diff mbox

drm/i915: make sure GPU freq drops to minimum after entering RC6 v3

Message ID 1366671953-12728-1-git-send-email-jbarnes@virtuousgeek.org (mailing list archive)
State New, archived
Headers show

Commit Message

Jesse Barnes April 22, 2013, 11:05 p.m. UTC
On VLV, the Punit doesn't automatically drop the GPU to it's minimum
voltage level when entering RC6, so we arm a timer to do it for us from
the RPS interrupt handler.  It'll generally only fire when we go idle
(or if for some reason there's a long delay between RPS interrupts), but
won't be re-armed again until the next RPS event, so shouldn't affect
power consumption after we go idle and it triggers.

v2: use delayed work instead of timer + work queue combo (Ville)
v3: fix up delayed work cancel (must be outside lock) (Daniel)
    fix up delayed work handling func for delayed work (Jesse)

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
---
 drivers/gpu/drm/i915/i915_drv.h |    2 ++
 drivers/gpu/drm/i915/i915_irq.c |   11 +++++++++++
 drivers/gpu/drm/i915/intel_pm.c |   22 ++++++++++++++++++++++
 3 files changed, 35 insertions(+)

Comments

Jani Nikula April 23, 2013, 8:45 a.m. UTC | #1
On Tue, 23 Apr 2013, Jesse Barnes <jbarnes@virtuousgeek.org> wrote:
> On VLV, the Punit doesn't automatically drop the GPU to it's minimum
> voltage level when entering RC6, so we arm a timer to do it for us from
> the RPS interrupt handler.  It'll generally only fire when we go idle
> (or if for some reason there's a long delay between RPS interrupts), but
> won't be re-armed again until the next RPS event, so shouldn't affect
> power consumption after we go idle and it triggers.
>
> v2: use delayed work instead of timer + work queue combo (Ville)
> v3: fix up delayed work cancel (must be outside lock) (Daniel)
>     fix up delayed work handling func for delayed work (Jesse)
>
> Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
> ---
>  drivers/gpu/drm/i915/i915_drv.h |    2 ++
>  drivers/gpu/drm/i915/i915_irq.c |   11 +++++++++++
>  drivers/gpu/drm/i915/intel_pm.c |   22 ++++++++++++++++++++++
>  3 files changed, 35 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 2557fc7..f563f25 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -660,6 +660,7 @@ struct i915_suspend_saved_registers {
>  
>  struct intel_gen6_power_mgmt {
>  	struct work_struct work;
> +	struct delayed_work vlv_work;
>  	u32 pm_iir;
>  	/* lock - irqsave spinlock that protectects the work_struct and
>  	 * pm_iir. */
> @@ -670,6 +671,7 @@ struct intel_gen6_power_mgmt {
>  	u8 cur_delay;
>  	u8 min_delay;
>  	u8 max_delay;
> +	u8 rpe_delay;
>  	u8 hw_max;
>  
>  	struct delayed_work delayed_resume_work;
> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> index 932e7f8..af385e2 100644
> --- a/drivers/gpu/drm/i915/i915_irq.c
> +++ b/drivers/gpu/drm/i915/i915_irq.c
> @@ -488,6 +488,17 @@ static void gen6_pm_rps_work(struct work_struct *work)
>  			gen6_set_rps(dev_priv->dev, new_delay);
>  	}
>  
> +	if (IS_VALLEYVIEW(dev_priv->dev)) {
> +		/*
> +		 * On VLV, when we enter RC6 we may not be at the minimum
> +		 * voltage level, so arm a timer to check.  It should only
> +		 * fire when there's activity or once after we've entered
> +		 * RC6, and then won't be re-armed until the next RPS interrupt.
> +		 */
> +		mod_delayed_work(dev_priv->wq, &dev_priv->rps.vlv_work,
> +				 jiffies + msecs_to_jiffies(100));

Hooray for conflicting semantics. Unlike mod_timer, mod_delayed_work
wants a delay, not an absolute time in jiffies.

