucas Stach <l.stach@pengutronix.de>, Russell King <linux+etnaviv@armlinux.org.uk>, 
	Christian Gmeiner <christian.gmeiner@gmail.com>, David Airlie <airlied@gmail.com>, 
	Daniel Vetter <daniel@ffwll.ch>, etnaviv@lists.freedesktop.org, 
	dri-devel@lists.freedesktop.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

Oded, Dave,

Do you have an opinion on this?

Thanks,

Tomeu

On Fri, Apr 26, 2024 at 8:10=E2=80=AFAM Tomeu Vizoso <tomeu@tomeuvizoso.net=
> wrote:
>
> On Thu, Apr 25, 2024 at 8:59=E2=80=AFPM Jeffrey Hugo <quic_jhugo@quicinc.=
com> wrote:
> >
> > On 4/24/2024 12:37 AM, Tomeu Vizoso wrote:
> > > If we expose a render node for NPUs without rendering capabilities, t=
he
> > > userspace stack will offer it to compositors and applications for
> > > rendering, which of course won't work.
> > >
> > > Userspace is probably right in not questioning whether a render node
> > > might not be capable of supporting rendering, so change it in the ker=
nel
> > > instead by exposing a /dev/accel node.
> > >
> > > Before we bring the device up we don't know whether it is capable of
> > > rendering or not (depends on the features of its blocks), so first tr=
y
> > > to probe a rendering node, and if we find out that there is no render=
ing
> > > hardware, abort and retry with an accel node.
> > >
> > > Signed-off-by: Tomeu Vizoso <tomeu@tomeuvizoso.net>
> > > Cc: Oded Gabbay <ogabbay@kernel.org>
> >
> > I hope Oded chimes in as Accel maintainer.  I think Airlie/Vetter had
> > also previously mentioned they'd have opinions on what is Accel vs DRM.
> >
> > This gets a nack from me in its current state.  This is not a strong
> > nack, and I don't want to discourage you.  I think there is a path forw=
ard.
> >
> > The Accel subsystem documentation says that accel drivers will reside i=
n
> > drivers/accel/ but this does not.
>
> Indeed, there is that code organization aspect.
>
> > Also, the commit text for "accel: add dedicated minor for accelerator
> > devices" mentions -
> >
> > "for drivers that
> > declare they handle compute accelerator, using a new driver feature
> > flag called DRIVER_COMPUTE_ACCEL. It is important to note that this
> > driver feature is mutually exclusive with DRIVER_RENDER. Devices that
> > want to expose both graphics and compute device char files should be
> > handled by two drivers that are connected using the auxiliary bus
> > framework."
> >
> > I don't see any of that happening here (two drivers connected by aux
> > bus, one in drivers/accel).
>
> Well, the text refers to devices, not drivers. The case we are talking
> about is a driver that wants to sometimes expose an accel node, and
> sometimes a render node, depending on the hardware it is dealing with.
> So there would either be a device exposing a single render node, or a
> device exposing a single accel node.
>
> Though by using the auxiliary bus we could in theory solve the code
> organization problem mentioned above, I'm not quite seeing how to do
> this in a clean way. The driver in /drivers/gpu/drm would have to be a
> DRM driver that doesn't register a DRM device, but registers a device
> in the auxiliary bus for the driver in /drivers/accel to bind to? Or
> are you seeing some possibility that would fit better in the current
> DRM framework?
>
> > I think this is the first case we've had of a combo DRM/Accel usecase,
> > and so there isn't an existing example to refer you to on how to
> > structure things.  I think you are going to be the first example where
> > we figure all of this out.
>
> Yep, I will be grateful for any ideas on how to structure this.
>
> > On a more implementation note, ioctls for Accel devices should not be
> > marked DRM_RENDER_ALLOW.  Seems like your attempt to reuse as much of
> > the code as possible trips over this.
>
> Indeed, thanks.
>
> Cheers,
>
> Tomeu
>
> > -Jeff