[Intel-gfx] [PATCH v2 11/11] drm/i915: Introduce separate status variable for RC6 and LLC ring frequency setup

[Chris Wilson]

Quoting Sagar Arun Kamble (2017-10-06 13:13:40)
> Defined new struct intel_rc6 to hold RC6 specific state and
> intel_ring_pstate to hold ring specific state.
> 
> v2: s/intel_ring_pstate/intel_llc_pstate and rebase. (Chris)
> 
> Signed-off-by: Sagar Arun Kamble <sagar.a.kamble at intel.com>
> Cc: Imre Deak <imre.deak at intel.com>
> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> Cc: Radoslaw Szwichtenberg <radoslaw.szwichtenberg at intel.com>
> ---
>  drivers/gpu/drm/i915/i915_drv.c |  2 +-
>  drivers/gpu/drm/i915/i915_drv.h | 10 ++++++++
>  drivers/gpu/drm/i915/intel_pm.c | 57 +++++++++++++++++++++++++++++++----------
>  3 files changed, 54 insertions(+), 15 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> index 470807c..154f231 100644
> --- a/drivers/gpu/drm/i915/i915_drv.c
> +++ b/drivers/gpu/drm/i915/i915_drv.c
> @@ -2502,7 +2502,7 @@ static int intel_runtime_suspend(struct device *kdev)
>         struct drm_i915_private *dev_priv = to_i915(dev);
>         int ret;
>  
> -       if (WARN_ON_ONCE(!(dev_priv->pm.rps.enabled && intel_rc6_enabled())))
> +       if (WARN_ON_ONCE(!(dev_priv->pm.rc6.enabled && intel_rc6_enabled())))
>                 return -ENODEV;
>  
>         if (WARN_ON_ONCE(!HAS_RUNTIME_PM(dev_priv)))
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 45944a8..a07aa71 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -1363,8 +1363,18 @@ struct intel_rps {
>         struct intel_rps_ei ei;
>  };
>  
> +struct intel_rc6 {
> +       bool enabled;
> +};
> +
> +struct intel_llc_pstate {
> +       bool configured;
> +};
> +
>  struct intel_gen6_power_mgmt {
>         struct intel_rps rps;
> +       struct intel_rc6 rc6;
> +       struct intel_llc_pstate llc_pstate;
>         struct delayed_work autoenable_work;
>  
>         /*
> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> index 03264fe..df36a6f 100644
> --- a/drivers/gpu/drm/i915/intel_pm.c
> +++ b/drivers/gpu/drm/i915/intel_pm.c
> @@ -7873,7 +7873,12 @@ static void intel_init_emon(struct drm_i915_private *dev_priv)
>  
>  static inline void intel_update_ring_freq(struct drm_i915_private *i915)
>  {
> +       if (READ_ONCE(i915->pm.llc_pstate.configured))
> +               return;

Tell me about how you expect the locking around this function to be.

The READ_ONCE() implies that we are doing a optimistic peek outside of a
lock, but then we set configured without acquiring a lock, so I assume
we are inside some lock.

That looks true for all, we don't need READ_ONCE() anymore as we only
inspect inside the mutex (and so READ_ONCE is giving the wrong
impression).

> +
>         gen6_update_ring_freq(i915);
> +
> +       i915->pm.llc_pstate.configured = true;
>  }
>  
>  void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
>  {
> -       if (!READ_ONCE(dev_priv->pm.rps.enabled))
> -               return;
> -
>         mutex_lock(&dev_priv->pm.pcu_lock);
>  
>         intel_disable_rc6(dev_priv);
>         intel_disable_rps(dev_priv);
> +       if (HAS_LLC(dev_priv))
> +               dev_priv->pm.llc_pstate.configured = false;

Always clear it? If no llc, it can never be configured.
Hmm, better if we just made it symmetrical with
s/intel_update_ring_freq/intel_enable_llc_pstate/ and
intel_disable_llc_pstate here.

>  
> -       dev_priv->pm.rps.enabled = false;
>         mutex_unlock(&dev_priv->pm.pcu_lock);
>  }

> @@ -8080,7 +8103,10 @@ static void __intel_autoenable_gt_powersave(struct work_struct *work)
>         struct intel_engine_cs *rcs;
>         struct drm_i915_gem_request *req;
>  
> -       if (READ_ONCE(dev_priv->pm.rps.enabled))
> +       if (READ_ONCE(dev_priv->pm.rps.enabled) &&
> +           READ_ONCE(dev_priv->pm.rc6.enabled) &&
> +           !(HAS_LLC(dev_priv) ^
> +             READ_ONCE(dev_priv->pm.llc_pstate.configured)))
>                 goto out;

This optimisation has lost its appeal :)

Kill it, if we need something like it we can try again later.
-Chris