[Intel-gfx] [RFC PATCH] drm/i915/skl: Add DC6 disabling as a power well

[Ville Syrjälä]

On Thu, Sep 24, 2015 at 02:50:16PM +0200, Patrik Jakobsson wrote:
> On Wed, Sep 23, 2015 at 01:18:00PM +0200, Patrik Jakobsson wrote:
> > On Wed, Sep 23, 2015 at 10:43:00AM +0200, Daniel Vetter wrote:
> > > On Mon, Sep 21, 2015 at 10:00:45AM +0200, Patrik Jakobsson wrote:
> > > > On Wed, Sep 16, 2015 at 11:10:07PM +0300, Ville Syrjälä wrote:
> > > > > On Fri, Sep 11, 2015 at 01:55:22PM +0200, Patrik Jakobsson wrote:
> > > > > > We need to be able to control if DC6 is allowed or not. Much like
> > > > > > requesting power to a specific piece of the hardware we need to be able
> > > > > > to request that we don't enter DC6 during certain hw access.
> > > > > > 
> > > > > > To solve this without introducing too much infrastructure I'm hooking
> > > > > > into the power well / power domain framework. DC6 prevention is modeled
> > > > > > much like an enabled power well. Thus I'm using the terminology on/off
> > > > > > for DC states instead of enable/disable.
> > > > > > 
> > > > > > The problem that started this work is the need for DC6 to be disabled
> > > > > > when accessing DP_AUX_A during CRTC on/off. That is also fixed in this
> > > > > > patch.
> > > > > > 
> > > > > > This is posted as an RFC since DMC and DC state handling is being
> > > > > > reworked and will possibly affect the outcome of this patch. The patch
> > > > > > has known warnings.
> > > > > > 
> > > > > > Signed-off-by: Patrik Jakobsson <patrik.jakobsson at linux.intel.com>
> > > > > > ---
> > > > > >  drivers/gpu/drm/i915/intel_ddi.c        |  9 +++++
> > > > > >  drivers/gpu/drm/i915/intel_drv.h        |  2 +
> > > > > >  drivers/gpu/drm/i915/intel_runtime_pm.c | 69 +++++++++++++++++++++++++--------
> > > > > >  3 files changed, 64 insertions(+), 16 deletions(-)
> > > > > > 
> > > > > > diff --git a/drivers/gpu/drm/i915/intel_ddi.c b/drivers/gpu/drm/i915/intel_ddi.c
> > > > > > index 4823184..c2c1ad2 100644
> > > > > > --- a/drivers/gpu/drm/i915/intel_ddi.c
> > > > > > +++ b/drivers/gpu/drm/i915/intel_ddi.c
> > > > > > @@ -2288,6 +2288,8 @@ static void intel_ddi_pre_enable(struct intel_encoder *intel_encoder)
> > > > > >  	if (type == INTEL_OUTPUT_DISPLAYPORT || type == INTEL_OUTPUT_EDP) {
> > > > > >  		struct intel_dp *intel_dp = enc_to_intel_dp(encoder);
> > > > > >  
> > > > > > +		intel_display_power_get(dev_priv, POWER_DOMAIN_AUX_A);
> > > > > > +
> > > > > 
> > > > > These I think shouldn't be necessary with my
> > > > > intel_display_port_aux_power_domain() stuff since intel_dp_aux_ch() will
> > > > > itself grab the appropriate power domain.
> > > > > 
> > > > > That's of course assuming that AUX is the only reason why we need to
> > > > > keep DC6 disabled here.
> > > > > 
> > > > 
> > > > The upside with having get/put around bigger aux transfers is that we don't get
> > > > tons of enable/disable lines in the log. My vote is that we keep this but also
> > > > have your fine-grained get/puts.
> > > 
> > > Imo the correct solution to avoid this is by adding a slight bit of
> > > hystersis to the power well code. Which means that yes, we reinvent
> > > another feature of the core power_domain code in our home-grown solution -
> > > I hate it when my years old predictions come true ;-)
> > > 
> > > Sprinkling higher-level get/put calls all over the place is imo just
> > > leaking the abstraction, which isn't good.
> > > -Daniel
> > 
> > With Ville's patches the problem is not as bad as I first thought. We can add
> > hysteresis later if needed.
> > 
> > -Patrik
> 
> So I discovered that we cannot have DC5 and DC6 as seperate power wells since
> they are mutually exclusive. As Ville pointed out we don't use DC5 for anything
> so we could get away for now with just DC6 as a power well.
> 
> As I see it there are three ways to go about this:
> 
> A. Add AUX A to Power well 2.
> This would power up PW2 during DP Aux A even though we don't need it but since
> we get the side effect of DC6 being disabled it should work.
> 
> B. Add DC6 off as a power well and remove DC5 off.
> Fairly straight forward but assumes we don't need DC5 control.
> 
> C. Add multi-state support for the DC power well.
> Would be the nice way but perhaps a bit overkill. Good thing about this would be
> that we can incorporate DC9 control for BXT and unify more of the DC code.
> 
> So A, B or C?

After some spitballing with Imre we came up with the following power well layout,
which I think matches the documented seuqences in the spec the best:

display well:
    domains: 
        all
    enable:
        init display sequence
            enable pch reset handshake
            enable pw1 and miscio
            enable cdclk pll
            enable dbuf
        dc_state_en = up_to_dc6
    disable:
        dc_state_en = disable
        uninit display sequence
            disable dbuf
            disable cdclk pll
            disable pw1 and miscio

dc6 off well:
    domains:
        gmbus, dc5 off well domains
    enable:
        dc_state_en = up_to_dc5
    disable:
        dc_state_en = up_to_dc6

dc5 off well:
    domains:
        aux A, pw2 well domains
    enable:
        dc_state_en = disable
    disable:
        dc_state_en = up_to_dc5

pw2 well:
    domains:
        ddi B/C/D/E, aux B/C/D/E, vga, audio, pipe B/C, pfit B/C
    enable:
        enable pw2
    disable:
        disable pw2

-- 
Ville Syrjälä
Intel OTC