> +	}
> +
>  	mutex_unlock(&dev_priv->rps.hw_lock);
>  }
>  
> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> index 2557926..c106187 100644
> --- a/drivers/gpu/drm/i915/intel_pm.c
> +++ b/drivers/gpu/drm/i915/intel_pm.c
> @@ -2822,6 +2822,23 @@ int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
>  	return val & 0xff;
>  }
>  
> +static void vlv_rps_timer_work(struct work_struct *work)
> +{
> +	drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
> +						    rps.vlv_work.work);
> +
> +	/*
> +	 * Timer fired, we must be idle.  Drop to min voltage state.
> +	 * Note: we use RPe here since it should match the
> +	 * Vmin we were shooting for.  That should give us better
> +	 * perf when we come back out of RC6 than if we used the
> +	 * min freq available.
> +	 */
> +	mutex_lock(&dev_priv->rps.hw_lock);
> +	valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
> +	mutex_unlock(&dev_priv->rps.hw_lock);
> +}
> +
>  static void valleyview_enable_rps(struct drm_device *dev)
>  {
>  	struct drm_i915_private *dev_priv = dev->dev_private;
> @@ -2886,6 +2903,7 @@ static void valleyview_enable_rps(struct drm_device *dev)
>  	rpe = valleyview_rps_rpe_freq(dev_priv);
>  	DRM_DEBUG_DRIVER("RPe GPU freq: %d\n",
>  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> +	dev_priv->rps.rpe_delay = rpe;
>  
>  	val = valleyview_rps_min_freq(dev_priv);
>  	DRM_DEBUG_DRIVER("min GPU freq: %d\n", vlv_gpu_freq(dev_priv->mem_freq,
> @@ -2895,6 +2913,8 @@ static void valleyview_enable_rps(struct drm_device *dev)
>  	DRM_DEBUG_DRIVER("setting GPU freq to %d\n",
>  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
>  
> +	INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
> +
>  	valleyview_set_rps(dev_priv->dev, rpe);
>  
>  	/* requires MSI enabled */
> @@ -3640,6 +3660,8 @@ void intel_disable_gt_powersave(struct drm_device *dev)
>  		mutex_lock(&dev_priv->rps.hw_lock);
>  		gen6_disable_rps(dev);
>  		mutex_unlock(&dev_priv->rps.hw_lock);

Theoretically the work could get executed here, between unlock and
cancel. Does that matter?

