[PATCH 3/3] present: Check for an invalid target CRTC before flipping

[Chris Wilson]

On Sun, Feb 08, 2015 at 02:10:07AM -0800, Kenneth Graunke wrote:
> On Friday, February 06, 2015 08:25:44 AM Chris Wilson wrote:
> > Once we have chosen the target CRTC, check for an error when querying
> > the current MSC so we can report the error before attempting to record
> > the uninitialised MSC/UST, and subsequently feeding garbage values to
> > drmWaitVBlank.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  present/present.c | 4 +++-
> >  1 file changed, 3 insertions(+), 1 deletion(-)
> > 
> > diff --git a/present/present.c b/present/present.c
> > index 1ce587e..4a1cf0c 100644
> > --- a/present/present.c
> > +++ b/present/present.c
> > @@ -727,7 +727,9 @@ present_pixmap(WindowPtr window,
> >      else if (!target_crtc)
> >  	target_crtc = present_get_crtc(window);
> >  
> > -    present_get_ust_msc(screen, target_crtc, &ust, &crtc_msc);
> > +    ret = present_get_ust_msc(screen, target_crtc, &ust, &crtc_msc);
> > +    if (ret)
> > +	    return ret;
> >  
> >      target_msc = present_window_to_crtc_msc(window, target_crtc, window_msc, crtc_msc);
> 
> You've definitely found a bug here. Some drivers' get_ust_msc hooks
> (UXA, modesetting) return BadMatch and leave crtc_msc uninitialized.  
> At which point we'd happily continue and use rubbish values.
> 
> I discovered this today as well, when using valgrind to try and debug
> some crashes in xf86-video-modesetting using my 'pageflip' branch:
> [term1]# valgrind --vgdb-error=1 X
> [term2]# gdb
> [term2](gdb) target remote | vgdb
> [term2](gdb) c
> [term3]$ DISPLAY=:0 glxgears -fullscreen &
> [term3]$ DISPLAY=:0 xset force dpms off
> <hit a key to DPMS on>
> 
> valgrind will complain about the uninitialized values in the crtc_msc >=
> target_msc branch.
> 
> But...is an early return really the right thing to do?  I tested this
> with xf86-video-modesetting and my 'pageflip' branch, and it caused
> "glxgears -fullscreen" to die from a BadMatch error after a single DPMS
> cycle.  I tried changing it to 'return Success', and that caused it to
> stop drawing altogether.  I also tried it with UXA+DRI3, and it seemed
> to break there as well.  So it seems like we need to do something more.

The previous patch fixes that bug.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre