[Intel-gfx] [PATCH 9/9] drm/i915: Make intel_display_power_put_unchecked() an internal-only function

[Imre Deak]

On Tue, Dec 01, 2020 at 11:13:38AM +0000, Chris Wilson wrote:
> Quoting Imre Deak (2020-12-01 10:52:22)
> > On Tue, Dec 01, 2020 at 08:14:40AM +0000, Chris Wilson wrote:
> > > Quoting Imre Deak (2020-11-30 22:47:08)
> > > > On Mon, Nov 30, 2020 at 10:07:01PM +0000, Chris Wilson wrote:
> > > > > Quoting Imre Deak (2020-11-30 21:22:00)
> > > > > > All the display power domain references are wakeref tracked now, so we
> > > > > > can mark intel_display_power_put_unchecked() as an internal function
> > > > > > (for suppressing wakeref tracking in non-debug builds).
> > > > > > 
> > > > > > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > > > > > Signed-off-by: Imre Deak <imre.deak at intel.com>
> > > > > 
> > > > > Ok, after all previous patches it will only be called from the header
> > > > > after throwing away the wakeref.
> > > > > 
> > > > > I have a sneaky suspicion you might take another path after reviewing
> > > > > the danger caused by the debug build, 
> > > > 
> > > > Yes, how about also adding:
> > > > 
> > > > +static inline void
> > > > +____intel_display_power_put(struct drm_i915_private *i915,
> > > > +                           enum intel_display_power_domain domain,
> > > > +                           intel_wakeref_t wakeref)
> > > > +{
> > > > +       intel_display_power_put_unchecked(i915, domain);
> > > > +}
> > > > +
> > > >  static inline void
> > > >  intel_display_power_put(struct drm_i915_private *i915,
> > > >                         enum intel_display_power_domain domain,
> > > >                         intel_wakeref_t wakeref)
> > > >  {
> > > > -       intel_display_power_put_unchecked(i915, domain);
> > > > +       ____intel_display_power_put(i915, domain, wakeref);
> > > >  }
> > > > 
> > > > (and similar change for intel_display_power_put_async()) ?
> > > 
> > > Hmm. The compiler shouldn't DCE the wakeref since it has a side-effect.
> > > We can but see.
> > 
> > Yes, arguments passed to functions are evaluated exactly once. The above
> > extra call doesn't make sense anyway.
> > 
> > Are you ok with patch 4 then?
> 
> If you've done a quick non-debug test run, then I'm convinced I was
> barking up the wrong tree.

Yes, I was testing w/o CONFIG_DRM_I915_DEBUG_RUNTIME_PM, it worked as
expected.

> The remaining patches are
> Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>
> (since they were all blocked by the uncertainty in behaviour of the
> fetch_and_zero).

Thanks. After this discussion, still pondering about a generic policy
wrt. allowing side-effects in function arguments. I decided that
put_ref(..., fetch_and_zero(&wakeref)) is practical and frequent enough
to allow it there.

> -Chris