> +		if (IS_VALLEYVIEW(dev))
> +			cancel_delayed_work_sync(&dev_priv->rps.vlv_work);
>  	}
>  }
>  
> -- 
> 1.7.10.4
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Ville Syrjälä April 23, 2013, 1:28 p.m. UTC | #2
On Tue, Apr 23, 2013 at 11:45:03AM +0300, Jani Nikula wrote:
> On Tue, 23 Apr 2013, Jesse Barnes <jbarnes@virtuousgeek.org> wrote:
> > On VLV, the Punit doesn't automatically drop the GPU to it's minimum
> > voltage level when entering RC6, so we arm a timer to do it for us from
> > the RPS interrupt handler.  It'll generally only fire when we go idle
> > (or if for some reason there's a long delay between RPS interrupts), but
> > won't be re-armed again until the next RPS event, so shouldn't affect
> > power consumption after we go idle and it triggers.
> >
> > v2: use delayed work instead of timer + work queue combo (Ville)
> > v3: fix up delayed work cancel (must be outside lock) (Daniel)
> >     fix up delayed work handling func for delayed work (Jesse)
> >
> > Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
> > ---
> >  drivers/gpu/drm/i915/i915_drv.h |    2 ++
> >  drivers/gpu/drm/i915/i915_irq.c |   11 +++++++++++
> >  drivers/gpu/drm/i915/intel_pm.c |   22 ++++++++++++++++++++++
> >  3 files changed, 35 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > index 2557fc7..f563f25 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.h
> > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > @@ -660,6 +660,7 @@ struct i915_suspend_saved_registers {
> >  
> >  struct intel_gen6_power_mgmt {
> >  	struct work_struct work;
> > +	struct delayed_work vlv_work;
> >  	u32 pm_iir;
> >  	/* lock - irqsave spinlock that protectects the work_struct and
> >  	 * pm_iir. */
> > @@ -670,6 +671,7 @@ struct intel_gen6_power_mgmt {
> >  	u8 cur_delay;
> >  	u8 min_delay;
> >  	u8 max_delay;
> > +	u8 rpe_delay;
> >  	u8 hw_max;
> >  
> >  	struct delayed_work delayed_resume_work;
> > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> > index 932e7f8..af385e2 100644
> > --- a/drivers/gpu/drm/i915/i915_irq.c
> > +++ b/drivers/gpu/drm/i915/i915_irq.c
> > @@ -488,6 +488,17 @@ static void gen6_pm_rps_work(struct work_struct *work)
> >  			gen6_set_rps(dev_priv->dev, new_delay);
> >  	}
> >  
> > +	if (IS_VALLEYVIEW(dev_priv->dev)) {
> > +		/*
> > +		 * On VLV, when we enter RC6 we may not be at the minimum
> > +		 * voltage level, so arm a timer to check.  It should only
> > +		 * fire when there's activity or once after we've entered
> > +		 * RC6, and then won't be re-armed until the next RPS interrupt.
> > +		 */
> > +		mod_delayed_work(dev_priv->wq, &dev_priv->rps.vlv_work,
> > +				 jiffies + msecs_to_jiffies(100));
> 
> Hooray for conflicting semantics. Unlike mod_timer, mod_delayed_work
> wants a delay, not an absolute time in jiffies.
> 
> > +	}
> > +
> >  	mutex_unlock(&dev_priv->rps.hw_lock);
> >  }
> >  
> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > index 2557926..c106187 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > @@ -2822,6 +2822,23 @@ int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
> >  	return val & 0xff;
> >  }
> >  
> > +static void vlv_rps_timer_work(struct work_struct *work)
> > +{
> > +	drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
> > +						    rps.vlv_work.work);
> > +
> > +	/*
> > +	 * Timer fired, we must be idle.  Drop to min voltage state.
> > +	 * Note: we use RPe here since it should match the
> > +	 * Vmin we were shooting for.  That should give us better
> > +	 * perf when we come back out of RC6 than if we used the
> > +	 * min freq available.
> > +	 */
> > +	mutex_lock(&dev_priv->rps.hw_lock);
> > +	valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
> > +	mutex_unlock(&dev_priv->rps.hw_lock);
> > +}
> > +
> >  static void valleyview_enable_rps(struct drm_device *dev)
> >  {
> >  	struct drm_i915_private *dev_priv = dev->dev_private;
> > @@ -2886,6 +2903,7 @@ static void valleyview_enable_rps(struct drm_device *dev)
> >  	rpe = valleyview_rps_rpe_freq(dev_priv);
> >  	DRM_DEBUG_DRIVER("RPe GPU freq: %d\n",
> >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> > +	dev_priv->rps.rpe_delay = rpe;
> >  
> >  	val = valleyview_rps_min_freq(dev_priv);
> >  	DRM_DEBUG_DRIVER("min GPU freq: %d\n", vlv_gpu_freq(dev_priv->mem_freq,
> > @@ -2895,6 +2913,8 @@ static void valleyview_enable_rps(struct drm_device *dev)
> >  	DRM_DEBUG_DRIVER("setting GPU freq to %d\n",
> >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> >  
> > +	INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
> > +
> >  	valleyview_set_rps(dev_priv->dev, rpe);
> >  
> >  	/* requires MSI enabled */
> > @@ -3640,6 +3660,8 @@ void intel_disable_gt_powersave(struct drm_device *dev)
> >  		mutex_lock(&dev_priv->rps.hw_lock);
> >  		gen6_disable_rps(dev);
> >  		mutex_unlock(&dev_priv->rps.hw_lock);
> 
> Theoretically the work could get executed here, between unlock and
> cancel. Does that matter?

AFAICS even gen6_pm_rps_work() might be executing even after
gen6_disable_rps() was called.

