[Intel-gfx] [PATCH 2/5] drm/i915: Always prepare planes at the start of an atomic commit

[Chris Wilson]

On Tue, Nov 15, 2016 at 11:10:26AM +0100, Daniel Vetter wrote:
> On Tue, Nov 15, 2016 at 09:57:37AM +0000, Chris Wilson wrote:
> > On Tue, Nov 15, 2016 at 10:37:09AM +0100, Daniel Vetter wrote:
> > > On Tue, Nov 15, 2016 at 08:58:14AM +0000, Chris Wilson wrote:
> > > > The generic atomic helper likes to skip a prepare_plane_fb() if it
> > > > decides that the plane->fb is unchanged. This is wrong for us for a
> > > > couple of reasons:
> > > > 
> > > >  - if the pipe is reconfigured (i.e. a size change) but the framebuffer
> > > >    is untouched, we still have to flush any rendering prior to the
> > > >    reconfiguration to prevent wait-for-scanline GPU hangs
> > > > 
> > > >  - if the framebuffer is rotated, it remains the same but has a
> > > >    different view and a different address.
> > > > 
> > > > Finally, even if the framebuffer is unchanged the flip/modeset should be
> > > > ordered with respect to rendering to the frontbuffer.
> > > > 
> > > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > 
> > > This is directly against
> > > 
> > > commit fcc60b413d14dd06ddbd79ec50e83c4fb2a097ba
> > > Author: Keith Packard <keithp at keithp.com>
> > > Date:   Sat Jun 4 01:16:22 2016 -0700
> > > 
> > >     drm: Don't prepare or cleanup unchanging frame buffers [v3]
> > > 
> > > and I'm pretty sure that was tested on i915.
> > 
> > That patch is bogus. But if we want plane granularity, the helpers are
> > not sufficient either.
> 
> Hm, what do you mean with plane granularity? If you just mean async
> commits, then yes the current helper machinery won't work. But I don't
> think extending them is a good idea either, per-plane async updates
> probably need something entirely different on the back-end. We could still
> use the atomic ioctl just as the transport, but that's about it.

I mean an atomic commit that only modifies the cursor plane (no wm, no
other crtc or global state updates) does not impede an atomic commit that
only modifies the sprite/primary plane (again no wm, no other crtc or
global state updates).

crtc granularity (i.e. a single timeline per crtc) simply doesn't work
when the existing ABI is plane granularity.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre