Adding set_blob ioctl to DRM

[Bob Paauwe]

On Tue, 4 Mar 2014 09:35:57 +0100
Daniel Vetter <daniel at ffwll.ch> wrote:

> On Thu, Feb 20, 2014 at 11:24:40AM +1000, Dave Airlie wrote:
> > > I am working on enabling a Color Enhancement block for primary display
> > > for Exynos SoC. I need to
> > > set a bunch of parameters like Color Conversion matrix, Contrast
> > > Improvement parameters etc ~ 30 parameters from User Space.
> > >
> > > I am planning to use KDS blob property to receive these parameters.
> > > Currently drivers are not allowed to create a blob property inside
> > > drm. Neither, user space can set the blob. There is no ioctl provided
> > > for same.
> > 
> > I don't really like the idea of sticking unstructured data into an
> > ioctl, for the driver to interpret,
> > 
> > it opens the door to all kinds of ugly driver hacks, so I think we
> > should have writable blobs,
> > but with well defined structures inside them, not per-driver ones.
> > 
> > Per-driver structures will lead to binary userspace drivers that start
> > sticking a pll timings blob on the end and require it to set a mode,
> > because I know driver developers will abuse any interface in the worst
> > way possible.
> > 
> > Currently the only blob we really have is EDID and its well defined,
> > so if we are going to add writable blobs, they need to be well defined
> > and as little as possible driver specific, just to avoid driver
> > writings doing what driver writers do.
> 
> tbh I don't see a use for structured blobs at all and would much more
> prefer a pile of properties. With the atomic modeset ioctl proposals
> floating around we could then pass in an entire sets of properties, which
> is essentially what the blob prop would be doing.
> 
> The upshot of going all-in with explicitly named properties is that we can
> shovel kms configuration descriptions into different formats. I'm thinking
> of shovelling an initial config into DT or something similar as an
> example.  For i915 we have some experimental patches which load a DT blob
> as a firmware file and use the property values to set things up.

I have most of this working with i915 now.  Using a DT configuration I
can set module parameters, select which crtc's to enable, which
connectors to enable, set connector properties at init time.  I'm
working on some plane property code now.

It's close to being ready for initial review/feedback but there's one
problem, you can't currently build an x86_64 (32 bit is fine) kernel
with DT enabled because of an unresolved symbol in the apic code. 

> 
> So imo a blob property needs some _really_ good justification. My default
> stance is to oppose it ;-)
> 
> Cheers, Daniel