I'm not quite sure if allowing gen6_pm_rps_work() to do stuff after
gen6_disable_rps() was called is already a problem on it's own. At
least it can overwrite RPNSWREQ on !VLV. BTW that register doesn't seem
to exist on VLV and yet we're poking at it from gen6_disable_rps()...

> 
> > +		if (IS_VALLEYVIEW(dev))
> > +			cancel_delayed_work_sync(&dev_priv->rps.vlv_work);
> >  	}
> >  }
> >  
> > -- 
> > 1.7.10.4
> >
> > _______________________________________________
> > Intel-gfx mailing list
> > Intel-gfx@lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Daniel Vetter April 23, 2013, 2:46 p.m. UTC | #3
On Tue, Apr 23, 2013 at 04:28:17PM +0300, Ville Syrjälä wrote:
> On Tue, Apr 23, 2013 at 11:45:03AM +0300, Jani Nikula wrote:
> > On Tue, 23 Apr 2013, Jesse Barnes <jbarnes@virtuousgeek.org> wrote:
> > > @@ -2895,6 +2913,8 @@ static void valleyview_enable_rps(struct drm_device *dev)
> > >  	DRM_DEBUG_DRIVER("setting GPU freq to %d\n",
> > >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> > >  
> > > +	INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
> > > +
> > >  	valleyview_set_rps(dev_priv->dev, rpe);
> > >  
> > >  	/* requires MSI enabled */
> > > @@ -3640,6 +3660,8 @@ void intel_disable_gt_powersave(struct drm_device *dev)
> > >  		mutex_lock(&dev_priv->rps.hw_lock);
> > >  		gen6_disable_rps(dev);
> > >  		mutex_unlock(&dev_priv->rps.hw_lock);
> > 
> > Theoretically the work could get executed here, between unlock and
> > cancel. Does that matter?
> 
> AFAICS even gen6_pm_rps_work() might be executing even after
> gen6_disable_rps() was called.
> 
> I'm not quite sure if allowing gen6_pm_rps_work() to do stuff after
> gen6_disable_rps() was called is already a problem on it's own. At
> least it can overwrite RPNSWREQ on !VLV. BTW that register doesn't seem
> to exist on VLV and yet we're poking at it from gen6_disable_rps()...

Yeah, ordering in our rps teardown code seems screwed-up, since we disable
interrupts a bit too late. Now disable_rps does clear PM_IIR and disable
all interrupts, but like you've said it doesn't bother to stop the
rps_work, so disable_rps can still get rearmed.

Two things:
- I do not know how much we really should care here for now, since ...

- Setup/teardown ordering is already a giant mess and did blow up a lot of
  times recently. I wonder whether we shouldn't just start to massively
  sprinkle asserts all over the place to double/triple check ordering
  sequence constrainst, similarly to how we currently handle the modeset
  sequence.

Another tool to (partially) alleviate these issues (especially failure
path handling) is devres.c, which I've just recently stumbled over.

Definitely an area to ponder, since imo setup/teardown is the currently
worst part (in terms of likeliness to break unnoticed) in our driver.

Cheers, Daniel
Jesse Barnes April 23, 2013, 4:36 p.m. UTC | #4
On Tue, 23 Apr 2013 16:28:17 +0300
Ville Syrjälä <ville.syrjala@linux.intel.com> wrote:

