[Intel-gfx] [PATCH 00/23] drm: Eliminate plane->fb/crtc usage for atomic drivers

[Ville Syrjälä]

On Thu, Mar 22, 2018 at 05:51:35PM +0100, Noralf Trønnes wrote:
> tinydrm is also using plane->fb:
> 
> $ grep -r "plane\.fb" drivers/gpu/drm/tinydrm/
> drivers/gpu/drm/tinydrm/repaper.c:      if (tdev->pipe.plane.fb != fb)
> drivers/gpu/drm/tinydrm/mipi-dbi.c:     if (tdev->pipe.plane.fb != fb)
> drivers/gpu/drm/tinydrm/mipi-dbi.c:     struct drm_framebuffer *fb = 
> mipi->tinydrm.pipe.plane.fb;

Oh dear, and naturally it's the most annoying one of the bunch. So we
want to take the plane lock in the fb.dirty() hook to look at the fb,
but mipi-dbi.c calls that directly from the bowels of its
.atomic_enable() hook. So that would deadlock pretty neatly if the
commit happens while already holding the lock.

So we'd either need to plumb an acquire context into fb.dirty(),
or maybe have tinydrm provide a lower level lockless dirty() hook
that gets called by both (there are just too many layers in this
particular cake to immediately see where such a hook would be best
placed). Or maybe there's some other solution I didn't think of.

Looking at this I'm also left wondering how this is supposed
handle fb.dirty() getting called concurrently with a modeset.
The dirty_mutex only seems to offer any protection for
fb.dirty() against another fb.dirty()...

-- 
Ville Syrjälä
Intel OTC