[Intel-gfx] [PATCH 1/3] drm/i915: Convert intel_tv connector properties to atomic.

[Ville Syrjälä]

On Mon, Mar 13, 2017 at 10:22:51AM +0100, Maarten Lankhorst wrote:
> Op 09-03-17 om 18:36 schreef Ville Syrjälä:
> > On Thu, Mar 09, 2017 at 02:06:15PM +0100, Maarten Lankhorst wrote:
> >> As a proof of concept, first try to convert intel_tv, which is a rarely
> >> used connector. It has 5 properties, tv format and 4 margins.
> > Since it's so rare, if you want someone to actually test the code
> > it'll probably make sense to pick another connector ;)
> Yeah but the properties are among the most annoying, with the self modifying code and using state in mode_detect().
> >> I'm less certain about the state behavior itself, should we pass a size
> >> parameter to intel_connector_alloc instead, so duplicate_state
> >> can be done globally if it can be blindly copied?
> >>
> >> Can we also have a atomic_check function for connectors, so the
> >> crtc_state->connectors_changed can be set there? It would be cleaner
> >> and more atomic-like.
> > Hmm. I think it migth be really useful only if we have some
> > interactions between multiple properties that really need to be
> > checked. We might have those already I suppose but we don't seem
> > to check any of it currently. So as a first step I guess we can
> > just keep ignoring any such issues.
> Well it might be, for example not all properties may be set yet so you can only do a sane check in a separate step.
> >> To match the legacy behavior, format can be changed by probing just like
> >> in legacy mode.
> > Self modifying state irks me, but it's what we've been doing so I guess
> > we should keep it.
> Yeah, I hate it too.
> >> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> >> ---
> >>  drivers/gpu/drm/i915/intel_tv.c | 238 +++++++++++++++++++++++-----------------
> >>  1 file changed, 136 insertions(+), 102 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/i915/intel_tv.c b/drivers/gpu/drm/i915/intel_tv.c
> >> index 6ed1a3ce47b7..0fb1d8621fe8 100644
> >> --- a/drivers/gpu/drm/i915/intel_tv.c
> >> +++ b/drivers/gpu/drm/i915/intel_tv.c
> >> @@ -48,8 +48,6 @@ struct intel_tv {
> >>  	struct intel_encoder base;
> >>  
> >>  	int type;
> >> -	const char *tv_format;
> >> -	int margin[4];
> >>  	u32 save_TV_H_CTL_1;
> >>  	u32 save_TV_H_CTL_2;
> >>  	u32 save_TV_H_CTL_3;
> >> @@ -85,6 +83,16 @@ struct intel_tv {
> >>  	u32 save_TV_CTL;
> >>  };
> >>  
> >> +struct intel_tv_connector_state {
> >> +	struct drm_connector_state base;
> >> +
> >> +	int format;
> >> +	int margin[4];
> >> +};
> >> +
> >> +#define to_intel_tv_connector_state(state) \
> >> +	container_of((state), struct intel_tv_connector_state, base)
> >> +
> >>  struct video_levels {
> >>  	u16 blank, black;
> >>  	u8 burst;
> >> @@ -873,32 +881,18 @@ intel_disable_tv(struct intel_encoder *encoder,
> >>  	I915_WRITE(TV_CTL, I915_READ(TV_CTL) & ~TV_ENC_ENABLE);
> >>  }
> >>  
> >> -static const struct tv_mode *
> >> -intel_tv_mode_lookup(const char *tv_format)
> >> +static const struct tv_mode *intel_tv_mode_find(struct drm_connector_state *conn_state)
> >>  {
> >> -	int i;
> >> +	int format = to_intel_tv_connector_state(conn_state)->format;
> >>  
> >> -	for (i = 0; i < ARRAY_SIZE(tv_modes); i++) {
> >> -		const struct tv_mode *tv_mode = &tv_modes[i];
> >> -
> >> -		if (!strcmp(tv_format, tv_mode->name))
> >> -			return tv_mode;
> >> -	}
> >> -	return NULL;
> >> -}
> >> -
> >> -static const struct tv_mode *
> >> -intel_tv_mode_find(struct intel_tv *intel_tv)
> >> -{
> >> -	return intel_tv_mode_lookup(intel_tv->tv_format);
> >> +	return &tv_modes[format];
> >>  }
> >>  
> >>  static enum drm_mode_status
> >>  intel_tv_mode_valid(struct drm_connector *connector,
> >>  		    struct drm_display_mode *mode)
> >>  {
> >> -	struct intel_tv *intel_tv = intel_attached_tv(connector);
> >> -	const struct tv_mode *tv_mode = intel_tv_mode_find(intel_tv);
> >> +	const struct tv_mode *tv_mode = intel_tv_mode_find(connector->state);
> > It feels a bit fishy to use the state here. Generally that's a no-no.
> > But in this case I wonder if it's the right choice after all. 
> >
> > Not sure if some kind of "automatic" enum value might also work. It
> > would at least avoid the self modifying property problem. Although I
> > wonder if the user would still like to know what was actually used
> > if they chose they automatic mode, so we might need a self modifying
> > RO property for the current mode anyway.
> >
> > But that still leaves the problem of how the user would know which modes
> > they should be able to use if .get_modes()/.mode_valid() doesn't respect
> > the users choice of the tv format. Hmm, tricky. Might be the self
> > modifying property is the only good choice.
> >
> > But if we would use the state here, what's the story with locking going
> > to be? connection_mutex is what protects this stuff, but we're not
> > holding that during mode enumeration.
> Yeah locking is tricky, honestly I have no idea what would be the right thing to do here..
> 
> I don't really see a good solution, or at least one that would work correctly with atomic properties.

Maybe we need to keep the format information in both intel_tv and the
state. We'd use the intel_tv->format during detect, and during
duplicate_state we'd do 'state->format = intel_tv->format' and during
commit we'd do 'intel_tv->format = state->format' ?

Still self modifying, and somewhat racy still, but at least we
shouldn't explode on account of the connector->state dereference.

-- 
Ville Syrjälä
Intel OTC