> On Tue, Apr 23, 2013 at 11:45:03AM +0300, Jani Nikula wrote:
> > On Tue, 23 Apr 2013, Jesse Barnes <jbarnes@virtuousgeek.org> wrote:
> > > On VLV, the Punit doesn't automatically drop the GPU to it's minimum
> > > voltage level when entering RC6, so we arm a timer to do it for us from
> > > the RPS interrupt handler.  It'll generally only fire when we go idle
> > > (or if for some reason there's a long delay between RPS interrupts), but
> > > won't be re-armed again until the next RPS event, so shouldn't affect
> > > power consumption after we go idle and it triggers.
> > >
> > > v2: use delayed work instead of timer + work queue combo (Ville)
> > > v3: fix up delayed work cancel (must be outside lock) (Daniel)
> > >     fix up delayed work handling func for delayed work (Jesse)
> > >
> > > Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
> > > ---
> > >  drivers/gpu/drm/i915/i915_drv.h |    2 ++
> > >  drivers/gpu/drm/i915/i915_irq.c |   11 +++++++++++
> > >  drivers/gpu/drm/i915/intel_pm.c |   22 ++++++++++++++++++++++
> > >  3 files changed, 35 insertions(+)
> > >
> > > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > > index 2557fc7..f563f25 100644
> > > --- a/drivers/gpu/drm/i915/i915_drv.h
> > > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > > @@ -660,6 +660,7 @@ struct i915_suspend_saved_registers {
> > >  
> > >  struct intel_gen6_power_mgmt {
> > >  	struct work_struct work;
> > > +	struct delayed_work vlv_work;
> > >  	u32 pm_iir;
> > >  	/* lock - irqsave spinlock that protectects the work_struct and
> > >  	 * pm_iir. */
> > > @@ -670,6 +671,7 @@ struct intel_gen6_power_mgmt {
> > >  	u8 cur_delay;
> > >  	u8 min_delay;
> > >  	u8 max_delay;
> > > +	u8 rpe_delay;
> > >  	u8 hw_max;
> > >  
> > >  	struct delayed_work delayed_resume_work;
> > > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> > > index 932e7f8..af385e2 100644
> > > --- a/drivers/gpu/drm/i915/i915_irq.c
> > > +++ b/drivers/gpu/drm/i915/i915_irq.c
> > > @@ -488,6 +488,17 @@ static void gen6_pm_rps_work(struct work_struct *work)
> > >  			gen6_set_rps(dev_priv->dev, new_delay);
> > >  	}
> > >  
> > > +	if (IS_VALLEYVIEW(dev_priv->dev)) {
> > > +		/*
> > > +		 * On VLV, when we enter RC6 we may not be at the minimum
> > > +		 * voltage level, so arm a timer to check.  It should only
> > > +		 * fire when there's activity or once after we've entered
> > > +		 * RC6, and then won't be re-armed until the next RPS interrupt.
> > > +		 */
> > > +		mod_delayed_work(dev_priv->wq, &dev_priv->rps.vlv_work,
> > > +				 jiffies + msecs_to_jiffies(100));
> > 
> > Hooray for conflicting semantics. Unlike mod_timer, mod_delayed_work
> > wants a delay, not an absolute time in jiffies.
> > 
> > > +	}
> > > +
> > >  	mutex_unlock(&dev_priv->rps.hw_lock);
> > >  }
> > >  
> > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > index 2557926..c106187 100644
> > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > @@ -2822,6 +2822,23 @@ int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
> > >  	return val & 0xff;
> > >  }
> > >  
> > > +static void vlv_rps_timer_work(struct work_struct *work)
> > > +{
> > > +	drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
> > > +						    rps.vlv_work.work);
> > > +
> > > +	/*
> > > +	 * Timer fired, we must be idle.  Drop to min voltage state.
> > > +	 * Note: we use RPe here since it should match the
> > > +	 * Vmin we were shooting for.  That should give us better
> > > +	 * perf when we come back out of RC6 than if we used the
> > > +	 * min freq available.
> > > +	 */
> > > +	mutex_lock(&dev_priv->rps.hw_lock);
> > > +	valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
> > > +	mutex_unlock(&dev_priv->rps.hw_lock);
> > > +}
> > > +
> > >  static void valleyview_enable_rps(struct drm_device *dev)
> > >  {
> > >  	struct drm_i915_private *dev_priv = dev->dev_private;
> > > @@ -2886,6 +2903,7 @@ static void valleyview_enable_rps(struct drm_device *dev)
> > >  	rpe = valleyview_rps_rpe_freq(dev_priv);
> > >  	DRM_DEBUG_DRIVER("RPe GPU freq: %d\n",
> > >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> > > +	dev_priv->rps.rpe_delay = rpe;
> > >  
> > >  	val = valleyview_rps_min_freq(dev_priv);
> > >  	DRM_DEBUG_DRIVER("min GPU freq: %d\n", vlv_gpu_freq(dev_priv->mem_freq,
> > > @@ -2895,6 +2913,8 @@ static void valleyview_enable_rps(struct drm_device *dev)
> > >  	DRM_DEBUG_DRIVER("setting GPU freq to %d\n",
> > >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> > >  
> > > +	INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
> > > +
> > >  	valleyview_set_rps(dev_priv->dev, rpe);
> > >  
> > >  	/* requires MSI enabled */
> > > @@ -3640,6 +3660,8 @@ void intel_disable_gt_powersave(struct drm_device *dev)
> > >  		mutex_lock(&dev_priv->rps.hw_lock);
> > >  		gen6_disable_rps(dev);
> > >  		mutex_unlock(&dev_priv->rps.hw_lock);
> > 
> > Theoretically the work could get executed here, between unlock and
> > cancel. Does that matter?
> 
> AFAICS even gen6_pm_rps_work() might be executing even after
> gen6_disable_rps() was called.
> 
> I'm not quite sure if allowing gen6_pm_rps_work() to do stuff after
> gen6_disable_rps() was called is already a problem on it's own. At
> least it can overwrite RPNSWREQ on !VLV. BTW that register doesn't seem
> to exist on VLV and yet we're poking at it from gen6_disable_rps()...

