[Intel-gfx] [PATCH 15/15] drm/i915: Fix up PSR frontbuffer tracking

[Daniel Vetter]

On Tue, Jun 17, 2014 at 08:07:28AM +0100, Chris Wilson wrote:
> On Mon, Jun 16, 2014 at 07:51:35PM +0200, Daniel Vetter wrote:
> > +void intel_edp_psr_invalidate(struct drm_device *dev,
> > +			      unsigned frontbuffer_bits)
> > +{
> > +	struct drm_i915_private *dev_priv = dev->dev_private;
> > +	struct drm_crtc *crtc;
> > +	enum pipe pipe;
> > +
> > +	if (!HAS_PSR(dev))
> > +		return;
> 
> if (!psr.enabled)
>    return;

Already fixed.
> 
> > +
> > +
> > +	mutex_lock(&dev_priv->psr.lock);
> > +	crtc = dp_to_dig_port(dev_priv->psr.enabled)->base.base.crtc;
> > +	pipe = to_intel_crtc(crtc)->pipe;
> > +
> > +	intel_edp_psr_do_exit(dev);
> > +
> > +	frontbuffer_bits &= INTEL_FRONTBUFFER_ALL_MASK(pipe);
> > +
> > +	dev_priv->psr.busy_frontbuffer_bits |= frontbuffer_bits;
> > +	mutex_unlock(&dev_priv->psr.lock);
> > +}
> > +
> > +void intel_edp_psr_flush(struct drm_device *dev,
> > +			 unsigned frontbuffer_bits)
> > +{
> > +	struct drm_i915_private *dev_priv = dev->dev_private;
> > +	struct drm_crtc *crtc;
> > +	enum pipe pipe;
> > +
> > +	if (!HAS_PSR(dev))
> > +		return;
> 
> if (!psr.enabled)
>    return;
> 
> > +
> > +
> > +	mutex_lock(&dev_priv->psr.lock);
> > +	crtc = dp_to_dig_port(dev_priv->psr.enabled)->base.base.crtc;
> > +	pipe = to_intel_crtc(crtc)->pipe;
> > +	dev_priv->psr.busy_frontbuffer_bits &= ~frontbuffer_bits;
> > +
> > +	if (IS_HASWELL(dev) &&
> > +	    (frontbuffer_bits & INTEL_FRONTBUFFER_SPRITE(pipe)))
> > +		intel_edp_psr_do_exit(dev);
> 
> This deserves comment explaining/referencing the discrepancy.

Agreed.
> 
> > +	if (!dev_priv->psr.busy_frontbuffer_bits)
> > +		schedule_delayed_work(&dev_priv->psr.work,
> > +				      msecs_to_jiffies(100));
> 
> 100ms made sense when queueing the task at the start of the write cycle,
> but on flush, it only wants to be ~2frames. As we know we have an active
> pipe here we could queue it for exactly N frames...

Well I don't have real hw to test so figured I'll leave the actual hw
touching code unchanged. What we really want is a function which uploads 1
final frame (starting from the next vblank) and then goes into psr
immediately. But getting there is a 2nd step.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch