[Intel-gfx] [PATCH v2 3b/3] drm/i915: Kill i915_gem_execbuffer_wait_for_flips()

[Chris Wilson]

On Fri, 2 Nov 2012 14:24:12 -0400, Kristian Høgsberg <krh at bitplanet.net> wrote:
> On Fri, Nov 2, 2012 at 9:29 AM, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > On Thu,  1 Nov 2012 20:06:03 +0200, ville.syrjala at linux.intel.com wrote:
> >> From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> >>
> >> As per Chris Wilson's suggestion make
> >> i915_gem_execbuffer_wait_for_flips() go away.
> >>
> >> This was used to stall the GPU ring while there are pending
> >> page flips involving the relevant BO. Ie. while the BO is still
> >> being scanned out by the display controller.
> >>
> >> The recommended alternative is to use the page flip events to
> >> wait for the page flips to fully complete before reusing the BO
> >> of the old front buffer. Or use more buffers.
> >>
> >> Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>
> >
> > Needs an ack from either Jesse or Kristian.
> 
> I wouldn't mind seeing this go; we don't rely on this behavior in
> Weston, we use the events.  However, this is a user visible change in
> behavior of the pageflip ioctl.  I don't remember if X needs this for
> the case where it's pageflipping to fullscreen client buffers... I
> think it might (or at least might have), since we did it this way so
> that the client wouldn't have to wait for a protocol event before
> starting the next frame.  With this the client can start rendering and
> submit the first batch before it has to block.  It also minimizes the
> time from pageflip done to the client can start rendering, since the
> kernel now just unblocks the client directly.  Of course you can argue
> that we can fix that with more buffers, and maybe it's fixed in newer
> servers / ddx drivers, but it doesn't change that this is how it used
> to work.  Without this, the  client doesn't know when it's new
> backbuffer done scanning out.

We never relied on this in userspace, as you provided the flip-completed
event at the same time as you introduced pageflipping and that was used
in the original ddx to throttle the client. And from experience, I can say
that this blocking is not sufficient to prevent userspace from being
stupid...
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre