[Intel-gfx] [RFC 3/4] drm/i915: valleyview: Make intel_set_rps get FORCEWAKE_MEDIA

[Ville Syrjälä]

On Mon, Jan 02, 2017 at 02:21:58PM +0000, Chris Wilson wrote:
> On Mon, Jan 02, 2017 at 04:10:49PM +0200, Ville Syrjälä wrote:
> > On Mon, Jan 02, 2017 at 11:37:59AM +0000, Chris Wilson wrote:
> > > On Sun, Jan 01, 2017 at 09:48:53PM +0100, Hans de Goede wrote:
> > > > Hi,
> > > > 
> > > > On 01-01-17 21:24, Chris Wilson wrote:
> > > > >On Sun, Jan 01, 2017 at 09:14:02PM +0100, Hans de Goede wrote:
> > > > >>All callers of valleyview_set_rps() get at least FORCEWAKE_MEDIA, except
> > > > >>for intel_set_rps(). Since intel_set_rps can for example be called from
> > > > >>sysfs store functions, there is no guarantee this is already done, so add
> > > > >>an intel_uncore_forcewake_get(FORCEWAKE_MEDIA) call to intel_set_rps.
> > > > >>
> > > > >>Signed-off-by: Hans de Goede <hdegoede at redhat.com>
> > > > >>---
> > > > >> drivers/gpu/drm/i915/intel_pm.c | 9 +++++++--
> > > > >> 1 file changed, 7 insertions(+), 2 deletions(-)
> > > > >>
> > > > >>diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > > >>index 4b12637..cc4fbd7 100644
> > > > >>--- a/drivers/gpu/drm/i915/intel_pm.c
> > > > >>+++ b/drivers/gpu/drm/i915/intel_pm.c
> > > > >>@@ -5096,9 +5096,14 @@ void gen6_rps_boost(struct drm_i915_private *dev_priv,
> > > > >>
> > > > >> void intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
> > > > >> {
> > > > >>-	if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
> > > > >>+	if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
> > > > >>+		/* Wake up the media well, as that takes a lot less
> > > > >>+		 * power than the Render well.
> > > > >>+		 */
> > > > >>+		intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
> > > > >> 		valleyview_set_rps(dev_priv, val);
> > > > >
> > > > >Both powerwells are woken for rps. (Taking one but not the other has no
> > > > >benefit, and very misleading.)
> > > > 
> > > > The comment on why FORCEWAKE_MEDIA is used + code is copy pasted from the
> > > > existing code in vlv_set_rps_idle().
> > > 
> > > It's not correct there either...
> > 
> > Why not? If the render well is in rc6 already we don't want to waste
> > power by waking it. The only reason we have to wake up *something* is
> > so that the gpll frequency actually gets changed.
> 
> If the register write requires the powerwell, the mmio access will take
> the powerwell.

The register write doesn't require a power well. It's a sideband access.
The punit will simply not change the GPLL frequency if the GPLL is
currently not running (which will/can happen when all power wells are
asleep). That in itself doesn't sound too back (why change the
frequency if the thing isn't even running, right?). But the real problem
is that the punit will not let the voltage on the rail to drop
either until the new frequency gets programmed into the GPLL. Hence if
the GPU goes idle before we've dropped the GPLL frequency to the
minimum value, we will waste power by having a needlessly high voltage.

Originally we tried to avoid this problem via vlv_force_gfx_clock(),
which should just force the GPLL to turn on without powering on
any power wells. But that caused some spurious WARNs or something
IIRC, so we had to come up with another workaround. And since powering
either well is sufficient we chose to use the cheaper media well.

> Manually taking the wakelock has to be for a reason, the
> one given is bogus. (That is if the register write only requires meda
> access that is taken, but iirc, rps requires both, so both are taken
> inside.) If the claimed benefit is indeed real, and there is indeed
> freedom of choice, we want that inside the mmio accessor to choose the
> cheaper well.
> -Chris
> 
> -- 
> Chris Wilson, Intel Open Source Technology Centre

-- 
Ville Syrjälä
Intel OTC