Hm that could cause a problem, I know some of the reserved regs in that
region have bad side effects.

I'll cook up a patch to make a valleyview_disable_rps.

Having the work done after rps disable is probably ok, but it's still
messy.  I'll move the bits up above the rps disable call.  Since
they're sync that should serialize things.

And I'll fix up the mod timer; converting from timers and work queues
to delayed work isn't as mindless as it ought to be (I already caught
the INIT_DELAYED_WORK bug I had before too, groan).
Jesse Barnes April 23, 2013, 5:10 p.m. UTC | #5
On Tue, 23 Apr 2013 16:28:17 +0300
Ville Syrjälä <ville.syrjala@linux.intel.com> wrote:

> On Tue, Apr 23, 2013 at 11:45:03AM +0300, Jani Nikula wrote:
> > On Tue, 23 Apr 2013, Jesse Barnes <jbarnes@virtuousgeek.org> wrote:
> > > On VLV, the Punit doesn't automatically drop the GPU to it's minimum
> > > voltage level when entering RC6, so we arm a timer to do it for us from
> > > the RPS interrupt handler.  It'll generally only fire when we go idle
> > > (or if for some reason there's a long delay between RPS interrupts), but
> > > won't be re-armed again until the next RPS event, so shouldn't affect
> > > power consumption after we go idle and it triggers.
> > >
> > > v2: use delayed work instead of timer + work queue combo (Ville)
> > > v3: fix up delayed work cancel (must be outside lock) (Daniel)
> > >     fix up delayed work handling func for delayed work (Jesse)
> > >
> > > Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
> > > ---
> > >  drivers/gpu/drm/i915/i915_drv.h |    2 ++
> > >  drivers/gpu/drm/i915/i915_irq.c |   11 +++++++++++
> > >  drivers/gpu/drm/i915/intel_pm.c |   22 ++++++++++++++++++++++
> > >  3 files changed, 35 insertions(+)
> > >
> > > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > > index 2557fc7..f563f25 100644
> > > --- a/drivers/gpu/drm/i915/i915_drv.h
> > > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > > @@ -660,6 +660,7 @@ struct i915_suspend_saved_registers {
> > >  
> > >  struct intel_gen6_power_mgmt {
> > >  	struct work_struct work;
> > > +	struct delayed_work vlv_work;
> > >  	u32 pm_iir;
> > >  	/* lock - irqsave spinlock that protectects the work_struct and
> > >  	 * pm_iir. */
> > > @@ -670,6 +671,7 @@ struct intel_gen6_power_mgmt {
> > >  	u8 cur_delay;
> > >  	u8 min_delay;
> > >  	u8 max_delay;
> > > +	u8 rpe_delay;
> > >  	u8 hw_max;
> > >  
> > >  	struct delayed_work delayed_resume_work;
> > > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> > > index 932e7f8..af385e2 100644
> > > --- a/drivers/gpu/drm/i915/i915_irq.c
> > > +++ b/drivers/gpu/drm/i915/i915_irq.c
> > > @@ -488,6 +488,17 @@ static void gen6_pm_rps_work(struct work_struct *work)
> > >  			gen6_set_rps(dev_priv->dev, new_delay);
> > >  	}
> > >  
> > > +	if (IS_VALLEYVIEW(dev_priv->dev)) {
> > > +		/*
> > > +		 * On VLV, when we enter RC6 we may not be at the minimum
> > > +		 * voltage level, so arm a timer to check.  It should only
