RFC on upstreaming of a Mediatek DRM modesetting driver

[Daniel Vetter]

On Tue, Dec 02, 2014 at 09:21:51AM +0000, Frank Binns wrote:
> On 01/12/14 15:28, Daniel Vetter wrote:
> > On Mon, Dec 01, 2014 at 10:01:37AM +0000, Frank Binns wrote:
> >> Hi,
> >>
> >> We are currently in negotiations with one of our customers (Mediatek) on
> >> a strategy that will allow them to push a DRM modesetting driver into
> >> the upstream kernel. We are writing to get people's opinions and
> >> feedback on our proposed approach.
> >>
> >> Currently, our driver is structured in such a way that the display
> >> driver is more tightly integrated with the GPU driver than we would
> >> like. Although our kernel driver has been shipped with a GPL license for
> >> a long time, it is not in a form that would be considered acceptable
> >> upstream. Unfortunately, it is going to be a long process to get this
> >> part of the driver into a reasonable state. However, in the meantime, we
> >> don't want to prevent customer portions of the driver from being
> >> upstreamed. With the work done on recent kernels, and with a willing
> >> partner in Mediatek, we now think that we can restructure our driver in
> >> such a way as to allow this to happen.
> >>
> >> We see two basic approaches to achieving this:
> >> 1) Two independent DRM drivers, i.e. modesetting and render node drivers
> >> 2) A single componentised DRM driver
> >>
> >> Our (IMG's) preferred approach is to have a single componentised DRM
> >> driver. This is due to the following reasons:
> >>
> >> - Existing user space is not fully prepared to handle render nodes.
> >>
> >> - There is concern that any IMG DRM render node driver will need
> >> knowledge about multiple SoCs, each one being from a different vendor.
> >> Would this be deemed acceptable?
> >>
> >> - There is a trend, at least for DRM SoC drivers, towards using the
> >> component interface. Although there appears to be very few (one?)
> >> examples of GPU component drivers.
> >>
> >> To give some high level details on how we expect the componentised DRM
> >> driver model to work, each vendor (in this case Mediatek) will write
> >> their own DRM driver (supporting modesetting, dumb buffers, GEM, prime,
> >> etc) and IMG will provide an almost entirely independent component
> >> driver that adds in GPU support. Until our GPU driver is in a suitable
> >> state this will most likely necessitate a small kernel patch to wire up
> >> support, e.g. GPU specific ioctls.
> >>
> >> Cross-device and cross-process memory allocations will be made using the
> >> DRM driver. In order for this memory to be shared with the GPU component
> >> driver it will be necessary, at least for the time being, to export it
> >> via prime and import it via a GPU ioctl. Synchronisation between the
> >> display and GPU will be performed via reservation objects.
> >>
> >> Does this sound like a sane approach? Questions and/or feedback is very
> >> welcome.
> > Rule of thumb is that if it's an externally licensed IP block it should be
> > a separate driver. Which is the case here. The idea is that the mostly
> > generic IMG driver could be reused on other platforms that ship the same
> > IP-block, while linking up with the respective display controller driver.
> > The end result is 2 drm drivers:
> > - Display block drm driver which expose KMS objects for modesetting, but
> >   only very basic gem (just enough to allocate dumb framebuffers and
> >   import/export dma-bufs).
> > - Full-blown gem driver for the img render IP block.
> >
> > For an example look at the tegra/nouveau combo which can run on TK1.
> >
> > Plugggin in an IMG driver into each display block like it's currently done
> > with all the armsoc stuff on android is imo completely no-go.
> >
> > Note that the component interface is completely irrelevant wrt the
> > interface you expose to userspace. It's just an driver-internal helper
> > library useful in certain situation. Not even the drm core really cares
> > whether you use component helpers or not.
> >
> > Thanks, Daniel
> 
> OK, so it seems the consensus is that IMG should provide a separate
> render-node only DRM driver.
> 
> Having not worked directly on the core DRM code I'm not completely
> familiar with it but it seems to me that the DRIVER_MODESET flag has a
> dual meaning. Firstly it means that the driver supports KMS and secondly
> it means that a lot of the legacy stuff isn't supported. It also changes
> the way in which driver initialisation is performed. Would it make sense
> for the DRIVER_RENDER flag to have a similar effect? In other words,
> should it turn off legacy stuff and use the newer method of driver
> initialisation?

DRIVER_MODESET means you have a modern driver which binds to the device.
We should probably RENAME it to DRIVER_LEGACY and invert it's sense, but
no one has stepped up to the taks.

DRIVER_RENDER just means you support render nodes. And new drm driver
which exposes driver-specific ioctls should set this flag. Again we should
probably rename this to DRIVER_IOCTLS_BROKEN_CANT_DO_RENDER and invert the
sense ;-) Ofc with a better neame.

So for the IMG render-only driver you'd need to set both. Both flags just
hide legacy behaviour if not set, but that's the way you need to do it
without changing every driver.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch