[Intel-gfx] [PATCH v4] drm/i915/chv: Set min freq to efficient frequency on chv

[Ville Syrjälä]

On Thu, Apr 30, 2015 at 03:42:42PM +0530, Deepak S wrote:
> 
> 
> On Thursday 30 April 2015 01:23 AM, Ville Syrjälä wrote:
> > On Wed, Apr 29, 2015 at 06:31:56PM +0300, Ville Syrjälä wrote:
> >> On Wed, Apr 29, 2015 at 08:23:21AM +0530, deepak.s at linux.intel.com wrote:
> >>> From: Deepak S <deepak.s at linux.intel.com>
> >>>
> >>> After feedback from the hardware team, now we set the GPU min/idel freq to RPe.
> >>> Punit is expecting us to operate GPU between Rpe & Rp0. If we drop the
> >>> frequency to RPn, punit is failing to change the input voltage to
> >>> minimum :(
> >> So far I can't reproduce this problem on my BSW. In fact what I see
> >> that the voltage at RPn is lower than the voltage at RPe, even while
> >> we're in rc6.
> >>
> >> without forcewake:
> >> RPn -> 0x66
> >> RPe -> 0x67
> >> RP0 -> 0x69
> >>
> >> with forcewake:
> >> RPn -> 0x66
> >> RPe -> 0x76
> >> RP0 -> 0x9d
> >>
> >> Also asking Punit to change the frequency after the GPU has gone to
> >> rc6 does absolutely nothing (remind anyone of VLV?). I think I need to
> >> retest my VLV C0 to see if my earlier observations there were accurate.
> >> The shared Vnn rail does make it harder to observe this stuff on
> >> VLV though.
> > I went back to my VLVs (had a B3 and C0 actually). And I'm seeing the
> > exact same behaviour on both, ie. requesting a new frequency from Punit
> > does nothing when the GPU is in rc6, and if I let it enter rc6 with a
> > high frequency Vnn also remains high. Previously I had thought that C0
> > fixed this, but now it definitely shows the same problem here. I must
> > have had some accidental forcewake somewhere when I originally tested
> > it,
> >
> > So based on that, your other patch to remove the stepping check from
> > vlv_set_rps_idle() is in fact correct.
> >
> > The question remains however what should we do with CHV. According to my
> > testing to get the minimum voltage we should keep RPn around, and we
> > should also do the vlv_set_rps_idle() workaround on CHV.
> >
> > Oh and I also observed something else on VLV. Normally when entering rc6
> > the GPLL ref clock gets trunk gated at CCK (by Punit I assume). However
> > when using the vlv_gfx_clock_force() that doesn't happen. So I'm not
> > entirely sure the GPLL gets turned off properly in that case. Maybe we
> > should just use forcewake instead? Oh and BTW, CHV doesn't do the trunk
> > gating in either case. I'm not sure where to check if the GPLL is
> > actually running or not.
> >
> Hi Ville,
> 
> Thanks Ville for verifying on VLV and CHV. Its interesting to see when Idle,
> Voltage not dropping to Vnn on CHV :( This was supposed to be fixed
> in BSW/CHV :(.

Well it does drop, but not quite all the way if the current
frequency is above RPn.

I just tried to look at the energy numbers from RAPL a bit. Based
on that the difference in Vgg when forcewake is off is rather
insignificant, or the Vgg portion of the always-on well is so small
that it doesn't really matter.

With forcewake enabled I see a clear difference between RPn and
RPe though. But that might mostly matter for cases where would grab
forcewake without actually feeding any work to the GPU. That might
be rare enough to not make much of a difference in practice.

> As you suggested it would be better to extend the VLV WA to
> CHV also to make sure we drop the voltage when idle.
> 
> Below is the sequence I think we should follow (based  on your comments).
> 1. forcewake power wells
> 2. do gfx force clock on
> 3. request freq to punit
> 4. release gfx force clock on
> 5. release forcewake of power wells.
> 
> Please share your thoughts?

I'm thinking we shouldn't need the gfx clock force since forcewake
should already cause the clock to be enabled.

I've also not verified what happens if we drop the forcewake before
Punit has actually finished the frequency change. I'll try to hack
up some kind of test to see if I can make that happen.

> 
> Thanks
> Deepak
> 
> >> So based on my tests this patch feels a bit wrong.
> >>
> >>> Since Punit validates the rps range [RPe, RP0]. This patch
> >>> removes unused cherryview_rps_min_freq function.
> >>>
> >>> v2: Change commit message
> >>>
> >>> v3: set min_freq before idle_freq (chris)
> >>>
> >>> v4: Squash 'Remove unused rps min function' patch
> >>>
> >>> Signed-off-by: Deepak S <deepak.s at linux.intel.com>
> >>> Acked-by: Chris Wilson <chris at chris-wilson.co.uk>
> >>> ---
> >>>   drivers/gpu/drm/i915/intel_pm.c | 21 ++-------------------
> >>>   1 file changed, 2 insertions(+), 19 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> >>> index a7516ed..78c89ff 100644
> >>> --- a/drivers/gpu/drm/i915/intel_pm.c
> >>> +++ b/drivers/gpu/drm/i915/intel_pm.c
> >>> @@ -4715,24 +4715,6 @@ static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
> >>>   	return rp1;
> >>>   }
> >>>   
> >>> -static int cherryview_rps_min_freq(struct drm_i915_private *dev_priv)
> >>> -{
> >>> -	struct drm_device *dev = dev_priv->dev;
> >>> -	u32 val, rpn;
> >>> -
> >>> -	if (dev->pdev->revision >= 0x20) {
> >>> -		val = vlv_punit_read(dev_priv, FB_GFX_FMIN_AT_VMIN_FUSE);
> >>> -		rpn = ((val >> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT) &
> >>> -		       FB_GFX_FREQ_FUSE_MASK);
> >>> -	} else { /* For pre-production hardware */
> >>> -		val = vlv_punit_read(dev_priv, PUNIT_GPU_STATUS_REG);
> >>> -		rpn = ((val >> PUNIT_GPU_STATIS_GFX_MIN_FREQ_SHIFT) &
> >>> -		       PUNIT_GPU_STATUS_GFX_MIN_FREQ_MASK);
> >>> -	}
> >>> -
> >>> -	return rpn;
> >>> -}
> >>> -
> >>>   static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
> >>>   {
> >>>   	u32 val, rp1;
> >>> @@ -4984,7 +4966,8 @@ static void cherryview_init_gt_powersave(struct drm_device *dev)
> >>>   			 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
> >>>   			 dev_priv->rps.rp1_freq);
> >>>   
> >>> -	dev_priv->rps.min_freq = cherryview_rps_min_freq(dev_priv);
> >>> +	/* PUnit validated range is only [RPe, RP0] */
> >>> +	dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
> >>>   	DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
> >>>   			 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
> >>>   			 dev_priv->rps.min_freq);
> >>> -- 
> >>> 1.9.1
> >> -- 
> >> Ville Syrjälä
> >> Intel OTC
> >> _______________________________________________
> >> Intel-gfx mailing list
> >> Intel-gfx at lists.freedesktop.org
> >> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Ville Syrjälä
Intel OTC