> > > +		 * fire when there's activity or once after we've entered
> > > +		 * RC6, and then won't be re-armed until the next RPS interrupt.
> > > +		 */
> > > +		mod_delayed_work(dev_priv->wq, &dev_priv->rps.vlv_work,
> > > +				 jiffies + msecs_to_jiffies(100));
> > 
> > Hooray for conflicting semantics. Unlike mod_timer, mod_delayed_work
> > wants a delay, not an absolute time in jiffies.
> > 
> > > +	}
> > > +
> > >  	mutex_unlock(&dev_priv->rps.hw_lock);
> > >  }
> > >  
> > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > index 2557926..c106187 100644
> > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > @@ -2822,6 +2822,23 @@ int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
> > >  	return val & 0xff;
> > >  }
> > >  
> > > +static void vlv_rps_timer_work(struct work_struct *work)
> > > +{
> > > +	drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
> > > +						    rps.vlv_work.work);
> > > +
> > > +	/*
> > > +	 * Timer fired, we must be idle.  Drop to min voltage state.
> > > +	 * Note: we use RPe here since it should match the
> > > +	 * Vmin we were shooting for.  That should give us better
> > > +	 * perf when we come back out of RC6 than if we used the
> > > +	 * min freq available.
> > > +	 */
> > > +	mutex_lock(&dev_priv->rps.hw_lock);
> > > +	valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
> > > +	mutex_unlock(&dev_priv->rps.hw_lock);
> > > +}
> > > +
> > >  static void valleyview_enable_rps(struct drm_device *dev)
> > >  {
> > >  	struct drm_i915_private *dev_priv = dev->dev_private;
> > > @@ -2886,6 +2903,7 @@ static void valleyview_enable_rps(struct drm_device *dev)
> > >  	rpe = valleyview_rps_rpe_freq(dev_priv);
> > >  	DRM_DEBUG_DRIVER("RPe GPU freq: %d\n",
> > >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> > > +	dev_priv->rps.rpe_delay = rpe;
> > >  
> > >  	val = valleyview_rps_min_freq(dev_priv);
> > >  	DRM_DEBUG_DRIVER("min GPU freq: %d\n", vlv_gpu_freq(dev_priv->mem_freq,
> > > @@ -2895,6 +2913,8 @@ static void valleyview_enable_rps(struct drm_device *dev)
> > >  	DRM_DEBUG_DRIVER("setting GPU freq to %d\n",
> > >  			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
> > >  
> > > +	INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
> > > +
> > >  	valleyview_set_rps(dev_priv->dev, rpe);
> > >  
> > >  	/* requires MSI enabled */
> > > @@ -3640,6 +3660,8 @@ void intel_disable_gt_powersave(struct drm_device *dev)
> > >  		mutex_lock(&dev_priv->rps.hw_lock);
> > >  		gen6_disable_rps(dev);
> > >  		mutex_unlock(&dev_priv->rps.hw_lock);
> > 
> > Theoretically the work could get executed here, between unlock and
> > cancel. Does that matter?
> 
> AFAICS even gen6_pm_rps_work() might be executing even after
> gen6_disable_rps() was called.
> 
> I'm not quite sure if allowing gen6_pm_rps_work() to do stuff after
> gen6_disable_rps() was called is already a problem on it's own. At
> least it can overwrite RPNSWREQ on !VLV. BTW that register doesn't seem
> to exist on VLV and yet we're poking at it from gen6_disable_rps()...

