Re: [PATCH v4] drm/virtio: Add option to disable KMS support

[Gerd Hoffmann]

On Wed, Mar 01, 2023 at 03:37:24AM +0300, Dmitry Osipenko wrote:
> On 2/28/23 18:54, Rob Clark wrote:
> > From: Rob Clark <robdclark@xxxxxxxxxxxx>
> > 
> > Add a build option to disable modesetting support.  This is useful in
> > cases where the guest only needs to use the GPU in a headless mode, or
> > (such as in the CrOS usage) window surfaces are proxied to a host
> > compositor.
> > 
> > As the modesetting ioctls are a big surface area for potential security
> > bugs to be found (it's happened in the past, we should assume it will
> > again in the future), it makes sense to have a build option to disable
> > those ioctls in cases where they serve no legitimate purpose.
> > 
> > v2: Use more if (IS_ENABLED(...))
> > v3: Also permit the host to advertise no scanouts
> > v4: Spiff out commit msg
> > 
> > Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
> > Reviewed-by: Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx>
> > ---
> 
> Gerd, to give you some context on the v4.. we've chatted a bit more on
> the #dri-devel and concluded that config option is the most robust way
> of having KMS disabled from a security stand point. We would also want
> to have a per-driver option (and not global) because there are scenarios
> of using passthrough GPU + virtio-gpu in a guest, hence we would only
> want to toggle KMS for a particular driver.

IMHO both ways options to disable the KMS bits should work the same way.
With the current patch modeset_init() runs with num_scanouts == 0 but
doesn't with CONFIG_KMS=n.  There are also two different ways to tweak
driver_features.  Can we get rid of that please, for robustness reasons?

I'd suggest to have a is_kms_enabled() helper function (probably best as
inline so gcc can figure it is constant false for CONFIG_KMS=n and throw
away unreachable code).  Add "if (!is_kms_enabled()) return;" to
modeset_init() and modeset_fini() instead of stubbing them out.  Use the
drm_device->driver_features override in both cases.

Also the edid check can go away.  As already mentioned this is about a
device feature not a edid being present.

take care,
  Gerd