RADEON/AIGLX/DRM Problem

[vehemens]

On Monday 16 July 2007 03:02:53 am Kostik Belousov wrote:
> On Sun, Jul 15, 2007 at 02:31:02AM -0700, vehemens wrote:
> > The problem with the RADEON driver failing when run with the AIGLX
> > extension appears to be a problem with the DRM piece (library and/or
> > driver).
> >
> > What is happening is that the RADEON driver initializes DRM and sets a
> > lock using one file descriptor.   When the AIGLX extension loads, it
> > probes via DRM using open/close calls on another file descriptor.
> >
> > As both the RADEON driver and the AIGLX extension are part of the same
> > process, the AIGLX probe close calls drm_close which removes the DRI lock
> > causing the RADEON driver to fail during xserver initialization (i.e.
> > system hang).
> >
> > Just to complicate the issue, I've been told that this doesn't occur with
> > linux with the suggestion that the BSD driver is broken or that the close
> > semantics are different between the two operating systems.
> >
> > I'm looking for ideas here as I don't know what direction to go in a this
> > point.
>
> The following may change the behaviour of the drm_close() somewhat
> in needed direction. It does all cleanup only on last close. I am
> completely unsure whether this is right, and whether it would help. It
> seems that normal drm dev cloning is needed there.
>
> Patch only compile tested.
>
> diff --git a/sys/dev/drm/drm_drv.c b/sys/dev/drm/drm_drv.c
> index dda03c1..b7c9ab9 100644
> --- a/sys/dev/drm/drm_drv.c
> +++ b/sys/dev/drm/drm_drv.c
> @@ -711,6 +711,15 @@ int drm_close(struct cdev *kdev, int flags, int fmt,
> DRM_STRUCTPROC *p) return EINVAL;
>  	}
>
> +	atomic_inc( &dev->counts[_DRM_STAT_CLOSES] );
> +#ifdef __FreeBSD__
> +	device_unbusy(dev->device);
> +#endif
> +	if (--dev->open_count != 0) {
> +		DRM_UNLOCK();
> +		return (0);
> +	}
> +
>  	if (dev->driver.preclose != NULL)
>  		dev->driver.preclose(dev, filp);
>
> @@ -797,13 +806,8 @@ int drm_close(struct cdev *kdev, int flags, int fmt,
> DRM_STRUCTPROC *p) * End inline drm_release
>  	 */
>
> -	atomic_inc( &dev->counts[_DRM_STAT_CLOSES] );
> -#ifdef __FreeBSD__
> -	device_unbusy(dev->device);
> -#endif
> -	if (--dev->open_count == 0) {
> -		retcode = drm_lastclose(dev);
> -	}
> +
> +	retcode = drm_lastclose(dev);
>
>  	DRM_UNLOCK();

I did try your patch.  It allows me to remove all of my workarounds.  In 
addition, I no longer have the xserver exit hangs.

I get a white screen when I restart the xserver and beryl.  This is probably 
the driver not restoring the card state on exit.

I havn't looked at the driver code on the use of device_unbusy(dev->device) 
yet, but was wondering if it really should be per close?

It does appear to be the best solution and probably the right one at this 
point.