[PATCH] drm/i915: Align tiled scanouts from stolen memory to 256k in the GTT

[PATCH] drm/i915: Align tiled scanouts from stolen memory to 256k in the GTT

[Chris Wilson]

For unfathomable reasons this alignment appears to be required for tiled
scanouts being read from stolen memory. I can find no reference in the
w/a db to support this requirement, but the evidence of my own eyes says
this prevents many headaches.

Note that I have not tricked anything older than Sandybridge into using
stolen tiled scanouts, so the extra alignment may be required there as
well.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/intel_display.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
index 6b7ce06..a7573f2 100644
--- a/drivers/gpu/drm/i915/intel_display.c
+++ b/drivers/gpu/drm/i915/intel_display.c
@@ -1848,6 +1848,8 @@ intel_pin_and_fence_fb_obj(struct drm_device *dev,
 	case I915_TILING_X:
 		/* pin() will align the object as required by fence */
 		alignment = 0;
+		if (obj->stolen && INTEL_INFO(dev)->gen >= 6)
+			alignment = 256 * 1024;
 		break;
 	case I915_TILING_Y:
 		/* Despite that we check this in framebuffer_init userspace can
-- 
1.8.4.rc2

Re: [PATCH] drm/i915: Align tiled scanouts from stolen memory to 256k in the GTT

[Ville Syrjälä]

On Sun, Aug 11, 2013 at 11:17:28PM +0100, Chris Wilson wrote:
> For unfathomable reasons this alignment appears to be required for tiled
> scanouts being read from stolen memory. I can find no reference in the
> w/a db to support this requirement, but the evidence of my own eyes says
> this prevents many headaches.
> 
> Note that I have not tricked anything older than Sandybridge into using
> stolen tiled scanouts, so the extra alignment may be required there as
> well.

Strange. I can't find anything except async flips and vt-d which would
require 256k alignment.

> 
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/intel_display.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> index 6b7ce06..a7573f2 100644
> --- a/drivers/gpu/drm/i915/intel_display.c
> +++ b/drivers/gpu/drm/i915/intel_display.c
> @@ -1848,6 +1848,8 @@ intel_pin_and_fence_fb_obj(struct drm_device *dev,
>  	case I915_TILING_X:
>  		/* pin() will align the object as required by fence */
>  		alignment = 0;
> +		if (obj->stolen && INTEL_INFO(dev)->gen >= 6)
> +			alignment = 256 * 1024;
>  		break;
>  	case I915_TILING_Y:
>  		/* Despite that we check this in framebuffer_init userspace can
> -- 
> 1.8.4.rc2
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Ville Syrjälä
Intel OTC

Re: [PATCH] drm/i915: Align tiled scanouts from stolen memory to 256k in the GTT

[Chris Wilson]

On Mon, Aug 12, 2013 at 12:38:03PM +0300, Ville Syrjälä wrote:
> On Sun, Aug 11, 2013 at 11:17:28PM +0100, Chris Wilson wrote:
> > For unfathomable reasons this alignment appears to be required for tiled
> > scanouts being read from stolen memory. I can find no reference in the
> > w/a db to support this requirement, but the evidence of my own eyes says
> > this prevents many headaches.
> > 
> > Note that I have not tricked anything older than Sandybridge into using
> > stolen tiled scanouts, so the extra alignment may be required there as
> > well.
> 
> Strange. I can't find anything except async flips and vt-d which would
> require 256k alignment.

I can empathically state it is required though. It might not be 256k,
that's just the first value that stopped giving me headaches. :)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre