[Intel-gfx] [PATCH] drm/i915/vlv: Fix off-by-1 error in calculating num_levels.

[Ville Syrjälä]

On Mon, Jul 25, 2016 at 01:32:45PM +0200, Maarten Lankhorst wrote:
> Hey,
> 
> Op 19-07-16 om 18:21 schreef Ville Syrjälä:
> > On Tue, Jul 19, 2016 at 04:50:49PM +0100, Chris Wilson wrote:
> >> On Tue, Jul 19, 2016 at 06:25:42PM +0300, Ville Syrjälä wrote:
> >>> On Tue, Jul 19, 2016 at 05:14:23PM +0200, Maarten Lankhorst wrote:
> >>>> num_levels should be level+1, not level, else num_levels - 1 becomes
> >>>> negative. This resulted in bogus watermarks being written to the first
> >>>> 255 levels like below:
> >>>>
> >>>> [drm] Setting FIFO watermarks - C: plane=0, cursor=0, sprite0=0, sprite1=0, SR: plane=0, cursor=0 level=255 cxsr=0
> >>>> [drm:chv_set_memory_dvfs [i915]] *ERROR* timed out waiting for Punit DDR DVFS request
> >>>> [drm:intel_cpu_fifo_underrun_irq_handler [i915]] *ERROR* CPU pipe C FIFO underrun
> >>>> [drm:chv_set_memory_dvfs [i915]] *ERROR* timed out waiting for Punit DDR DVFS request
> >>>>
> >>>> Testcase: kms_atomic_transition
> >>>> Fixes: 262cd2e154c2 ("drm/i915: CHV DDR DVFS support and another watermark rewrite")
> >>>> Cc: stable at vger.kernel.org
> >>>> Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> >>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> >>>> ---
> >>>> Urgent fix for watermark support. This is definitely a pre-requisite for this series.
> >>>> With this I've noticed that patch "[RFC 3/8] drm/i915/vlv: Move fifo_size from
> >>>> intel_plane_wm_parameters to vlv_wm_state" introduces a regression with invalid FIFO split.
> >>>>
> >>>> I need to find out what's going wrong in that patch before this series can be applied.
> >>>>
> >>>>  drivers/gpu/drm/i915/intel_pm.c | 2 +-
> >>>>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> >>>> index 376c60b98515..8defdcc54529 100644
> >>>> --- a/drivers/gpu/drm/i915/intel_pm.c
> >>>> +++ b/drivers/gpu/drm/i915/intel_pm.c
> >>>> @@ -1148,7 +1148,7 @@ static void vlv_compute_wm(struct intel_crtc *crtc)
> >>>>  			}
> >>>>  		}
> >>>>  
> >>>> -		wm_state->num_levels = level;
> >>>> +		wm_state->num_levels = level + 1;
> >>> Nope. The loop above breaks when the current level is bad, hence level-1
> >>> is actually the higher usable level.
> >> Without knowing the limits of plane->wm.fifo_size, it looks like it can
> >> break on level == 0 though.
> > Hmm. That shouldn't be possible. So looks like a bug snuck in.
> >
> >> Might as well set that hack to paranoid levels:
> >>
> >> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> >> index 630b116988f6..e8c2874b8629 100644
> >> --- a/drivers/gpu/drm/i915/intel_pm.c
> >> +++ b/drivers/gpu/drm/i915/intel_pm.c
> >> @@ -1124,11 +1124,13 @@ static void vlv_compute_wm(struct intel_crtc *crtc)
> >>                 /* normal watermarks */
> >>                 for (level = 0; level < wm_state->num_levels; level++) {	
> >>                         int wm = vlv_compute_wm_level(plane, crtc, state, level);
> >> -                       int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
> >> -
> >>                         /* hack */
> >> -                       if (WARN_ON(level == 0 && wm > max_wm))
> >> -                               wm = max_wm;
> > Actually this should just be 
> >
> > if (WARN_ON(level == 0 && wm > fifo_size))
> > 	wm = fifo_size;
> >
> > assuming we want to keep the hack around for now.
> >
> > Eventually we'll want to make it just return an error though.
> Yes, that's what I came up with.
> 
> But we still need to clamp further, probably to plane->wm.fifo_size.
> However sr_fifo_size is also clamped to 511 because of the level calculations here.
> 
> Below it sets sr[level].plane = min(sr_fifo_size, wm[level].plane),
> which seems weird. How can this ever end up being something other than wm[level].plane?

There are three planes. CHV was supposed to have maxfifo with up to 2 active
planes. But I'm not sure that feature made it in. I never got around
testing it.

> Because sr_fifo_size >= max_wm is always true.
> 
> I guess vlv_invert_wms will invert it, but we could simply only set it there then, or remove the min()..
> 
> ~Maarten

-- 
Ville Syrjälä
Intel OTC