Ok just posted a new set addressing these issues, please check them out.

Thanks,
diff mbox

Patch

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 2557fc7..f563f25 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -660,6 +660,7 @@  struct i915_suspend_saved_registers {
 
 struct intel_gen6_power_mgmt {
 	struct work_struct work;
+	struct delayed_work vlv_work;
 	u32 pm_iir;
 	/* lock - irqsave spinlock that protectects the work_struct and
 	 * pm_iir. */
@@ -670,6 +671,7 @@  struct intel_gen6_power_mgmt {
 	u8 cur_delay;
 	u8 min_delay;
 	u8 max_delay;
+	u8 rpe_delay;
 	u8 hw_max;
 
 	struct delayed_work delayed_resume_work;
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index 932e7f8..af385e2 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -488,6 +488,17 @@  static void gen6_pm_rps_work(struct work_struct *work)
 			gen6_set_rps(dev_priv->dev, new_delay);
 	}
 
+	if (IS_VALLEYVIEW(dev_priv->dev)) {
+		/*
+		 * On VLV, when we enter RC6 we may not be at the minimum
+		 * voltage level, so arm a timer to check.  It should only
+		 * fire when there's activity or once after we've entered
+		 * RC6, and then won't be re-armed until the next RPS interrupt.
+		 */
+		mod_delayed_work(dev_priv->wq, &dev_priv->rps.vlv_work,
+				 jiffies + msecs_to_jiffies(100));
+	}
+
 	mutex_unlock(&dev_priv->rps.hw_lock);
 }
 
diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
index 2557926..c106187 100644
--- a/drivers/gpu/drm/i915/intel_pm.c
+++ b/drivers/gpu/drm/i915/intel_pm.c
@@ -2822,6 +2822,23 @@  int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
 	return val & 0xff;
 }
 
+static void vlv_rps_timer_work(struct work_struct *work)
+{
+	drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
+						    rps.vlv_work.work);
+
+	/*
+	 * Timer fired, we must be idle.  Drop to min voltage state.
+	 * Note: we use RPe here since it should match the
+	 * Vmin we were shooting for.  That should give us better
+	 * perf when we come back out of RC6 than if we used the
+	 * min freq available.
+	 */
+	mutex_lock(&dev_priv->rps.hw_lock);
+	valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
+	mutex_unlock(&dev_priv->rps.hw_lock);
+}
+
 static void valleyview_enable_rps(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -2886,6 +2903,7 @@  static void valleyview_enable_rps(struct drm_device *dev)
 	rpe = valleyview_rps_rpe_freq(dev_priv);
 	DRM_DEBUG_DRIVER("RPe GPU freq: %d\n",
 			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
+	dev_priv->rps.rpe_delay = rpe;
 
 	val = valleyview_rps_min_freq(dev_priv);
 	DRM_DEBUG_DRIVER("min GPU freq: %d\n", vlv_gpu_freq(dev_priv->mem_freq,
@@ -2895,6 +2913,8 @@  static void valleyview_enable_rps(struct drm_device *dev)
 	DRM_DEBUG_DRIVER("setting GPU freq to %d\n",
 			 vlv_gpu_freq(dev_priv->mem_freq, rpe));
 
+	INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
+
 	valleyview_set_rps(dev_priv->dev, rpe);
 
 	/* requires MSI enabled */
@@ -3640,6 +3660,8 @@  void intel_disable_gt_powersave(struct drm_device *dev)
 		mutex_lock(&dev_priv->rps.hw_lock);
 		gen6_disable_rps(dev);
 		mutex_unlock(&dev_priv->rps.hw_lock);
+		if (IS_VALLEYVIEW(dev))
+			cancel_delayed_work_sync(&dev_priv->rps.vlv_work);
 	}
 }