[RFC] drm: atomic mode set API

[Ville Syrjälä]

On Thu, Feb 16, 2012 at 12:12:49AM +0100, Daniel Vetter wrote:
> On Wed, Feb 15, 2012 at 05:59:45PM -0500, Adam Jackson wrote:
> > On 2/15/12 5:42 PM, Jesse Barnes wrote:
> > 
> > >+#define DRM_SET_CONFIG_TEST (1<<0) /* don't change the config, just test it for validity */

This would be OK for cases like checking whether the scene
can be composited with planes, of if you need to fall back
to full GPU composition.

For actaully trying to find out why the config is bad, this
is of course insufficient. For that we'd need something a lot
more fine grained for each logical part of the full state. And 
even that could be a bit vague, since it might not be any single
piece of the state that's wrong, it could be the combination of
some of the pieces.

> > >+
> > >+struct drm_mode_set_config {
> > >+	__u64 crtcs;
> > >+	__u64 crtc_fbs;
> > >+	__u64 crtc_xpos; /* array of x coords for crtcs */
> > >+	__u64 crtc_ypos; /* array of y coords for crtcs */
> > >+	__u32 count_crtcs;
> > >+
> > >+	__u64 plane_sets; /* array of set_plane structs */
> > >+
> > >+	__u32 count_planes;
> > >+
> > >+	__u64 connectors;
> > >+	__u64 connector_modes;
> > >+	__u32 count_connectors;
> > >+
> > >+	__u32 flags;
> > >+};
> > 
> > This appears to be missing some []s, but I think the intent is clear.
> > 
> > >  #define DRM_MODE_ENCODER_NONE	0
> > >  #define DRM_MODE_ENCODER_DAC	1
> > >  #define DRM_MODE_ENCODER_TMDS	2
> > >
> > >This allows you to bind a bunch of fbs to crtcs with independent
> > >positions, as well as set a bunch of planes to specific fbs and
> > >layouts.  Finally, it lets you change the connector config at the same
> > >time, with a flag to simply test a config instead of actually setting
> > >it.
> > >
> > >Any comments?  Do we also need to set gamma or other properties as part
> > >of this?  What about cursors?
> > 
> > I guess you might want to set gamma atomically, but I can't imagine
> > it being a factor in anyone's "can I do this" logic.
> > 
> > How do you pass in pixel format?  Do you just assume the existing fb
> > is already in the correct format?  That could work but it kind of
> > sucks for low-memory environments since you'd need to have enough
> > room to pre-create all the fbs.  You could still do the "tear
> > everything down first" approach to work around that, but then you'd
> > still have the possibility of having nothing lit up _and_ not being
> > able to set what was requested, and then needing to unwind in
> > userspace.
> > 
> > I'd sort of also want to see audio reflected in this (sigh), since
> > that's going to affect the bandwidth math.  DP 1.2 makes that even
> > worse.
> 
> Yeah, I think we should include any funky connector, crtc, plane
> properties

How about taking this idea to its ultimate conclusion: everything is a
property.

Then this ioctl would just be a collection of objects and a bunch of
changed properties for each.

It would require standardising the properties, or some of them at least
and having two property namespaces: one standard, one driver specific.

It would be easy to extend w/o ABI breaks, whereas in the current model
you either need to introduce a new "set everything v2" ioctl, or use
properties anyway.

It would also make it slightly easier to optimize use cases like page
flipping N planes/crtcs at the same time. Ideally you'd just do a few
sanity checks and program the new fb addresses into the hardware. With
the current structure you either need to redo all scaling calculations,
or check that none of the coordinates changed before you can take the
optimized path.

> (the latter don't exist yet, but I guess they will sooner or
> later)

That would be one way to implement color space conversion settings,
color adjustments etc., and I agree, they should be part of the atomic
mode set. The atomic mode set should include pretty much everything.

> because they all might affect how many and which hw resources we
> need (I'm thinking e.g. of plane setups for hw that reuses crtc engines as
> planes, but where you can use the crtc scanout engine for something else
> if it's completely occluded or set to just scan out the black borders with
> a parameter).

Another one of my favorite ideas would be to deprecate the crtc as a
scanout engine concept and just represent all scanout engines as planes.
As there will be a new API anyway, it might be a good time to make that
change.

-- 
Ville Syrjälä
Intel OTC