[Intel-gfx] [PATCH 21/23] drm/i915: Add support for FBs requiring a POT stride alignment

[Ville Syrjälä]

On Sat, Mar 13, 2021 at 04:36:47PM +0200, Imre Deak wrote:
> On Fri, Mar 12, 2021 at 08:02:36PM +0200, Ville Syrjälä wrote:
> > [...]
> > On Thu, Mar 11, 2021 at 12:17:34AM +0200, Imre Deak wrote:
> > > diff --git a/drivers/gpu/drm/i915/display/intel_display_types.h b/drivers/gpu/drm/i915/display/intel_display_types.h
> > > index fc02eca45e4d..08b348c9e3e1 100644
> > > --- a/drivers/gpu/drm/i915/display/intel_display_types.h
> > > +++ b/drivers/gpu/drm/i915/display/intel_display_types.h
> > > @@ -94,6 +94,7 @@ struct intel_framebuffer {
> > >  	struct drm_framebuffer base;
> > >  	struct intel_frontbuffer *frontbuffer;
> > >  	struct intel_rotation_info rot_info;
> > > +	struct intel_remapped_info rem_info;
> > >  
> > >  	/* for each plane in the normal GTT view */
> > >  	struct {
> > > @@ -101,6 +102,8 @@ struct intel_framebuffer {
> > >  	} normal[4];
> > >  	/* for each plane in the rotated GTT view for no-CCS formats */
> > >  	struct intel_fb_plane_remap_info rotated[2];
> > > +	/* for each plane in the remapped GTT view. TODO: CCS formats */
> > > +	struct intel_fb_plane_remap_info remapped[2];
> > 
> > We might want to look into restructuring this a it as a followup.
> > Maybe we can collect all the rotation vs. remapping stuff into
> > separate sub-structures. Not sure.
> 
> Do you mean
> 
>  struct intel_fb_plane_remap_info {
> +       struct intel_remapped_plane_info gtt;
>         unsigned int x, y;
>         unsigned int pitch; /* pixels */
>  };
> @@ -93,8 +94,6 @@ struct intel_fb_plane_remap_info {
>  struct intel_framebuffer {
>         struct drm_framebuffer base;
>         struct intel_frontbuffer *frontbuffer;
> -       struct intel_rotation_info rot_info;
> -       struct intel_remapped_info rem_info;
>  
>         /* for each plane in the normal GTT view */
>         struct {
> 
> resulting in
> https://github.com/ideak/linux/commit/0fd28d738f9
> 
> ?
> 
> Looks better to me.

Yeah something like that I guess.

> 
> > [...] 
> > > +static unsigned int plane_view_dst_stride(const struct intel_framebuffer *fb, int color_plane,
> > > +					  int pitch_tiles)
> > > +{
> > > +	unsigned int dst_stride;
> > > +
> > > +	if (!intel_fb_needs_pot_stride_remap(fb)) {
> > > +		dst_stride = pitch_tiles;
> > > +	} else {
> > > +		dst_stride = roundup_pow_of_two(pitch_tiles);
> > > +		drm_WARN_ON(fb->base.dev, dst_stride < pitch_tiles);
> > 
> > Dunno if that WARN is particularly useful. We're talking in tiles here
> > so seems extremely unlikely it could overflow.
> > 
> > So I'd probably just make this as simple as possible, like:
> > if (needs_pot)
> > 	return roundup(x);
> > else
> > 	return x;
> 
> Ok, maybe I was thinking of u16, but the overflow for that is checked
> later.
> 
> I'll also s/int pitch_tiles/unsigned int pitch_tiles/.
> 
> > > +	};
> > > +
> > > +	return dst_stride;
> > > +}
> > > +
> > [...]
> > > @@ -1280,9 +1282,27 @@ rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
> > >  			sg_dma_address(sg) =
> > >  				i915_gem_object_get_dma_address(obj, src_idx);
> > >  			sg_dma_len(sg) = I915_GTT_PAGE_SIZE;
> > > +
> > >  			sg = sg_next(sg);
> > > -			src_idx -= stride;
> > > +			src_idx -= src_stride;
> > >  		}
> > > +
> > > +		left = (dst_stride - height) * I915_GTT_PAGE_SIZE;
> > > +
> > > +		if (!left)
> > > +			continue;
> > > +
> > > +		st->nents++;
> > > +
> > > +		/*
> > > +		 * The DE ignores the PTEs for the padding tiles, the sg entry
> > > +		 * here is just a conenience to indicate how many padding PTEs
> > > +		 * to insert at this spot.
> > > +		 */
> > 
> > OK. That certainly makes this nice and simple.
> 
> This was only confirmed to be the case on ADL_P, but that's the only
> relevant place in any case.
> 
> > > +		sg_set_page(sg, NULL, left, 0);
> > > +		sg_dma_address(sg) = 0;
> > > +		sg_dma_len(sg) = left;
> > > +		sg = sg_next(sg);
> > 
> > Do we have enough sg entries for these extras? Ah, yeah we allocate
> > based on the worst case where each vma page needs its own entry.
> 
> Yes, and then i915_sg_trim() compacts it.
> 
> > Reviewed-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > 
> > >  	}
> > >  
> > >  	return sg;

-- 
Ville Syrjälä
Intel