[Intel-gfx] [PATCH] drm/i915: Consider plane rotation when calculating stride in skl_do_mmio_flip

[Ville Syrjälä]

On Tue, Oct 20, 2015 at 01:05:22PM +0100, Tvrtko Ursulin wrote:
> 
> On 20/10/15 12:53, Ville Syrjälä wrote:
> > On Tue, Oct 20, 2015 at 12:44:05PM +0100, Tvrtko Ursulin wrote:
> >>
> >> On 20/10/15 12:07, Ville Syrjälä wrote:
> >>> On Tue, Oct 20, 2015 at 10:06:58AM +0100, Tvrtko Ursulin wrote:
> >>>>
> >>>> On 20/10/15 08:42, Daniel Vetter wrote:
> >>>>> On Mon, Oct 19, 2015 at 09:20:36PM +0300, Ville Syrjälä wrote:
> >>>>>> On Wed, Oct 07, 2015 at 11:01:23AM +0100, Tvrtko Ursulin wrote:
> >>>>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >>>>>>>
> >>>>>>> Previously rotation was ignored and wrong stride programmed
> >>>>>>> into the plane registers resulting in a corrupt image on screen.
> >>>>>>>
> >>>>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >>>>>>> Cc: Sonika Jindal <sonika.jindal at intel.com>
> >>>>>>> ---
> >>>>>>>     drivers/gpu/drm/i915/intel_display.c | 16 ++++++++++++----
> >>>>>>>     1 file changed, 12 insertions(+), 4 deletions(-)
> >>>>>>>
> >>>>>>> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> >>>>>>> index 539c3737e823..6328788193e4 100644
> >>>>>>> --- a/drivers/gpu/drm/i915/intel_display.c
> >>>>>>> +++ b/drivers/gpu/drm/i915/intel_display.c
> >>>>>>> @@ -11126,9 +11126,10 @@ static void skl_do_mmio_flip(struct intel_crtc *intel_crtc)
> >>>>>>>     {
> >>>>>>>     	struct drm_device *dev = intel_crtc->base.dev;
> >>>>>>>     	struct drm_i915_private *dev_priv = dev->dev_private;
> >>>>>>> +	struct drm_plane *plane = intel_crtc->base.primary;
> >>>>>>>     	struct drm_framebuffer *fb = intel_crtc->base.primary->fb;
> >>>>>>>     	const enum pipe pipe = intel_crtc->pipe;
> >>>>>>> -	u32 ctl, stride;
> >>>>>>> +	u32 ctl, stride, tile_height;
> >>>>>>>
> >>>>>>>     	ctl = I915_READ(PLANE_CTL(pipe, 0));
> >>>>>>>     	ctl &= ~PLANE_CTL_TILED_MASK;
> >>>>>>> @@ -11152,9 +11153,16 @@ static void skl_do_mmio_flip(struct intel_crtc *intel_crtc)
> >>>>>>>     	 * The stride is either expressed as a multiple of 64 bytes chunks for
> >>>>>>>     	 * linear buffers or in number of tiles for tiled buffers.
> >>>>>>>     	 */
> >>>>>>> -	stride = fb->pitches[0] /
> >>>>>>> -		 intel_fb_stride_alignment(dev, fb->modifier[0],
> >>>>>>> -					   fb->pixel_format);
> >>>>>>> +	if (intel_rotation_90_or_270(plane->state->rotation)) {
> >>>>>>> +		/* stride = Surface height in tiles */
> >>>>>>> +		tile_height = intel_tile_height(dev, fb->pixel_format,
> >>>>>>> +						fb->modifier[0], 0);
> >>>>>>> +		stride = DIV_ROUND_UP(fb->height, tile_height);
> >>>>>>> +	} else {
> >>>>>>> +		stride = fb->pitches[0] /
> >>>>>>> +				intel_fb_stride_alignment(dev, fb->modifier[0],
> >>>>>>> +							  fb->pixel_format);
> >>>>>>> +	}
> >>>>>>
> >>>>>> I was wondering why we are allowing stride changes during page flip, but
> >>>>>> after looking at the history it seems we are not. The reason for
> >>>>>> updating the stride register is the fact that the units we specify it
> >>>>>> in change between different tiling modes on SKL+. We still reject actual
> >>>>>> stride changes during page flip, which is good because allowing it would
> >>>>>> cause problems for my fb->offsets[] stuff since the interpretation of the
> >>>>>> linear offset would change with the stride.
> >>>>>>
> >>>>>> We do allow changes to the rotated stride though since we don't reject
> >>>>>> changes to the fb height. I think I need to draw some pictures before I
> >>>>>> can say for sure whether that can cause problems or not. But we ca
> >>>>>> leave that for another patch if it turns out to need handling.
> >>>>>>
> >>>>>> One thing that's dodgy here is the plane->state->rotation check. I
> >>>>>> think currently we wait for pending flips during the atomic commit
> >>>>>> phase after we've swapped the state. So this may end up using the
> >>>>>> wrong rotation setting. It would be an even bigger problem if we
> >>>>>> already allowed queueing up or replaceing pending plane updates. I
> >>>>>> suppose the primary->fb thing doesn't suffer from this problem because
> >>>>>> we swap that pointer only after we've waited for pending flips.
> >>>>>
> >>>>> Current rule is that pageflip doesn't allow any change in any metadata.
> >>>>> There's some minor exception that on some platforms we can change the
> >>>>> tiling because someone asked for that specifically and it's possible.
> >>>>>
> >>>>> atomic flips will be able to cope with this. But for legacy pageflips imo
> >>>>> reject everything aggressively that changes metadata (stride, tiling,
> >>>>> rotation).
> >>>>
> >>>> I am not sure what is the conclusion. To re-iterate, idea is not to
> >>>> allow rotation changes between page flips, just to program PLANE_STRIDE
> >>>> accordingly, when rotation is already enabled. Since otherwise page flip
> >>>> will calculate it incorrectly.
> >>>>
> >>>> I though Ville was raising two concerns:
> >>>>
> >>>> 1. Will the plane state be swapped from under the pending page flips
> >>>> prematurely?
> >>>
> >>> The answer is yes.
> >>
> >> And is that OK?
> >
> > No, we could misprogram the stride and corrupt the display.
> 
> Oh I did not mean that but if it is OK to replace the state under queued 
> flips. Sounded like a bug to me but what do I know.

Would be perfectly normal if we allowed queuing up flips etc. I would
try to move the code towards that in any case.

Note that with CS flips all the state has been snaphost already and
emitted into the ring, so there it just works (or should at least).
With mmio flips it's a bit different naturally since we need to store
the snapshot somewhere else. I don't really like the plane->fb
situation either, so that too should get fixed to take an explicit
snapshot for the flip.

-- 
Ville Syrjälä
Intel OTC