[Intel-gfx] [PATCH 1/3] drm/i915: Add support for non-power-of-2 FB plane alignment

[Imre Deak]

On Mon, Dec 30, 2019 at 10:48:31AM +0000, Chris Wilson wrote:
> Quoting Imre Deak (2019-12-27 23:51:45)
> > At least one framebuffer plane on TGL - the UV plane of YUV semiplanar
> > FBs - requires a non-power-of-2 alignment, so add support for this. This
> > new alignment restriction applies only to an offset within an FB, so the
> > GEM buffer itself containing the FB must still be power-of-2 aligned.
> 
> It's worth talking about virtual memory alignment (in the GGTT) here and
> not the physical alignment of the backing store. The buffer itself plays
> no part here.

Yes, this new restriction is about the GGTT mapping and display
specific. In fact other engines have other restrictions when
reading/writing the same YUV surfaces - for instance via a PPGTT map.
And yes, the page physical addresses can be anything.

Will improve the commit log.

> > Add a check for this (in practice plane 0, since the plane 0 offset must
> > be 0).
> > 
> > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > Signed-off-by: Imre Deak <imre.deak at intel.com>
> > ---
> >  drivers/gpu/drm/i915/display/intel_display.c | 22 +++++++++++++-------
> >  1 file changed, 14 insertions(+), 8 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/display/intel_display.c b/drivers/gpu/drm/i915/display/intel_display.c
> > index 624ba9be7293..d8970198c77e 100644
> > --- a/drivers/gpu/drm/i915/display/intel_display.c
> > +++ b/drivers/gpu/drm/i915/display/intel_display.c
> > @@ -2194,6 +2194,7 @@ intel_pin_and_fence_fb_obj(struct drm_framebuffer *fb,
> >                 return ERR_PTR(-EINVAL);
> >  
> >         alignment = intel_surf_alignment(fb, 0);
> > +       WARN_ON(!is_power_of_2(alignment));
> 
> Handle the error, if you are going to the trouble to warn, add the
> return as well.

Ok.

> >  
> >         /* Note that the w/a also requires 64 PTE of padding following the
> >          * bo. We currently fill all unused PTE with the shadow page and so
> > @@ -2432,9 +2433,6 @@ static u32 intel_compute_aligned_offset(struct drm_i915_private *dev_priv,
> >         unsigned int cpp = fb->format->cpp[color_plane];
> >         u32 offset, offset_aligned;
> >  
> > -       if (alignment)
> > -               alignment--;
> > -
> >         if (!is_surface_linear(fb, color_plane)) {
> >                 unsigned int tile_size, tile_width, tile_height;
> >                 unsigned int tile_rows, tiles, pitch_tiles;
> > @@ -2456,17 +2454,24 @@ static u32 intel_compute_aligned_offset(struct drm_i915_private *dev_priv,
> >                 *x %= tile_width;
> >  
> >                 offset = (tile_rows * pitch_tiles + tiles) * tile_size;
> > -               offset_aligned = offset & ~alignment;
> > +
> > +               offset_aligned = offset;
> > +               if (alignment)
> > +                       offset_aligned = rounddown(offset_aligned, alignment);
> >  
> >                 intel_adjust_tile_offset(x, y, tile_width, tile_height,
> >                                          tile_size, pitch_tiles,
> >                                          offset, offset_aligned);
> >         } else {
> >                 offset = *y * pitch + *x * cpp;
> > -               offset_aligned = offset & ~alignment;
> > -
> > -               *y = (offset & alignment) / pitch;
> > -               *x = ((offset & alignment) - *y * pitch) / cpp;
> > +               offset_aligned = offset;
> > +               if (alignment) {
> > +                       offset_aligned = rounddown(offset_aligned, alignment);
> > +                       *y = (offset % alignment) / pitch;
> > +                       *x = ((offset % alignment) - *y * pitch) / cpp;
> > +               } else {
> > +                       *y = *x = 0;
> > +               }
> >         }
> >  
> >         return offset_aligned;
> > @@ -3738,6 +3743,7 @@ static int skl_check_main_surface(struct intel_plane_state *plane_state)
> >         intel_add_fb_offsets(&x, &y, plane_state, 0);
> >         offset = intel_plane_compute_aligned_offset(&x, &y, plane_state, 0);
> >         alignment = intel_surf_alignment(fb, 0);
> > +       WARN_ON(!is_power_of_2(alignment));
> 
> The other two are expected to handle !is_pot...

Not sure what the alignment of the address we write to the main surface
address register should be, the spec doesn't say anything about that. I
assume now that not wrapping around the right edge of the detiler fence
we add is enough (which is also checked in intel_fill_fb_info()).

> I would strongly suggest handling the WARNs, or else you may as well bug
> out for the programming error.

Ok, will change those.

> Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>
> -Chris