[Intel-gfx] [PATCH] i915: Increase *_latency array size

[Ville Syrjälä]

On Wed, May 05, 2021 at 07:18:30AM -0700, Andi Kleen wrote:
> > > Increase all the latency fields to 8 members, which is enough for SKL.
> > >
> > > I don't know if they are correctly initialized upto 8, but dev_priv
> > > should start out as zero, so presumably they will be zero.
> > 
> > Thanks, the warning should be fixed by commit
> > 
> > c6deb5e97ded ("drm/i915/pm: Make the wm parameter of print_wm_latency a pointer")
> > 
> > in drm-intel-next.
> 
> That's just hiding the problem.
> 
> > 
> > There doesn't actually seem to be a bug here,
> 
> Can you explain that please?
> 
> This is the loop in question
> 
>  max_level = ilk_wm_max_level(dev_priv);
> 
>         for (level = 0; level <= max_level; level++) {
>                 unsigned int latency = wm[level];
> 
>                 if (latency == 0) {
>                         drm_dbg_kms(&dev_priv->drm,
>                                     "%s WM%d latency not provided\n",
>                                     name, level);
>                         continue;
>                 }
> 
> 		...
> 	}
> 
> (no other loop termination condition)
> 
> and ilk_wm_max_level is
> 
> int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
> {
>         /* how many WM levels are we expecting */
>         if (INTEL_GEN(dev_priv) >= 9)
>                 return 7;
>         else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
>                 return 4;
>         else if (INTEL_GEN(dev_priv) >= 6)
>                 return 3;
>         else
>                 return 2;
> }
> 
> There is no loop termination in the loop above, it will always read
> every member through the max level reported. And on GEN>=9 it will be 7, while
> the input array for several of the cases has only 5 members.
> 
> So it will read beyond the array and gcc is correct in complaining.
> 
> What do I miss when you say there is no bug?

We always use dev_priv->wm.skl_latency[] for gen9+. See
{pri,spr,cur}_wm_latency_show(), skl_setup_wm_latency(), etc.

-- 
Ville Syrjälä
Intel