Waiting for vertical refresh on Intel 965 and 945 chips in recent git

[Jesse Barnes]

On Friday, June 8, 2007 8:24:46 Xavier Bestel wrote:
> On Fri, 2007-06-08 at 16:20 +0100, Torgeir Veimo wrote:
> > On 8 Jun 2007, at 12:18, Hamish Moffatt wrote:
> > > On Fri, Jun 08, 2007 at 08:23:05AM +0200, Michel Dänzer wrote:
> > >> On Thu, 2007-06-07 at 18:00 +0100, Simon Farnsworth wrote:
> > >>> In the 7.2 release, I could use the DRM directly to wait for
> > >>> vertical
> > >>> blanking, using the DRM_IOCTL_WAIT_VBLANK ioctl. The intel driver
> > >>> has
> > >>> been changed to only enable the interrupt if 3D is in use, so I
> > >>> reworked
> > >>> my code to use the GLX_SGI_video_sync extension to wait for VBlank.
> > >>>
> > >>> I am now finding that the OpenGL based wait for VBlank also fails
> > >>> if the
> > >>> window I create to get an OpenGL context is fully obscured;
> > >>> glXWaitVideoSyncSGI returns 5 (GLX_BAD_CONTEXT). My code
> > >>> automatically
> > >>> falls back to trying the DRM method, but DRM_IOCTL_WAIT_VBLANK
> > >>> returns
> > >>> 16 (EBUSY).
> > >>
> > >> Yes, this is because the X server DRI layer now counts DRI windows
> > >> according to visibility, to allow e.g. page flipping to work with
> > >> multiple windows as long as only one of them is visible. So the intel
> > >> driver can't tell the difference between a window being invisible
> > >> and it
> > >> not existing at all and disables the vblank interrupts.
> > >
> > > How is "3D in use" defined?
> > >
> > > MythTV uses DRM vertical blanking sync with Xv output; I wouldn't
> > > say it
> > > uses 3D by any stretch of the imagination.
> >
> > If the screensaver is disabled, as is often the case when a media
> > player runs full screen, can't that be a toggle to always turn on
> > vblank irqs?
>
> Why can't vblank be enabled only when the DRM_IOCTL_WAIT_VBLANK ioctl is
> called ? Is that because OpenGL hangs on to it ?

The problem is that GL exposes an absolute vblank event count to clients.  So 
we have to keep an accurate count or applications could hang or refresh on 
the wrong event.  Fortunately, some devices have internal vblank counters, so 
we can disable the interrupt and only enable it when some client is actively 
waiting, and use the register to keep track of the absolute number of events 
when interrupts were disabled.

Jesse