[Intel-gfx] [PATCH] drm/i915/vlv: assert and de-assert sideband reset at boot and resume v3

[Ville Syrjälä]

On Wed, May 21, 2014 at 11:11:15AM -0700, Jesse Barnes wrote:
> And to answer more specifically...
> 
> On Wed, 21 May 2014 20:54:03 +0300
> Ville Syrjälä <ville.syrjala at linux.intel.com> wrote:
> 
> > > +		__vlv_set_power_well(dev_priv, PUNIT_POWER_WELL_DPIO_CMN_BC,
> > > +				     false);
> > > +		__vlv_set_power_well(dev_priv, PUNIT_POWER_WELL_DPIO_CMN_BC,
> > > +				     true);
> > 
> > Hmm. I wonder how the power well hack in intel_uncore_sanitize() ties in
> > with this. We should definitely rip that out regardless.
> 
> Yeah we can rip that out.  That's just an ungate, and it assumes the
> BIOS has already done the reset toggle for us.

Well I'm assumign the system may boot/resume with the wells powered
down. And we definitely don't have the cri/ref clocks set up at this
point. So if they're needed to complete the resets the PHY may get stuck
or something. Also it just ungates everything at once, if the wells need
some ordering in bringup we have just messed things up here.

> 
> > Another thing I'm wodering is did the BIOS/hw really power on the
> > common lane, or did we do that outselves? If the latter, then I wonder
> > if we simply do that too early. Or more precisely do we need to make
> > sure the cri clock and/or refclock are enabled before we power on the
> > common lane?
> 
> Depends on the platform.  It looks like the right thing happens at boot
> time on most machines (i.e. the BIOS does a full toggle which causes a
> reset), but on suspend/resume it's up to us.  And of course on machines
> without video init at boot time, we need to do it ourselves as well.
> 
> I don't think the cri or refclk is required at this point, at least not
> if the docs we have are correct (the PHY reset is supposed to be the
> very first thing).

The timing diagrams do show some kind of relationship between the
cri/ref clocks and the reset signals getting deasserted.

> 
> > And third is that do we need to enable the TX wells before the CMN
> > well? Currently we do the opposite which could also explain why this
> > CMN well toggle fixes things.
> 
> I don't think that matters, but we should ask the PHY guys.  The lack
> of symmetry between the gate and ungate bothers me too.

My theory here is that the cmn and/or side reset needs to propagate to
the data lanes to propery initialize them. If the lanes are powered down
when the reset occurs things go bad.

And looks like the "Dynamic Power Gating" section in the phy cspec
pretty much says as much. That is you need to fully reset the entire
phy if you want to ungate even one of the blocks.

So based on that I think pretty much everything I said is needed to make
stuff work correctly. Also I think currently we power gate the TX blocks
whenever the port gets disabled. If we want to keep doing that, we need
to adjust valleyview_modeset_global_resources() to fully reset the phy
whenever previously power down wells need to be brought up. The
alternative would be to model the entire PHY as a single power well,
which would definitely be the simpler option.

-- 
Ville Syrjälä
Intel OTC