[PATCH] drm/fourcc: add LINEAR modifiers with an exact pitch alignment

[Brian Starkey]

On Tue, Dec 17, 2024 at 11:13:05AM +0000, Michel Dänzer wrote:
> On 2024-12-17 10:14, Brian Starkey wrote:
> > On Sun, Dec 15, 2024 at 03:53:14PM +0000, Marek Olšák wrote:
> >> The comment explains the problem with DRM_FORMAT_MOD_LINEAR.
> >>
> >> Signed-off-by: Marek Olšák <marek.olsak at amd.com>
> >>
> >> diff --git a/include/uapi/drm/drm_fourcc.h b/include/uapi/drm/drm_fourcc.h
> >> index 78abd819fd62e..8ec4163429014 100644
> >> --- a/include/uapi/drm/drm_fourcc.h
> >> +++ b/include/uapi/drm/drm_fourcc.h
> >> @@ -484,9 +484,27 @@ extern "C" {
> >>   * modifier (e.g. not setting DRM_MODE_FB_MODIFIERS in the DRM_ADDFB2
> >> ioctl),
> >>   * which tells the driver to also take driver-internal information into
> >> account
> >>   * and so might actually result in a tiled framebuffer.
> >> + *
> >> + * WARNING:
> >> + * There are drivers out there that expose DRM_FORMAT_MOD_LINEAR, but only
> >> + * support a certain pitch alignment and can't import images with this
> >> modifier
> >> + * if the pitch alignment isn't exactly the one supported. They can however
> >> + * allocate images with this modifier and other drivers can import them
> >> only
> >> + * if they support the same pitch alignment. Thus, DRM_FORMAT_MOD_LINEAR is
> >> + * fundamentically incompatible across devices and is the only modifier
> >> that
> >> + * has a chance of not working. The PITCH_ALIGN modifiers should be used
> >> + * instead.
> >>   */
> >>  #define DRM_FORMAT_MOD_LINEAR  fourcc_mod_code(NONE, 0)
> >>
> >> +/* Linear layout modifiers with an explicit pitch alignment in bytes.
> >> + * Exposing this modifier requires that the pitch alignment is exactly
> >> + * the number in the definition.
> >> + */
> >> +#define DRM_FORMAT_MOD_LINEAR_PITCH_ALIGN_64B fourcc_mod_code(NONE, 1)
> > 
> > Why do we want this to be a modifier? All (?) of the other modifiers
> > describe properties which the producer and consumer need to know in
> > order to correctly fill/interpret the data.
> > 
> > Framebuffers already have a pitch property which tells the
> > producer/consumer how to do that for linear buffers.
> 
> At this point, the entity which allocates a linear buffer on device
> A to be shared with another device B can't know the pitch
> restrictions of B. If it guesses incorrectly, accessing the buffer
> with B won't work, so any effort allocating the buffer and producing
> its contents will be wasted.

I do understand (and agree) the need for allocators to know about these
constraints.

> 
> 
> > Modifiers are meant to describe framebuffers, and this pitch alignment
> > requirement isn't really a framebuffer property - it's a device
> > constraint. It feels out of place to overload modifiers with it.
> > 
> > I'm not saying we don't need a way to describe constraints to
> > allocators, but I question if modifiers the right mechanism to
> > communicate them?
> While I agree with your concern in general, AFAIK there's no other
> solution for this even on the horizon, after years of talking about
> it. The solution proposed here seems like an acceptable stop gap,
> assuming it won't result in a gazillion linear modifiers.

UAPI is baked forever, so it's worth being a little wary IMO.

This sets a precedent for describing constraints via modifiers. The
reason no other proposal is on the horizon is because describing the
plethora of constraints across devices is a hard problem; and the
answer so far has been "userspace needs to know" (à la Android's
gralloc).

If we start down the road of describing constraints with modifiers, I
fear we'll end up in a mess. The full enumeration of modifiers is
already horrendous for parameterized types, please let's not
combinatorially multiply those by constraints.

Just thinking about HW I'm familiar with...

FORMAT_MOD_AFBC_16x16_ROTATABLE_ONLY_IF_LT_2048 (x5-ish variants)
FORMAT_MOD_AFBC_16x16_ROTATABLE_ONLY_IF_LT_1088 (x5-ish variants)
FORMAT_MOD_AFBC_16x16_USABLE_ONLY_IF_1_OTHER_AFBC_LAYER
	(x all AFBC modifiers, including multiply by the two ROTATABLE
	constraints above)
FORMAT_MOD_LINEAR_YUV420_MAX_2048_WIDE

That last one also highlights another problem with using modifiers for
constraints. That YUV420 restriction is orthogonal to the compression
scheme. So we'd need a FORMAT_MOD_LINEAR_YUV420_MAX_2048_WIDE *and* a
FORMAT_MOD_AFBC_YUV420_MAX_2048_WIDE (multiplied by all the AFBC
variants), and any other compression scheme multiplied by all its
variants. Not very nice.

Cheers,
-Brian

P.S. "is the only modifier that has a chance of not working" is
fundamentally false. Things can not work for an infinite number of
reasons, that's why we have TEST_ONLY. Allocating with the correct
pitch alignment is not a guarantee that you can display your
framebuffer.

> 
> 
> -- 
> Earthling Michel Dänzer       \        GNOME / Xwayland / Mesa developer
> https://redhat.com             \               Libre software enthusiast